qemu-e2k/hw/ppc/vof.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1063 lines
28 KiB
C
Raw Normal View History

spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
/*
* QEMU PowerPC Virtual Open Firmware.
*
* This implements client interface from OpenFirmware IEEE1275 on the QEMU
* side to leave only a very basic firmware in the VM.
*
* Copyright (c) 2021 IBM Corporation.
*
* SPDX-License-Identifier: GPL-2.0-or-later
*/
#include "qemu/osdep.h"
#include "qemu/timer.h"
#include "qemu/range.h"
#include "qemu/units.h"
#include "qemu/log.h"
#include "qapi/error.h"
#include "exec/address-spaces.h"
#include "hw/ppc/vof.h"
#include "hw/ppc/fdt.h"
#include "sysemu/runstate.h"
#include "qom/qom-qobject.h"
#include "trace.h"
#include <libfdt.h>
/*
* OF 1275 "nextprop" description suggests is it 32 bytes max but
* LoPAPR defines "ibm,query-interrupt-source-number" which is 33 chars long.
*/
#define OF_PROPNAME_LEN_MAX 64
#define VOF_MAX_PATH 256
#define VOF_MAX_SETPROPLEN 2048
#define VOF_MAX_METHODLEN 256
#define VOF_MAX_FORTHCODE 256
#define VOF_VTY_BUF_SIZE 256
typedef struct {
uint64_t start;
uint64_t size;
} OfClaimed;
typedef struct {
char *path; /* the path used to open the instance */
uint32_t phandle;
} OfInstance;
static int readstr(hwaddr pa, char *buf, int size)
{
if (VOF_MEM_READ(pa, buf, size) != MEMTX_OK) {
return -1;
}
if (strnlen(buf, size) == size) {
buf[size - 1] = '\0';
trace_vof_error_str_truncated(buf, size);
return -1;
}
return 0;
}
static bool cmpservice(const char *s, unsigned nargs, unsigned nret,
const char *s1, unsigned nargscheck, unsigned nretcheck)
{
if (strcmp(s, s1)) {
return false;
}
if ((nargscheck && (nargs != nargscheck)) ||
(nretcheck && (nret != nretcheck))) {
trace_vof_error_param(s, nargscheck, nretcheck, nargs, nret);
return false;
}
return true;
}
static void prop_format(char *tval, int tlen, const void *prop, int len)
{
int i;
const unsigned char *c;
char *t;
const char bin[] = "...";
for (i = 0, c = prop; i < len; ++i, ++c) {
if (*c == '\0' && i == len - 1) {
strncpy(tval, prop, tlen - 1);
return;
}
if (*c < 0x20 || *c >= 0x80) {
break;
}
}
for (i = 0, c = prop, t = tval; i < len; ++i, ++c) {
if (t >= tval + tlen - sizeof(bin) - 1 - 2 - 1) {
strcpy(t, bin);
return;
}
if (i && i % 4 == 0 && i != len - 1) {
strcat(t, " ");
++t;
}
t += sprintf(t, "%02X", *c & 0xFF);
}
}
static int get_path(const void *fdt, int offset, char *buf, int len)
{
int ret;
ret = fdt_get_path(fdt, offset, buf, len - 1);
if (ret < 0) {
return ret;
}
buf[len - 1] = '\0';
return strlen(buf) + 1;
}
static int phandle_to_path(const void *fdt, uint32_t ph, char *buf, int len)
{
int ret;
ret = fdt_node_offset_by_phandle(fdt, ph);
if (ret < 0) {
return ret;
}
return get_path(fdt, ret, buf, len);
}
static int path_offset(const void *fdt, const char *path)
{
g_autofree char *p = NULL;
char *at;
/*
* https://www.devicetree.org/open-firmware/bindings/ppc/release/ppc-2_1.html#HDR16
*
* "Conversion from numeric representation to text representation shall use
* the lower case forms of the hexadecimal digits in the range a..f,
* suppressing leading zeros".
*/
p = g_strdup(path);
for (at = strchr(p, '@'); at && *at; ) {
if (*at == '/') {
at = strchr(at, '@');
} else {
*at = tolower(*at);
++at;
}
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
return fdt_path_offset(fdt, p);
}
static uint32_t vof_finddevice(const void *fdt, uint32_t nodeaddr)
{
char fullnode[VOF_MAX_PATH];
uint32_t ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
int offset;
if (readstr(nodeaddr, fullnode, sizeof(fullnode))) {
return (uint32_t) ret;
}
offset = path_offset(fdt, fullnode);
if (offset >= 0) {
ret = fdt_get_phandle(fdt, offset);
}
trace_vof_finddevice(fullnode, ret);
return ret;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
static const void *getprop(const void *fdt, int nodeoff, const char *propname,
int *proplen, bool *write0)
{
const char *unit, *prop;
const void *ret = fdt_getprop(fdt, nodeoff, propname, proplen);
if (ret) {
if (write0) {
*write0 = false;
}
return ret;
}
if (strcmp(propname, "name")) {
return NULL;
}
/*
* We return a value for "name" from path if queried but property does not
* exist. @proplen does not include the unit part in this case.
*/
prop = fdt_get_name(fdt, nodeoff, proplen);
if (!prop) {
*proplen = 0;
return NULL;
}
unit = memchr(prop, '@', *proplen);
if (unit) {
*proplen = unit - prop;
}
*proplen += 1;
/*
* Since it might be cut at "@" and there will be no trailing zero
* in the prop buffer, tell the caller to write zero at the end.
*/
if (write0) {
*write0 = true;
}
return prop;
}
static uint32_t vof_getprop(const void *fdt, uint32_t nodeph, uint32_t pname,
uint32_t valaddr, uint32_t vallen)
{
char propname[OF_PROPNAME_LEN_MAX + 1];
uint32_t ret = 0;
int proplen = 0;
const void *prop;
char trval[64] = "";
int nodeoff = fdt_node_offset_by_phandle(fdt, nodeph);
bool write0;
if (nodeoff < 0) {
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
if (readstr(pname, propname, sizeof(propname))) {
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
prop = getprop(fdt, nodeoff, propname, &proplen, &write0);
if (prop) {
const char zero = 0;
int cb = MIN(proplen, vallen);
if (VOF_MEM_WRITE(valaddr, prop, cb) != MEMTX_OK ||
/* if that was "name" with a unit address, overwrite '@' with '0' */
(write0 &&
cb == proplen &&
VOF_MEM_WRITE(valaddr + cb - 1, &zero, 1) != MEMTX_OK)) {
ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
} else {
/*
* OF1275 says:
* "Size is either the actual size of the property, or -1 if name
* does not exist", hence returning proplen instead of cb.
*/
ret = proplen;
/* Do not format a value if tracepoint is silent, for performance */
if (trace_event_get_state(TRACE_VOF_GETPROP) &&
qemu_loglevel_mask(LOG_TRACE)) {
prop_format(trval, sizeof(trval), prop, ret);
}
}
} else {
ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
trace_vof_getprop(nodeph, propname, ret, trval);
return ret;
}
static uint32_t vof_getproplen(const void *fdt, uint32_t nodeph, uint32_t pname)
{
char propname[OF_PROPNAME_LEN_MAX + 1];
uint32_t ret = 0;
int proplen = 0;
const void *prop;
int nodeoff = fdt_node_offset_by_phandle(fdt, nodeph);
if (nodeoff < 0) {
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
if (readstr(pname, propname, sizeof(propname))) {
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
prop = getprop(fdt, nodeoff, propname, &proplen, NULL);
if (prop) {
ret = proplen;
} else {
ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
trace_vof_getproplen(nodeph, propname, ret);
return ret;
}
static uint32_t vof_setprop(MachineState *ms, void *fdt, Vof *vof,
uint32_t nodeph, uint32_t pname,
uint32_t valaddr, uint32_t vallen)
{
char propname[OF_PROPNAME_LEN_MAX + 1] = "";
uint32_t ret = PROM_ERROR;
int offset, rc;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
char trval[64] = "";
char nodepath[VOF_MAX_PATH] = "";
Object *vmo = object_dynamic_cast(OBJECT(ms), TYPE_VOF_MACHINE_IF);
VofMachineIfClass *vmc;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
g_autofree char *val = NULL;
if (vallen > VOF_MAX_SETPROPLEN) {
goto trace_exit;
}
if (readstr(pname, propname, sizeof(propname))) {
goto trace_exit;
}
offset = fdt_node_offset_by_phandle(fdt, nodeph);
if (offset < 0) {
goto trace_exit;
}
rc = get_path(fdt, offset, nodepath, sizeof(nodepath));
if (rc <= 0) {
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
goto trace_exit;
}
val = g_malloc0(vallen);
if (VOF_MEM_READ(valaddr, val, vallen) != MEMTX_OK) {
goto trace_exit;
}
if (!vmo) {
goto trace_exit;
}
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
vmc = VOF_MACHINE_GET_CLASS(vmo);
if (!vmc->setprop || !vmc->setprop(ms, nodepath, propname, val, vallen)) {
goto trace_exit;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
rc = fdt_setprop(fdt, offset, propname, val, vallen);
if (rc) {
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
goto trace_exit;
}
if (trace_event_get_state(TRACE_VOF_SETPROP) &&
qemu_loglevel_mask(LOG_TRACE)) {
prop_format(trval, sizeof(trval), val, vallen);
}
ret = vallen;
trace_exit:
trace_vof_setprop(nodeph, propname, trval, vallen, ret);
return ret;
}
static uint32_t vof_nextprop(const void *fdt, uint32_t phandle,
uint32_t prevaddr, uint32_t nameaddr)
{
int offset, nodeoff = fdt_node_offset_by_phandle(fdt, phandle);
char prev[OF_PROPNAME_LEN_MAX + 1];
const char *tmp;
if (readstr(prevaddr, prev, sizeof(prev))) {
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
fdt_for_each_property_offset(offset, fdt, nodeoff) {
if (!fdt_getprop_by_offset(fdt, offset, &tmp, NULL)) {
return 0;
}
if (prev[0] == '\0' || strcmp(prev, tmp) == 0) {
if (prev[0] != '\0') {
offset = fdt_next_property_offset(fdt, offset);
if (offset < 0) {
return 0;
}
}
if (!fdt_getprop_by_offset(fdt, offset, &tmp, NULL)) {
return 0;
}
if (VOF_MEM_WRITE(nameaddr, tmp, strlen(tmp) + 1) != MEMTX_OK) {
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
return 1;
}
}
return 0;
}
static uint32_t vof_peer(const void *fdt, uint32_t phandle)
{
uint32_t ret = 0;
int rc;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
if (phandle == 0) {
rc = fdt_path_offset(fdt, "/");
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
} else {
rc = fdt_next_subnode(fdt, fdt_node_offset_by_phandle(fdt, phandle));
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
if (rc >= 0) {
ret = fdt_get_phandle(fdt, rc);
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
return ret;
}
static uint32_t vof_child(const void *fdt, uint32_t phandle)
{
uint32_t ret = 0;
int rc = fdt_first_subnode(fdt, fdt_node_offset_by_phandle(fdt, phandle));
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
if (rc >= 0) {
ret = fdt_get_phandle(fdt, rc);
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
return ret;
}
static uint32_t vof_parent(const void *fdt, uint32_t phandle)
{
uint32_t ret = 0;
int rc = fdt_parent_offset(fdt, fdt_node_offset_by_phandle(fdt, phandle));
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
if (rc >= 0) {
ret = fdt_get_phandle(fdt, rc);
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
return ret;
}
static uint32_t vof_do_open(void *fdt, Vof *vof, int offset, const char *path)
{
uint32_t ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
OfInstance *inst = NULL;
if (vof->of_instance_last == 0xFFFFFFFF) {
/* We do not recycle ihandles yet */
goto trace_exit;
}
inst = g_new0(OfInstance, 1);
inst->phandle = fdt_get_phandle(fdt, offset);
g_assert(inst->phandle);
++vof->of_instance_last;
inst->path = g_strdup(path);
g_hash_table_insert(vof->of_instances,
GINT_TO_POINTER(vof->of_instance_last),
inst);
ret = vof->of_instance_last;
trace_exit:
trace_vof_open(path, inst ? inst->phandle : 0, ret);
return ret;
}
uint32_t vof_client_open_store(void *fdt, Vof *vof, const char *nodename,
const char *prop, const char *path)
{
int offset, node = fdt_path_offset(fdt, nodename);
uint32_t inst;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
offset = fdt_path_offset(fdt, path);
if (offset < 0) {
trace_vof_error_unknown_path(path);
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
inst = vof_do_open(fdt, vof, offset, path);
return fdt_setprop_cell(fdt, node, prop, inst) >= 0 ? 0 : PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
static uint32_t vof_open(void *fdt, Vof *vof, uint32_t pathaddr)
{
char path[VOF_MAX_PATH];
int offset;
if (readstr(pathaddr, path, sizeof(path))) {
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
offset = path_offset(fdt, path);
if (offset < 0) {
trace_vof_error_unknown_path(path);
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
return vof_do_open(fdt, vof, offset, path);
}
static void vof_close(Vof *vof, uint32_t ihandle)
{
if (!g_hash_table_remove(vof->of_instances, GINT_TO_POINTER(ihandle))) {
trace_vof_error_unknown_ihandle_close(ihandle);
}
}
static uint32_t vof_instance_to_package(Vof *vof, uint32_t ihandle)
{
gpointer instp = g_hash_table_lookup(vof->of_instances,
GINT_TO_POINTER(ihandle));
uint32_t ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
if (instp) {
ret = ((OfInstance *)instp)->phandle;
}
trace_vof_instance_to_package(ihandle, ret);
return ret;
}
static uint32_t vof_package_to_path(const void *fdt, uint32_t phandle,
uint32_t buf, uint32_t len)
{
int rc;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
char tmp[VOF_MAX_PATH] = "";
rc = phandle_to_path(fdt, phandle, tmp, sizeof(tmp));
if (rc > 0) {
if (VOF_MEM_WRITE(buf, tmp, rc) != MEMTX_OK) {
rc = -1;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
}
trace_vof_package_to_path(phandle, tmp, rc);
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
return rc > 0 ? (uint32_t)rc : PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
static uint32_t vof_instance_to_path(void *fdt, Vof *vof, uint32_t ihandle,
uint32_t buf, uint32_t len)
{
int rc = -1;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
uint32_t phandle = vof_instance_to_package(vof, ihandle);
char tmp[VOF_MAX_PATH] = "";
if (phandle != -1) {
rc = phandle_to_path(fdt, phandle, tmp, sizeof(tmp));
if (rc > 0) {
if (VOF_MEM_WRITE(buf, tmp, rc) != MEMTX_OK) {
rc = -1;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
}
}
trace_vof_instance_to_path(ihandle, phandle, tmp, rc);
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
return rc > 0 ? (uint32_t)rc : PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
static uint32_t vof_write(Vof *vof, uint32_t ihandle, uint32_t buf,
uint32_t len)
{
char tmp[VOF_VTY_BUF_SIZE];
unsigned cb;
OfInstance *inst = (OfInstance *)
g_hash_table_lookup(vof->of_instances, GINT_TO_POINTER(ihandle));
if (!inst) {
trace_vof_error_write(ihandle);
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
for ( ; len > 0; len -= cb) {
cb = MIN(len, sizeof(tmp) - 1);
if (VOF_MEM_READ(buf, tmp, cb) != MEMTX_OK) {
return PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
/* FIXME: there is no backend(s) yet so just call a trace */
if (trace_event_get_state(TRACE_VOF_WRITE) &&
qemu_loglevel_mask(LOG_TRACE)) {
tmp[cb] = '\0';
trace_vof_write(ihandle, cb, tmp);
}
}
return len;
}
static void vof_claimed_dump(GArray *claimed)
{
int i;
OfClaimed c;
if (trace_event_get_state(TRACE_VOF_CLAIMED) &&
qemu_loglevel_mask(LOG_TRACE)) {
for (i = 0; i < claimed->len; ++i) {
c = g_array_index(claimed, OfClaimed, i);
trace_vof_claimed(c.start, c.start + c.size, c.size);
}
}
}
static bool vof_claim_avail(GArray *claimed, uint64_t virt, uint64_t size)
{
int i;
OfClaimed c;
for (i = 0; i < claimed->len; ++i) {
c = g_array_index(claimed, OfClaimed, i);
if (ranges_overlap(c.start, c.size, virt, size)) {
return false;
}
}
return true;
}
static void vof_claim_add(GArray *claimed, uint64_t virt, uint64_t size)
{
OfClaimed newclaim;
newclaim.start = virt;
newclaim.size = size;
g_array_append_val(claimed, newclaim);
}
static gint of_claimed_compare_func(gconstpointer a, gconstpointer b)
{
return ((OfClaimed *)a)->start - ((OfClaimed *)b)->start;
}
static void vof_dt_memory_available(void *fdt, GArray *claimed, uint64_t base)
{
int i, n, offset, proplen = 0, sc, ac;
target_ulong mem0_end;
const uint8_t *mem0_reg;
g_autofree uint8_t *avail = NULL;
uint8_t *availcur;
if (!fdt || !claimed) {
return;
}
offset = fdt_path_offset(fdt, "/");
_FDT(offset);
ac = fdt_address_cells(fdt, offset);
g_assert(ac == 1 || ac == 2);
sc = fdt_size_cells(fdt, offset);
g_assert(sc == 1 || sc == 2);
offset = fdt_path_offset(fdt, "/memory@0");
_FDT(offset);
mem0_reg = fdt_getprop(fdt, offset, "reg", &proplen);
g_assert(mem0_reg && proplen == sizeof(uint32_t) * (ac + sc));
if (sc == 2) {
mem0_end = be64_to_cpu(*(uint64_t *)(mem0_reg + sizeof(uint32_t) * ac));
} else {
mem0_end = be32_to_cpu(*(uint32_t *)(mem0_reg + sizeof(uint32_t) * ac));
}
g_array_sort(claimed, of_claimed_compare_func);
vof_claimed_dump(claimed);
/*
* VOF resides in the first page so we do not need to check if there is
* available memory before the first claimed block
*/
g_assert(claimed->len && (g_array_index(claimed, OfClaimed, 0).start == 0));
avail = g_malloc0(sizeof(uint32_t) * (ac + sc) * claimed->len);
for (i = 0, n = 0, availcur = avail; i < claimed->len; ++i) {
OfClaimed c = g_array_index(claimed, OfClaimed, i);
uint64_t start, size;
start = c.start + c.size;
if (i < claimed->len - 1) {
OfClaimed cn = g_array_index(claimed, OfClaimed, i + 1);
size = cn.start - start;
} else {
size = mem0_end - start;
}
if (ac == 2) {
*(uint64_t *) availcur = cpu_to_be64(start);
} else {
*(uint32_t *) availcur = cpu_to_be32(start);
}
availcur += sizeof(uint32_t) * ac;
if (sc == 2) {
*(uint64_t *) availcur = cpu_to_be64(size);
} else {
*(uint32_t *) availcur = cpu_to_be32(size);
}
availcur += sizeof(uint32_t) * sc;
if (size) {
trace_vof_avail(c.start + c.size, c.start + c.size + size, size);
++n;
}
}
_FDT((fdt_setprop(fdt, offset, "available", avail, availcur - avail)));
}
/*
* OF1275:
* "Allocates size bytes of memory. If align is zero, the allocated range
* begins at the virtual address virt. Otherwise, an aligned address is
* automatically chosen and the input argument virt is ignored".
*
* In other words, exactly one of @virt and @align is non-zero.
*/
uint64_t vof_claim(Vof *vof, uint64_t virt, uint64_t size,
uint64_t align)
{
uint64_t ret;
if (size == 0) {
ret = -1;
} else if (align == 0) {
if (!vof_claim_avail(vof->claimed, virt, size)) {
ret = -1;
} else {
ret = virt;
}
} else {
vof->claimed_base = QEMU_ALIGN_UP(vof->claimed_base, align);
while (1) {
if (vof->claimed_base >= vof->top_addr) {
error_report("Out of RMA memory for the OF client");
return -1;
}
if (vof_claim_avail(vof->claimed, vof->claimed_base, size)) {
break;
}
vof->claimed_base += size;
}
ret = vof->claimed_base;
}
if (ret != -1) {
vof->claimed_base = MAX(vof->claimed_base, ret + size);
vof_claim_add(vof->claimed, ret, size);
}
trace_vof_claim(virt, size, align, ret);
return ret;
}
static uint32_t vof_release(Vof *vof, uint64_t virt, uint64_t size)
{
uint32_t ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
int i;
GArray *claimed = vof->claimed;
OfClaimed c;
for (i = 0; i < claimed->len; ++i) {
c = g_array_index(claimed, OfClaimed, i);
if (c.start == virt && c.size == size) {
g_array_remove_index(claimed, i);
ret = 0;
break;
}
}
trace_vof_release(virt, size, ret);
return ret;
}
static void vof_instantiate_rtas(Error **errp)
{
error_setg(errp, "The firmware should have instantiated RTAS");
}
static uint32_t vof_call_method(MachineState *ms, Vof *vof, uint32_t methodaddr,
uint32_t ihandle, uint32_t param1,
uint32_t param2, uint32_t param3,
uint32_t param4, uint32_t *ret2)
{
uint32_t ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
char method[VOF_MAX_METHODLEN] = "";
OfInstance *inst;
if (!ihandle) {
goto trace_exit;
}
inst = (OfInstance *)g_hash_table_lookup(vof->of_instances,
GINT_TO_POINTER(ihandle));
if (!inst) {
goto trace_exit;
}
if (readstr(methodaddr, method, sizeof(method))) {
goto trace_exit;
}
if (strcmp(inst->path, "/") == 0) {
if (strcmp(method, "ibm,client-architecture-support") == 0) {
Object *vmo = object_dynamic_cast(OBJECT(ms), TYPE_VOF_MACHINE_IF);
if (vmo) {
VofMachineIfClass *vmc = VOF_MACHINE_GET_CLASS(vmo);
g_assert(vmc->client_architecture_support);
ret = (uint32_t)vmc->client_architecture_support(ms, first_cpu,
param1);
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
*ret2 = 0;
}
} else if (strcmp(inst->path, "/rtas") == 0) {
if (strcmp(method, "instantiate-rtas") == 0) {
vof_instantiate_rtas(&error_fatal);
ret = 0;
*ret2 = param1; /* rtas-base */
}
} else {
trace_vof_error_unknown_method(method);
}
trace_exit:
trace_vof_method(ihandle, method, param1, ret, *ret2);
return ret;
}
static uint32_t vof_call_interpret(uint32_t cmdaddr, uint32_t param1,
uint32_t param2, uint32_t *ret2)
{
uint32_t ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
char cmd[VOF_MAX_FORTHCODE] = "";
/* No interpret implemented so just call a trace */
readstr(cmdaddr, cmd, sizeof(cmd));
trace_vof_interpret(cmd, param1, param2, ret, *ret2);
return ret;
}
static void vof_quiesce(MachineState *ms, void *fdt, Vof *vof)
{
Object *vmo = object_dynamic_cast(OBJECT(ms), TYPE_VOF_MACHINE_IF);
/* After "quiesce", no change is expected to the FDT, pack FDT to ensure */
int rc = fdt_pack(fdt);
assert(rc == 0);
if (vmo) {
VofMachineIfClass *vmc = VOF_MACHINE_GET_CLASS(vmo);
if (vmc->quiesce) {
vmc->quiesce(ms);
}
}
vof_claimed_dump(vof->claimed);
}
static uint32_t vof_client_handle(MachineState *ms, void *fdt, Vof *vof,
const char *service,
uint32_t *args, unsigned nargs,
uint32_t *rets, unsigned nrets)
{
uint32_t ret = 0;
/* @nrets includes the value which this function returns */
#define cmpserv(s, a, r) \
cmpservice(service, nargs, nrets, (s), (a), (r))
if (cmpserv("finddevice", 1, 1)) {
ret = vof_finddevice(fdt, args[0]);
} else if (cmpserv("getprop", 4, 1)) {
ret = vof_getprop(fdt, args[0], args[1], args[2], args[3]);
} else if (cmpserv("getproplen", 2, 1)) {
ret = vof_getproplen(fdt, args[0], args[1]);
} else if (cmpserv("setprop", 4, 1)) {
ret = vof_setprop(ms, fdt, vof, args[0], args[1], args[2], args[3]);
} else if (cmpserv("nextprop", 3, 1)) {
ret = vof_nextprop(fdt, args[0], args[1], args[2]);
} else if (cmpserv("peer", 1, 1)) {
ret = vof_peer(fdt, args[0]);
} else if (cmpserv("child", 1, 1)) {
ret = vof_child(fdt, args[0]);
} else if (cmpserv("parent", 1, 1)) {
ret = vof_parent(fdt, args[0]);
} else if (cmpserv("open", 1, 1)) {
ret = vof_open(fdt, vof, args[0]);
} else if (cmpserv("close", 1, 0)) {
vof_close(vof, args[0]);
} else if (cmpserv("instance-to-package", 1, 1)) {
ret = vof_instance_to_package(vof, args[0]);
} else if (cmpserv("package-to-path", 3, 1)) {
ret = vof_package_to_path(fdt, args[0], args[1], args[2]);
} else if (cmpserv("instance-to-path", 3, 1)) {
ret = vof_instance_to_path(fdt, vof, args[0], args[1], args[2]);
} else if (cmpserv("write", 3, 1)) {
ret = vof_write(vof, args[0], args[1], args[2]);
} else if (cmpserv("claim", 3, 1)) {
uint64_t ret64 = vof_claim(vof, args[0], args[1], args[2]);
if (ret64 < 0x100000000UL) {
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
vof_dt_memory_available(fdt, vof->claimed, vof->claimed_base);
ret = (uint32_t)ret64;
} else {
if (ret64 != -1) {
vof_release(vof, ret, args[1]);
}
ret = PROM_ERROR;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
} else if (cmpserv("release", 2, 0)) {
ret = vof_release(vof, args[0], args[1]);
if (ret != PROM_ERROR) {
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
vof_dt_memory_available(fdt, vof->claimed, vof->claimed_base);
}
} else if (cmpserv("call-method", 0, 0)) {
ret = vof_call_method(ms, vof, args[0], args[1], args[2], args[3],
args[4], args[5], rets);
} else if (cmpserv("interpret", 0, 0)) {
ret = vof_call_interpret(args[0], args[1], args[2], rets);
} else if (cmpserv("milliseconds", 0, 1)) {
ret = qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL);
} else if (cmpserv("quiesce", 0, 0)) {
vof_quiesce(ms, fdt, vof);
} else if (cmpserv("exit", 0, 0)) {
error_report("Stopped as the VM requested \"exit\"");
vm_stop(RUN_STATE_PAUSED);
} else {
trace_vof_error_unknown_service(service, nargs, nrets);
ret = -1;
}
#undef cmpserv
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
return ret;
}
/* Defined as Big Endian */
struct prom_args {
uint32_t service;
uint32_t nargs;
uint32_t nret;
uint32_t args[10];
} QEMU_PACKED;
int vof_client_call(MachineState *ms, Vof *vof, void *fdt,
target_ulong args_real)
{
struct prom_args args_be;
uint32_t args[ARRAY_SIZE(args_be.args)];
uint32_t rets[ARRAY_SIZE(args_be.args)] = { 0 }, ret;
char service[64];
unsigned nargs, nret, i;
if (VOF_MEM_READ(args_real, &args_be, sizeof(args_be)) != MEMTX_OK) {
return -EINVAL;
}
nargs = be32_to_cpu(args_be.nargs);
if (nargs >= ARRAY_SIZE(args_be.args)) {
return -EINVAL;
}
if (VOF_MEM_READ(be32_to_cpu(args_be.service), service, sizeof(service)) !=
MEMTX_OK) {
return -EINVAL;
}
if (strnlen(service, sizeof(service)) == sizeof(service)) {
/* Too long service name */
return -EINVAL;
}
for (i = 0; i < nargs; ++i) {
args[i] = be32_to_cpu(args_be.args[i]);
}
nret = be32_to_cpu(args_be.nret);
if (nret > ARRAY_SIZE(args_be.args) - nargs) {
return -EINVAL;
}
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
ret = vof_client_handle(ms, fdt, vof, service, args, nargs, rets, nret);
if (!nret) {
return 0;
}
/* @nrets includes the value which this function returns */
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
args_be.args[nargs] = cpu_to_be32(ret);
for (i = 1; i < nret; ++i) {
args_be.args[nargs + i] = cpu_to_be32(rets[i - 1]);
}
if (VOF_MEM_WRITE(args_real + offsetof(struct prom_args, args[nargs]),
args_be.args + nargs, sizeof(args_be.args[0]) * nret) !=
MEMTX_OK) {
return -EINVAL;
}
return 0;
}
static void vof_instance_free(gpointer data)
{
OfInstance *inst = (OfInstance *)data;
g_free(inst->path);
g_free(inst);
}
void vof_init(Vof *vof, uint64_t top_addr, Error **errp)
{
vof_cleanup(vof);
vof->of_instances = g_hash_table_new_full(g_direct_hash, g_direct_equal,
NULL, vof_instance_free);
vof->claimed = g_array_new(false, false, sizeof(OfClaimed));
/* Keep allocations in 32bit as CLI ABI can only return cells==32bit */
vof->top_addr = MIN(top_addr, 4 * GiB);
if (vof_claim(vof, 0, vof->fw_size, 0) == -1) {
error_setg(errp, "Memory for firmware is in use");
}
}
void vof_cleanup(Vof *vof)
{
if (vof->claimed) {
g_array_unref(vof->claimed);
}
if (vof->of_instances) {
g_hash_table_unref(vof->of_instances);
}
vof->claimed = NULL;
vof->of_instances = NULL;
vof->of_instance_last = 0;
vof->claimed_base = 0;
spapr: Implement Open Firmware client interface The PAPR platform describes an OS environment that's presented by a combination of a hypervisor and firmware. The features it specifies require collaboration between the firmware and the hypervisor. Since the beginning, the runtime component of the firmware (RTAS) has been implemented as a 20 byte shim which simply forwards it to a hypercall implemented in qemu. The boot time firmware component is SLOF - but a build that's specific to qemu, and has always needed to be updated in sync with it. Even though we've managed to limit the amount of runtime communication we need between qemu and SLOF, there's some, and it has become increasingly awkward to handle as we've implemented new features. This implements a boot time OF client interface (CI) which is enabled by a new "x-vof" pseries machine option (stands for "Virtual Open Firmware). When enabled, QEMU implements the custom H_OF_CLIENT hcall which implements Open Firmware Client Interface (OF CI). This allows using a smaller stateless firmware which does not have to manage the device tree. The new "vof.bin" firmware image is included with source code under pc-bios/. It also includes RTAS blob. This implements a handful of CI methods just to get -kernel/-initrd working. In particular, this implements the device tree fetching and simple memory allocator - "claim" (an OF CI memory allocator) and updates "/memory@0/available" to report the client about available memory. This implements changing some device tree properties which we know how to deal with, the rest is ignored. To allow changes, this skips fdt_pack() when x-vof=on as not packing the blob leaves some room for appending. In absence of SLOF, this assigns phandles to device tree nodes to make device tree traversing work. When x-vof=on, this adds "/chosen" every time QEMU (re)builds a tree. This adds basic instances support which are managed by a hash map ihandle -> [phandle]. Before the guest started, the used memory is: 0..e60 - the initial firmware 8000..10000 - stack 400000.. - kernel 3ea0000.. - initramdisk This OF CI does not implement "interpret". Unlike SLOF, this does not format uninitialized nvram. Instead, this includes a disk image with pre-formatted nvram. With this basic support, this can only boot into kernel directly. However this is just enough for the petitboot kernel and initradmdisk to boot from any possible source. Note this requires reasonably recent guest kernel with: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=df5be5be8735 The immediate benefit is much faster booting time which especially crucial with fully emulated early CPU bring up environments. Also this may come handy when/if GRUB-in-the-userspace sees light of the day. This separates VOF and sPAPR in a hope that VOF bits may be reused by other POWERPC boards which do not support pSeries. This assumes potential support for booting from QEMU backends such as blockdev or netdev without devices/drivers used. Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru> Message-Id: <20210625055155.2252896-1-aik@ozlabs.ru> Reviewed-by: BALATON Zoltan <balaton@eik.bme.hu> [dwg: Adjusted some includes which broke compile in some more obscure compilation setups] Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
2021-06-25 07:51:55 +02:00
}
void vof_build_dt(void *fdt, Vof *vof)
{
uint32_t phandle = fdt_get_max_phandle(fdt);
int offset, proplen = 0;
const void *prop;
/* Assign phandles to nodes without predefined phandles (like XICS/XIVE) */
for (offset = fdt_next_node(fdt, -1, NULL);
offset >= 0;
offset = fdt_next_node(fdt, offset, NULL)) {
prop = fdt_getprop(fdt, offset, "phandle", &proplen);
if (prop) {
continue;
}
++phandle;
_FDT(fdt_setprop_cell(fdt, offset, "phandle", phandle));
}
vof_dt_memory_available(fdt, vof->claimed, vof->claimed_base);
}
static const TypeInfo vof_machine_if_info = {
.name = TYPE_VOF_MACHINE_IF,
.parent = TYPE_INTERFACE,
.class_size = sizeof(VofMachineIfClass),
};
static void vof_machine_if_register_types(void)
{
type_register_static(&vof_machine_if_info);
}
type_init(vof_machine_if_register_types)