2008-07-30 07:33:58 +02:00
|
|
|
/*
|
|
|
|
* SN Platform GRU Driver
|
|
|
|
*
|
|
|
|
* DRIVER TABLE MANAGER + GRU CONTEXT LOAD/UNLOAD
|
|
|
|
*
|
2009-06-18 01:28:36 +02:00
|
|
|
* Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved.
|
2008-07-30 07:33:58 +02:00
|
|
|
*
|
2009-06-18 01:28:36 +02:00
|
|
|
* This program is free software; you can redistribute it and/or modify
|
|
|
|
* it under the terms of the GNU General Public License as published by
|
|
|
|
* the Free Software Foundation; either version 2 of the License, or
|
|
|
|
* (at your option) any later version.
|
|
|
|
*
|
|
|
|
* This program is distributed in the hope that it will be useful,
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
|
|
* GNU General Public License for more details.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU General Public License
|
|
|
|
* along with this program; if not, write to the Free Software
|
|
|
|
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
|
2008-07-30 07:33:58 +02:00
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/kernel.h>
|
|
|
|
#include <linux/slab.h>
|
|
|
|
#include <linux/mm.h>
|
|
|
|
#include <linux/spinlock.h>
|
|
|
|
#include <linux/sched.h>
|
|
|
|
#include <linux/device.h>
|
|
|
|
#include <linux/list.h>
|
2009-12-16 01:48:08 +01:00
|
|
|
#include <linux/err.h>
|
2008-07-30 07:33:58 +02:00
|
|
|
#include <asm/uv/uv_hub.h>
|
|
|
|
#include "gru.h"
|
|
|
|
#include "grutables.h"
|
|
|
|
#include "gruhandles.h"
|
|
|
|
|
2008-07-30 07:34:02 +02:00
|
|
|
unsigned long gru_options __read_mostly;
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
static struct device_driver gru_driver = {
|
|
|
|
.name = "gru"
|
|
|
|
};
|
|
|
|
|
|
|
|
static struct device gru_device = {
|
2009-01-06 19:44:37 +01:00
|
|
|
.init_name = "",
|
2008-07-30 07:33:58 +02:00
|
|
|
.driver = &gru_driver,
|
|
|
|
};
|
|
|
|
|
|
|
|
struct device *grudev = &gru_device;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Select a gru fault map to be used by the current cpu. Note that
|
|
|
|
* multiple cpus may be using the same map.
|
|
|
|
* ZZZ should be inline but did not work on emulator
|
|
|
|
*/
|
|
|
|
int gru_cpu_fault_map_id(void)
|
|
|
|
{
|
2009-12-16 01:48:15 +01:00
|
|
|
#ifdef CONFIG_IA64
|
|
|
|
return uv_blade_processor_id() % GRU_NUM_TFM;
|
|
|
|
#else
|
2009-12-16 01:48:11 +01:00
|
|
|
int cpu = smp_processor_id();
|
|
|
|
int id, core;
|
|
|
|
|
|
|
|
core = uv_cpu_core_number(cpu);
|
|
|
|
id = core + UV_MAX_INT_CORES * uv_cpu_socket_number(cpu);
|
|
|
|
return id;
|
2009-12-16 01:48:15 +01:00
|
|
|
#endif
|
2008-07-30 07:33:58 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/*--------- ASID Management -------------------------------------------
|
|
|
|
*
|
|
|
|
* Initially, assign asids sequentially from MIN_ASID .. MAX_ASID.
|
|
|
|
* Once MAX is reached, flush the TLB & start over. However,
|
|
|
|
* some asids may still be in use. There won't be many (percentage wise) still
|
|
|
|
* in use. Search active contexts & determine the value of the first
|
|
|
|
* asid in use ("x"s below). Set "limit" to this value.
|
|
|
|
* This defines a block of assignable asids.
|
|
|
|
*
|
|
|
|
* When "limit" is reached, search forward from limit+1 and determine the
|
|
|
|
* next block of assignable asids.
|
|
|
|
*
|
|
|
|
* Repeat until MAX_ASID is reached, then start over again.
|
|
|
|
*
|
|
|
|
* Each time MAX_ASID is reached, increment the asid generation. Since
|
|
|
|
* the search for in-use asids only checks contexts with GRUs currently
|
|
|
|
* assigned, asids in some contexts will be missed. Prior to loading
|
|
|
|
* a context, the asid generation of the GTS asid is rechecked. If it
|
|
|
|
* doesn't match the current generation, a new asid will be assigned.
|
|
|
|
*
|
|
|
|
* 0---------------x------------x---------------------x----|
|
|
|
|
* ^-next ^-limit ^-MAX_ASID
|
|
|
|
*
|
|
|
|
* All asid manipulation & context loading/unloading is protected by the
|
|
|
|
* gs_lock.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/* Hit the asid limit. Start over */
|
|
|
|
static int gru_wrap_asid(struct gru_state *gru)
|
|
|
|
{
|
2009-04-03 01:59:05 +02:00
|
|
|
gru_dbg(grudev, "gid %d\n", gru->gs_gid);
|
2008-07-30 07:33:58 +02:00
|
|
|
STAT(asid_wrap);
|
|
|
|
gru->gs_asid_gen++;
|
|
|
|
return MIN_ASID;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Find the next chunk of unused asids */
|
|
|
|
static int gru_reset_asid_limit(struct gru_state *gru, int asid)
|
|
|
|
{
|
|
|
|
int i, gid, inuse_asid, limit;
|
|
|
|
|
2009-04-03 01:59:05 +02:00
|
|
|
gru_dbg(grudev, "gid %d, asid 0x%x\n", gru->gs_gid, asid);
|
2008-07-30 07:33:58 +02:00
|
|
|
STAT(asid_next);
|
|
|
|
limit = MAX_ASID;
|
|
|
|
if (asid >= limit)
|
|
|
|
asid = gru_wrap_asid(gru);
|
2009-04-03 01:59:08 +02:00
|
|
|
gru_flush_all_tlb(gru);
|
2008-07-30 07:33:58 +02:00
|
|
|
gid = gru->gs_gid;
|
|
|
|
again:
|
|
|
|
for (i = 0; i < GRU_NUM_CCH; i++) {
|
2009-06-18 01:28:22 +02:00
|
|
|
if (!gru->gs_gts[i] || is_kernel_context(gru->gs_gts[i]))
|
2008-07-30 07:33:58 +02:00
|
|
|
continue;
|
|
|
|
inuse_asid = gru->gs_gts[i]->ts_gms->ms_asids[gid].mt_asid;
|
2009-04-03 01:59:05 +02:00
|
|
|
gru_dbg(grudev, "gid %d, gts %p, gms %p, inuse 0x%x, cxt %d\n",
|
|
|
|
gru->gs_gid, gru->gs_gts[i], gru->gs_gts[i]->ts_gms,
|
|
|
|
inuse_asid, i);
|
2008-07-30 07:33:58 +02:00
|
|
|
if (inuse_asid == asid) {
|
|
|
|
asid += ASID_INC;
|
|
|
|
if (asid >= limit) {
|
|
|
|
/*
|
|
|
|
* empty range: reset the range limit and
|
|
|
|
* start over
|
|
|
|
*/
|
|
|
|
limit = MAX_ASID;
|
|
|
|
if (asid >= MAX_ASID)
|
|
|
|
asid = gru_wrap_asid(gru);
|
|
|
|
goto again;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if ((inuse_asid > asid) && (inuse_asid < limit))
|
|
|
|
limit = inuse_asid;
|
|
|
|
}
|
|
|
|
gru->gs_asid_limit = limit;
|
|
|
|
gru->gs_asid = asid;
|
2009-04-03 01:59:05 +02:00
|
|
|
gru_dbg(grudev, "gid %d, new asid 0x%x, new_limit 0x%x\n", gru->gs_gid,
|
|
|
|
asid, limit);
|
2008-07-30 07:33:58 +02:00
|
|
|
return asid;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Assign a new ASID to a thread context. */
|
|
|
|
static int gru_assign_asid(struct gru_state *gru)
|
|
|
|
{
|
|
|
|
int asid;
|
|
|
|
|
|
|
|
gru->gs_asid += ASID_INC;
|
|
|
|
asid = gru->gs_asid;
|
|
|
|
if (asid >= gru->gs_asid_limit)
|
|
|
|
asid = gru_reset_asid_limit(gru, asid);
|
|
|
|
|
2009-04-03 01:59:05 +02:00
|
|
|
gru_dbg(grudev, "gid %d, asid 0x%x\n", gru->gs_gid, asid);
|
2008-07-30 07:33:58 +02:00
|
|
|
return asid;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Clear n bits in a word. Return a word indicating the bits that were cleared.
|
|
|
|
* Optionally, build an array of chars that contain the bit numbers allocated.
|
|
|
|
*/
|
|
|
|
static unsigned long reserve_resources(unsigned long *p, int n, int mmax,
|
|
|
|
char *idx)
|
|
|
|
{
|
|
|
|
unsigned long bits = 0;
|
|
|
|
int i;
|
|
|
|
|
2009-06-18 01:28:23 +02:00
|
|
|
while (n--) {
|
2008-07-30 07:33:58 +02:00
|
|
|
i = find_first_bit(p, mmax);
|
|
|
|
if (i == mmax)
|
|
|
|
BUG();
|
|
|
|
__clear_bit(i, p);
|
|
|
|
__set_bit(i, &bits);
|
|
|
|
if (idx)
|
|
|
|
*idx++ = i;
|
2009-06-18 01:28:23 +02:00
|
|
|
}
|
2008-07-30 07:33:58 +02:00
|
|
|
return bits;
|
|
|
|
}
|
|
|
|
|
2008-07-30 07:34:02 +02:00
|
|
|
unsigned long gru_reserve_cb_resources(struct gru_state *gru, int cbr_au_count,
|
2008-07-30 07:33:58 +02:00
|
|
|
char *cbmap)
|
|
|
|
{
|
|
|
|
return reserve_resources(&gru->gs_cbr_map, cbr_au_count, GRU_CBR_AU,
|
|
|
|
cbmap);
|
|
|
|
}
|
|
|
|
|
2008-07-30 07:34:02 +02:00
|
|
|
unsigned long gru_reserve_ds_resources(struct gru_state *gru, int dsr_au_count,
|
2008-07-30 07:33:58 +02:00
|
|
|
char *dsmap)
|
|
|
|
{
|
|
|
|
return reserve_resources(&gru->gs_dsr_map, dsr_au_count, GRU_DSR_AU,
|
|
|
|
dsmap);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void reserve_gru_resources(struct gru_state *gru,
|
|
|
|
struct gru_thread_state *gts)
|
|
|
|
{
|
|
|
|
gru->gs_active_contexts++;
|
|
|
|
gts->ts_cbr_map =
|
2008-07-30 07:34:02 +02:00
|
|
|
gru_reserve_cb_resources(gru, gts->ts_cbr_au_count,
|
2008-07-30 07:33:58 +02:00
|
|
|
gts->ts_cbr_idx);
|
|
|
|
gts->ts_dsr_map =
|
2008-07-30 07:34:02 +02:00
|
|
|
gru_reserve_ds_resources(gru, gts->ts_dsr_au_count, NULL);
|
2008-07-30 07:33:58 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
static void free_gru_resources(struct gru_state *gru,
|
|
|
|
struct gru_thread_state *gts)
|
|
|
|
{
|
|
|
|
gru->gs_active_contexts--;
|
|
|
|
gru->gs_cbr_map |= gts->ts_cbr_map;
|
|
|
|
gru->gs_dsr_map |= gts->ts_dsr_map;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Check if a GRU has sufficient free resources to satisfy an allocation
|
|
|
|
* request. Note: GRU locks may or may not be held when this is called. If
|
|
|
|
* not held, recheck after acquiring the appropriate locks.
|
|
|
|
*
|
|
|
|
* Returns 1 if sufficient resources, 0 if not
|
|
|
|
*/
|
|
|
|
static int check_gru_resources(struct gru_state *gru, int cbr_au_count,
|
|
|
|
int dsr_au_count, int max_active_contexts)
|
|
|
|
{
|
|
|
|
return hweight64(gru->gs_cbr_map) >= cbr_au_count
|
|
|
|
&& hweight64(gru->gs_dsr_map) >= dsr_au_count
|
|
|
|
&& gru->gs_active_contexts < max_active_contexts;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* TLB manangment requires tracking all GRU chiplets that have loaded a GSEG
|
|
|
|
* context.
|
|
|
|
*/
|
2009-04-03 01:59:05 +02:00
|
|
|
static int gru_load_mm_tracker(struct gru_state *gru,
|
|
|
|
struct gru_thread_state *gts)
|
2008-07-30 07:33:58 +02:00
|
|
|
{
|
2009-04-03 01:59:05 +02:00
|
|
|
struct gru_mm_struct *gms = gts->ts_gms;
|
2008-07-30 07:33:58 +02:00
|
|
|
struct gru_mm_tracker *asids = &gms->ms_asids[gru->gs_gid];
|
2009-04-03 01:59:05 +02:00
|
|
|
unsigned short ctxbitmap = (1 << gts->ts_ctxnum);
|
2008-07-30 07:33:58 +02:00
|
|
|
int asid;
|
|
|
|
|
|
|
|
spin_lock(&gms->ms_asid_lock);
|
|
|
|
asid = asids->mt_asid;
|
|
|
|
|
2009-04-03 01:59:08 +02:00
|
|
|
spin_lock(&gru->gs_asid_lock);
|
|
|
|
if (asid == 0 || (asids->mt_ctxbitmap == 0 && asids->mt_asid_gen !=
|
|
|
|
gru->gs_asid_gen)) {
|
2008-07-30 07:33:58 +02:00
|
|
|
asid = gru_assign_asid(gru);
|
|
|
|
asids->mt_asid = asid;
|
|
|
|
asids->mt_asid_gen = gru->gs_asid_gen;
|
|
|
|
STAT(asid_new);
|
|
|
|
} else {
|
|
|
|
STAT(asid_reuse);
|
|
|
|
}
|
2009-04-03 01:59:08 +02:00
|
|
|
spin_unlock(&gru->gs_asid_lock);
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
BUG_ON(asids->mt_ctxbitmap & ctxbitmap);
|
|
|
|
asids->mt_ctxbitmap |= ctxbitmap;
|
|
|
|
if (!test_bit(gru->gs_gid, gms->ms_asidmap))
|
|
|
|
__set_bit(gru->gs_gid, gms->ms_asidmap);
|
|
|
|
spin_unlock(&gms->ms_asid_lock);
|
|
|
|
|
|
|
|
gru_dbg(grudev,
|
2009-04-03 01:59:05 +02:00
|
|
|
"gid %d, gts %p, gms %p, ctxnum %d, asid 0x%x, asidmap 0x%lx\n",
|
|
|
|
gru->gs_gid, gts, gms, gts->ts_ctxnum, asid,
|
|
|
|
gms->ms_asidmap[0]);
|
2008-07-30 07:33:58 +02:00
|
|
|
return asid;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void gru_unload_mm_tracker(struct gru_state *gru,
|
2009-04-03 01:59:05 +02:00
|
|
|
struct gru_thread_state *gts)
|
2008-07-30 07:33:58 +02:00
|
|
|
{
|
2009-04-03 01:59:05 +02:00
|
|
|
struct gru_mm_struct *gms = gts->ts_gms;
|
2008-07-30 07:33:58 +02:00
|
|
|
struct gru_mm_tracker *asids;
|
|
|
|
unsigned short ctxbitmap;
|
|
|
|
|
|
|
|
asids = &gms->ms_asids[gru->gs_gid];
|
2009-04-03 01:59:05 +02:00
|
|
|
ctxbitmap = (1 << gts->ts_ctxnum);
|
2008-07-30 07:33:58 +02:00
|
|
|
spin_lock(&gms->ms_asid_lock);
|
2009-04-03 01:59:08 +02:00
|
|
|
spin_lock(&gru->gs_asid_lock);
|
2008-07-30 07:33:58 +02:00
|
|
|
BUG_ON((asids->mt_ctxbitmap & ctxbitmap) != ctxbitmap);
|
|
|
|
asids->mt_ctxbitmap ^= ctxbitmap;
|
2009-04-03 01:59:05 +02:00
|
|
|
gru_dbg(grudev, "gid %d, gts %p, gms %p, ctxnum 0x%d, asidmap 0x%lx\n",
|
|
|
|
gru->gs_gid, gts, gms, gts->ts_ctxnum, gms->ms_asidmap[0]);
|
2009-04-03 01:59:08 +02:00
|
|
|
spin_unlock(&gru->gs_asid_lock);
|
2008-07-30 07:33:58 +02:00
|
|
|
spin_unlock(&gms->ms_asid_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Decrement the reference count on a GTS structure. Free the structure
|
|
|
|
* if the reference count goes to zero.
|
|
|
|
*/
|
|
|
|
void gts_drop(struct gru_thread_state *gts)
|
|
|
|
{
|
|
|
|
if (gts && atomic_dec_return(>s->ts_refcnt) == 0) {
|
2009-12-16 01:48:08 +01:00
|
|
|
if (gts->ts_gms)
|
|
|
|
gru_drop_mmu_notifier(gts->ts_gms);
|
2008-07-30 07:33:58 +02:00
|
|
|
kfree(gts);
|
|
|
|
STAT(gts_free);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Locate the GTS structure for the current thread.
|
|
|
|
*/
|
|
|
|
static struct gru_thread_state *gru_find_current_gts_nolock(struct gru_vma_data
|
|
|
|
*vdata, int tsid)
|
|
|
|
{
|
|
|
|
struct gru_thread_state *gts;
|
|
|
|
|
|
|
|
list_for_each_entry(gts, &vdata->vd_head, ts_next)
|
|
|
|
if (gts->ts_tsid == tsid)
|
|
|
|
return gts;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Allocate a thread state structure.
|
|
|
|
*/
|
2009-06-18 01:28:20 +02:00
|
|
|
struct gru_thread_state *gru_alloc_gts(struct vm_area_struct *vma,
|
2009-12-16 01:48:13 +01:00
|
|
|
int cbr_au_count, int dsr_au_count,
|
|
|
|
unsigned char tlb_preload_count, int options, int tsid)
|
2008-07-30 07:33:58 +02:00
|
|
|
{
|
|
|
|
struct gru_thread_state *gts;
|
2009-12-16 01:48:08 +01:00
|
|
|
struct gru_mm_struct *gms;
|
2008-07-30 07:33:58 +02:00
|
|
|
int bytes;
|
|
|
|
|
2009-06-18 01:28:20 +02:00
|
|
|
bytes = DSR_BYTES(dsr_au_count) + CBR_BYTES(cbr_au_count);
|
2008-07-30 07:33:58 +02:00
|
|
|
bytes += sizeof(struct gru_thread_state);
|
2009-06-18 01:28:24 +02:00
|
|
|
gts = kmalloc(bytes, GFP_KERNEL);
|
2008-07-30 07:33:58 +02:00
|
|
|
if (!gts)
|
2009-12-16 01:48:08 +01:00
|
|
|
return ERR_PTR(-ENOMEM);
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
STAT(gts_alloc);
|
2009-06-18 01:28:24 +02:00
|
|
|
memset(gts, 0, sizeof(struct gru_thread_state)); /* zero out header */
|
2008-07-30 07:33:58 +02:00
|
|
|
atomic_set(>s->ts_refcnt, 1);
|
|
|
|
mutex_init(>s->ts_ctxlock);
|
2009-06-18 01:28:20 +02:00
|
|
|
gts->ts_cbr_au_count = cbr_au_count;
|
|
|
|
gts->ts_dsr_au_count = dsr_au_count;
|
2009-12-16 01:48:13 +01:00
|
|
|
gts->ts_tlb_preload_count = tlb_preload_count;
|
2009-06-18 01:28:20 +02:00
|
|
|
gts->ts_user_options = options;
|
2009-12-16 01:48:04 +01:00
|
|
|
gts->ts_user_blade_id = -1;
|
|
|
|
gts->ts_user_chiplet_id = -1;
|
2008-07-30 07:33:58 +02:00
|
|
|
gts->ts_tsid = tsid;
|
|
|
|
gts->ts_ctxnum = NULLCTX;
|
|
|
|
gts->ts_tlb_int_select = -1;
|
2009-06-18 01:28:33 +02:00
|
|
|
gts->ts_cch_req_slice = -1;
|
2009-04-03 01:59:12 +02:00
|
|
|
gts->ts_sizeavail = GRU_SIZEAVAIL(PAGE_SHIFT);
|
2009-06-18 01:28:20 +02:00
|
|
|
if (vma) {
|
|
|
|
gts->ts_mm = current->mm;
|
|
|
|
gts->ts_vma = vma;
|
2009-12-16 01:48:08 +01:00
|
|
|
gms = gru_register_mmu_notifier();
|
|
|
|
if (IS_ERR(gms))
|
2009-06-18 01:28:20 +02:00
|
|
|
goto err;
|
2009-12-16 01:48:08 +01:00
|
|
|
gts->ts_gms = gms;
|
2009-06-18 01:28:20 +02:00
|
|
|
}
|
2008-07-30 07:33:58 +02:00
|
|
|
|
2009-06-18 01:28:20 +02:00
|
|
|
gru_dbg(grudev, "alloc gts %p\n", gts);
|
2008-07-30 07:33:58 +02:00
|
|
|
return gts;
|
|
|
|
|
|
|
|
err:
|
|
|
|
gts_drop(gts);
|
2009-12-16 01:48:08 +01:00
|
|
|
return ERR_CAST(gms);
|
2008-07-30 07:33:58 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Allocate a vma private data structure.
|
|
|
|
*/
|
|
|
|
struct gru_vma_data *gru_alloc_vma_data(struct vm_area_struct *vma, int tsid)
|
|
|
|
{
|
|
|
|
struct gru_vma_data *vdata = NULL;
|
|
|
|
|
|
|
|
vdata = kmalloc(sizeof(*vdata), GFP_KERNEL);
|
|
|
|
if (!vdata)
|
|
|
|
return NULL;
|
|
|
|
|
2009-12-16 01:48:12 +01:00
|
|
|
STAT(vdata_alloc);
|
2008-07-30 07:33:58 +02:00
|
|
|
INIT_LIST_HEAD(&vdata->vd_head);
|
|
|
|
spin_lock_init(&vdata->vd_lock);
|
|
|
|
gru_dbg(grudev, "alloc vdata %p\n", vdata);
|
|
|
|
return vdata;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Find the thread state structure for the current thread.
|
|
|
|
*/
|
|
|
|
struct gru_thread_state *gru_find_thread_state(struct vm_area_struct *vma,
|
|
|
|
int tsid)
|
|
|
|
{
|
|
|
|
struct gru_vma_data *vdata = vma->vm_private_data;
|
|
|
|
struct gru_thread_state *gts;
|
|
|
|
|
|
|
|
spin_lock(&vdata->vd_lock);
|
|
|
|
gts = gru_find_current_gts_nolock(vdata, tsid);
|
|
|
|
spin_unlock(&vdata->vd_lock);
|
|
|
|
gru_dbg(grudev, "vma %p, gts %p\n", vma, gts);
|
|
|
|
return gts;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Allocate a new thread state for a GSEG. Note that races may allow
|
|
|
|
* another thread to race to create a gts.
|
|
|
|
*/
|
|
|
|
struct gru_thread_state *gru_alloc_thread_state(struct vm_area_struct *vma,
|
|
|
|
int tsid)
|
|
|
|
{
|
|
|
|
struct gru_vma_data *vdata = vma->vm_private_data;
|
|
|
|
struct gru_thread_state *gts, *ngts;
|
|
|
|
|
2009-12-16 01:48:13 +01:00
|
|
|
gts = gru_alloc_gts(vma, vdata->vd_cbr_au_count,
|
|
|
|
vdata->vd_dsr_au_count,
|
|
|
|
vdata->vd_tlb_preload_count,
|
2009-06-18 01:28:20 +02:00
|
|
|
vdata->vd_user_options, tsid);
|
2009-12-16 01:48:08 +01:00
|
|
|
if (IS_ERR(gts))
|
|
|
|
return gts;
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
spin_lock(&vdata->vd_lock);
|
|
|
|
ngts = gru_find_current_gts_nolock(vdata, tsid);
|
|
|
|
if (ngts) {
|
|
|
|
gts_drop(gts);
|
|
|
|
gts = ngts;
|
|
|
|
STAT(gts_double_allocate);
|
|
|
|
} else {
|
|
|
|
list_add(>s->ts_next, &vdata->vd_head);
|
|
|
|
}
|
|
|
|
spin_unlock(&vdata->vd_lock);
|
|
|
|
gru_dbg(grudev, "vma %p, gts %p\n", vma, gts);
|
|
|
|
return gts;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Free the GRU context assigned to the thread state.
|
|
|
|
*/
|
|
|
|
static void gru_free_gru_context(struct gru_thread_state *gts)
|
|
|
|
{
|
|
|
|
struct gru_state *gru;
|
|
|
|
|
|
|
|
gru = gts->ts_gru;
|
2009-04-03 01:59:05 +02:00
|
|
|
gru_dbg(grudev, "gts %p, gid %d\n", gts, gru->gs_gid);
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
spin_lock(&gru->gs_lock);
|
|
|
|
gru->gs_gts[gts->ts_ctxnum] = NULL;
|
|
|
|
free_gru_resources(gru, gts);
|
|
|
|
BUG_ON(test_bit(gts->ts_ctxnum, &gru->gs_context_map) == 0);
|
|
|
|
__clear_bit(gts->ts_ctxnum, &gru->gs_context_map);
|
|
|
|
gts->ts_ctxnum = NULLCTX;
|
|
|
|
gts->ts_gru = NULL;
|
2009-04-03 01:59:08 +02:00
|
|
|
gts->ts_blade = -1;
|
2008-07-30 07:33:58 +02:00
|
|
|
spin_unlock(&gru->gs_lock);
|
|
|
|
|
|
|
|
gts_drop(gts);
|
|
|
|
STAT(free_context);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Prefetching cachelines help hardware performance.
|
2008-07-30 07:34:02 +02:00
|
|
|
* (Strictly a performance enhancement. Not functionally required).
|
2008-07-30 07:33:58 +02:00
|
|
|
*/
|
|
|
|
static void prefetch_data(void *p, int num, int stride)
|
|
|
|
{
|
|
|
|
while (num-- > 0) {
|
|
|
|
prefetchw(p);
|
|
|
|
p += stride;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline long gru_copy_handle(void *d, void *s)
|
|
|
|
{
|
|
|
|
memcpy(d, s, GRU_HANDLE_BYTES);
|
|
|
|
return GRU_HANDLE_BYTES;
|
|
|
|
}
|
|
|
|
|
2009-04-03 01:59:04 +02:00
|
|
|
static void gru_prefetch_context(void *gseg, void *cb, void *cbe,
|
|
|
|
unsigned long cbrmap, unsigned long length)
|
2008-07-30 07:33:58 +02:00
|
|
|
{
|
|
|
|
int i, scr;
|
|
|
|
|
|
|
|
prefetch_data(gseg + GRU_DS_BASE, length / GRU_CACHE_LINE_BYTES,
|
|
|
|
GRU_CACHE_LINE_BYTES);
|
|
|
|
|
|
|
|
for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
|
|
|
|
prefetch_data(cb, 1, GRU_CACHE_LINE_BYTES);
|
|
|
|
prefetch_data(cbe + i * GRU_HANDLE_STRIDE, 1,
|
|
|
|
GRU_CACHE_LINE_BYTES);
|
|
|
|
cb += GRU_HANDLE_STRIDE;
|
|
|
|
}
|
2008-10-16 07:05:13 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
static void gru_load_context_data(void *save, void *grubase, int ctxnum,
|
2009-06-18 01:28:24 +02:00
|
|
|
unsigned long cbrmap, unsigned long dsrmap,
|
|
|
|
int data_valid)
|
2008-10-16 07:05:13 +02:00
|
|
|
{
|
|
|
|
void *gseg, *cb, *cbe;
|
|
|
|
unsigned long length;
|
|
|
|
int i, scr;
|
2008-07-30 07:33:58 +02:00
|
|
|
|
2008-10-16 07:05:13 +02:00
|
|
|
gseg = grubase + ctxnum * GRU_GSEG_STRIDE;
|
2008-07-30 07:33:58 +02:00
|
|
|
cb = gseg + GRU_CB_BASE;
|
2008-10-16 07:05:13 +02:00
|
|
|
cbe = grubase + GRU_CBE_BASE;
|
|
|
|
length = hweight64(dsrmap) * GRU_DSR_AU_BYTES;
|
|
|
|
gru_prefetch_context(gseg, cb, cbe, cbrmap, length);
|
|
|
|
|
2008-07-30 07:33:58 +02:00
|
|
|
for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
|
2009-06-18 01:28:24 +02:00
|
|
|
if (data_valid) {
|
|
|
|
save += gru_copy_handle(cb, save);
|
|
|
|
save += gru_copy_handle(cbe + i * GRU_HANDLE_STRIDE,
|
|
|
|
save);
|
|
|
|
} else {
|
|
|
|
memset(cb, 0, GRU_CACHE_LINE_BYTES);
|
|
|
|
memset(cbe + i * GRU_HANDLE_STRIDE, 0,
|
|
|
|
GRU_CACHE_LINE_BYTES);
|
|
|
|
}
|
2009-12-16 01:48:11 +01:00
|
|
|
/* Flush CBE to hide race in context restart */
|
|
|
|
mb();
|
|
|
|
gru_flush_cache(cbe + i * GRU_HANDLE_STRIDE);
|
2008-07-30 07:33:58 +02:00
|
|
|
cb += GRU_HANDLE_STRIDE;
|
|
|
|
}
|
|
|
|
|
2009-06-18 01:28:24 +02:00
|
|
|
if (data_valid)
|
|
|
|
memcpy(gseg + GRU_DS_BASE, save, length);
|
|
|
|
else
|
|
|
|
memset(gseg + GRU_DS_BASE, 0, length);
|
2008-07-30 07:33:58 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
static void gru_unload_context_data(void *save, void *grubase, int ctxnum,
|
|
|
|
unsigned long cbrmap, unsigned long dsrmap)
|
|
|
|
{
|
|
|
|
void *gseg, *cb, *cbe;
|
|
|
|
unsigned long length;
|
|
|
|
int i, scr;
|
|
|
|
|
|
|
|
gseg = grubase + ctxnum * GRU_GSEG_STRIDE;
|
|
|
|
cb = gseg + GRU_CB_BASE;
|
|
|
|
cbe = grubase + GRU_CBE_BASE;
|
2008-10-16 07:05:13 +02:00
|
|
|
length = hweight64(dsrmap) * GRU_DSR_AU_BYTES;
|
2009-12-16 01:48:11 +01:00
|
|
|
|
|
|
|
/* CBEs may not be coherent. Flush them from cache */
|
|
|
|
for_each_cbr_in_allocation_map(i, &cbrmap, scr)
|
|
|
|
gru_flush_cache(cbe + i * GRU_HANDLE_STRIDE);
|
|
|
|
mb(); /* Let the CL flush complete */
|
|
|
|
|
2008-10-16 07:05:13 +02:00
|
|
|
gru_prefetch_context(gseg, cb, cbe, cbrmap, length);
|
|
|
|
|
2008-07-30 07:33:58 +02:00
|
|
|
for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
|
|
|
|
save += gru_copy_handle(save, cb);
|
|
|
|
save += gru_copy_handle(save, cbe + i * GRU_HANDLE_STRIDE);
|
|
|
|
cb += GRU_HANDLE_STRIDE;
|
|
|
|
}
|
|
|
|
memcpy(save, gseg + GRU_DS_BASE, length);
|
|
|
|
}
|
|
|
|
|
|
|
|
void gru_unload_context(struct gru_thread_state *gts, int savestate)
|
|
|
|
{
|
|
|
|
struct gru_state *gru = gts->ts_gru;
|
|
|
|
struct gru_context_configuration_handle *cch;
|
|
|
|
int ctxnum = gts->ts_ctxnum;
|
|
|
|
|
2009-06-18 01:28:22 +02:00
|
|
|
if (!is_kernel_context(gts))
|
|
|
|
zap_vma_ptes(gts->ts_vma, UGRUADDR(gts), GRU_GSEG_PAGESIZE);
|
2008-07-30 07:33:58 +02:00
|
|
|
cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
|
|
|
|
|
2009-12-16 01:48:12 +01:00
|
|
|
gru_dbg(grudev, "gts %p, cbrmap 0x%lx, dsrmap 0x%lx\n",
|
|
|
|
gts, gts->ts_cbr_map, gts->ts_dsr_map);
|
2008-07-30 07:33:58 +02:00
|
|
|
lock_cch_handle(cch);
|
|
|
|
if (cch_interrupt_sync(cch))
|
|
|
|
BUG();
|
|
|
|
|
2009-06-18 01:28:22 +02:00
|
|
|
if (!is_kernel_context(gts))
|
|
|
|
gru_unload_mm_tracker(gru, gts);
|
2009-06-18 01:28:24 +02:00
|
|
|
if (savestate) {
|
2008-07-30 07:33:58 +02:00
|
|
|
gru_unload_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr,
|
|
|
|
ctxnum, gts->ts_cbr_map,
|
|
|
|
gts->ts_dsr_map);
|
2009-06-18 01:28:24 +02:00
|
|
|
gts->ts_data_valid = 1;
|
|
|
|
}
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
if (cch_deallocate(cch))
|
|
|
|
BUG();
|
|
|
|
unlock_cch_handle(cch);
|
|
|
|
|
|
|
|
gru_free_gru_context(gts);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Load a GRU context by copying it from the thread data structure in memory
|
|
|
|
* to the GRU.
|
|
|
|
*/
|
2009-06-18 01:28:20 +02:00
|
|
|
void gru_load_context(struct gru_thread_state *gts)
|
2008-07-30 07:33:58 +02:00
|
|
|
{
|
|
|
|
struct gru_state *gru = gts->ts_gru;
|
|
|
|
struct gru_context_configuration_handle *cch;
|
2009-06-18 01:28:21 +02:00
|
|
|
int i, err, asid, ctxnum = gts->ts_ctxnum;
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
|
|
|
|
lock_cch_handle(cch);
|
|
|
|
cch->tfm_fault_bit_enable =
|
|
|
|
(gts->ts_user_options == GRU_OPT_MISS_FMM_POLL
|
|
|
|
|| gts->ts_user_options == GRU_OPT_MISS_FMM_INTR);
|
|
|
|
cch->tlb_int_enable = (gts->ts_user_options == GRU_OPT_MISS_FMM_INTR);
|
|
|
|
if (cch->tlb_int_enable) {
|
|
|
|
gts->ts_tlb_int_select = gru_cpu_fault_map_id();
|
|
|
|
cch->tlb_int_select = gts->ts_tlb_int_select;
|
|
|
|
}
|
2009-06-18 01:28:33 +02:00
|
|
|
if (gts->ts_cch_req_slice >= 0) {
|
|
|
|
cch->req_slice_set_enable = 1;
|
|
|
|
cch->req_slice = gts->ts_cch_req_slice;
|
|
|
|
} else {
|
|
|
|
cch->req_slice_set_enable =0;
|
|
|
|
}
|
2008-07-30 07:33:58 +02:00
|
|
|
cch->tfm_done_bit_enable = 0;
|
2009-06-18 01:28:21 +02:00
|
|
|
cch->dsr_allocation_map = gts->ts_dsr_map;
|
|
|
|
cch->cbr_allocation_map = gts->ts_cbr_map;
|
2009-06-18 01:28:22 +02:00
|
|
|
|
|
|
|
if (is_kernel_context(gts)) {
|
|
|
|
cch->unmap_enable = 1;
|
2009-06-18 01:28:25 +02:00
|
|
|
cch->tfm_done_bit_enable = 1;
|
|
|
|
cch->cb_int_enable = 1;
|
2009-12-16 01:48:11 +01:00
|
|
|
cch->tlb_int_select = 0; /* For now, ints go to cpu 0 */
|
2009-06-18 01:28:22 +02:00
|
|
|
} else {
|
|
|
|
cch->unmap_enable = 0;
|
2009-06-18 01:28:25 +02:00
|
|
|
cch->tfm_done_bit_enable = 0;
|
|
|
|
cch->cb_int_enable = 0;
|
2009-06-18 01:28:22 +02:00
|
|
|
asid = gru_load_mm_tracker(gru, gts);
|
|
|
|
for (i = 0; i < 8; i++) {
|
|
|
|
cch->asid[i] = asid + i;
|
|
|
|
cch->sizeavail[i] = gts->ts_sizeavail;
|
|
|
|
}
|
2009-06-18 01:28:21 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
err = cch_allocate(cch);
|
2008-07-30 07:33:58 +02:00
|
|
|
if (err) {
|
|
|
|
gru_dbg(grudev,
|
|
|
|
"err %d: cch %p, gts %p, cbr 0x%lx, dsr 0x%lx\n",
|
|
|
|
err, cch, gts, gts->ts_cbr_map, gts->ts_dsr_map);
|
|
|
|
BUG();
|
|
|
|
}
|
|
|
|
|
|
|
|
gru_load_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr, ctxnum,
|
2009-06-18 01:28:24 +02:00
|
|
|
gts->ts_cbr_map, gts->ts_dsr_map, gts->ts_data_valid);
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
if (cch_start(cch))
|
|
|
|
BUG();
|
|
|
|
unlock_cch_handle(cch);
|
2009-12-16 01:48:12 +01:00
|
|
|
|
|
|
|
gru_dbg(grudev, "gid %d, gts %p, cbrmap 0x%lx, dsrmap 0x%lx, tie %d, tis %d\n",
|
|
|
|
gts->ts_gru->gs_gid, gts, gts->ts_cbr_map, gts->ts_dsr_map,
|
|
|
|
(gts->ts_user_options == GRU_OPT_MISS_FMM_INTR), gts->ts_tlb_int_select);
|
2008-07-30 07:33:58 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Update fields in an active CCH:
|
|
|
|
* - retarget interrupts on local blade
|
2009-04-03 01:59:12 +02:00
|
|
|
* - update sizeavail mask
|
2008-07-30 07:33:58 +02:00
|
|
|
*/
|
2009-12-16 01:48:06 +01:00
|
|
|
int gru_update_cch(struct gru_thread_state *gts)
|
2008-07-30 07:33:58 +02:00
|
|
|
{
|
|
|
|
struct gru_context_configuration_handle *cch;
|
|
|
|
struct gru_state *gru = gts->ts_gru;
|
|
|
|
int i, ctxnum = gts->ts_ctxnum, ret = 0;
|
|
|
|
|
|
|
|
cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
|
|
|
|
|
|
|
|
lock_cch_handle(cch);
|
|
|
|
if (cch->state == CCHSTATE_ACTIVE) {
|
|
|
|
if (gru->gs_gts[gts->ts_ctxnum] != gts)
|
|
|
|
goto exit;
|
|
|
|
if (cch_interrupt(cch))
|
|
|
|
BUG();
|
2009-12-16 01:48:06 +01:00
|
|
|
for (i = 0; i < 8; i++)
|
|
|
|
cch->sizeavail[i] = gts->ts_sizeavail;
|
|
|
|
gts->ts_tlb_int_select = gru_cpu_fault_map_id();
|
|
|
|
cch->tlb_int_select = gru_cpu_fault_map_id();
|
|
|
|
cch->tfm_fault_bit_enable =
|
|
|
|
(gts->ts_user_options == GRU_OPT_MISS_FMM_POLL
|
|
|
|
|| gts->ts_user_options == GRU_OPT_MISS_FMM_INTR);
|
2008-07-30 07:33:58 +02:00
|
|
|
if (cch_start(cch))
|
|
|
|
BUG();
|
|
|
|
ret = 1;
|
|
|
|
}
|
|
|
|
exit:
|
|
|
|
unlock_cch_handle(cch);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Update CCH tlb interrupt select. Required when all the following is true:
|
|
|
|
* - task's GRU context is loaded into a GRU
|
|
|
|
* - task is using interrupt notification for TLB faults
|
|
|
|
* - task has migrated to a different cpu on the same blade where
|
|
|
|
* it was previously running.
|
|
|
|
*/
|
|
|
|
static int gru_retarget_intr(struct gru_thread_state *gts)
|
|
|
|
{
|
|
|
|
if (gts->ts_tlb_int_select < 0
|
|
|
|
|| gts->ts_tlb_int_select == gru_cpu_fault_map_id())
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
gru_dbg(grudev, "retarget from %d to %d\n", gts->ts_tlb_int_select,
|
|
|
|
gru_cpu_fault_map_id());
|
2009-12-16 01:48:06 +01:00
|
|
|
return gru_update_cch(gts);
|
2008-07-30 07:33:58 +02:00
|
|
|
}
|
|
|
|
|
2009-12-16 01:48:15 +01:00
|
|
|
/*
|
|
|
|
* Check if a GRU context is allowed to use a specific chiplet. By default
|
|
|
|
* a context is assigned to any blade-local chiplet. However, users can
|
|
|
|
* override this.
|
|
|
|
* Returns 1 if assignment allowed, 0 otherwise
|
|
|
|
*/
|
|
|
|
static int gru_check_chiplet_assignment(struct gru_state *gru,
|
|
|
|
struct gru_thread_state *gts)
|
|
|
|
{
|
|
|
|
int blade_id;
|
|
|
|
int chiplet_id;
|
|
|
|
|
|
|
|
blade_id = gts->ts_user_blade_id;
|
|
|
|
if (blade_id < 0)
|
|
|
|
blade_id = uv_numa_blade_id();
|
|
|
|
|
|
|
|
chiplet_id = gts->ts_user_chiplet_id;
|
|
|
|
return gru->gs_blade_id == blade_id &&
|
|
|
|
(chiplet_id < 0 || chiplet_id == gru->gs_chiplet_id);
|
|
|
|
}
|
|
|
|
|
2009-12-16 01:48:05 +01:00
|
|
|
/*
|
|
|
|
* Unload the gru context if it is not assigned to the correct blade or
|
|
|
|
* chiplet. Misassignment can occur if the process migrates to a different
|
|
|
|
* blade or if the user changes the selected blade/chiplet.
|
|
|
|
*/
|
|
|
|
void gru_check_context_placement(struct gru_thread_state *gts)
|
|
|
|
{
|
|
|
|
struct gru_state *gru;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the current task is the context owner, verify that the
|
|
|
|
* context is correctly placed. This test is skipped for non-owner
|
|
|
|
* references. Pthread apps use non-owner references to the CBRs.
|
|
|
|
*/
|
|
|
|
gru = gts->ts_gru;
|
|
|
|
if (!gru || gts->ts_tgid_owner != current->tgid)
|
|
|
|
return;
|
|
|
|
|
2009-12-16 01:48:15 +01:00
|
|
|
if (!gru_check_chiplet_assignment(gru, gts)) {
|
2009-12-16 01:48:05 +01:00
|
|
|
STAT(check_context_unload);
|
|
|
|
gru_unload_context(gts, 1);
|
|
|
|
} else if (gru_retarget_intr(gts)) {
|
|
|
|
STAT(check_context_retarget_intr);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Insufficient GRU resources available on the local blade. Steal a context from
|
|
|
|
* a process. This is a hack until a _real_ resource scheduler is written....
|
|
|
|
*/
|
|
|
|
#define next_ctxnum(n) ((n) < GRU_NUM_CCH - 2 ? (n) + 1 : 0)
|
|
|
|
#define next_gru(b, g) (((g) < &(b)->bs_grus[GRU_CHIPLETS_PER_BLADE - 1]) ? \
|
|
|
|
((g)+1) : &(b)->bs_grus[0])
|
|
|
|
|
2009-06-18 01:28:22 +02:00
|
|
|
static int is_gts_stealable(struct gru_thread_state *gts,
|
|
|
|
struct gru_blade_state *bs)
|
|
|
|
{
|
|
|
|
if (is_kernel_context(gts))
|
|
|
|
return down_write_trylock(&bs->bs_kgts_sema);
|
|
|
|
else
|
|
|
|
return mutex_trylock(>s->ts_ctxlock);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void gts_stolen(struct gru_thread_state *gts,
|
|
|
|
struct gru_blade_state *bs)
|
|
|
|
{
|
|
|
|
if (is_kernel_context(gts)) {
|
|
|
|
up_write(&bs->bs_kgts_sema);
|
|
|
|
STAT(steal_kernel_context);
|
|
|
|
} else {
|
|
|
|
mutex_unlock(>s->ts_ctxlock);
|
|
|
|
STAT(steal_user_context);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2009-12-16 01:48:05 +01:00
|
|
|
void gru_steal_context(struct gru_thread_state *gts)
|
2008-07-30 07:33:58 +02:00
|
|
|
{
|
|
|
|
struct gru_blade_state *blade;
|
|
|
|
struct gru_state *gru, *gru0;
|
|
|
|
struct gru_thread_state *ngts = NULL;
|
|
|
|
int ctxnum, ctxnum0, flag = 0, cbr, dsr;
|
2009-12-16 01:48:15 +01:00
|
|
|
int blade_id;
|
2008-07-30 07:33:58 +02:00
|
|
|
|
2009-12-16 01:48:15 +01:00
|
|
|
blade_id = gts->ts_user_blade_id;
|
2009-12-16 01:48:05 +01:00
|
|
|
if (blade_id < 0)
|
|
|
|
blade_id = uv_numa_blade_id();
|
2008-07-30 07:33:58 +02:00
|
|
|
cbr = gts->ts_cbr_au_count;
|
|
|
|
dsr = gts->ts_dsr_au_count;
|
|
|
|
|
2009-06-18 01:28:20 +02:00
|
|
|
blade = gru_base[blade_id];
|
2008-07-30 07:33:58 +02:00
|
|
|
spin_lock(&blade->bs_lock);
|
|
|
|
|
|
|
|
ctxnum = next_ctxnum(blade->bs_lru_ctxnum);
|
|
|
|
gru = blade->bs_lru_gru;
|
|
|
|
if (ctxnum == 0)
|
|
|
|
gru = next_gru(blade, gru);
|
2009-12-16 01:48:05 +01:00
|
|
|
blade->bs_lru_gru = gru;
|
|
|
|
blade->bs_lru_ctxnum = ctxnum;
|
2008-07-30 07:33:58 +02:00
|
|
|
ctxnum0 = ctxnum;
|
|
|
|
gru0 = gru;
|
|
|
|
while (1) {
|
2009-12-16 01:48:15 +01:00
|
|
|
if (gru_check_chiplet_assignment(gru, gts)) {
|
2009-12-16 01:48:05 +01:00
|
|
|
if (check_gru_resources(gru, cbr, dsr, GRU_NUM_CCH))
|
2008-07-30 07:33:58 +02:00
|
|
|
break;
|
2009-12-16 01:48:05 +01:00
|
|
|
spin_lock(&gru->gs_lock);
|
|
|
|
for (; ctxnum < GRU_NUM_CCH; ctxnum++) {
|
|
|
|
if (flag && gru == gru0 && ctxnum == ctxnum0)
|
|
|
|
break;
|
|
|
|
ngts = gru->gs_gts[ctxnum];
|
|
|
|
/*
|
|
|
|
* We are grabbing locks out of order, so trylock is
|
|
|
|
* needed. GTSs are usually not locked, so the odds of
|
|
|
|
* success are high. If trylock fails, try to steal a
|
|
|
|
* different GSEG.
|
|
|
|
*/
|
|
|
|
if (ngts && is_gts_stealable(ngts, blade))
|
|
|
|
break;
|
|
|
|
ngts = NULL;
|
|
|
|
}
|
|
|
|
spin_unlock(&gru->gs_lock);
|
|
|
|
if (ngts || (flag && gru == gru0 && ctxnum == ctxnum0))
|
2008-07-30 07:33:58 +02:00
|
|
|
break;
|
|
|
|
}
|
2009-12-16 01:48:05 +01:00
|
|
|
if (flag && gru == gru0)
|
2008-07-30 07:33:58 +02:00
|
|
|
break;
|
2009-12-16 01:48:05 +01:00
|
|
|
flag = 1;
|
2008-07-30 07:33:58 +02:00
|
|
|
ctxnum = 0;
|
|
|
|
gru = next_gru(blade, gru);
|
|
|
|
}
|
|
|
|
spin_unlock(&blade->bs_lock);
|
|
|
|
|
|
|
|
if (ngts) {
|
2009-06-18 01:28:30 +02:00
|
|
|
gts->ustats.context_stolen++;
|
2008-07-30 07:33:58 +02:00
|
|
|
ngts->ts_steal_jiffies = jiffies;
|
2009-06-18 01:28:22 +02:00
|
|
|
gru_unload_context(ngts, is_kernel_context(ngts) ? 0 : 1);
|
|
|
|
gts_stolen(ngts, blade);
|
2008-07-30 07:33:58 +02:00
|
|
|
} else {
|
|
|
|
STAT(steal_context_failed);
|
|
|
|
}
|
|
|
|
gru_dbg(grudev,
|
2009-04-03 01:59:05 +02:00
|
|
|
"stole gid %d, ctxnum %d from gts %p. Need cb %d, ds %d;"
|
2008-07-30 07:33:58 +02:00
|
|
|
" avail cb %ld, ds %ld\n",
|
|
|
|
gru->gs_gid, ctxnum, ngts, cbr, dsr, hweight64(gru->gs_cbr_map),
|
|
|
|
hweight64(gru->gs_dsr_map));
|
|
|
|
}
|
|
|
|
|
2009-12-16 01:48:05 +01:00
|
|
|
/*
|
|
|
|
* Assign a gru context.
|
|
|
|
*/
|
|
|
|
static int gru_assign_context_number(struct gru_state *gru)
|
|
|
|
{
|
|
|
|
int ctxnum;
|
|
|
|
|
|
|
|
ctxnum = find_first_zero_bit(&gru->gs_context_map, GRU_NUM_CCH);
|
|
|
|
__set_bit(ctxnum, &gru->gs_context_map);
|
|
|
|
return ctxnum;
|
|
|
|
}
|
|
|
|
|
2008-07-30 07:33:58 +02:00
|
|
|
/*
|
|
|
|
* Scan the GRUs on the local blade & assign a GRU context.
|
|
|
|
*/
|
2009-12-16 01:48:05 +01:00
|
|
|
struct gru_state *gru_assign_gru_context(struct gru_thread_state *gts)
|
2008-07-30 07:33:58 +02:00
|
|
|
{
|
|
|
|
struct gru_state *gru, *grux;
|
|
|
|
int i, max_active_contexts;
|
2009-12-16 01:48:05 +01:00
|
|
|
int blade_id = gts->ts_user_blade_id;
|
2008-07-30 07:33:58 +02:00
|
|
|
|
2009-12-16 01:48:05 +01:00
|
|
|
if (blade_id < 0)
|
|
|
|
blade_id = uv_numa_blade_id();
|
2008-07-30 07:33:58 +02:00
|
|
|
again:
|
|
|
|
gru = NULL;
|
|
|
|
max_active_contexts = GRU_NUM_CCH;
|
2009-12-16 01:48:05 +01:00
|
|
|
for_each_gru_on_blade(grux, blade_id, i) {
|
2009-12-16 01:48:15 +01:00
|
|
|
if (!gru_check_chiplet_assignment(grux, gts))
|
2009-12-16 01:48:05 +01:00
|
|
|
continue;
|
2008-07-30 07:33:58 +02:00
|
|
|
if (check_gru_resources(grux, gts->ts_cbr_au_count,
|
|
|
|
gts->ts_dsr_au_count,
|
|
|
|
max_active_contexts)) {
|
|
|
|
gru = grux;
|
|
|
|
max_active_contexts = grux->gs_active_contexts;
|
|
|
|
if (max_active_contexts == 0)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (gru) {
|
|
|
|
spin_lock(&gru->gs_lock);
|
|
|
|
if (!check_gru_resources(gru, gts->ts_cbr_au_count,
|
|
|
|
gts->ts_dsr_au_count, GRU_NUM_CCH)) {
|
|
|
|
spin_unlock(&gru->gs_lock);
|
|
|
|
goto again;
|
|
|
|
}
|
|
|
|
reserve_gru_resources(gru, gts);
|
|
|
|
gts->ts_gru = gru;
|
2009-04-03 01:59:08 +02:00
|
|
|
gts->ts_blade = gru->gs_blade_id;
|
2009-12-16 01:48:05 +01:00
|
|
|
gts->ts_ctxnum = gru_assign_context_number(gru);
|
2008-07-30 07:33:58 +02:00
|
|
|
atomic_inc(>s->ts_refcnt);
|
|
|
|
gru->gs_gts[gts->ts_ctxnum] = gts;
|
|
|
|
spin_unlock(&gru->gs_lock);
|
|
|
|
|
|
|
|
STAT(assign_context);
|
|
|
|
gru_dbg(grudev,
|
2009-04-03 01:59:05 +02:00
|
|
|
"gseg %p, gts %p, gid %d, ctx %d, cbr %d, dsr %d\n",
|
2008-07-30 07:33:58 +02:00
|
|
|
gseg_virtual_address(gts->ts_gru, gts->ts_ctxnum), gts,
|
|
|
|
gts->ts_gru->gs_gid, gts->ts_ctxnum,
|
|
|
|
gts->ts_cbr_au_count, gts->ts_dsr_au_count);
|
|
|
|
} else {
|
|
|
|
gru_dbg(grudev, "failed to allocate a GTS %s\n", "");
|
|
|
|
STAT(assign_context_failed);
|
|
|
|
}
|
|
|
|
|
|
|
|
return gru;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* gru_nopage
|
|
|
|
*
|
|
|
|
* Map the user's GRU segment
|
2008-07-30 07:34:02 +02:00
|
|
|
*
|
|
|
|
* Note: gru segments alway mmaped on GRU_GSEG_PAGESIZE boundaries.
|
2008-07-30 07:33:58 +02:00
|
|
|
*/
|
|
|
|
int gru_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
|
|
|
|
{
|
|
|
|
struct gru_thread_state *gts;
|
|
|
|
unsigned long paddr, vaddr;
|
|
|
|
|
|
|
|
vaddr = (unsigned long)vmf->virtual_address;
|
|
|
|
gru_dbg(grudev, "vma %p, vaddr 0x%lx (0x%lx)\n",
|
|
|
|
vma, vaddr, GSEG_BASE(vaddr));
|
|
|
|
STAT(nopfn);
|
|
|
|
|
2008-07-30 07:34:02 +02:00
|
|
|
/* The following check ensures vaddr is a valid address in the VMA */
|
2008-07-30 07:33:58 +02:00
|
|
|
gts = gru_find_thread_state(vma, TSID(vaddr, vma));
|
|
|
|
if (!gts)
|
|
|
|
return VM_FAULT_SIGBUS;
|
|
|
|
|
|
|
|
again:
|
|
|
|
mutex_lock(>s->ts_ctxlock);
|
2009-04-03 01:59:04 +02:00
|
|
|
preempt_disable();
|
2009-06-18 01:28:20 +02:00
|
|
|
|
2009-12-16 01:48:05 +01:00
|
|
|
gru_check_context_placement(gts);
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
if (!gts->ts_gru) {
|
2009-06-18 01:28:22 +02:00
|
|
|
STAT(load_user_context);
|
2009-12-16 01:48:05 +01:00
|
|
|
if (!gru_assign_gru_context(gts)) {
|
2008-07-30 07:33:58 +02:00
|
|
|
preempt_enable();
|
2009-06-18 01:28:20 +02:00
|
|
|
mutex_unlock(>s->ts_ctxlock);
|
|
|
|
set_current_state(TASK_INTERRUPTIBLE);
|
2008-07-30 07:33:58 +02:00
|
|
|
schedule_timeout(GRU_ASSIGN_DELAY); /* true hack ZZZ */
|
|
|
|
if (gts->ts_steal_jiffies + GRU_STEAL_DELAY < jiffies)
|
2009-12-16 01:48:05 +01:00
|
|
|
gru_steal_context(gts);
|
2008-07-30 07:33:58 +02:00
|
|
|
goto again;
|
|
|
|
}
|
|
|
|
gru_load_context(gts);
|
|
|
|
paddr = gseg_physical_address(gts->ts_gru, gts->ts_ctxnum);
|
|
|
|
remap_pfn_range(vma, vaddr & ~(GRU_GSEG_PAGESIZE - 1),
|
|
|
|
paddr >> PAGE_SHIFT, GRU_GSEG_PAGESIZE,
|
|
|
|
vma->vm_page_prot);
|
|
|
|
}
|
|
|
|
|
|
|
|
preempt_enable();
|
2009-06-18 01:28:20 +02:00
|
|
|
mutex_unlock(>s->ts_ctxlock);
|
2008-07-30 07:33:58 +02:00
|
|
|
|
|
|
|
return VM_FAULT_NOPAGE;
|
|
|
|
}
|
|
|
|
|