Enable the pmem driver to handle PFN device instances. Attaching a pmem
namespace to a pfn device triggers the driver to allocate and initialize
struct page entries for pmem. Memory capacity for this allocation comes
exclusively from RAM for now which is suitable for low PMEM to RAM
ratios. This mechanism will be expanded later for setting an "allocate
from PMEM" policy.
Cc: Boaz Harrosh <boaz@plexistor.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Implement the base infrastructure for libnvdimm PFN devices. Similar to
BTT devices they take a namespace as a backing device and layer
functionality on top. In this case the functionality is reserving space
for an array of 'struct page' entries to be handed out through
pfn_to_page(). For now this is just the basic libnvdimm-device-model for
configuring the base PFN device.
As the namespace claiming mechanism for PFN devices is mostly identical
to BTT devices drivers/nvdimm/claim.c is created to house the common
bits.
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
This should result in a pretty sizeable performance gain for reads. For
rough comparison I did some simple read testing using PMEM to compare
reads of write combining (WC) mappings vs write-back (WB). This was
done on a random lab machine.
PMEM reads from a write combining mapping:
# dd of=/dev/null if=/dev/pmem0 bs=4096 count=100000
100000+0 records in
100000+0 records out
409600000 bytes (410 MB) copied, 9.2855 s, 44.1 MB/s
PMEM reads from a write-back mapping:
# dd of=/dev/null if=/dev/pmem0 bs=4096 count=1000000
1000000+0 records in
1000000+0 records out
4096000000 bytes (4.1 GB) copied, 3.44034 s, 1.2 GB/s
To be able to safely support a write-back aperture I needed to add
support for the "read flush" _DSM flag, as outlined in the DSM spec:
http://pmem.io/documents/NVDIMM_DSM_Interface_Example.pdf
This flag tells the ND BLK driver that it needs to flush the cache lines
associated with the aperture after the aperture is moved but before any
new data is read. This ensures that any stale cache lines from the
previous contents of the aperture will be discarded from the processor
cache, and the new data will be read properly from the DIMM. We know
that the cache lines are clean and will be discarded without any
writeback because either a) the previous aperture operation was a read,
and we never modified the contents of the aperture, or b) the previous
aperture operation was a write and we must have written back the dirtied
contents of the aperture to the DIMM before the I/O was completed.
In order to add support for the "read flush" flag I needed to add a
generic routine to invalidate cache lines, mmio_flush_range(). This is
protected by the ARCH_HAS_MMIO_FLUSH Kconfig variable, and is currently
only supported on x86.
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
We currently register a platform device for e820 type-12 memory and
register a nvdimm bus beneath it. Registering the platform device
triggers the device-core machinery to probe for a driver, but that
search currently comes up empty. Building the nvdimm-bus registration
into the e820_pmem platform device registration in this way forces
libnvdimm to be built-in. Instead, convert the built-in portion of
CONFIG_X86_PMEM_LEGACY to simply register a platform device and move the
rest of the logic to the driver for e820_pmem, for the following
reasons:
1/ Letting e820_pmem support be a module allows building and testing
libnvdimm.ko changes without rebooting
2/ All the normal policy around modules can be applied to e820_pmem
(unbind to disable and/or blacklisting the module from loading by
default)
3/ Moving the driver to a generic location and converting it to scan
"iomem_resource" rather than "e820.map" means any other architecture can
take advantage of this simple nvdimm resource discovery mechanism by
registering a resource named "Persistent Memory (legacy)"
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
[djbw: tools/testing/nvdimm/ and memunmap_pmem support]
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Kill arch_memremap_pmem() and just let the architecture specify the
flags to be passed to memremap(). Default to writethrough by default.
Suggested-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Add support for the three ARS DSM commands:
- Query ARS Capabilities - Queries the firmware to check if a given
range supports scrub, and if so, which type (persistent vs. volatile)
- Start ARS - Starts a scrub for a given range/type
- Query ARS Status - Checks status of a previously started scrub, and
provides the error logs if any.
The commands are described by the example DSM spec at:
http://pmem.io/documents/NVDIMM_DSM_Interface_Example.pdf
Also add these commands to the nfit_test test framework, and return
canned data.
Signed-off-by: Vishal Verma <vishal.l.verma@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
In preparation for fixing the BLK path to properly use "directed
pcommit" enable the unit test infrastructure to emit mock "flush"
tables. Writes to these flush addresses trigger a memory controller to
flush its internal buffers to persistent media, similar to the x86
"pcommit" instruction.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
The implementation for the new "DIMM Flags" DSM relies on the -ENOTTY
return code to indicate that the flags are unimplimented and to fall
back to a safe default. As is the -ENXIO error code erroneoously
indicates to fail enabling a BLK region.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
In the 4.2-rc1 merge the default_memremap_pmem() implementation switched
from ioremap_nocache() to ioremap_wt(). Add it to the list of mocked
routines to restore the ability to run the unit tests.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Upon detection of an unarmed dimm in a region, arrange for descendant
BTT, PMEM, or BLK instances to be read-only. A dimm is primarily marked
"unarmed" via flags passed by platform firmware (NFIT).
The flags in the NFIT memory device sub-structure indicate the state of
the data on the nvdimm relative to its energy source or last "flush to
persistence". For the most part there is nothing the driver can do but
advertise the state of these flags in sysfs and emit a message if
firmware indicates that the contents of the device may be corrupted.
However, for the case of ACPI_NFIT_MEM_ARMED, the driver can arrange for
the block devices incorporating that nvdimm to be marked read-only.
This is a safe default as the data is still available and new writes are
held off until the administrator either forces read-write mode, or the
energy source becomes armed.
A 'read_only' attribute is added to REGION devices to allow for
overriding the default read-only policy of all descendant block devices.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
'libnvdimm' is the first driver sub-system in the kernel to implement
mocking for unit test coverage. The nfit_test module gets built as an
external module and arranges for external module replacements of nfit,
libnvdimm, nd_pmem, and nd_blk. These replacements use the linker
--wrap option to redirect calls to ioremap() + request_mem_region() to
custom defined unit test resources. The end result is a fully
functional nvdimm_bus, as far as userspace is concerned, but with the
capability to perform otherwise destructive tests on emulated resources.
Q: Why not use QEMU for this emulation?
QEMU is not suitable for unit testing. QEMU's role is to faithfully
emulate the platform. A unit test's role is to unfaithfully implement
the platform with the goal of triggering bugs in the corners of the
sub-system implementation. As bugs are discovered in platforms, or the
sub-system itself, the unit tests are extended to backstop a fix with a
reproducer unit test.
Another problem with QEMU is that it would require coordination of 3
software projects instead of 2 (kernel + libndctl [1]) to maintain and
execute the tests. The chances for bit rot and the difficulty of
getting the tests running goes up non-linearly the more components
involved.
Q: Why submit this to the kernel tree instead of external modules in
libndctl?
Simple, to alleviate the same risk that out-of-tree external modules
face. Updates to drivers/nvdimm/ can be immediately evaluated to see if
they have any impact on tools/testing/nvdimm/.
Q: What are the negative implications of merging this?
It is a unique maintenance burden because the purpose of mocking an
interface to enable a unit test is to purposefully short circuit the
semantics of a routine to enable testing. For example
__wrap_ioremap_cache() fakes the pmem driver into "ioremap()'ing" a test
resource buffer allocated by dma_alloc_coherent(). The future
maintenance burden hits when someone changes the semantics of
ioremap_cache() and wonders what the implications are for the unit test.
[1]: https://github.com/pmem/ndctl
Cc: <linux-acpi@vger.kernel.org>
Cc: Lv Zheng <lv.zheng@intel.com>
Cc: Robert Moore <robert.moore@intel.com>
Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Dan Williams
Ross Zwisler
Christoph Hellwig
Vishal Verma