New Changes:
- Refactor to a common struct for DRAM and general media CXL events
- Add abstract distance calculation support for CXL
- Add CXL maturity map documentation to detail current state of CXL enabling
- Add warning on mixed CXL VH and RCH/RCD hierachy to inform unsupported config
- Replace ENXIO with EBUSY for inject poison limit reached via debugfs
- Replace ENXIO with EBUSY for inject poison cxl-test support
- XOR math fixup for DPA to SPA translation. Current math works for MODULO arithmetic
where HPA==SPA, however not for XOR decode.
- Move pci config read in cxl_dvsec_rr_decode() to avoid unnecessary acess
Fixes:
- Add a fix to address race condition in CXL memory hotplug notifier
- Add missing MODULE_DESCRIPTION() for CXL modules
- Fix incorrect vendor debug UUID define
-----BEGIN PGP SIGNATURE-----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=HaEG
-----END PGP SIGNATURE-----
Merge tag 'cxl-for-6.11' of git://git.kernel.org/pub/scm/linux/kernel/git/cxl/cxl
Pull CXL updates from Dave Jiang:
"Core:
- A CXL maturity map has been added to the documentation to detail
the current state of CXL enabling.
It provides the status of the current state of various CXL features
to inform current and future contributors of where things are and
which areas need contribution.
- A notifier handler has been added in order for a newly created CXL
memory region to trigger the abstract distance metrics calculation.
This should bring parity for CXL memory to the same level vs
hotplugged DRAM for NUMA abstract distance calculation. The
abstract distance reflects relative performance used for memory
tiering handling.
- An addition for XOR math has been added to address the CXL DPA to
SPA translation.
CXL address translation did not support address interleave math
with XOR prior to this change.
Fixes:
- Fix to address race condition in the CXL memory hotplug notifier
- Add missing MODULE_DESCRIPTION() for CXL modules
- Fix incorrect vendor debug UUID define
Misc:
- A warning has been added to inform users of an unsupported
configuration when mixing CXL VH and RCH/RCD hierarchies
- The ENXIO error code has been replaced with EBUSY for inject poison
limit reached via debugfs and cxl-test support
- Moving the PCI config read in cxl_dvsec_rr_decode() to avoid
unnecessary PCI config reads
- A refactor to a common struct for DRAM and general media CXL
events"
* tag 'cxl-for-6.11' of git://git.kernel.org/pub/scm/linux/kernel/git/cxl/cxl:
cxl/core/pci: Move reading of control register to immediately before usage
cxl: Remove defunct code calculating host bridge target positions
cxl/region: Verify target positions using the ordered target list
cxl: Restore XOR'd position bits during address translation
cxl/core: Fold cxl_trace_hpa() into cxl_dpa_to_hpa()
cxl/test: Replace ENXIO with EBUSY for inject poison limit reached
cxl/memdev: Replace ENXIO with EBUSY for inject poison limit reached
cxl/acpi: Warn on mixed CXL VH and RCH/RCD Hierarchy
cxl/core: Fix incorrect vendor debug UUID define
Documentation: CXL Maturity Map
cxl/region: Simplify cxl_region_nid()
cxl/region: Support to calculate memory tier abstract distance
cxl/region: Fix a race condition in memory hotplug notifier
cxl: add missing MODULE_DESCRIPTION() macros
cxl/events: Use a common struct for DRAM and General Media events
Series to fix XOR math for DPA to SPA translation
- Refactor and fold cxl_trace_hpa() into cxl_dpa_to_hpa()
- Complete DPA->HPA->SPA translation and correct XOR translation issue
- Add new method to verify a CXL target position
- Remove old method of CXL target position verifiation
When a root decoder is configured the interleave target list is read
from the BIOS populated CFMWS structure. Per the CXL spec 3.1 Table
9-22 the target list is in interleave order. The CXL driver populates
its decoder target list in the same order and stores it in 'struct
cxl_switch_decoder' field "@target: active ordered target list in
current decoder configuration"
Given the promise of an ordered list, the driver can stop duplicating
the work of BIOS and simply check target positions against the ordered
list during region configuration.
The simplified check against the ordered list is presented here.
A follow-on patch will remove the unused code.
For Modulo arithmetic this is not a fix, only a simplification.
For XOR arithmetic this is a fix for HB IW of 3,6,12.
Fixes: f9db85bfec ("cxl/acpi: Support CXL XOR Interleave Math (CXIMS)")
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://patch.msgid.link/35d08d3aba08fee0f9b86ab1cef0c25116ca8a55.1719980933.git.alison.schofield@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
When a device reports a DPA in events like poison, general_media,
and dram, the driver translates that DPA back to an HPA. Presently,
the CXL driver translation only considers the Modulo position and
will report the wrong HPA for XOR configured root decoders.
Add a helper function that restores the XOR'd bits during DPA->HPA
address translation. Plumb a root decoder callback to the new helper
when XOR interleave arithmetic is in use. For Modulo arithmetic, just
let the callback be NULL - as in no extra work required.
Upon completion of a DPA->HPA translation a couple of checks are
performed on the result. One simply confirms that the calculated
HPA is within the address range of the region. That test is useful
for both Modulo and XOR interleave arithmetic decodes.
A second check confirms that the HPA is within an expected chunk
based on the endpoints position in the region and the region
granularity. An XOR decode disrupts the Modulo pattern making the
chunk check useless.
To align the checks with the proper decode, pull the region range
check inline and use the helper to do the chunk check for Modulo
decodes only.
A cxl-test unit test is posted for upstream review here:
https://lore.kernel.org/20240624210644.495563-1-alison.schofield@intel.com/
Fixes: 28a3ae4ff6 ("cxl/trace: Add an HPA to cxl_poison trace events")
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Tested-by: Diego Garcia Rodriguez <diego.garcia.rodriguez@intel.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Link: https://patch.msgid.link/1a1ac880d9f889bd6384e657e810431b9a0a72e5.1719980933.git.alison.schofield@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Although cxl_trace_hpa() is used to populate TRACE EVENTs with HPA
addresses the work it performs is a DPA to HPA translation not a
trace. Tidy up this naming by moving the minimal work done in
cxl_trace_hpa() into cxl_dpa_to_hpa() and use cxl_dpa_to_hpa()
for trace event callbacks.
Suggested-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Robert Richter <rrichter@amd.com>
Link: https://patch.msgid.link/452a9b0c525b774c72d9d5851515ffa928750132.1719980933.git.alison.schofield@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
The node ID of the region can be gotten via resource start address
directly. This simplifies the implementation of cxl_region_nid().
Signed-off-by: Huang Ying <ying.huang@intel.com>
Suggested-by: Alison Schofield <alison.schofield@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Bharata B Rao <bharata@amd.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://patch.msgid.link/20240618084639.1419629-4-ying.huang@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
An abstract distance value must be assigned by the driver that makes
the memory available to the system. It reflects relative performance
and is used to place memory nodes backed by CXL regions in the appropriate
memory tiers allowing promotion/demotion within the existing memory tiering
mechanism.
The abstract distance is calculated based on the memory access latency
and bandwidth of CXL regions.
Signed-off-by: Huang, Ying <ying.huang@intel.com>
Acked-by: Dan Williams <dan.j.williams@intel.com>
Cc: Alison Schofield <alison.schofield@intel.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Bharata B Rao <bharata@amd.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://patch.msgid.link/20240618084639.1419629-3-ying.huang@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
In the memory hotplug notifier function of the CXL region,
cxl_region_perf_attrs_callback(), the node ID is obtained by checking
the host address range of the region. However, the address range
information is not available when the region is registered in
devm_cxl_add_region(). Additionally, this information may be removed
or added under the protection of cxl_region_rwsem during runtime. If
the memory notifier is called for nodes other than that backed by the
region, a race condition may occur, potentially leading to a NULL
dereference or an invalid address range.
The race condition is addressed by checking the availability of the
address range information under the protection of cxl_region_rwsem. To
enhance code readability and use guard(), the relevant code has been
moved into a newly added function: cxl_region_nid().
Fixes: 067353a46d ("cxl/region: Add memory hotplug notifier for cxl region")
Signed-off-by: Huang, Ying <ying.huang@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Alison Schofield <alison.schofield@intel.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Bharata B Rao <bharata@amd.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://patch.msgid.link/20240618084639.1419629-2-ying.huang@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Since interleave capability is not verified, if the interleave
capability of a target does not match the region need, committing decoder
should have failed at the device end.
In order to checkout this error as quickly as possible, driver needs
to check the interleave capability of target during attaching it to
region.
Per CXL specification r3.1(8.2.4.20.1 CXL HDM Decoder Capability Register),
bits 11 and 12 indicate the capability to establish interleaving in 3, 6,
12 and 16 ways. If these bits are not set, the target cannot be attached to
a region utilizing such interleave ways.
Additionally, bits 8 and 9 represent the capability of the bits used for
interleaving in the address, Linux tracks this in the cxl_port
interleave_mask.
Per CXL specification r3.1(8.2.4.20.13 Decoder Protection):
eIW means encoded Interleave Ways.
eIG means encoded Interleave Granularity.
in HPA:
if eIW is 0 or 8 (interleave ways: 1, 3), all the bits of HPA are used,
the interleave bits are none, the following check is ignored.
if eIW is less than 8 (interleave ways: 2, 4, 8, 16), the interleave bits
start at bit position eIG + 8 and end at eIG + eIW + 8 - 1.
if eIW is greater than 8 (interleave ways: 6, 12), the interleave bits
start at bit position eIG + 8 and end at eIG + eIW - 1.
if the interleave mask is insufficient to cover the required interleave
bits, the target cannot be attached to the region.
Fixes: 384e624bb2 ("cxl/region: Attach endpoint decoders")
Signed-off-by: Yao Xingtao <yaoxt.fnst@fujitsu.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://patch.msgid.link/20240614084755.59503-2-yaoxt.fnst@fujitsu.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
cxl_dpa_to_region() looks up a region based on a memdev and DPA.
It wrongly assumes an endpoint found mapping the DPA is also of
a fully assembled region. When not true it leads to a null pointer
dereference looking up the region name.
This appears during testing of region lookup after a failure to
assemble a BIOS defined region or if the lookup raced with the
assembly of the BIOS defined region.
Failure to clean up BIOS defined regions that fail assembly is an
issue in itself and a fix to that problem will alleviate some of
the impact. It will not alleviate the race condition so let's harden
this path.
The behavior change is that the kernel oops due to a null pointer
dereference is replaced with a dev_dbg() message noting that an
endpoint was mapped.
Additional comments are added so that future users of this function
can more clearly understand what it provides.
Fixes: 0a105ab28a ("cxl/memdev: Warn of poison inject or clear to a mapped region")
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://patch.msgid.link/20240604003609.202682-1-alison.schofield@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
When CXL subsystem is auto-assembling a pmem region during cxl
endpoint port probing, always hit below calltrace.
BUG: kernel NULL pointer dereference, address: 0000000000000078
#PF: supervisor read access in kernel mode
#PF: error_code(0x0000) - not-present page
RIP: 0010:cxl_pmem_region_probe+0x22e/0x360 [cxl_pmem]
Call Trace:
<TASK>
? __die+0x24/0x70
? page_fault_oops+0x82/0x160
? do_user_addr_fault+0x65/0x6b0
? exc_page_fault+0x7d/0x170
? asm_exc_page_fault+0x26/0x30
? cxl_pmem_region_probe+0x22e/0x360 [cxl_pmem]
? cxl_pmem_region_probe+0x1ac/0x360 [cxl_pmem]
cxl_bus_probe+0x1b/0x60 [cxl_core]
really_probe+0x173/0x410
? __pfx___device_attach_driver+0x10/0x10
__driver_probe_device+0x80/0x170
driver_probe_device+0x1e/0x90
__device_attach_driver+0x90/0x120
bus_for_each_drv+0x84/0xe0
__device_attach+0xbc/0x1f0
bus_probe_device+0x90/0xa0
device_add+0x51c/0x710
devm_cxl_add_pmem_region+0x1b5/0x380 [cxl_core]
cxl_bus_probe+0x1b/0x60 [cxl_core]
The cxl_nvd of the memdev needs to be available during the pmem region
probe. Currently the cxl_nvd is registered after the endpoint port probe.
The endpoint probe, in the case of autoassembly of regions, can cause a
pmem region probe requiring the not yet available cxl_nvd. Adjust the
sequence so this dependency is met.
This requires adding a port parameter to cxl_find_nvdimm_bridge() that
can be used to query the ancestor root port. The endpoint port is not
yet available, but will share a common ancestor with its parent, so
start the query from there instead.
Fixes: f17b558d66 ("cxl/pmem: Refactor nvdimm device registration, delete the workqueue")
Co-developed-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Li Ming <ming4.li@intel.com>
Tested-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Alison Schofield <alison.schofield@intel.com>
Link: https://patch.msgid.link/20240612064423.2567625-1-ming4.li@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Move the mode verification to __create_region() before allocating the
memregion to avoid the memregion leaks.
Fixes: 6e09926418 ("cxl/region: Add volatile region creation support")
Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Link: https://lore.kernel.org/r/20240507053421.456439-1-lizhijian@fujitsu.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
A recent bugfix to cxl_pmem_region_alloc() to fix an
error-unwind-memleak [1], highlighted a use case for scope-based resource
management.
Delete the goto for releasing @cxl_region_rwsem, and return error codes
directly from error condition paths.
The caller, devm_cxl_add_pmem_region(), is no longer given @cxlr_pmem
directly it must retrieve it from @cxlr->cxlr_pmem. This retrieval from
@cxlr was already in place for @cxlr->cxl_nvb, and converting
cxl_pmem_region_alloc() to return an int makes it less awkward to handle
no_free_ptr().
Cc: Li Zhijian <lizhijian@fujitsu.com>
Reported-by: Jonathan Cameron <Jonathan.Cameron@Huawei.com>
Closes: http://lore.kernel.org/r/20240430174540.000039ce@Huawei.com
Link: http://lore.kernel.org/r/20240428030748.318985-1-lizhijian@fujitsu.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/171451430965.1147997.15782562063090960666.stgit@dwillia2-xfh.jf.intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Before this error path, cxlr_pmem pointed to a kzalloc() memory, free
it to avoid this memory leaking.
Fixes: f17b558d66 ("cxl/pmem: Refactor nvdimm device registration, delete the workqueue")
Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20240428030748.318985-1-lizhijian@fujitsu.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
This work belongs in the region driver as it is only useful with
CONFIG_CXL_REGION. Add a stub in core.h for when the region driver
is not built.
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Link: https://lore.kernel.org/r/183222631f11a43c5e6debc42ec22fe1bd4b818a.1714496730.git.alison.schofield@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
This helper belongs in the region driver as it is only useful
with CONFIG_CXL_REGION. Add a stub in core.h for when the region
driver is not built.
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Link: https://lore.kernel.org/r/05e30f788d62b3dd398aff2d2ea50a6aaa7c3313.1714496730.git.alison.schofield@intel.com
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
For the numa nodes that are not created by SRAT, no memory_target is
allocated and is not managed by the HMAT_REPORTING code. Therefore
hmat_callback() memory hotplug notifier will exit early on those NUMA
nodes. The CXL memory hotplug notifier will need to call
node_set_perf_attrs() directly in order to setup the access sysfs
attributes.
In acpi_numa_init(), the last proximity domain (pxm) id created by SRAT is
stored. Add a helper function acpi_node_backed_by_real_pxm() in order to
check if a NUMA node id is defined by SRAT or created by CFMWS.
node_set_perf_attrs() symbol is exported to allow update of perf attribs
for a node. The sysfs path of
/sys/devices/system/node/nodeX/access0/initiators/* is created by
node_set_perf_attrs() for the various attributes where nodeX is matched
to the NUMA node of the CXL region.
Cc: Rafael J. Wysocki <rafael@kernel.org>
Reviewed-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/20240308220055.2172956-13-dave.jiang@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
When the CXL region is formed, the driver computes the performance data
for the region. However this data is not available at the node data
collection that has been populated by the HMAT during kernel
initialization. Add a memory hotplug notifier to update the access
coordinates to the 'struct memory_target' context kept by the
HMAT_REPORTING code.
Add CXL_CALLBACK_PRI for a memory hotplug callback priority. Set the
priority number to be called before HMAT_CALLBACK_PRI. The CXL update must
happen before hmat_callback().
A new HMAT_REPORTING helper hmat_update_target_coordinates() is added in
order to allow CXL to update the memory_target access coordinates.
A new ext_updated member is added to the memory_target to indicate that
the access coordinates within the memory_target has been updated by an
external agent such as CXL. This prevents data being overwritten by the
hmat_update_target_attrs() triggered by hmat_callback().
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Rafael J. Wysocki <rafael@kernel.org>
Reviewed-by: Huang, Ying <ying.huang@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/20240308220055.2172956-12-dave.jiang@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Add read/write latencies and bandwidth sysfs attributes for the enabled CXL
region. The bandwidth is the aggregated bandwidth of all devices that
contribute to the CXL region. The latency is the worst latency of the
device amongst all the devices that contribute to the CXL region.
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/20240308220055.2172956-11-dave.jiang@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Calculate and store the performance data for a CXL region. Find the worst
read and write latency for all the included ranges from each of the devices
that attributes to the region and designate that as the latency data. Sum
all the read and write bandwidth data for each of the device region and
that is the total bandwidth for the region.
The perf list is expected to be constructed before the endpoint decoders
are registered and thus there should be no early reading of the entries
from the region assemble action. The calling of the region qos calculate
function is under the protection of cxl_dpa_rwsem and will ensure that
all DPA associated work has completed.
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Tested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/20240308220055.2172956-10-dave.jiang@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Autodiscovered regions can fail to assemble if they are not discovered
in HPA decode order. The user will see failure messages like:
[] cxl region0: endpoint5: HPA order violation region1
[] cxl region0: endpoint5: failed to allocate region reference
The check that is causing the failure helps the CXL driver enforce
a CXL spec mandate that decoders be committed in HPA order. The
check is needless for autodiscovered regions since their decoders
are already programmed. Trying to enforce order in the assembly of
these regions is useless because they are assembled once all their
member endpoints arrive, and there is no guarantee on the order in
which endpoints are discovered during probe.
Keep the existing check, but for autodiscovered regions, allow the
out of order assembly after a sanity check that the lesser numbered
decoder has the lesser HPA starting address.
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Tested-by: Wonjae Lee <wj28.lee@samsung.com>
Link: https://lore.kernel.org/r/3dec69ee97524ab229a20c6739272c3000b18408.1706736863.git.alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
In preparation for adding a new caller of cxl_region_find_decoders()
teach it to find a decoder from a cxl_endpoint_decoder structure.
Combining switch and endpoint decoder lookup in one function prevents
code duplication in call sites.
Update the existing caller.
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Tested-by: Wonjae Lee <wj28.lee@samsung.com>
Link: https://lore.kernel.org/r/79ae6d72978ef9f3ceec9722e1cb793820553c8e.1706736863.git.alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Creating a region with 16 memory devices caused a problem. The div_u64_rem
function, used for dividing an unsigned 64-bit number by a 32-bit one,
faced an issue when SZ_256M * p->interleave_ways. The result surpassed
the maximum limit of the 32-bit divisor (4G), leading to an overflow
and a remainder of 0.
note: At this point, p->interleave_ways is 16, meaning 16 * 256M = 4G
To fix this issue, I replaced the div_u64_rem function with div64_u64_rem
and adjusted the type of the remainder.
Signed-off-by: Quanquan Cao <caoqq@fujitsu.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Fixes: 23a22cd1c9 ("cxl/region: Allocate HPA capacity to regions")
Cc: <stable@vger.kernel.org>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
CXL supports x3, x6 and x12 - not x9.
Fixes: 80d10a6cee ("cxl/region: Add interleave geometry attributes")
Signed-off-by: Jim Harris <jim.harris@samsung.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Fan Ni <fan.ni@samsung.com>
Link: https://lore.kernel.org/r/169904271254.204936.8580772404462743630.stgit@ubuntu
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Use "%pap" to print a resource_size_t (phys_addr_t derived type)
to prevent build warnings on 32-bit arches (seen on i386 and
riscv-32).
../drivers/cxl/core/region.c: In function 'alloc_hpa':
../drivers/cxl/core/region.c:556:25: warning: format '%llx' expects argument of type 'long long unsigned int', but argument 5 has type 'resource_size_t' {aka 'unsigned int'} [-Wformat=]
556 | "HPA allocation error (%ld) for size:%#llx in %s %pr\n",
Fixes: 7984d22f13 ("cxl/region: Add dev_dbg() detail on failure to allocate HPA space")
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Fan Ni <fan.ni@samsung.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Jonathan Cameron <jonathan.cameron@huawei.com>
Cc: Dave Jiang <dave.jiang@intel.com>
Cc: Alison Schofield <alison.schofield@intel.com>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Cc: Ira Weiny <ira.weiny@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: <linux-cxl@vger.kernel.org>
Link: https://lore.kernel.org/r/20240102173917.19718-1-rdunlap@infradead.org
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
When the region driver fails while allocating HPA space for a
new region it can be because the parent resource, the CXL Window,
has no more available space.
In that case, the debug user sees this message:
cxl_core:alloc_hpa:555: cxl region2: failed to allocate HPA: -34
Expand the message like this:
cxl_core:alloc_hpa:555: cxl region8: HPA allocation error (-34) for size:0x20000000 in CXL Window 0 [mem 0xf010000000-0xf04fffffff flags 0x200]
Now the debug user can examine /proc/iomem and consider actions
like removing other allocations in that space or reducing the
size of their region request.
Suggested-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Vishal Verma <vishal.l.verma@intel.com>
Reviewed-by: Fan Ni <fan.ni@samsung.com>
Link: https://lore.kernel.org/r/20231223004740.1401858-1-alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
In devm_cxl_add_region(), devm_add_action_or_reset() is called by
passing in unregister_region() with data ptr of 'cxlr'. However, in
unregister_region(), the passed in parameter is incorrectly assumed to
be a 'struct device' rather than the 'cxlr' pointer. The code has been
working because 'struct device' is the first member of 'struct
cxl_region'. Issue found by inspection. Fix the assignment so that cxlr
is pointing directly to the passed in parameter.
Not flagged for -stable since there is no functional impact of this fix.
Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Link: https://lore.kernel.org/r/170258123810.952211.3907381447996426480.stgit@djiang5-mobl3
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
A read of a device poison list is triggered via a sysfs attribute
and the results are logged as kernel trace events of type cxl_poison.
The work is managed by either: a) the region driver when one of more
regions map the device, or by b) the memdev driver when no regions
map the device.
In the case of a) the region driver holds the region_rwsem while
reading the poison by committed endpoint decoder mappings and for
any unmapped resources. This makes sure that the cxl_poison trace
event trace reports valid region info. (Region name, HPA, and UUID).
In the case of b) the memdev driver holds the dpa_rwsem preventing
new DPA resources from being attached to a region. However, it leaves
a gap between region attach and decoder commit actions. If a DPA in
the gap is in the poison list, the cxl_poison trace event will omit
the region info.
Close the gap by holding the region_rwsem and the dpa_rwsem when
reading poison per memdev. Since both methods now hold both locks,
down_read both from the caller. Doing so also addresses the lockdep
assert that found this issue:
Commit 458ba8189c ("cxl: Add cxl_decoders_committed() helper")
Fixes: f0832a5863 ("cxl/region: Provide region info to the cxl_poison trace event")
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/08e8e7ec9a3413b91d51de39e385653494b1eed0.1701041440.git.alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Dan reports that cxl_decoder_commit() potentially leaks a hold of
cxl_dpa_rwsem. The potential error case is a "should not" happen
scenario, turn it into a "can not" happen scenario by adding the error
check to cxl_port_setup_targets() where other setting validation occurs.
Reported-by: Dan Carpenter <dan.carpenter@linaro.org>
Closes: http://lore.kernel.org/r/63295673-5d63-4919-b851-3b06d48734c0@moroto.mountain
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Fixes: 176baefb2e ("cxl/hdm: Commit decoder state to hardware")
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Root decoder granularity must match value from CFWMS, which may not
be the region's granularity for non-interleaved root decoders.
So when calculating granularities for host bridge decoders, use the
region's granularity instead of the root decoder's granularity to ensure
the correct granularities are set for the host bridge decoders and any
downstream switch decoders.
Test configuration is 1 host bridge * 2 switches * 2 endpoints per switch.
Region created with 2048 granularity using following command line:
cxl create-region -m -d decoder0.0 -w 4 mem0 mem2 mem1 mem3 \
-g 2048 -s 2048M
Use "cxl list -PDE | grep granularity" to get a view of the granularity
set at each level of the topology.
Before this patch:
"interleave_granularity":2048,
"interleave_granularity":2048,
"interleave_granularity":512,
"interleave_granularity":2048,
"interleave_granularity":2048,
"interleave_granularity":512,
"interleave_granularity":256,
After:
"interleave_granularity":2048,
"interleave_granularity":2048,
"interleave_granularity":4096,
"interleave_granularity":2048,
"interleave_granularity":2048,
"interleave_granularity":4096,
"interleave_granularity":2048,
Fixes: 27b3f8d138 ("cxl/region: Program target lists")
Cc: <stable@vger.kernel.org>
Signed-off-by: Jim Harris <jim.harris@samsung.com>
Link: https://lore.kernel.org/r/169824893473.1403938.16110924262989774582.stgit@bgt-140510-bm03.eng.stellus.in
[djbw: fixup the prebuilt cxl_test region]
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Fix a missed "goto out" to unlock on error to cleanup this splat:
WARNING: lock held when returning to user space!
6.6.0-rc3-lizhijian+ #213 Not tainted
------------------------------------------------
cxl/673 is leaving the kernel with locks still held!
1 lock held by cxl/673:
#0: ffffffffa013b9d0 (cxl_region_rwsem){++++}-{3:3}, at: commit_store+0x7d/0x3e0 [cxl_core]
In terms of user visible impact of this bug for backports:
cxl_region_invalidate_memregion() on x86 invokes wbinvd which is a
problematic instruction for virtualized environments. So, on virtualized
x86, cxl_region_invalidate_memregion() returns an error. This failure
case got missed because CXL memory-expander device passthrough is not a
production use case, and emulation of CXL devices is typically limited
to kernel development builds with CONFIG_CXL_REGION_INVALIDATION_TEST=y,
that makes cxl_region_invalidate_memregion() succeed.
In other words, the expected exposure of this bug is limited to CXL
subsystem development environments using QEMU that neglected
CONFIG_CXL_REGION_INVALIDATION_TEST=y.
Fixes: d1257d098a ("cxl/region: Move cache invalidation before region teardown, and before setup")
Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Link: https://lore.kernel.org/r/20231025085450.2514906-1-lizhijian@fujitsu.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
For auto-discovered regions the driver must assign each target to
a valid position in the region interleave set based on the decoder
topology.
The current implementation fails to parse valid decode topologies,
as it does not consider the child offset into a parent port. The sort
put all targets of one port ahead of another port when an interleave
was expected, causing the region assembly to fail.
Replace the existing relative sort with cxl_calc_interleave_pos() that
finds the exact position in a region interleave for an endpoint based
on a walk up the ancestral tree from endpoint to root decoder.
cxl_calc_interleave_pos() was introduced in a prior patch, so the work
here is to use it in cxl_region_sort_targets().
Remove the obsoleted helper functions from the prior sort.
Testing passes on pre-production hardware with BIOS defined regions
that natively trigger this autodiscovery path of the region driver.
Testing passes a CXL unit test using the dev_dbg() calculation test
(see cxl_region_attach()) across an expanded set of region configs:
1, 1, 1+1, 1+1+1, 2, 2+2, 2+2+2, 2+2+2+2, 4, 4+4, where each number
represents the count of endpoints per host bridge.
Fixes: a32320b71f ("cxl/region: Add region autodiscovery")
Reported-by: Dmytro Adamenko <dmytro.adamenko@intel.com>
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Jim Harris <jim.harris@samsung.com>
Link: https://lore.kernel.org/r/3946cc55ddc19678733eddc9de2c317749f43f3b.1698263080.git.alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Introduce a calculation to find a target's position in a region
interleave. Perform a self-test of the calculation on user-defined
regions.
The region driver uses the kernel sort() function to put region
targets in relative order. Positions are assigned based on each
target's index in that sorted list. That relative sort doesn't
consider the offset of a port into its parent port which causes
some auto-discovered regions to fail creation. In one failure case,
a 2 + 2 config (2 host bridges each with 2 endpoints), the sort
puts all the targets of one port ahead of another port when they
were expected to be interleaved.
In preparation for repairing the autodiscovery region assembly,
introduce a new method for discovering a target position in the
region interleave.
cxl_calc_interleave_pos() adds a method to find the target position by
ascending from an endpoint to a root decoder. The calculation starts
with the endpoint's local position and position in the parent port. It
traverses towards the root decoder and examines both position and ways
in order to allow the position to be refined all the way to the root
decoder.
This calculation: position = position * parent_ways + parent_pos;
applied iteratively yields the correct position.
Include a self-test that exercises this new position calculation against
every successfully configured user-defined region.
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Link: https://lore.kernel.org/r/0ac32c75cf81dd8b86bf07d70ff139d33c2300bc.1698263080.git.alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
match_decoder_by_range() and decoder_match_range() both determine
if an HPA range matches a decoder. The first does it for root
decoders and the second one operates on switch decoders.
Tidy these up with clear naming and make the switch helper more
like the root decoder helper in style and functionality. Make it
take the actual range, rather than an endpoint decoder from which
it extracts the range. Require an exact match on switch decoders,
because unlike a root decoder that maps an entire region, Linux
only supports 1:1 mapping of switch to endpoint decoders. Note that
root-decoders are a super-set of switch-decoders and the range they
cover is a super-set of a region, hence the use of range_contains() for
that case.
Aside from aesthetics and maintainability, this is in preparation
for reuse.
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Jim Harris <jim.harris@samsung.com>
Link: https://lore.kernel.org/r/011b1f498e1758bb8df17c5951be00bd8d489e3b.1698263080.git.alison.schofield@intel.com
[djbw: fixup root decoder vs switch decoder range checks]
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Commit 5e42bcbc3f ("cxl/region: decrement ->nr_targets on error in
cxl_region_attach()") tried to avoid 'eiw' initialization errors when
->nr_targets exceeded 16, by just decrementing ->nr_targets when
cxl_region_setup_targets() failed.
Commit 86987c7662 ("cxl/region: Cleanup target list on attach error")
extended that cleanup to also clear cxled->pos and p->targets[pos]. The
initialization error was incidentally fixed separately by:
Commit 8d42854257 ("cxl/region: Fix port setup uninitialized variable
warnings") which was merged a few days after 5e42bcbc3f.
But now the original cleanup when cxl_region_setup_targets() fails
prevents endpoint and switch decoder resources from being reused:
1) the cleanup does not set the decoder's region to NULL, which results
in future dpa_size_store() calls returning -EBUSY
2) the decoder is not properly freed, which results in future commit
errors associated with the upstream switch
Now that the initialization errors were fixed separately, the proper
cleanup for this case is to just return immediately. Then the resources
associated with this target get cleanup up as normal when the failed
region is deleted.
The ->nr_targets decrement in the error case also helped prevent
a p->targets[] array overflow, so add a new check to prevent against
that overflow.
Tested by trying to create an invalid region for a 2 switch * 2 endpoint
topology, and then following up with creating a valid region.
Fixes: 5e42bcbc3f ("cxl/region: decrement ->nr_targets on error in cxl_region_attach()")
Cc: <stable@vger.kernel.org>
Signed-off-by: Jim Harris <jim.harris@samsung.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Acked-by: Dan Carpenter <dan.carpenter@linaro.org>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/169703589120.1202031.14696100866518083806.stgit@bgt-140510-bm03.eng.stellus.in
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
The sanitize operation is destructive and the expectation is that the
device is unmapped while in progress. The current implementation does a
lockless check for decoders being active, but then does nothing to
prevent decoders from racing to be committed. Introduce state tracking
to resolve this race.
This incidentally cleans up unpriveleged userspace from triggering mmio
read cycles by spinning on reading the 'security/state' attribute. Which
at a minimum is a waste since the kernel state machine can cache the
completion result.
Lastly cxl_mem_sanitize() was mistakenly marked EXPORT_SYMBOL() in the
original implementation, but an export was never required.
Fixes: 0c36b6ad43 ("cxl/mbox: Add sanitization handling machinery")
Cc: Davidlohr Bueso <dave@stgolabs.net>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
The cxl_test platform device CXL port hierarchy is useful for testing,
but throws warning messages of the form:
cxl_mem mem2: at cxl_root_port.1 no parent for dport: platform
cxl_mem mem3: at cxl_root_port.2 no parent for dport: platform
cxl_mem mem4: at cxl_root_port.3 no parent for dport: platform
cxl_mem mem5: at cxl_root_port.0 no parent for dport: platform
cxl_mem mem6: at cxl_root_port.1 no parent for dport: platform
cxl_mem mem7: at cxl_root_port.2 no parent for dport: platform
cxl_mem mem8: at cxl_root_port.3 no parent for dport: platform
cxl_mem mem9: at cxl_root_port.4 no parent for dport: platform
cxl_mem mem10: at cxl_root_port.4 no parent for dport: platform
...and this message when running testing in QEMU:
cxl_region region4: Bypassing cpu_cache_invalidate_memregion() for testing!
Noisy cxl_test warnings have caused other regressions to be missed. In
the interest of using cxl_test for early detection of dev_err() and
dev_warn() messages, silence platform device topology and
cache-invalidation messages.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
In cxl_port_setup_targets() the region driver validates the
configuration of auto-discovered region decoders, as well
as decoders the driver is preparing to program.
The existing calculations use the encoded interleave granularity
value to create an interleave granularity that properly fans out
when routing an x1 interleave to a greater than x1 interleave.
That all worked well, until this config came along:
Host Bridge: 2 way at 256 granularity
Switch Decoder_A: 1 way at 512
Endpoint_X: 2 way at 256
Switch Decoder_B: 1 way at 512
Endpoint_Y: 2 way at 256
When the Host Bridge interleave is greater than 1 and the root
decoder interleave is exactly 1, the region driver needs to
consider the number of targets in the region when calculating
the expected granularity.
While examining the existing logic, and trying to cover the case
above, a couple of simplifications appeared, hence this proposed
refactoring.
The first simplification is to apply the logic to the nominal
values and use the existing helper function granularity_to_eig() to
translate the desired granularity to the encoded form. This means
the comment and code regarding setting address bits is discarded.
Although that logic is not wrong, it adds a level of complexity that
is not required in the granularity selection. The eig and eiw are
indeed part of the routing instructions programmed into the decoders.
Up-level the discussion to nominal ways and granularity for clearer
analysis.
The second simplification reduces the logic to a single granularity
calculation that works for all cases. The new calculation doesn't
care if parent_iw => 1 because parent_iw is used as a multiplier.
The refactor cleans up a useless assignment of eiw made after the
iw is already calculated.
Regression testing included an examination of all of the ways and
granularity selections made during a run of the cxl_test unit tests.
There were no differences in selections before and after this patch.
Fixes: ("27b3f8d13830 cxl/region: Program target lists")
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/20230822180928.117596-1-alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Currently, when the region driver attaches a region to a port, it
selects the ports next available decoder to program.
With the addition of auto-discovered regions, a port decoder has
already been programmed so grabbing the next available decoder can
be a mismatch when there is more than one region using the port.
The failure appears like this with CXL DEBUG enabled:
[] cxl_core:alloc_region_ref:754: cxl region0: endpoint9: HPA order violation region0:[mem 0x14780000000-0x1478fffffff flags 0x200] vs [mem 0x880000000-0x185fffffff flags 0x200]
[] cxl_core:cxl_port_attach_region:972: cxl region0: endpoint9: failed to allocate region reference
When CXL DEBUG is not enabled, there is no failure message. The region
just never materializes. Users can suspect this issue if they know their
firmware has programmed decoders so that more than one region is using
a port. Note that the problem may appear intermittently, ie not on
every reboot.
Add a matching method for auto-discovered regions that finds a decoder
based on an HPA range. The decoder range must exactly match the region
resource parameter.
Fixes: a32320b71f ("cxl/region: Add region autodiscovery")
Signed-off-by: Alison Schofield <alison.schofield@intel.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20230905211007.256385-1-alison.schofield@intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Pick up the first half of the RCH error handling series. The back half
needs some fixups for test regressions. Small conflicts with the PMU
work around register enumeration and setup helpers.
Pick up the recent fixes to how CPU caches are managed relative to
region setup / teardown, and make sure that all decoders transition
successfully before updating the region state from COMMIT => ACTIVE.
Switch-level (mid-level) decoders between the platform root and an
endpoint can dynamically switch modes between HDM-H and HDM-D[B]
depending on which region they target. Use the region type to fixup each
decoder that gets allocated to map the given region.
Note that endpoint decoders are meant to determine the region type, so
warn if those ever need to be fixed up, but since it is possible to
continue do so.
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/168679262543.3436160.13053831955768440312.stgit@dwillia2-xfh.jf.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
In preparation for support for HDM-D and HDM-DB configuration
(device-memory, and device-memory with back-invalidate). Rename the current
type designators to use HOSTONLYMEM and DEVMEM as a suffix.
HDM-DB can be supported by devices that are not accelerators, so DEVMEM is
a more generic term for that case.
Fixup one location where this type value was open coded.
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/168679261369.3436160.7042443847605280593.stgit@dwillia2-xfh.jf.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Jonathan reports that failed attempts to reset a region (teardown its
HDM decoder configuration) mistakenly advance the state of the region
to "not committed". Revert to the previous state of the region on reset
failure so that the reset can be re-attempted.
Reported-by: Jonathan Cameron <Jonathan.Cameron@Huawei.com>
Closes: http://lore.kernel.org/r/20230316171441.0000205b@Huawei.com
Fixes: 176baefb2e ("cxl/hdm: Commit decoder state to hardware")
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/168696507968.3590522.14484000711718573626.stgit@dwillia2-xfh.jf.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
cxl_region_decode_reset() walks all the decoders associated with a given
region and disables them. Due to decoder ordering rules it is possible
that a switch in the topology notices that a given decoder can not be
shutdown before another region with a higher HPA is shutdown first. That
can leave the region in a partially committed state.
Capture that state in a new CXL_REGION_F_NEEDS_RESET flag and require
that a successful cxl_region_decode_reset() attempt must be completed
before cxl_region_probe() accepts the region.
This is a corollary for the bug that Jonathan identified in "CXL/region
: commit reset of out of order region appears to succeed." [1].
Cc: Jonathan Cameron <Jonathan.Cameron@Huawei.com>
Link: http://lore.kernel.org/r/20230316171441.0000205b@Huawei.com [1]
Fixes: 176baefb2e ("cxl/hdm: Commit decoder state to hardware")
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/168696507423.3590522.16254212607926684429.stgit@dwillia2-xfh.jf.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Vikram raised a concern with the theoretical case of a CPU sending
MemClnEvict to a device that is not prepared to receive. MemClnEvict is
a message that is sent after a CPU has taken ownership of a cacheline
from accelerator memory (HDM-DB). In the case of hotplug or HDM decoder
reconfiguration it is possible that the CPU is holding old contents for
a new device that has taken over the physical address range being cached
by the CPU.
To avoid this scenario, invalidate caches prior to tearing down an HDM
decoder configuration.
Now, this poses another problem that it is possible for something to
speculate into that space while the decode configuration is still up, so
to close that gap also invalidate prior to establish new contents behind
a given physical address range.
With this change the cache invalidation is now explicit and need not be
checked in cxl_region_probe(), and that obviates the need for
CXL_REGION_F_INCOHERENT.
Cc: Jonathan Cameron <Jonathan.Cameron@Huawei.com>
Fixes: d18bc74ace ("cxl/region: Manage CPU caches relative to DPA invalidation events")
Reported-by: Vikram Sethi <vsethi@nvidia.com>
Closes: http://lore.kernel.org/r/BYAPR12MB33364B5EB908BF7239BB996BBD53A@BYAPR12MB3336.namprd12.prod.outlook.com
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Reviewed-by: Dave Jiang <dave.jiang@intel.com>
Link: https://lore.kernel.org/r/168696506886.3590522.4597053660991916591.stgit@dwillia2-xfh.jf.intel.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
For symmetry with the recent rename of ->dport_dev for a 'struct
cxl_dport', add the "_dev" suffix to the ->uport property of a 'struct
cxl_port'. These devices represent the downstream-port-device and
upstream-port-device respectively in the CXL/PCIe topology.
Signed-off-by: Terry Bowman <terry.bowman@amd.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Link: https://lore.kernel.org/r/20230622205523.85375-6-terry.bowman@amd.com
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
