Some infrastructures below,
Implement SdmaTimePacket which records the global GPU timestamp.
Introduce class AsyncMPSQ and AsyncMPMQ.
AsyncMPSQ is aka async multiple packet single queue. It takes a set of
packet when create and submits them to a GPU to run. While AsyncMPMQ is
aka async multiple packet multiple queue. It manages a set of AsyncMPSQ,
and use a forloop to do operations of AsyncMPSQ.
Implement sdma_multicopy helper functions.
Change-Id: I47e1d2ca9630113b2a1d85a0055f3f8ee629fb5f
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
For following test cases:
- KFDQMTest.QueueLatency
- KFDQMTest.BasicCuMaskingLinear
- KFDQMTest.BasicCuMaskingEven
- KFDMemoryTest.MMBandWidth
- KFDMemoryTest.MMapLarge
- KFDMemoryTest.MMBench
v2: xml element cannot start with a number, so change the key name of
MMBandWidth and MMBench accordingly
xml element cannot contain whitespaces, so trim whitespaces in "VRAM "
v3: introduce KFDLog-like way to use KFDRecord
Change-Id: Ifc3ed5657621252a7b39dccf1ef4f50a92593f77
Signed-off-by: Xiaojie Yuan <xiaojie.yuan@amd.com>
This change is from commit 62f7dc2a("kfdtest: Do not set GTEST_FLAG
throw_on_failure").
But it is unexpected to reverted by commit 414042ab("kfdtest: Clean up
comments"). So add this change back.
Fix: 414042ab
Change-Id: Ia9e99c9ca17b99aab62b4db55017018ddae43dfb
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
The timestamp written by releaseMemory packet might still not be visible
when we fetch it.
To fix this bug, use event-based wait.
Change-Id: If2324eb3b3a632c711ee4dff4d03a93d5306c289
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
Handle the case that svm.dgpu_aperture does not exist in vm_find_object.
Change-Id: Ic0983d4f321f1b6248514f2fa25162976e90bd75
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Use the NodeFrom returned by hsaKmtGetNodeIoLinkProperties() to check
its correctness.
Change-Id: I6ce436dc7c5d5b192bee21156292bd3eff77f916
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Some nodes are unavailable based on the task's cgroup hierarchy. Handle
this situation by ignoring those nodes
Change-Id: I72f9e822d2ec8cf15732df95e427d5549a75b55d
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
With GPU resource management, some nodes are unavailable based on the
cgroup hierarchy of the task. Kernel via sysfs specifies all the
iolinks. Skip the links which are not accessible.
Also iolinks specified by the kernel refer to sysfs Node IDs. Map it to
relevant user Node IDs
v2: NodeFrom mapped from sysfs Node to User Node
Change-Id: I95312ee6ca51b89fe9e6ca2a9185c2ea1e94afc4
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Wait4PacketConsumption now can accept an event to wait all packets subbmitted
to be processed.
Change-Id: I1497b7704e892b04d05811b8d3e4742237c1be57
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
This reverts commit ab181c46c0.
This fixes ambiguity when looking up GPU addresses with
hsaKmtQueryPointerInfo.
hsa_amd_agents_allow_access uses hsaKmtQueryPointerInfo, and
depends on finding the correct object from a GPU address. Finding
the wrong userptr object based on its CPU address leads to
incorrect GPU mappings and results in VM faults.
Change-Id: I7c5f571ee6e1f9d32687aa3eab6d96944ad032be
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Removed some tests from the blacklist that are now passing. Added two
new tests that hang the GPU.
Change-Id: I09e729590e5181311375058be492d387342ba2fe
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Added a helper vm_find_object that can be used everywhere we need to
lookup objects by their address and optionally size. This unifies
all subtly different, partially incomplete, or broken ways of doing
this in various functions:
* map
* unmap
* register
* deregister
* free
* get_mem_info
* set_mem_user_data
At the same time fix some subtle problems for userptr lookup that
got a bit more complex when the userptr address can match the GPU
address.
Change-Id: I98572d1734fc7688a1d68f6a784e02c8dee90af5
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
PCIe P2P (indirect) IOLinks should only be created if the remote GPU
is large-BAR
Change-Id: I55cbb5e37c5d41267583e07aca6bdcc708403029
Signed-off-by: shaoyunl <Shaoyun.Liu@amd.com>
As it will alloc as much as small system memory to reach the allocation limit.
We can try to alloc memory several times to see if any allocation in
the previous step cause memory leak.
Also we test if GPU can access these memory correctly or not.
Change-Id: I309f9821b6bc99c212a6bfbc21fe3086ab589fd3
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
Add XGMI related defines in thunk according to the document
HSAKMT library interface specification v1.16
Change-Id: Ib25ff0ddf7380c97d06bd76fb730915e7c634270
Signed-off-by: Shaoyun Liu <Shaoyun.Liu@amd.com>
Similar with SdmaEventInterrupt, verify event interrupt on pm4 queue.
Change-Id: I0e43f26fd0d965126985820704215d2ef5e52c1a
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
Simulate some workload there to verify the sDMA event interrupt.
Change-Id: Ib5ad0c238cc66898f7835e765df50427ef106b04
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
It should have PASS/FAIL report for the vram allocated size.
Change-Id: I546c02c2ed02f1cfb5278e0dfd7b18ade39faafb
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
Currently, all HSA nodes are exposed to user. So the existing
implementation assumes a one to one mapping between user
NodeId and sysfs nodeId.
GPU Resource Management will provide control over the exposed
HSA nodes. This means not all HSA nodes will be exposed to the user.
Decouple it.
The mapping from user NodeId to sysfs NodeId will be local
to topology.c and topology helper functions. For others NodeId
should be sequential from 0 to Number of Nodes exposed to user.
v1: initial implementation
v2: map node id within the topology_* functions
v3: remove two static globals
v4: add bounds check got node id
Change-Id: Id12147ece41d682430f398944bbb339ca906eb1b
Signed-off-by: Mike Li <Tianxinmike.Li@amd.com>
ASSERT failures result in immediate termination of the test. EXPECT
returns a failure but continues execution. Reserve ASSERT for required
functionality (node initialization, queue creation, etc) where the rest
of the test cannot run if that call fails. Use EXPECT everywhere else
Change-Id: I1c11326fc3ae22b50fa83b07b3b49af1e1f4e69e
Add initial support for the kfd debugger trap support
for GFX9 chips.
- Adding support for Enable/Disable trap support
- Setting debug trap support data
- Setting wave launch trap override
- Setting wave launch mode
Change-Id: If39f2395c4b6cf56249cf76f1c44cfcbdcef891c
Signed-off-by: Philip Cox <Philip.Cox@amd.com>
AMDKFD_IOC_WAIT_EVENTS with multiple events and wait_for_all = 0
returns success after any of the events have signaled. So we can't
blindly assume that a memory fault event that was in the list has
actually signaled. Check the gpu_id as an indicator whether there
really was a memory fault before processing it further.
Change-Id: I6cc311bfc184c631beaf684027176a6ca42e05c1
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
If the CPU addr of a userptr is accessible by the GPU, try to use it
instead of allocating a different GPU address. If something else is
already registered with an overlapping address range, we still need to
allocate a GPU address, because KFD does not support overlapping GPUVM
mappings.
Change-Id: I452963ee45a454f735755a0b43122b9aee5d55be
Signed-off-by: Felix Kuehling <felix.kuehling@gmail.com>
If the GPU virtual address space is >= 47 bits, don't reserve virtual
address space at startup and use mmap to allocate virtual addresses.
Change-Id: Ic935b03c8e78271829fc8e6cfd0e543184aff818
Signed-off-by: Felix Kuehling <felix.kuehling@gmail.com>
This should fix gtest compile errors.
code like below has trouble,
typedef char char8;
typedef unsigned char uchar8;
ASSERT_NE((uchar8)1, 0);
ASSERT_NE((unsigned char8)1, 0); // compile error here
or
ASSERT_NE((unsigned char8)1, 0);
ASSERT_NE((uchar8)1, 0); // compile error here
HSA[u]int64 are alias. So ASSERT_XX((unsigned HSAint64)..)
with ASSERT_XX((HSAuint64)..) fail to compile.
Change-Id: I4c24bc699a69bd4f37c4bc8aaaa9f1a92a24a33e
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
The flag makes EXPECT_* to behave like ASSERT_*, which actually work against
our favor, so disable the flag.
Change-Id: I2ea1dfeaf916b396593a504d081148abdac0fc70
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
So far we have assumed that userptrs are always memory outside
reserved SVM apertures that are mapped into the SVM aperture for
GPU access.
With an unreserved SVM aperture that covers the entire virtual
address range, this distinction will no longer be true. Userptrs
will generally be inside the unreserved SVM aperture. Take that
into consideration when registering, mapping and unmapping virtual
addresses.
We now need a retry logic when looking up buffers from addresses.
If it is not found by its GPU address, try it as a userptr.
We also need to consider the new possibility that a userptr is
registered at the same address for CPU and GPU access. So a buffer
found by its GPU address may also turn out to be a userptr. In
that case use a stricter lookup using the userptr and size (if
the size is known), to identify the correct one of multiple
overlapping objects.
Change-Id: Ia43633aaa40f9fd2a74918ae969a631d2ff68419
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Make dgpu_aperture and dgpu_alt_aperture pointers that can point to
the same actual aperture. This will be useful on GFXv9 and later,
where the MType is not defined by the aperture and we want to have
a single aperture covering the entire virtual address space.
aperture->is_coherent can no longer be a reliable indicator of
coherency. Replace it with different conditions based on mem flags
and svm.disable_cache (from HSA_DISABLE_CACHE environment).
Change-Id: Iefc415b87b8abd96e3916586485a0a55d9b27c19
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
This prepares the code for an alternative aperture management method
that needs to unmap memory differently.
Change-Id: I5494aa5420f85edb8f7857f00c17e1d2e6479a51
Signed-off-by: Felix Kuehling <felix.kuehling@gmail.com>
Only scratch_physical, for scratch-backing memory is managed by the Thunk.
Change-Id: I4716981aa908d9569584dc35f40ffd270a2f9014
Signed-off-by: Felix Kuehling <felix.kuehling@gmail.com>
This parameter was used for non-canonical GPUVM allocations on GFX7/8 APUs
only, to prevent getting NULL pointers from valid allocation after
subtracting the aperture base. The same can be achieved less intrusively
by reserving address space at the start of the aperture during
initialization.
Change-Id: I0aae773f069c2b228824ba464b0612a4d8b489ce
Signed-off-by: Felix Kuehling <felix.kuehling@gmail.com>
A lot of tests were disabled on gfx900 for historical reasons that
are no longer valid. The only remaining one that won't work on
gfx900 is BasicAddressWatch.
Change-Id: I11507de0dfd31262713127d6cb15cc09c14b8b9f
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
When skipping a test, the output should be:
Skipping test: <reason>.
This will allow for easier identification, automation and general readability
Change-Id: I98bda1c068f9dbc83aeea74f642b6101121f234d
Make indentation consistent, which is that subsequent lines are aligned
with the variables declared above
Change-Id: I590f7768d93565145b986ad1fb6ac8e82f9c0d58
Clean up the KFDTest style via CPPLint. Some warnings remain regarding
volatile variables being cast to void*. This is the command used:
cpplint.py --linelength=120
--filter=-readability/multiline_string,-readability/todo,-build/include,-runtime/references
multiline_string is due to using ISA code
todo is to avoid errors that we don't have TODO(username) instead of TODO
include is about including the folder in the header includes
references is regarding non-const references '&' being const or using
pointers. That can be addressed later
Change-Id: I3c6622da0a13dd33ab29b2bfff48be25e763b750
When mapMemoryToGpu fails, we need unregister it with user address as
the gpu address is not available.
Change-Id: I4418eeaa7aa37008f5bffa144e2c2171f0d238fd
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
Hit queue create failure when do kfdtest with --gtest_repet=-1
fix: 4bb90d04("Remove the use of IS_DGPU()")
Change-Id: I04fa73f90cef13a5517dbaceb89c41dc0f821a79
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
Because gfx700 has local memory but other APUs don't, we should reflect
that in the code. Meanwhile, fix a bug that on gfx902 svm aperture is not
added when calling hsaKmtGetNodeMemoryProperties().
Change-Id: Id840f2db0b14fda9ee713b219a9474c15f8a9771
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
On some asics, like tonga, the memory alignment size is as big as 0x8000.
fmm_allocate* alloc vm area with size passed in which is not aligned mostly.
But __fmm_release free vm area with vm_object_t->size which is aligned.
That might cause aperture_release_area fail to free the vm area as the
size might be bigger than zone itself or it just free another vm area
nearby unexpected.
This patch somehow will alloc more space than it needed on tonga.
gfx900+ is not affected.
Change-Id: I5a88c92b08c4e6f6bc05881798f769b55d6debe9
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
Previously we used the first dgpu mem, but after careful examination, we
found it only needs to be a GPU, so we modify the code to reflect that as
well.
Change-Id: I069d9b8e247aed55c1f885b79f743ea8e03ddf93
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Implement sDMA copy packet broadcast.
Each time sDMA will copy its local vram to sysbuf and next GPU's vram.
That will verify where the p2p link is broken.
Currently we just test push of p2p.
test result on 2 cpus, 4 gpus, numa enabled system.
[ RUN ] KFDQMTest.P2PTest
[ ] Test 2 -> 3
[ ] PASS 2 -> 3
[ ] Test 3 -> 4
[ ] PASS 3 -> 4
[ ] Test 4 -> 5
[ ] PASS 4 -> 5
[ ] Test 5 -> 0
[ ] PASS 5 -> 0
[ OK ] KFDQMTest.P2PTest (190 ms)
Change-Id: Ie6fb2604109e39465b8a873b3bb42abc6259825a
The information can be obtained directly from node id. Also improve the
whole logic for future compatibility.
Change-Id: I130733be4e7930d5953d5e81409905e60c2ec35e
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Unset ret_addr when unmapping the address space reservation. Otherwise
it may try to unmap it again later.
Remember the actual map_size and use it instead of len outside the
reservation loops.
Change-Id: I1a6b3fecfb59e22a713e5ed49c3ed37914cb6fb5
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Both the include and libpath were incorrect after recent build
system changes. Use the proper GNUInstallDirs definitions in
libhsakmt.pc.in to write the proper locations.
This is needed for end users building KFDTest, which depends on
correct pkg-config information.
Change-Id: Ia45f36f054c2a607a77e7ecbcbd9eb7edd067348
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
This test has been intermittently failing for various reasons and
was already disabled on all chips except Ellesmere. It stresses
memory management in unusual ways by having lots of memory allocated
but +# not mapped, which is not relevant to compute applications over
ROCr.
Change-Id: I6b791ca7e2e0fcfe93fc720063b4b56acfded751
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
xinhui pan
Xiaojie Yuan
Felix Kuehling
Harish Kasiviswanathan
shaoyunl
Mike Li
Kent Russell
Philip Cox
Felix Kuehling
Yong Zhao