Update MES API to support oversubscription without aggregated doorbell
for usermode queues.
v2: Change oversubscription_no_aggregated_en to is_kfd_process (align
with MES)
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Jack Xiao <Jack.Xiao@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Starting with GFX11, MES requires wptr BOs to be GTT allocated/mapped to
GART for usermode queues in order to support oversubscription. In the
case that work is submitted to an unmapped queue, MES must have a GART
wptr address to determine whether the queue should be mapped.
This change is accompanied with changes in MES and is applicable for
MES_API_VERSION >= 2.
v3:
- Use amdgpu_vm_bo_lookup_mapping for wptr_bo mapping lookup
- Move wptr_bo refcount increment to amdgpu_amdkfd_map_gtt_bo_to_gart
- Remove list_del_init from amdgpu_amdkfd_map_gtt_bo_to_gart
- Cleanup/fix create_queue wptr_bo error handling
v4:
- Add MES version shift/mask defines to amdgpu_mes.h
- Change version check from MES_VERSION to MES_API_VERSION
- Add check in kfd_ioctl_create_queue before wptr bo pin/GART map to
ensure bo is a single page.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
After queue unmap or remove from MES successfully, free queue sysfs
entries, doorbell and remove from queue list. Otherwise, application may
destroy queue again, cause below kernel warning or crash backtrace.
For outstanding queues, either application forget to destroy or failed
to destroy, kfd_process_notifier_release will remove queue sysfs
entries, kfd_process_wq_release will free queue doorbell.
v2: decrement_queue_count for MES queue
refcount_t: underflow; use-after-free.
WARNING: CPU: 7 PID: 3053 at lib/refcount.c:28
Call Trace:
kobject_put+0xd6/0x1a0
kfd_procfs_del_queue+0x27/0x30 [amdgpu]
pqm_destroy_queue+0xeb/0x240 [amdgpu]
kfd_ioctl_destroy_queue+0x32/0x70 [amdgpu]
kfd_ioctl+0x27d/0x500 [amdgpu]
do_syscall_64+0x35/0x80
WARNING: CPU: 2 PID: 3053 at drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_device_queue_manager.c:400
Call Trace:
deallocate_doorbell.isra.0+0x39/0x40 [amdgpu]
destroy_queue_cpsch+0xb3/0x270 [amdgpu]
pqm_destroy_queue+0x108/0x240 [amdgpu]
kfd_ioctl_destroy_queue+0x32/0x70 [amdgpu]
kfd_ioctl+0x27d/0x500 [amdgpu]
general protection fault, probably for non-canonical address
0xdead000000000108:
Call Trace:
pqm_destroy_queue+0xf0/0x200 [amdgpu]
kfd_ioctl_destroy_queue+0x2f/0x60 [amdgpu]
kfd_ioctl+0x19b/0x600 [amdgpu]
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Graham Sider <Graham.Sider@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We remove the user queue from MES scheduler to update queue properties.
If the queue becomes active after updating, add the user queue to MES
scheduler, to be able to handle command packet submission.
v2: don't break pqm_set_gws
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Graham Sider <Graham.Sider@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_topology.c:1542:11:
warning: variable 'i' set but not used [-Wunused-but-set-variable]
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add a new KFD ioctl to return the largest possible memory size that
can be allocated as a buffer object using
kfd_ioctl_alloc_memory_of_gpu. It attempts to use exactly the same
accept/reject criteria as that function so that allocating a new
buffer object of the size returned by this new ioctl is guaranteed to
succeed, barring races with other allocating tasks.
This IOCTL will be used by libhsakmt:
https://www.mail-archive.com/amd-gfx@lists.freedesktop.org/msg75743.html
Signed-off-by: Daniel Phillips <Daniel.Phillips@amd.com>
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Extend KFD device topology to surface peer-to-peer links among
GPU devices connected over PCIe or xGMI. Enabling HSA_AMD_P2P is
REQUIRED to surface peer-to-peer links.
Prior to this KFD did not expose to user mode any P2P links or
indirect links that go over two or more direct hops. Old versions
of the Thunk used to make up their own P2P and indirect links without
the information about peer-accessibility and chipset support available
to the kernel mode driver. In this patch we expose P2P links in a new
sysfs directory to provide more reliable P2P link information to user
mode.
Old versions of the Thunk will continue to work as before and ignore
the new directory. This avoids conflicts between P2P links exposed by
KFD and P2P links created by the Thunk itself. New versions of the Thunk
will use only the P2P links provided in the new p2p_links directory, if
it exists, or fall back to the old code path on older KFDs that don't
expose p2p_links.
Signed-off-by: Ramesh Errabolu <Ramesh.Errabolu@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Extend current kernel config requirements of amdgpu by adding config
HSA_AMD_P2P. Enabling HSA_AMD_P2P is REQUIRED to support peer-to-peer
communication between AMD GPU devices over PCIe bus
Signed-off-by: Ramesh Errabolu <Ramesh.Errabolu@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Migration range from system memory to VRAM, if system page can not be
locked or unmapped, we do partial migration and leave some pages in
system memory. Several bugs found to copy pages and update GPU mapping
for this situation:
1. copy to vram should use migrate->npage which is total pages of range
as npages, not migrate->cpages which is number of pages can be migrated.
2. After partial copy, set VRAM res cursor as j + 1, j is number of
system pages copied plus 1 page to skip copy.
3. copy to ram, should collect all continuous VRAM pages and copy
together.
4. Call amdgpu_vm_update_range, should pass in offset as bytes, not
as number of pages.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
MMU notifier callback may pass in mm with mm->mm_users==0 when process
is exiting, use mmget_no_zero to avoid accessing invalid mm in deferred
list work after mm is gone.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add a AMDGPU_GEM_CREATE_DISCARDABLE flag to note that the content of a BO
doesn't needs to be preserved during eviction.
KFD was already using a similar functionality for SVM BOs so replace the
internal flag with the new UAPI.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Based on gfx10 with following changes:
- GPR_ALLOC.VGPR_SIZE field moved (and size corrected in gfx10)
- s_sendmsg_rtn_b64 replaces some s_sendmsg/s_getreg
- Buffer instructions no longer have direct-to-LDS modifier
Signed-off-by: Jay Cornwall <jay.cornwall@amd.com>
Reviewed-by: Laurent Morichetti <laurent.morichetti@amd.com>
Signed-off-by: Eric Huang <jinhuieric.huang@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
CPU hive assignment currently assumes when a GPU hive is connected_to_cpu,
there is only one hive in the system.
Only assign CPUs to the hive if they are explicitly directly connected to
the GPU hive to get rid of the need for this assumption.
It's more efficient to do this when querying IO links since other non-CRAT
info has to be filled in anyways. Also, stop re-assigning the
same CPU to the same GPU hive if it has already been done before.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add amdgpu_no_queue_eviction_on_vm_fault condition to
event_interrupt_isr_v11 return. If no queue eviction on vm fault
specified, function should return false for client/source ids specifying
vm fault.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Reviewed-by: Mukul Joshi <mukul.joshi@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Return boolean values ("true" or "false") instead of 1 or 0 from bool
functions. This fixes the following warnings from coccicheck:
./drivers/gpu/drm/amd/amdkfd/kfd_int_process_v11.c:244:9-10: WARNING:
return of 0/1 in function 'event_interrupt_isr_v11' with return type
bool
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add initial support for soc21 in KFD compute
driver (Mukul)
- Add new definition for soc21 device.
- Add new file for amdgpu-kfd interface for GFX11 family.
- Add new file for queue management, interrupt handling,
mqd management for GFX11 family in KFD driver.
- Related changes/updates for soc21 device in
KFD driver.
- Repurpose last 2 entries of SDMA MQD for driver use.
v2: Add an optional argument into update queue operation (Mukul)
v3: Switch to ip version check, replace kgd_dev with
amdgpu_device (Hawking)
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Rather than using hardcoded tables, we can use the gfx and
gmc config pulled from the IP discovery table to generate the
cache configuration.
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The logic to update the IO links when a KFD device
is removed was not correct as it would miss updating
the proximity domain values for some nodes where the
node_from and node_to both were greater values than the
proximity domain value of the KFD device being removed
from topology.
Fixes: 46d18d510d ("drm/amdkfd: Cleanup IO links during KFD device removal")
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
All uses of the 'kfd->gtt_sa_bitmap' bitmap are protected with the
'kfd->gtt_sa_lock' mutex.
So:
- prefer the non-atomic '__set_bit()' function
- use the non-atomic 'bitmap_[set|clear]()' functions instead of
equivalent 'for' loops. These functions can work on several bits at a
time
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
'kfd->gtt_sa_bitmap' is a bitmap. So use 'bitmap_zalloc()' to simplify
code, improve the semantic and avoid some open-coded arithmetic in
allocator arguments.
Also change the corresponding 'kfree()' into 'bitmap_free()' to keep
consistency.
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Change SVM range mapping flags or access attributes don't trigger
migration, if range is already mapped on GPUs we should update GPU
mapping and pass flush_tlb flag true to amdgpu vm.
Change SVM range preferred_loc or migration granularity don't need
update GPU mapping, skip the validate_and_map.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
To avoid unnecessary unmap SVM range from GPUs if range is not mapped on
GPUs when migrating the range. This flag will also be used to flush TLB
when updating the existing mapping on GPUs.
It is protected by prange->migrate_mutex and mmap read lock in MMU
notifier callback.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
MEC firmware sometimes sends signal interrupts without a valid context ID
on end of pipe events that don't intend to signal any HSA signals.
This triggers the slow path in kfd_signal_event_interrupt that scans the
entire event page for signaled events. Detect these signals in the top
half interrupt handler to stop processing them as early as possible.
Because we now always treat event ID 0 as invalid, reserve that ID during
process initialization.
v2: Update firmware version checks to support more GPUs
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This reverts commit 36bf93216e.
It causes SVM regressions on Vega10 with XNACK-ON. Just revert it
at the moment.
./kfdtest --gtest_filter=KFDSVMRangeTest.MigratePolicyTest
Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Reviewed-by: Philip Yang<Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
dqm->gws_queue_count and pdd->qpd.mapped_gws_queue need to be updated
each time the queue gets evicted.
Fixes: b8020b0304 ("drm/amdkfd: Enable over-subscription with >1 GWS queue")
Signed-off-by: David Yat Sin <david.yatsin@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The idea is from
commit a50fe70780 ("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran")
and
commit f61c40c075 ("drm/amdkfd: enable heavy-weight TLB flush on Arcturus").
At the moment, heavy-weight TLB could cause problems on ASICs except
Aldebaran and Arcturus.
A simple hipMallocManaged/hipFree program could trigger this issue.
[ 97.787657] amdgpu 0000:01:00.0: amdgpu: wait for kiq fence error: 0.
[ 106.868758] amdgpu: qcm fence wait loop timeout expired
[ 106.868966] amdgpu: The cp might be in an unrecoverable state due to an unsuccessful queues preemption
[ 106.869203] amdgpu: Failed to evict process queues
[ 106.869261] amdgpu: Failed to quiesce KFD
Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
To make kfd_flush_tlb_after_unmap visible in kfd_svm.c,
move it into kfd_priv.h. And change it to an inline function.
Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add the waiters to the wait queue during initialization, while holding the
event spinlock. Otherwise the waiter will not get activated if the event
signals before being added to the wait queue.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang<Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If lookup_event_by_id() returns a NULL "ev" pointer then the
spin_lock(&ev->lock) will crash. This was detected by Smatch:
drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_events.c:644 kfd_set_event()
error: we previously assumed 'ev' could be null (see line 639)
Fixes: 5273e82c5f ("drm/amdkfd: Improve concurrency of event handling")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Currently, the IO-links to the device being removed from topology,
are not cleared. As a result, there would be dangling links left in
the KFD topology. This patch aims to fix the following:
1. Cleanup all IO links to the device being removed.
2. Ensure that node numbering in sysfs and nodes proximity domain
values are consistent after the device is removed:
a. Adding a device and removing a GPU device are made mutually
exclusive.
b. The global proximity domain counter is no longer required to be
an atomic counter. A normal 32-bit counter can be used instead.
3. Update generation_count to let user-mode know that topology has
changed due to device removal.
CC: Shuotao Xu <shuotaoxu@microsoft.com>
Reviewed-by: Shuotao Xu <shuotaoxu@microsoft.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The synchronize_rcu call in destroy_events can take several ms, which
noticeably slows down applications destroying many events. Use kfree_rcu
to free the event structure asynchronously and eliminate the
synchronize_rcu call in the user thread.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
