| Age | Commit message (Collapse) | Author |
|---|
|
The header usr/linux/kfd_ioctl.h is a duplicate of uapi/linux/kfd_ioctl.h.
And it is actually not a file in the source code tree.
Ideally, the usr version should be updated whenever the source code is recompiled.
However, I have noticed a discrepancy between the two headers
even after rebuilding the kernel.
This commit modifies kfd_priv.h to always include the header from uapi to ensure
the latest changes are reflected. We should always include the source
code header other than the duplication.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
For CPX mode, each KFD node has interrupt worker to process ih_fifo to
send events to user space. Currently all interrupt workers of same adev
queue to same CPU, all workers execution are actually serialized and
this cause KFD ih_fifo overflow when CPU usage is high.
Use per-GPU unbounded highpri queue with number of workers equals to
number of partitions, let queue_work select the next CPU round robin
among the local CPUs of same NUMA.
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>
|
|
After GPU page fault, there are lots of page fault interrupts generated
at short period even with CAM filter enabled because the fault address
is different. Each page fault copy to KFD ih fifo to send event to user
space by KFD interrupt worker, this could cause KFD ih fifo overflow
while other processes generate events at same time.
KFD process is aborted after GPU page fault, we only need one GPU page
fault interrupt sent to KFD ih fifo to send memory exception event to
user space.
Incease KFD ih fifo size to 2 times of IH primary ring size, to handle
the burst events case.
This patch handle the gfx v9 path, cover retry on/off and CAM filter
on/off cases.
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>
|
|
Initial support for GC 9.5.0.
v2: squash in pqm_clean_queue_resource() fix from Lijo
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Process device data pdd->vram_usage is read by rocm-smi via sysfs, this
is currently missing the svm_bo usage accounting, so "rocm-smi
--showpids" per process VRAM usage report is incorrect.
Add pdd->vram_usage accounting when svm_bo allocation and release,
change to atomic64_t type because it is updated outside process mutex
now.
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>
|
|
struct file *f is unused in queue creation, remove it.
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>
|
|
Enables users to update SVM's default granularity, used in
buffer migration and handling of recoverable page faults.
Param value is set in terms of log(numPages(buffer)),
e.g. 9 for a 2 MIB buffer
Signed-off-by: Ramesh Errabolu <Ramesh.Errabolu@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
When app unmap vm ranges(munmap) kfd/svm starts drain pending page fault and
not handle any incoming pages fault of this process until a deferred work item
got executed by default system wq. The time period of "not handle page fault"
can be long and is unpredicable. That is advese to kfd performance on page
faults recovery.
This patch uses time stamp of incoming page fault to decide to drop or recover
page fault. When app unmap vm ranges kfd records each gpu device's ih ring
current time stamp. These time stamps are used at kfd page fault recovery
routine.
Any page fault happened on unmapped ranges after unmap events is application
bug that accesses vm range after unmap. It is not driver work to cover that.
By using time stamp of page fault do not need drain page faults at deferred
work. So, the time period that kfd does not handle page faults is reduced
and can be controlled.
Signed-off-by: Xiaogang.Chen <Xiaogang.Chen@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Based on the recommendation of MEC FW, update BadOpcode interrupt
handling by unmapping all queues, removing the queue that got the
interrupt from scheduling and remapping rest of the queues back when
using MES scheduler. This is done to prevent the case where unmapping
of the bad queue can fail thereby causing a GPU reset.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Acked-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Add helper function kfd_queue_unreference_buffers to reduce queue buffer
refcount, separate it from release queue buffers.
Because it is circular locking to hold dqm_lock to take vm lock,
kfd_ioctl_destroy_queue should take vm lock, unreference queue buffers
first, but not release queue buffers, to handle error in case failed to
hold vm lock. Then hold dqm_lock to remove queue from queue list and
then release queue buffers.
Restore process worker restore queue hold dqm_lock, will always find
the queue with valid queue buffers.
v2 (Felix):
- renamed kfd_queue_unreference_buffer(s) to kfd_queue_unref_bo_va(s)
- added two FIXME comments for follow up
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
The number of watchpoints should be set and constrained per logical
partition device, not by the socket device.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Harish Kasiviswanathan <harish.kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Support per-queue reset for GFX9. The recommendation is for the driver
to target reset the HW queue via a SPI MMIO register write.
Since this requires pipe and HW queue info and MEC FW is limited to
doorbell reports of hung queues after an unmap failure, scan the HW
queue slots defined by SET_RESOURCES first to identify the user queue
candidates to reset.
Only signal reset events to processes that have had a queue reset.
If queue reset fails, fall back to GPU reset.
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>
|
|
Certain GPUs have better copy performance over xGMI on specific
SDMA engines depending on the source and destination GPU.
Allow users to create SDMA queues on these recommended engines.
Close to 2x overall performance has been observed with this
optimization.
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>
|
|
Use the queue eop buffer size, cwsr area size, ctl stack size
calculation from Thunk, store the value to KFD node properties.
Those will be used to validate queue eop buffer size, cwsr area size,
ctl stack size when creating KFD user compute queue.
Those will be exposed to user space via sysfs KFD node properties, to
remove the duplicate calculation code from Thunk.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Add atomic queue_refcount to struct bo_va, return -EBUSY to fail unmap
BO from the GPU if the bo_va queue_refcount is not zero.
Create queue to increase the bo_va queue_refcount, destroy queue to
decrease the bo_va queue_refcount, to ensure the queue buffers mapped on
the GPU when queue is active.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Find user queue rptr, ring buf, eop buffer and cwsr area BOs, and
check BOs are mapped on the GPU with correct size and take the BO
reference.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Add helper function kfd_queue_acquire_buffers to get queue wptr_bo
reference from queue write_ptr if it is mapped to the KFD node with
expected size.
Add wptr_bo to structure queue_properties because structure queue is
allocated after queue buffers are validated, then we can remove wptr_bo
parameter from pqm_create_queue.
Rename structure queue wptr_bo_gart to hold wptr_bo reference for GART
mapping and umapping. Move MES wptr_bo_gart mapping to init_user_queue,
the same location with queue ctx_bo GART mapping.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
is_hws_hang and is_resetting serves pretty much the same purpose and
they all duplicates the work of the reset_domain lock, just check that
directly instead. This also eliminate a few bugs listed below and get
rid of dqm->ops.pre_reset.
kfd_hws_hang did not need to avoid scheduling another reset. If the
on-going reset decided to skip GPU reset we have a bad time, otherwise
the extra reset will get cancelled anyway.
remove_queue_mes forgot to check is_resetting flag compared to the
pre-MES path unmap_queue_cpsch, so it did not block hw access during
reset correctly.
Signed-off-by: Yunxiang Li <Yunxiang.Li@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Initial implementation, based on GFX11.
v2: Removed dbg_wa code as not needed on GFX12.
v3: squash in SDMA queue fixes (Alex)
v4: rebase (Alex)
Signed-off-by: David Belanger <david.belanger@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Add gfx v9_4_4 ip block support
Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Le Ma <le.ma@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
TLB flush after unmap accidentially was removed on
gfx9.4.2. It is to add it back.
Signed-off-by: Eric Huang <jinhuieric.huang@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
|
|
In a future commit, the cwsr trap handler code size for gfx10.1 will
increase to slightly above the one page mark. Since the TMA does not
need to be page aligned, and only 2 pointers are stored in it, push
the TMA offset by 2 KiB and keep the TBA+TMA reserved memory size
to two pages.
Signed-off-by: Laurent Morichetti <laurent.morichetti@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
In certain cooperative group dispatch scenarios the default SPI resource
allocation may cause reduced per-CU workgroup occupancy. Set
COMPUTE_RESOURCE_LIMITS.FORCE_SIMD_DIST=1 to mitigate soft hang
scenarions.
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Suggested-by: Joseph Greathouse <Joseph.Greathouse@amd.com>
Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
On GFX 9.4.3, for a given KFD node, fetch the correct drm device from
XCP manager when checking for cgroup permissions.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Properly mark kfd_process->ef as __rcu and consistently use the right
accessor functions.
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202312052245.yFpBSgNH-lkp@intel.com/
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
dbg_flags looks to be defined with incorrect data type; to process
multiple debug flag options, and hence defined dbg_flags as u32.
Fixes the below:
drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_packet_manager_v9.c:117 pm_map_process_aldebaran() warn: maybe use && instead of &
Fixes: 0de4ec9a0353 ("drm/amdgpu: prepare map process for multi-process debug devices")
Suggested-by: Lijo Lazar <lijo.lazar@amd.com>
Cc: Felix Kuehling <Felix.Kuehling@amd.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Srinivasan Shanmugam <srinivasan.shanmugam@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
KFD_GC_VERSION was recently updated to use a new function
for IP version checks. As a result, use KFD_GC_VERSION as
the common function for all IP version checks in KFD.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
This will make it possible for amdgpu GEM ioctls to flush TLBs on compute
VMs.
This removes VMID-based TLB flushing and always uses PASID-based
flushing. This still works because it scans the VMID-PASID mapping
registers to find the right VMID. It's only slightly less efficient. This
is not a production use case.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Use an inline function for version check. Gives more flexibility to
handle any format changes.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Heavy-weight TLB flush is required after unmap on all GPUs for
correctness and security.
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Replace pr_err with dev_err to show the bus-id of
failing device with kfd queue errors
Signed-off-by: Asad Kamal <asad.kamal@amd.com>
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
Now that we use the dGPU path for all APUs, drop the
IOMMUv2 support.
v2: drop the now unused queue manager functions for gfx7/8 APUs
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Tested-by: Mike Lothian <mike@fireburn.co.uk>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
This patch removes some variables and functions from KFD
doorbell handling code, which are no more required since
doorbell manager is handling doorbell calculations.
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Christian Koenig <christian.koenig@amd.com>
Cc: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Shashank Sharma <shashank.sharma@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
This patch:
- adds a doorbell object in kfd pdd structure.
- allocates doorbells for a process while creating its queue.
- frees the doorbells with pdd destroy.
- moves doorbell bitmap init function to kfd_doorbell.c
PS: This patch ensures that we don't break the existing KFD
functionality, but now KFD userspace library should also
create doorbell pages as AMDGPU GEM objects using libdrm
functions in userspace. The reference code for the same
is available with AMDGPU Usermode queue libdrm MR. Once
this is done, we will not need to create process doorbells
in kernel.
V2: - Do not use doorbell wrapper API, use amdgpu_bo_create_kernel
instead (Alex).
- Do not use custom doorbell structure, instead use separate
variables for bo and doorbell_bitmap (Alex)
V3:
- Do not allocate doorbell page with PDD, delay doorbell process
page allocation until really needed (Felix)
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Christian Koenig <christian.koenig@amd.com>
Cc: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felilx.Kuehling@amd.com>
Signed-off-by: Shashank Sharma <shashank.sharma@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
This patch:
- adds a doorbell bo in kfd device structure.
- creates doorbell page for kfd kernel usages.
- updates the get_kernel_doorbell and free_kernel_doorbell functions
accordingly
V2: Do not use wrapper API, use direct amdgpu_create_kernel(Alex)
V3:
- Move single variable declaration below (Christian)
- Add a to-do item to reuse the KGD kernel level doorbells for
KFD for non-MES cases, instead of reserving one page (Felix)
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Christian Koenig <christian.koenig@amd.com>
Cc: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Shashank Sharma <shashank.sharma@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
For GFX 9.4.3, interrupt handling needs to be updated for:
- Interrupt cookie will have a NodeId field. Each KFD
node needs to check the NodeId before processing the
interrupt.
- For CPX mode, there are additional checks of client ID
needed to process the interrupt.
- Add NodeId to the process drain interrupt.
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>
|
|
This patch fixes a regression caused by a bad merge where
the handling of reserved SDMA queues was accidentally removed.
With the fix, the reserved SDMA queues are again correctly
marked as unavailable for allocation.
Fixes: a805889a1531 ("drm/amdkfd: Update SDMA queue management for GFX9.4.3")
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>
|
|
Allow the debugger to get a snapshot of a specified number of queues
containing various queue property information that is copied to the
debugger.
Since the debugger doesn't know how many queues exist at any given time,
allow the debugger to pass the requested number of snapshots as 0 to get
the actual number of potential snapshots to use for a subsequent snapshot
request for actual information.
To prevent future ABI breakage, pass in the requested entry_size.
The KFD will return it's own entry_size in case the debugger still wants
log the information in a core dump on sizing failure.
Also allow the debugger to clear exceptions when doing a snapshot.
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>
|
|
Shader read, write and atomic memory operations can be alerted to the
debugger as an address watch exception.
Allow the debugger to pass in a watch point to a particular memory
address per device.
Note that there exists only 4 watch points per devices to date, so have
the KFD keep track of what watch points are allocated or not.
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>
|
|
In order to inspect waves from the saved context at any point during a
debug session, the debugger must be able to preempt queues to trigger
context save by suspending them.
On queue suspend, the KFD will copy the context save header information
so that the debugger can correctly crawl the appropriate size of the saved
context. The debugger must then also be allowed to resume suspended queues.
A queue that is newly created cannot be suspended because queue ids are
recycled after destruction so the debugger needs to know that this has
occurred. Query functions will be later added that will clear a given
queue of its new queue status.
A queue cannot be destroyed while it is suspended to preserve its saved
context during debugger inspection. Have queue destruction block while
a queue is suspended and unblocked when it is resumed. Likewise, if a
queue is about to be destroyed, it cannot be suspended.
Return the number of queues successfully suspended or resumed along with
a per queue status array where the upper bits per queue status show that
the request was invalid (new/destroyed queue suspend request, missing
queue) or an error occurred (HWS in a fatal state so it can't suspend or
resume queues).
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>
|
|
The debugger must be notified by any debugger subscribed exception
that comes from hardware interrupts.
If a debugger session exits, any exceptions it subscribed to may still
have interrupts in the interrupt ring buffer or KGD/KFD pipeline.
To prevent a new session from inheriting stale interrupts, when a new
queue is created, open an interrupt drain and allow the IH ring to drain
from a timestamped checkpoint. Then inject a custom IV so that once
the custom IV is picked up by the KFD, it's safe to close the drain
and proceed with queue creation.
The drain must also be on debug disable as SW interrupts may still
be processed. Drain at this time and clear all the exception status.
The debugger may also not be attached nor subscibed to certain
exceptions so forward them directly to the runtime.
GFX10 also requires its own IV processing, hence the creation of
kfd_int_process_v10.c. This is because the IV from SQ interrupts are
packed into a new continguous format unlike GFX9. To make this clear,
a separate interrupting handling code file was created.
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>
|
|
Trap handler behavior will differ when a debugger is attached.
Make the debug trap flag available in the trap handler TMA.
Update it when the debug trap ioctl is invoked.
Signed-off-by: Jay Cornwall <jay.cornwall@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
|
|
The debugger can attach to a process prior to HSA enablement (i.e.
inferior is spawned by the debugger and attached to immediately before
target process has been enabled for HSA dispatches) or it
can attach to a running target that is already HSA enabled. Either
way, the debugger needs to know the enablement status to know when
it can inspect queues.
For the scenario where the debugger spawns the target process,
it will have to wait for ROCr's runtime enable request from the target.
The runtime enable request will be able to see that its process has been
debug attached. ROCr raises an EC_PROCESS_RUNTIME signal to the
debugger then blocks the target process while waiting the debugger's
response. Once the debugger has received the runtime signal, it will
unblock the target process.
For the scenario where the debugger attaches to a running target
process, ROCr will set the target process' runtime status as enabled so
that on an attach request, the debugger will be able to see this
status and will continue with debug enablement as normal.
A secondary requirement is to conditionally enable the trap tempories only
if the user requests it (env var HSA_ENABLE_DEBUG=1) or if the debugger
attaches with HSA runtime enabled. This is because setting up the trap
temporaries incurs a performance overhead that is unacceptable for
microbench performance in normal mode for certain customers.
In the scenario where the debugger spawns the target process, when ROCr
detects that the debugger has attached during the runtime enable
request, it will enable the trap temporaries before it blocks the target
process while waiting for the debugger to respond.
In the scenario where the debugger attaches to a running target process,
it will enable to trap temporaries itself.
Finally, there is an additional restriction that is required to be
enforced with runtime enable and HW debug mode setting. The debugger must
first ensure that HW debug mode has been enabled before permitting HW debug
mode operations.
With single process debug devices, allowing the debugger to set debug
HW modes prior to trap activation means that debug HW mode setting can
occur before the KFD has reserved the debug VMID (0xf) from the hardware
scheduler's VMID allocation resource pool. This can result in the
hardware scheduler assigning VMID 0xf to a non-debugged process and
having that process inherit debug HW mode settings intended for the
debugged target process instead, which is both incorrect and potentially
fatal for normal mode operation.
With multi process debug devices, allowing the debugger to set debug
HW modes prior to trap activation means that non-debugged processes
migrating to a new VMID could inherit unintended debug settings.
All debug operations that touch HW settings must require trap activation
where trap activation is triggered by both debug attach and runtime
enablement (target has KFD opened and is ready to dispatch work).
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 a debug operation that allows the debugger to send an exception
directly to runtime through a payload address.
For memory violations, normal vmfault signals will be applied to
notify runtime instead after passing in the saved exception data
when a memory violation was raised to the debugger.
For runtime exceptions, this will unblock the runtime enable
function which will be explained and implemented in a follow up
patch.
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>
|
|
Exception events can be generated from interrupts or queue activitity.
The raise event function will save exception status of a queue, device
or process then notify the debugger of the status change by writing to
a debugger polled file descriptor that the debugger provides during
debug attach.
For memory violation exceptions, extra exception data will be saved.
The debugger will be able to query the saved exception states by query
operation that will be provided by follow up patches.
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>
|
|
Due to a HW bug, waves in only half the shader arrays can enter trap.
When starting a debug session, relocate all waves to the first shader
array of each shader engine and mask off the 2nd shader array as
unavailable.
When ending a debug session, re-enable the 2nd shader array per
shader engine.
User CU masking per queue cannot be guaranteed to remain functional
if requested during debugging (e.g. user cu mask requests only 2nd shader
array as an available resource leading to zero HW resources available)
nor can runtime be alerted of any of these changes during execution.
Make user CU masking and debugging mutual exclusive with respect to
availability.
If the debugger tries to attach to a process with a user cu masked
queue, return the runtime status as enabled but busy.
If the debugger tries to attach and fails to reallocate queue waves to
the first shader array of each shader engine, return the runtime status
as enabled but with an error.
In addition, like any other mutli-process debug supported devices,
disable trap temporary setup per-process to avoid performance impact from
setup overhead.
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>
|
|
Unlike single process debug devices, multi-process debug devices allow
debug mode setting per-VMID (non-device-global).
Because the HWS manages PASID-VMID mapping, the new MAP_PROCESS API allows
the KFD to forward the required SPI debug register write requests.
To request a new debug mode setting change, the KFD must be able to
preempt all queues then remap all queues with these new setting
requests for MAP_PROCESS to take effect.
Note that by default, trap enablement in non-debug mode must be disabled
for performance reasons for multi-process debug devices due to setup
overhead in FW.
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>
|
|
The HWS schedule allows a grace period for wave completion prior to
preemption for better performance by avoiding CWSR on waves that can
potentially complete quickly. The debugger, on the other hand, will
want to inspect wave status immediately after it actively triggers
preemption (a suspend function to be provided).
To minimize latency between preemption and debugger wave inspection, allow
immediate preemption by setting the grace period to 0.
Note that setting the preepmtion grace period to 0 will result in an
infinite grace period being set due to a CP FW bug so set it to 1 for now.
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>
|
|
The ROCm debugger will attach to a process to debug by PTRACE and will
expect the KFD to prepare a process for the target PID, whether the
target PID has opened the KFD device or not.
This patch is to explicity handle this requirement. Further HW mode
setting and runtime coordination requirements will be handled in
following patches.
In the case where the target process has not opened the KFD device,
a new KFD process must be created for the target PID.
The debugger as well as the target process for this case will have not
acquired any VMs so handle process restoration to correctly account for
this.
To coordinate with HSA runtime, the debugger must be aware of the target
process' runtime enablement status and will copy the runtime status
information into the debugged KFD process for later query.
On enablement, the debugger will subscribe to a set of exceptions where
each exception events will notify the debugger through a pollable FIFO
file descriptor that the debugger provides to the KFD to manage.
Finally on process termination of either the debugger or the target,
debugging must be disabled if it has not been done so.
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>
|
|
We need to track memory usage on a per partition basis. To do
that, store the local memory information in KFD node instead
of kfd device.
v2: squash in fix ("amdkfd: Use mem_id to access mem_partition info")
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>
|
