Currently this is very broken if someone attempts to create a bind
queue and share it across multiple VMs. For example currently we assume
it is safe to acquire the user VM lock to protect some of the bind queue
state, but if allow sharing the bind queue with multiple VMs then this
quickly breaks down.
To fix this reject using a bind queue with any VM that is not the same
VM that was originally passed when creating the bind queue. This a uAPI
change, however this was more of an oversight on kernel side that we
didn't reject this, and expectation is that userspace shouldn't be using
bind queues in this way, so in theory this change should go unnoticed.
Based on a patch from Matt Brost.
v2 (Matt B):
- Hold the vm lock over queue create, to ensure it can't be closed as
we attach the user_vm to the queue.
- Make sure we actually check for NULL user_vm in destruction path.
v3:
- Fix error path handling.
Fixes: dd08ebf6c3 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
Reported-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: José Roberto de Souza <jose.souza@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Michal Mrozek <michal.mrozek@intel.com>
Cc: Carl Zhang <carl.zhang@intel.com>
Cc: <stable@vger.kernel.org> # v6.8+
Acked-by: José Roberto de Souza <jose.souza@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Arvind Yadav <arvind.yadav@intel.com>
Acked-by: Michal Mrozek <michal.mrozek@intel.com>
Link: https://patch.msgid.link/20260120110609.77958-3-matthew.auld@intel.com
(cherry picked from commit 9dd08fdecc)
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
If an exec queue is idle, there is no need to issue a schedule disable
to the GuC when suspending the queue’s execution. Opportunistically skip
this step if the queue is idle and not a parallel queue. Parallel queues
must have their scheduling state flipped in the GuC due to limitations
in how submission is implemented in run_job().
Also if all pagefault queues can skip the schedule disable during a
switch to dma-fence mode, do not schedule a resume for the pagefault
queues after the next submission.
v2:
- Don't touch the LRC tail is queue is suspended but enabled in run_job
(CI)
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patch.msgid.link/20251212182847.1683222-5-matthew.brost@intel.com
Multi Queue is a new mode of execution supported by the compute and
blitter copy command streamers (CCS and BCS, respectively). It is an
enhancement of the existing hardware architecture and leverages the
same submission model. It enables support for efficient, parallel
execution of multiple queues within a single context. All the queues
of a group must use the same address space (VM).
The new DRM_XE_EXEC_QUEUE_SET_PROPERTY_MULTI_QUEUE execution queue
property supports creating a multi queue group and adding queues to
a queue group. All queues of a multi queue group share the same
context.
A exec queue create ioctl call with above property specified with value
DRM_XE_SUPER_GROUP_CREATE will create a new multi queue group with the
queue being created as the primary queue (aka q0) of the group. To add
secondary queues to the group, they need to be created with the above
property with id of the primary queue as the value. The properties of
the primary queue (like priority, timeslice) applies to the whole group.
So, these properties can't be set for secondary queues of a group.
Once destroyed, the secondary queues of a multi queue group can't be
replaced. However, they can be dynamically added to the group up to a
total of 64 queues per group. Once the primary queue is destroyed,
secondary queues can't be added to the queue group.
v2: Remove group->lock, fix xe_exec_queue_group_add()/delete()
function semantics, add additional comments, remove unused
group->list_lock, add XE_BO_FLAG_GGTT_INVALIDATE for cgp bo,
Assert LRC is valid, update uapi kernel doc.
(Matt Brost)
v3: Use XE_BO_FLAG_PINNED_LATE_RESTORE/USER_VRAM/GGTT_INVALIDATE
flags for cgp bo (Matt)
v4: Ensure queue is not a vm_bind queue
uapi change due to rebase
Signed-off-by: Stuart Summers <stuart.summers@intel.com>
Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://patch.msgid.link/20251211010249.1647839-21-niranjana.vishwanathapura@intel.com
Add support for attaching the last fence to TLB invalidation job queues
to address serialization issues during bursts of unbind jobs. Ensure
that user fence signaling for a bind job reflects both the bind job
itself and the last fences of all related TLB invalidations. Maintain
submission order based solely on the state of the bind and TLB
invalidation queues.
Introduce support functions for last fence attachment to TLB
invalidation queues.
v3:
- Fix assert in xe_exec_queue_tlb_inval_last_fence_set (CI)
- Ensure migrate lock held for migrate queues (Testing)
v5:
- Style nits (Thomas)
- Rewrite commit message (Thomas)
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patch.msgid.link/20251031234050.3043507-3-matthew.brost@intel.com
VF migration requires jobs to remain pending so they can be replayed
after the VF comes back. Previously, LR job fences were intentionally
signaled immediately after submission to avoid the risk of exporting
them, as these fences do not naturally signal in a timely manner and
could break dma-fence contracts. A side effect of this approach was that
LR jobs were never added to the DRM scheduler’s pending list, preventing
them from being tracked for later resubmission.
We now avoid signaling LR job fences and ensure they are never exported;
Xe already guards against exporting these internal fences. With that
guarantee in place, we can safely track LR jobs in the scheduler’s
pending list so they are eligible for resubmission during VF
post-migration recovery (and similar recovery paths).
An added benefit is that LR queues now gain the DRM scheduler’s built-in
flow control over ring usage rather than rejecting new jobs in the exec
IOCTL if the ring is full.
v2:
- Ensure DRM scheduler TDR doesn't run for LR jobs
- Stack variable for killed_or_banned_or_wedged
v4:
- Clarify commit message (Tomasz)
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Tomasz Lis <tomasz.lis@intel.com>
Link: https://lore.kernel.org/r/20251008214532.3442967-5-matthew.brost@intel.com
The WA buffer we use to capture context utilization contains GGTT
references. This means its instructions have to be either fixed or
re-emitted during VF post-migration recovery.
This patch adds re-emitting content of the utilization WA BB during
the recovery.
The way we write to vram requires scratch buffer to be used before
the whole block is memcopied. We are re-using a scratch buffer
introduced in earlier part of the recovery. This is not a performance
optimization, but a necessity to avoid creating dependencies between
locks.
v2: Notable rebase after "Prepare WA BB setup for more users" patch
v3: Added error propagation
Signed-off-by: Tomasz Lis <tomasz.lis@intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
Cc: Michal Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Winiarski <michal.winiarski@intel.com>
Link: https://lore.kernel.org/r/20250802031045.1127138-8-tomasz.lis@intel.com
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
All contexts require an update of state data, as the data includes
GGTT references to memirq-related buffers.
Default contexts need these references updated as well, because they
are not refreshed when a new context is created from them.
The way we write to vram requires scratch buffer to be used
before the whole block is memcopied. Since using kalloc() within
specific recovery functions would lead to unintended relations
between locks, we are allocating the buffer earlier, before
any locks are taken. The same buffer will be used for other steps
of the recovery.
v2: Update addresses by xe_lrc_write_ctx_reg() rather than
set_memory_based_intr()
v3: Renamed parameter, reordered parameters in some functs
v4: Check if have MEMIRQ, move `xe_gt*` funct to proper file
v5: Revert back to requiring scratch buffer, but allocate it
earlier this time
Signed-off-by: Tomasz Lis <tomasz.lis@intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
Cc: Michal Winiarski <michal.winiarski@intel.com>
Acked-by: Satyanarayana K V P <satyanarayana.k.v.p@intel.com>
Reviewed-by: Michal Winiarski <michal.winiarski@intel.com>
Link: https://lore.kernel.org/r/20250802031045.1127138-6-tomasz.lis@intel.com
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Userspace is required to mark a queue as using PXP to guarantee that the
PXP instructions will work. In addition to managing the PXP sessions,
when a PXP queue is created the driver will set the relevant bits in
its context control register.
On submission of a valid PXP queue, the driver will validate all
encrypted objects mapped to the VM to ensured they were encrypted with
the current key.
v2: Remove pxp_types include outside of PXP code (Jani), better comments
and code cleanup (John)
v3: split the internal PXP management to a separate patch for ease of
review. re-order ioctl checks to always return -EINVAL if parameters are
invalid, rebase on msix changes.
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20250129174140.948829-9-daniele.ceraolospurio@intel.com
Ensure we can safely take a ref of the exec queue's last fence from the
context of resuming jobs from the hw engine group. The locking requirements
differ from the general case, hence the introduction of this new function.
v2: Add kernel doc, rework the code to prevent code duplication
v3: Fix kernel doc, remove now unnecessary lockdep variants (Matt Brost)
v4: Remove new put function (Matt Brost)
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240809155156.1955925-7-francois.dugast@intel.com
Add a helper to accumulate per-client runtime of all its
exec queues. This is called every time a sched job is finished.
v2:
- Use guc_exec_queue_free_job() and execlist_job_free() to accumulate
runtime when job is finished since xe_sched_job_completed() is not a
notification that job finished.
- Stop trying to update runtime from xe_exec_queue_fini() - that is
redundant and may happen after xef is closed, leading to a
use-after-free
- Do not special case the first timestamp read: the default LRC sets
CTX_TIMESTAMP to zero, so even the first sample should be a valid
one.
- Handle the parallel submission case by multiplying the runtime by
width.
v3: Update comments
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240517204310.88854-6-lucas.demarchi@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
In many validation situations when debugging GPU Hangs,
it is useful to preserve the GT situation from the moment
that the timeout occurred.
This patch introduces a module parameter that could be used
on situations like this.
If xe.wedged module parameter is set to 2, Xe will be declared
wedged on every single execution timeout (a.k.a. GPU hang) right
after devcoredump snapshot capture and without attempting any
kind of GT reset and blocking entirely any kind of execution.
v2: Really block gt_reset from guc side. (Lucas)
s/wedged/busted (Lucas)
v3: - s/busted/wedged
- Really use global_flags (Dafna)
- More robust timeout handling when wedging it.
v4: A really robust clean exit done by Matt Brost.
No more kernel warns on unbind.
v5: Simplify error message (Lucas)
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Dafna Hirschfeld <dhirschfeld@habana.ai>
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Cc: Alan Previn <alan.previn.teres.alexis@intel.com>
Cc: Himanshu Somaiya <himanshu.somaiya@intel.com>
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240423221817.1285081-3-rodrigo.vivi@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Setting of exec_queue user extensions is moved from the end of the ioctl
function earlier, into __xe_exec_queue_alloc().
This fixes bug in that the USM attributes for access counters were being
applied too late, and effectively were ignored.
However, in order to apply user extensions this early, we can no longer
call q->ops functions. Instead, make it more efficient. The user extension
functions can simply update the q->sched_props values and they will be
applied by the backend during q->ops->init().
v2: minor changes for readability (Matt)
Signed-off-by: Brian Welty <brian.welty@intel.com>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
A case existed where an out-sync of a later VM bind operation could
signal before a previous one if the later operation results in a NOP
(e.g. a unbind or prefetch to a VA range without any mappings). This
breaks the ordering rules, fix this. This patch also lays the groundwork
for users to pass in num_binds == 0 and out-syncs.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
