linux.git/drivers/gpu/drm/xe/xe_exec_queue.c, branch v6.8-rc2

drm/xe: Fix modifying exec_queue priority in xe_migrate_init

2024-01-15T14:36:50+00:00

After exec_queue has been created, we cannot simply modify q->priority.
This needs to be done by the backend via q->ops.  However in this case,
it would be more efficient to simply pass a flag when creating the
exec_queue and set the desired priority upfront during queue creation.

To that end: new flag EXEC_QUEUE_FLAG_HIGH_PRIORITY is introduced.
The priority field is moved to be with other scheduling properties and
is now exec_queue.sched_props.priority. This is no longer set to initial
value by the backend, but is now set within __xe_exec_queue_create().

Fixes: b4eecedc75c1 ("drm/xe: Fix potential deadlock handling page faults")
Signed-off-by: Brian Welty 
Signed-off-by: Matthew Brost 
Reviewed-by: Matthew Brost 
(cherry picked from commit a8004af338f6b3319476ecbed63ea49bf393fc1f)
Signed-off-by: Thomas Hellström

drm/xe/uapi: Remove sync binds

2023-12-21T16:46:59+00:00

Remove concept of async vs sync VM bind queues, rather make all binds
async.

The following bits have dropped from the uAPI:
DRM_XE_ENGINE_CLASS_VM_BIND_ASYNC
DRM_XE_ENGINE_CLASS_VM_BIND_SYNC
DRM_XE_VM_CREATE_FLAG_ASYNC_DEFAULT
DRM_XE_VM_BIND_FLAG_ASYNC

To implement sync binds the UMD is expected to use the out-fence
interface.

v2: Send correct version
v3: Drop drm_xe_syncs

Cc: Rodrigo Vivi 
Cc: Thomas Hellström 
Cc: Francois Dugast 
Signed-off-by: Matthew Brost 
Reviewed-by: Thomas Hellström 
Acked-by: José Roberto de Souza 
Acked-by: Mateusz Naklicki 
Signed-off-by: Rodrigo Vivi

drm/xe/uapi: Ensure every uapi struct has drm_xe prefix

2023-12-21T16:46:59+00:00

To ensure consistency and avoid possible later conflicts,
let's add drm_xe prefix to xe_user_extension struct.

Cc: Francois Dugast 
Suggested-by: Lucas De Marchi 
Signed-off-by: Rodrigo Vivi 
Reviewed-by: Matthew Brost 
Acked-by: José Roberto de Souza 
Acked-by: Mateusz Naklicki

drm/xe: Allow num_batch_buffer / num_binds == 0 in IOCTLs

2023-12-21T16:46:09+00:00

The idea being out-syncs can signal indicating all previous operations
on the bind queue are complete. An example use case of this would be
support for implementing vkQueueWaitIdle easily.

All in-syncs are waited on before signaling out-syncs. This is
implemented by forming a composite software fence of in-syncs and
installing this fence in the out-syncs and exec queue last fence slot.

The last fence must be added as a dependency for jobs on user exec
queues as it is possible for the last fence to be a composite software
fence (unordered, ioctl with zero bb or binds) rather than hardware
fence (ordered, previous job on queue).

Cc: Thomas Hellström 
Signed-off-by: Matthew Brost 
Reviewed-by: Thomas Hellström 
Signed-off-by: Rodrigo Vivi

drm/xe: Rename info.supports_* to info.has_*

2023-12-21T16:45:27+00:00

Rename supports_mmio_ext and supports_usm to use a has_ prefix so the
flags are grouped together. This settles on just one variant for
positive info matching ("has_") and one for negative ("skip_").

Also make sure the has_* flags are grouped together in xe_pci.c.

Reviewed-by: Koby Elbaz 
Reviewed-by: Gustavo Sousa 
Link: https://lore.kernel.org/r/20231205145235.2114761-1-lucas.demarchi@intel.com
Signed-off-by: Lucas De Marchi 
Signed-off-by: Rodrigo Vivi

drm/xe/uapi: Kill exec_queue_set_property

2023-12-21T16:45:23+00:00

All the properties should be immutable and set upon exec_queue creation
using the existent extension. So, let's kill this useless and dangerous
uapi.

Cc: Francois Dugast 
Cc: José Roberto de Souza 
Cc: Matthew Brost 
Signed-off-by: Rodrigo Vivi 
Reviewed-by: José Roberto de Souza 
Signed-off-by: Francois Dugast

drm/xe: fix mem_access for early lrc generation

2023-12-21T16:45:07+00:00

We spawn some hw queues during device probe to generate the default LRC
for every engine type, however the queue destruction step is typically
async. Queue destruction needs to do stuff like GuC context deregister
which requires GuC CT, which in turn requires an active mem_access ref.
The caller during probe is meant to hold the mem_access token, however
due to the async destruction it might have already been dropped if we
are unlucky.

Similar to how we already handle migrate VMs for which there is no
mem_access ref, fix this by keeping the callers token alive, releasing
it only when destroying the queue. We can treat a NULL vm as indication
that we need to grab our own extra ref.

Fixes the following splat sometimes seen during load:

[ 1682.899930] WARNING: CPU: 1 PID: 8642 at drivers/gpu/drm/xe/xe_device.c:537 xe_device_assert_mem_access+0x27/0x30 [xe]
[ 1682.900209] CPU: 1 PID: 8642 Comm: kworker/u24:97 Tainted: G     U  W   E    N 6.6.0-rc3+ #6
[ 1682.900214] Workqueue: submit_wq xe_sched_process_msg_work [xe]
[ 1682.900303] RIP: 0010:xe_device_assert_mem_access+0x27/0x30 [xe]
[ 1682.900388] Code: 90 90 90 66 0f 1f 00 0f 1f 44 00 00 53 48 89 fb e8 1e 6c 03 00 48 85 c0 74 06 5b c3 cc cc cc cc 8b 83 28 23 00 00 85 c0 75 f0 <0f> 0b 5b c3 cc cc cc cc 90 90 90 90 90 90 90 90 90 90 90 90 90 90
[ 1682.900390] RSP: 0018:ffffc900021cfb68 EFLAGS: 00010246
[ 1682.900394] RAX: 0000000000000000 RBX: ffff8886a96d8000 RCX: 0000000000000000
[ 1682.900396] RDX: 0000000000000001 RSI: ffff8886a6311a00 RDI: ffff8886a96d8000
[ 1682.900398] RBP: ffffc900021cfcc0 R08: 0000000000000001 R09: 0000000000000000
[ 1682.900400] R10: ffffc900021cfcd0 R11: 0000000000000002 R12: 0000000000000004
[ 1682.900402] R13: 0000000000000000 R14: ffff8886a6311990 R15: ffffc900021cfd74
[ 1682.900405] FS:  0000000000000000(0000) GS:ffff888829880000(0000) knlGS:0000000000000000
[ 1682.900407] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 1682.900409] CR2: 000055f70bad3fb0 CR3: 000000025243a004 CR4: 00000000003706e0
[ 1682.900412] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ 1682.900413] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ 1682.900415] Call Trace:
[ 1682.900418]  
[ 1682.900420]  ? xe_device_assert_mem_access+0x27/0x30 [xe]
[ 1682.900504]  ? __warn+0x85/0x170
[ 1682.900510]  ? xe_device_assert_mem_access+0x27/0x30 [xe]
[ 1682.900596]  ? report_bug+0x171/0x1a0
[ 1682.900604]  ? handle_bug+0x3c/0x80
[ 1682.900608]  ? exc_invalid_op+0x17/0x70
[ 1682.900612]  ? asm_exc_invalid_op+0x1a/0x20
[ 1682.900621]  ? xe_device_assert_mem_access+0x27/0x30 [xe]
[ 1682.900706]  ? xe_device_assert_mem_access+0x12/0x30 [xe]
[ 1682.900790]  guc_ct_send_locked+0xb9/0x1550 [xe]
[ 1682.900882]  ? lock_acquire+0xca/0x2b0
[ 1682.900885]  ? guc_ct_send+0x3c/0x1a0 [xe]
[ 1682.900977]  ? lock_is_held_type+0x9b/0x110
[ 1682.900984]  ? __mutex_lock+0xc0/0xb90
[ 1682.900989]  ? __pfx___drm_printfn_info+0x10/0x10
[ 1682.900999]  guc_ct_send+0x53/0x1a0 [xe]
[ 1682.901090]  ? __lock_acquire+0xf22/0x21b0
[ 1682.901097]  ? process_one_work+0x1a0/0x500
[ 1682.901109]  xe_guc_ct_send+0x19/0x50 [xe]
[ 1682.901202]  set_min_preemption_timeout+0x75/0xa0 [xe]
[ 1682.901294]  disable_scheduling_deregister+0x55/0x250 [xe]
[ 1682.901383]  ? xe_sched_process_msg_work+0x76/0xd0 [xe]
[ 1682.901467]  ? lock_release+0xc9/0x260
[ 1682.901474]  xe_sched_process_msg_work+0x82/0xd0 [xe]
[ 1682.901559]  process_one_work+0x20a/0x500

v2: Add the splat

Signed-off-by: Matthew Auld 
Cc: Vinay Belgaumkar 
Cc: Matthew Brost 
Cc: Rodrigo Vivi 
Reviewed-by: Matthew Brost 
Signed-off-by: Rodrigo Vivi

drm/xe: Internally change the compute_mode and no_dma_fence mode naming

2023-12-21T16:44:58+00:00

The name "compute_mode" can be confusing since compute uses either this
mode or fault_mode to achieve the long-running semantics, and compute_mode
can, moving forward, enable fault_mode under the hood to work around
hardware limitations.

Also the name no_dma_fence_mode really refers to what we elsewhere call
long-running mode and the mode contrary to what its name suggests allows
dma-fences as in-fences.

So in an attempt to be more consistent, rename
no_dma_fence_mode -> lr_mode
compute_mode      -> preempt_fence_mode

And adjust flags so that

preempt_fence_mode sets XE_VM_FLAG_LR_MODE
fault_mode sets XE_VM_FLAG_LR_MODE | XE_VM_FLAG_FAULT_MODE

v2:
- Fix a typo in the commit message (Oak Zeng)

Signed-off-by: Thomas Hellström 
Reviewed-by: Oak Zeng 
Link: https://patchwork.freedesktop.org/patch/msgid/20231127123349.23698-1-thomas.hellstrom@linux.intel.com
Signed-off-by: Rodrigo Vivi

drm/xe/uapi: Add missing DRM_ prefix in uAPI constants

2023-12-21T16:44:33+00:00

Most constants defined in xe_drm.h use DRM_XE_ as prefix which is
helpful to identify the name space. Make this systematic and add
this prefix where it was missing.

v2:
- fix vertical alignment of define values
- remove double DRM_ in some variables (José Roberto de Souza)

v3: Rebase

Signed-off-by: Francois Dugast 
Reviewed-by: Matthew Brost 
Signed-off-by: Rodrigo Vivi

drm/xe/xe_exec_queue: Add check for access counter granularity

2023-12-21T16:43:31+00:00

Add conditional check for access counter granularity.
This check will return -EINVAL if granularity is beyond 64M
which is a hardware limitation.

v2: Defined
XE_ACC_GRANULARITY_128K 0
XE_ACC_GRANULARITY_2M 1
XE_ACC_GRANULARITY_16M 2
XE_ACC_GRANULARITY_64M 3
as part of uAPI.
So, that user can also use it.(Oak)

v3: Move uAPI to proper location and give proper
documentation.(Brian, Oak)

Cc: Oak Zeng 
Cc: Janga Rahul Kumar 
Cc: Brian Welty 
Signed-off-by: Priyanka Dandamudi 
Reviewed-by: Oak Zeng 
Reviewed-by: Oak Zeng 
Signed-off-by: Rodrigo Vivi