Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
Merge tag 'drm-xe-next-fixes-2025-12-04' of https://gitlab.freedesktop.org/drm/xe/kernel into drm-next
Driver Changes:
- Fix a memory leak (Mika)
- Fix a 64-bit division (Michal Wajdeczko)
- vf migration fix (Matt Brost)
- LRC pause Fix (Tomasz lis)
Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Thomas Hellstrom <thomas.hellstrom@linux.intel.com>
Link: https://patch.msgid.link/aTIGiHJnnMtqbDOO@fedora
+36
-23
+3
-2
drivers/gpu/drm/xe/xe_gpu_scheduler.h
+3
-2
drivers/gpu/drm/xe/xe_gpu_scheduler.h
···
54
54
static inline void xe_sched_resubmit_jobs(struct xe_gpu_scheduler *sched)
55
55
{
56
56
struct drm_sched_job *s_job;
57
+
bool restore_replay = false;
57
58
58
59
list_for_each_entry(s_job, &sched->base.pending_list, list) {
59
60
struct drm_sched_fence *s_fence = s_job->s_fence;
60
61
struct dma_fence *hw_fence = s_fence->parent;
61
62
62
-
if (to_xe_sched_job(s_job)->skip_emit ||
63
-
(hw_fence && !dma_fence_is_signaled(hw_fence)))
63
+
restore_replay |= to_xe_sched_job(s_job)->restore_replay;
64
+
if (restore_replay || (hw_fence && !dma_fence_is_signaled(hw_fence)))
64
65
sched->base.ops->run_job(s_job);
65
66
}
66
67
}
+1
-1
drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
+1
-1
drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
+30
-17
drivers/gpu/drm/xe/xe_guc_submit.c
+30
-17
drivers/gpu/drm/xe/xe_guc_submit.c
···
822
822
823
823
xe_gt_assert(guc_to_gt(guc), exec_queue_registered(q));
824
824
825
-
if (!job->skip_emit || job->last_replay) {
825
+
if (!job->restore_replay || job->last_replay) {
826
826
if (xe_exec_queue_is_parallel(q))
827
827
wq_item_append(q);
828
828
else
···
881
881
if (!killed_or_banned_or_wedged && !xe_sched_job_is_error(job)) {
882
882
if (!exec_queue_registered(q))
883
883
register_exec_queue(q, GUC_CONTEXT_NORMAL);
884
-
if (!job->skip_emit)
884
+
if (!job->restore_replay)
885
885
q->ring_ops->emit_job(job);
886
886
submit_exec_queue(q, job);
887
-
job->skip_emit = false;
887
+
job->restore_replay = false;
888
888
}
889
889
890
890
/*
···
2112
2112
q->guc->resume_time = 0;
2113
2113
}
2114
2114
2115
+
static void lrc_parallel_clear(struct xe_lrc *lrc)
2116
+
{
2117
+
struct xe_device *xe = gt_to_xe(lrc->gt);
2118
+
struct iosys_map map = xe_lrc_parallel_map(lrc);
2119
+
int i;
2120
+
2121
+
for (i = 0; i < WQ_SIZE / sizeof(u32); ++i)
2122
+
parallel_write(xe, map, wq[i],
2123
+
FIELD_PREP(WQ_TYPE_MASK, WQ_TYPE_NOOP) |
2124
+
FIELD_PREP(WQ_LEN_MASK, 0));
2125
+
}
2126
+
2115
2127
/*
2116
2128
* This function is quite complex but only real way to ensure no state is lost
2117
2129
* during VF resume flows. The function scans the queue state, make adjustments
···
2147
2135
guc_exec_queue_revert_pending_state_change(guc, q);
2148
2136
2149
2137
if (xe_exec_queue_is_parallel(q)) {
2150
-
struct xe_device *xe = guc_to_xe(guc);
2151
-
struct iosys_map map = xe_lrc_parallel_map(q->lrc[0]);
2138
+
/* Pairs with WRITE_ONCE in __xe_exec_queue_init */
2139
+
struct xe_lrc *lrc = READ_ONCE(q->lrc[0]);
2152
2140
2153
2141
/*
2154
2142
* NOP existing WQ commands that may contain stale GGTT
···
2156
2144
* seems to get confused if the WQ head/tail pointers are
2157
2145
* adjusted.
2158
2146
*/
2159
-
for (i = 0; i < WQ_SIZE / sizeof(u32); ++i)
2160
-
parallel_write(xe, map, wq[i],
2161
-
FIELD_PREP(WQ_TYPE_MASK, WQ_TYPE_NOOP) |
2162
-
FIELD_PREP(WQ_LEN_MASK, 0));
2147
+
if (lrc)
2148
+
lrc_parallel_clear(lrc);
2163
2149
}
2164
2150
2165
2151
job = xe_sched_first_pending_job(sched);
2166
2152
if (job) {
2153
+
job->restore_replay = true;
2154
+
2167
2155
/*
2168
2156
* Adjust software tail so jobs submitted overwrite previous
2169
2157
* position in ring buffer with new GGTT addresses.
···
2253
2241
struct xe_exec_queue *q)
2254
2242
{
2255
2243
struct xe_gpu_scheduler *sched = &q->guc->sched;
2256
-
struct drm_sched_job *s_job;
2257
2244
struct xe_sched_job *job = NULL;
2245
+
bool restore_replay = false;
2258
2246
2259
-
list_for_each_entry(s_job, &sched->base.pending_list, list) {
2260
-
job = to_xe_sched_job(s_job);
2247
+
list_for_each_entry(job, &sched->base.pending_list, drm.list) {
2248
+
restore_replay |= job->restore_replay;
2249
+
if (restore_replay) {
2250
+
xe_gt_dbg(guc_to_gt(guc), "Replay JOB - guc_id=%d, seqno=%d",
2251
+
q->guc->id, xe_sched_job_seqno(job));
2261
2252
2262
-
xe_gt_dbg(guc_to_gt(guc), "Replay JOB - guc_id=%d, seqno=%d",
2263
-
q->guc->id, xe_sched_job_seqno(job));
2264
-
2265
-
q->ring_ops->emit_job(job);
2266
-
job->skip_emit = true;
2253
+
q->ring_ops->emit_job(job);
2254
+
job->restore_replay = true;
2255
+
}
2267
2256
}
2268
2257
2269
2258
if (job)
-1
drivers/gpu/drm/xe/xe_pagefault.c
-1
drivers/gpu/drm/xe/xe_pagefault.c
+2
-2
drivers/gpu/drm/xe/xe_sched_job_types.h
+2
-2
drivers/gpu/drm/xe/xe_sched_job_types.h
···
63
63
bool ring_ops_flush_tlb;
64
64
/** @ggtt: mapped in ggtt. */
65
65
bool ggtt;
66
-
/** @skip_emit: skip emitting the job */
67
-
bool skip_emit;
66
+
/** @restore_replay: job being replayed for restore */
67
+
bool restore_replay;
68
68
/** @last_replay: last job being replayed */
69
69
bool last_replay;
70
70
/** @ptrs: per instance pointers. */