OSDN Git Service

drm/i915: Do not share hwsp across contexts any more, v8.
authorMaarten Lankhorst <maarten.lankhorst@linux.intel.com>
Tue, 23 Mar 2021 15:49:50 +0000 (16:49 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Wed, 24 Mar 2021 10:38:56 +0000 (11:38 +0100)
Instead of sharing pages with breadcrumbs, give each timeline a
single page. This allows unrelated timelines not to share locks
any more during command submission.

As an additional benefit, seqno wraparound no longer requires
i915_vma_pin, which means we no longer need to worry about a
potential -EDEADLK at a point where we are ready to submit.

Changes since v1:
- Fix erroneous i915_vma_acquire that should be a i915_vma_release (ickle).
- Extra check for completion in intel_read_hwsp().
Changes since v2:
- Fix inconsistent indent in hwsp_alloc() (kbuild)
- memset entire cacheline to 0.
Changes since v3:
- Do same in intel_timeline_reset_seqno(), and clflush for good measure.
Changes since v4:
- Use refcounting on timeline, instead of relying on i915_active.
- Fix waiting on kernel requests.
Changes since v5:
- Bump amount of slots to maximum (256), for best wraparounds.
- Add hwsp_offset to i915_request to fix potential wraparound hang.
- Ensure timeline wrap test works with the changes.
- Assign hwsp in intel_timeline_read_hwsp() within the rcu lock to
  fix a hang.
Changes since v6:
- Rename i915_request_active_offset to i915_request_active_seqno(),
  and elaborate the function. (tvrtko)
Changes since v7:
- Move hunk to where it belongs. (jekstrand)
- Replace CACHELINE_BYTES with TIMELINE_SEQNO_BYTES. (jekstrand)

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@intel.com> #v1
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20210323155059.628690-2-maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/gt/gen2_engine_cs.c
drivers/gpu/drm/i915/gt/gen6_engine_cs.c
drivers/gpu/drm/i915/gt/gen8_engine_cs.c
drivers/gpu/drm/i915/gt/intel_engine_cs.c
drivers/gpu/drm/i915/gt/intel_gt_types.h
drivers/gpu/drm/i915/gt/intel_timeline.c
drivers/gpu/drm/i915/gt/intel_timeline_types.h
drivers/gpu/drm/i915/gt/selftest_engine_cs.c
drivers/gpu/drm/i915/gt/selftest_timeline.c
drivers/gpu/drm/i915/i915_request.c
drivers/gpu/drm/i915/i915_request.h

index b491a64..9646200 100644 (file)
@@ -143,7 +143,7 @@ static u32 *__gen2_emit_breadcrumb(struct i915_request *rq, u32 *cs,
                                   int flush, int post)
 {
        GEM_BUG_ON(i915_request_active_timeline(rq)->hwsp_ggtt != rq->engine->status_page.vma);
-       GEM_BUG_ON(offset_in_page(i915_request_active_timeline(rq)->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
+       GEM_BUG_ON(offset_in_page(rq->hwsp_seqno) != I915_GEM_HWS_SEQNO_ADDR);
 
        *cs++ = MI_FLUSH;
 
index ce38d1b..b388cee 100644 (file)
@@ -161,7 +161,7 @@ u32 *gen6_emit_breadcrumb_rcs(struct i915_request *rq, u32 *cs)
                 PIPE_CONTROL_DC_FLUSH_ENABLE |
                 PIPE_CONTROL_QW_WRITE |
                 PIPE_CONTROL_CS_STALL);
-       *cs++ = i915_request_active_timeline(rq)->hwsp_offset |
+       *cs++ = i915_request_active_seqno(rq) |
                PIPE_CONTROL_GLOBAL_GTT;
        *cs++ = rq->fence.seqno;
 
@@ -359,7 +359,7 @@ u32 *gen7_emit_breadcrumb_rcs(struct i915_request *rq, u32 *cs)
                 PIPE_CONTROL_QW_WRITE |
                 PIPE_CONTROL_GLOBAL_GTT_IVB |
                 PIPE_CONTROL_CS_STALL);
-       *cs++ = i915_request_active_timeline(rq)->hwsp_offset;
+       *cs++ = i915_request_active_seqno(rq);
        *cs++ = rq->fence.seqno;
 
        *cs++ = MI_USER_INTERRUPT;
@@ -374,7 +374,7 @@ u32 *gen7_emit_breadcrumb_rcs(struct i915_request *rq, u32 *cs)
 u32 *gen6_emit_breadcrumb_xcs(struct i915_request *rq, u32 *cs)
 {
        GEM_BUG_ON(i915_request_active_timeline(rq)->hwsp_ggtt != rq->engine->status_page.vma);
-       GEM_BUG_ON(offset_in_page(i915_request_active_timeline(rq)->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
+       GEM_BUG_ON(offset_in_page(rq->hwsp_seqno) != I915_GEM_HWS_SEQNO_ADDR);
 
        *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
        *cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
@@ -394,7 +394,7 @@ u32 *gen7_emit_breadcrumb_xcs(struct i915_request *rq, u32 *cs)
        int i;
 
        GEM_BUG_ON(i915_request_active_timeline(rq)->hwsp_ggtt != rq->engine->status_page.vma);
-       GEM_BUG_ON(offset_in_page(i915_request_active_timeline(rq)->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
+       GEM_BUG_ON(offset_in_page(rq->hwsp_seqno) != I915_GEM_HWS_SEQNO_ADDR);
 
        *cs++ = MI_FLUSH_DW | MI_INVALIDATE_TLB |
                MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
index 07ba524..4b69ede 100644 (file)
@@ -338,15 +338,14 @@ static u32 preempt_address(struct intel_engine_cs *engine)
 
 static u32 hwsp_offset(const struct i915_request *rq)
 {
-       const struct intel_timeline_cacheline *cl;
+       const struct intel_timeline *tl;
 
-       /* Before the request is executed, the timeline/cachline is fixed */
+       /* Before the request is executed, the timeline is fixed */
+       tl = rcu_dereference_protected(rq->timeline,
+                                      !i915_request_signaled(rq));
 
-       cl = rcu_dereference_protected(rq->hwsp_cacheline, 1);
-       if (cl)
-               return cl->ggtt_offset;
-
-       return rcu_dereference_protected(rq->timeline, 1)->hwsp_offset;
+       /* See the comment in i915_request_active_seqno(). */
+       return page_mask_bits(tl->hwsp_offset) + offset_in_page(rq->hwsp_seqno);
 }
 
 int gen8_emit_init_breadcrumb(struct i915_request *rq)
index 40024f4..36d7370 100644 (file)
@@ -763,6 +763,7 @@ static int measure_breadcrumb_dw(struct intel_context *ce)
        frame->rq.engine = engine;
        frame->rq.context = ce;
        rcu_assign_pointer(frame->rq.timeline, ce->timeline);
+       frame->rq.hwsp_seqno = ce->timeline->hwsp_seqno;
 
        frame->ring.vaddr = frame->cs;
        frame->ring.size = sizeof(frame->cs);
index a83d3e1..f7dab4f 100644 (file)
@@ -39,10 +39,6 @@ struct intel_gt {
        struct intel_gt_timelines {
                spinlock_t lock; /* protects active_list */
                struct list_head active_list;
-
-               /* Pack multiple timelines' seqnos into the same page */
-               spinlock_t hwsp_lock;
-               struct list_head hwsp_free_list;
        } timelines;
 
        struct intel_gt_requests {
index 037b0e3..0bed879 100644 (file)
 #include "intel_ring.h"
 #include "intel_timeline.h"
 
-#define ptr_set_bit(ptr, bit) ((typeof(ptr))((unsigned long)(ptr) | BIT(bit)))
-#define ptr_test_bit(ptr, bit) ((unsigned long)(ptr) & BIT(bit))
+#define TIMELINE_SEQNO_BYTES 8
 
-#define CACHELINE_BITS 6
-#define CACHELINE_FREE CACHELINE_BITS
-
-struct intel_timeline_hwsp {
-       struct intel_gt *gt;
-       struct intel_gt_timelines *gt_timelines;
-       struct list_head free_link;
-       struct i915_vma *vma;
-       u64 free_bitmap;
-};
-
-static struct i915_vma *__hwsp_alloc(struct intel_gt *gt)
+static struct i915_vma *hwsp_alloc(struct intel_gt *gt)
 {
        struct drm_i915_private *i915 = gt->i915;
        struct drm_i915_gem_object *obj;
@@ -45,220 +33,59 @@ static struct i915_vma *__hwsp_alloc(struct intel_gt *gt)
        return vma;
 }
 
-static struct i915_vma *
-hwsp_alloc(struct intel_timeline *timeline, unsigned int *cacheline)
-{
-       struct intel_gt_timelines *gt = &timeline->gt->timelines;
-       struct intel_timeline_hwsp *hwsp;
-
-       BUILD_BUG_ON(BITS_PER_TYPE(u64) * CACHELINE_BYTES > PAGE_SIZE);
-
-       spin_lock_irq(&gt->hwsp_lock);
-
-       /* hwsp_free_list only contains HWSP that have available cachelines */
-       hwsp = list_first_entry_or_null(&gt->hwsp_free_list,
-                                       typeof(*hwsp), free_link);
-       if (!hwsp) {
-               struct i915_vma *vma;
-
-               spin_unlock_irq(&gt->hwsp_lock);
-
-               hwsp = kmalloc(sizeof(*hwsp), GFP_KERNEL);
-               if (!hwsp)
-                       return ERR_PTR(-ENOMEM);
-
-               vma = __hwsp_alloc(timeline->gt);
-               if (IS_ERR(vma)) {
-                       kfree(hwsp);
-                       return vma;
-               }
-
-               GT_TRACE(timeline->gt, "new HWSP allocated\n");
-
-               vma->private = hwsp;
-               hwsp->gt = timeline->gt;
-               hwsp->vma = vma;
-               hwsp->free_bitmap = ~0ull;
-               hwsp->gt_timelines = gt;
-
-               spin_lock_irq(&gt->hwsp_lock);
-               list_add(&hwsp->free_link, &gt->hwsp_free_list);
-       }
-
-       GEM_BUG_ON(!hwsp->free_bitmap);
-       *cacheline = __ffs64(hwsp->free_bitmap);
-       hwsp->free_bitmap &= ~BIT_ULL(*cacheline);
-       if (!hwsp->free_bitmap)
-               list_del(&hwsp->free_link);
-
-       spin_unlock_irq(&gt->hwsp_lock);
-
-       GEM_BUG_ON(hwsp->vma->private != hwsp);
-       return hwsp->vma;
-}
-
-static void __idle_hwsp_free(struct intel_timeline_hwsp *hwsp, int cacheline)
-{
-       struct intel_gt_timelines *gt = hwsp->gt_timelines;
-       unsigned long flags;
-
-       spin_lock_irqsave(&gt->hwsp_lock, flags);
-
-       /* As a cacheline becomes available, publish the HWSP on the freelist */
-       if (!hwsp->free_bitmap)
-               list_add_tail(&hwsp->free_link, &gt->hwsp_free_list);
-
-       GEM_BUG_ON(cacheline >= BITS_PER_TYPE(hwsp->free_bitmap));
-       hwsp->free_bitmap |= BIT_ULL(cacheline);
-
-       /* And if no one is left using it, give the page back to the system */
-       if (hwsp->free_bitmap == ~0ull) {
-               i915_vma_put(hwsp->vma);
-               list_del(&hwsp->free_link);
-               kfree(hwsp);
-       }
-
-       spin_unlock_irqrestore(&gt->hwsp_lock, flags);
-}
-
-static void __rcu_cacheline_free(struct rcu_head *rcu)
-{
-       struct intel_timeline_cacheline *cl =
-               container_of(rcu, typeof(*cl), rcu);
-
-       /* Must wait until after all *rq->hwsp are complete before removing */
-       i915_gem_object_unpin_map(cl->hwsp->vma->obj);
-       __idle_hwsp_free(cl->hwsp, ptr_unmask_bits(cl->vaddr, CACHELINE_BITS));
-
-       i915_active_fini(&cl->active);
-       kfree(cl);
-}
-
-static void __idle_cacheline_free(struct intel_timeline_cacheline *cl)
-{
-       GEM_BUG_ON(!i915_active_is_idle(&cl->active));
-       call_rcu(&cl->rcu, __rcu_cacheline_free);
-}
-
 __i915_active_call
-static void __cacheline_retire(struct i915_active *active)
+static void __timeline_retire(struct i915_active *active)
 {
-       struct intel_timeline_cacheline *cl =
-               container_of(active, typeof(*cl), active);
+       struct intel_timeline *tl =
+               container_of(active, typeof(*tl), active);
 
-       i915_vma_unpin(cl->hwsp->vma);
-       if (ptr_test_bit(cl->vaddr, CACHELINE_FREE))
-               __idle_cacheline_free(cl);
+       i915_vma_unpin(tl->hwsp_ggtt);
+       intel_timeline_put(tl);
 }
 
-static int __cacheline_active(struct i915_active *active)
+static int __timeline_active(struct i915_active *active)
 {
-       struct intel_timeline_cacheline *cl =
-               container_of(active, typeof(*cl), active);
+       struct intel_timeline *tl =
+               container_of(active, typeof(*tl), active);
 
-       __i915_vma_pin(cl->hwsp->vma);
+       __i915_vma_pin(tl->hwsp_ggtt);
+       intel_timeline_get(tl);
        return 0;
 }
 
-static struct intel_timeline_cacheline *
-cacheline_alloc(struct intel_timeline_hwsp *hwsp, unsigned int cacheline)
-{
-       struct intel_timeline_cacheline *cl;
-       void *vaddr;
-
-       GEM_BUG_ON(cacheline >= BIT(CACHELINE_BITS));
-
-       cl = kmalloc(sizeof(*cl), GFP_KERNEL);
-       if (!cl)
-               return ERR_PTR(-ENOMEM);
-
-       vaddr = i915_gem_object_pin_map(hwsp->vma->obj, I915_MAP_WB);
-       if (IS_ERR(vaddr)) {
-               kfree(cl);
-               return ERR_CAST(vaddr);
-       }
-
-       cl->hwsp = hwsp;
-       cl->vaddr = page_pack_bits(vaddr, cacheline);
-
-       i915_active_init(&cl->active, __cacheline_active, __cacheline_retire);
-
-       return cl;
-}
-
-static void cacheline_acquire(struct intel_timeline_cacheline *cl,
-                             u32 ggtt_offset)
-{
-       if (!cl)
-               return;
-
-       cl->ggtt_offset = ggtt_offset;
-       i915_active_acquire(&cl->active);
-}
-
-static void cacheline_release(struct intel_timeline_cacheline *cl)
-{
-       if (cl)
-               i915_active_release(&cl->active);
-}
-
-static void cacheline_free(struct intel_timeline_cacheline *cl)
-{
-       if (!i915_active_acquire_if_busy(&cl->active)) {
-               __idle_cacheline_free(cl);
-               return;
-       }
-
-       GEM_BUG_ON(ptr_test_bit(cl->vaddr, CACHELINE_FREE));
-       cl->vaddr = ptr_set_bit(cl->vaddr, CACHELINE_FREE);
-
-       i915_active_release(&cl->active);
-}
-
 static int intel_timeline_init(struct intel_timeline *timeline,
                               struct intel_gt *gt,
                               struct i915_vma *hwsp,
                               unsigned int offset)
 {
        void *vaddr;
+       u32 *seqno;
 
        kref_init(&timeline->kref);
        atomic_set(&timeline->pin_count, 0);
 
        timeline->gt = gt;
 
-       timeline->has_initial_breadcrumb = !hwsp;
-       timeline->hwsp_cacheline = NULL;
-
-       if (!hwsp) {
-               struct intel_timeline_cacheline *cl;
-               unsigned int cacheline;
-
-               hwsp = hwsp_alloc(timeline, &cacheline);
+       if (hwsp) {
+               timeline->hwsp_offset = offset;
+               timeline->hwsp_ggtt = i915_vma_get(hwsp);
+       } else {
+               timeline->has_initial_breadcrumb = true;
+               hwsp = hwsp_alloc(gt);
                if (IS_ERR(hwsp))
                        return PTR_ERR(hwsp);
-
-               cl = cacheline_alloc(hwsp->private, cacheline);
-               if (IS_ERR(cl)) {
-                       __idle_hwsp_free(hwsp->private, cacheline);
-                       return PTR_ERR(cl);
-               }
-
-               timeline->hwsp_cacheline = cl;
-               timeline->hwsp_offset = cacheline * CACHELINE_BYTES;
-
-               vaddr = page_mask_bits(cl->vaddr);
-       } else {
-               timeline->hwsp_offset = offset;
-               vaddr = i915_gem_object_pin_map(hwsp->obj, I915_MAP_WB);
-               if (IS_ERR(vaddr))
-                       return PTR_ERR(vaddr);
+               timeline->hwsp_ggtt = hwsp;
        }
 
-       timeline->hwsp_seqno =
-               memset(vaddr + timeline->hwsp_offset, 0, CACHELINE_BYTES);
+       vaddr = i915_gem_object_pin_map(hwsp->obj, I915_MAP_WB);
+       if (IS_ERR(vaddr))
+               return PTR_ERR(vaddr);
+
+       timeline->hwsp_map = vaddr;
+       seqno = vaddr + timeline->hwsp_offset;
+       WRITE_ONCE(*seqno, 0);
+       timeline->hwsp_seqno = seqno;
 
-       timeline->hwsp_ggtt = i915_vma_get(hwsp);
        GEM_BUG_ON(timeline->hwsp_offset >= hwsp->size);
 
        timeline->fence_context = dma_fence_context_alloc(1);
@@ -269,6 +96,7 @@ static int intel_timeline_init(struct intel_timeline *timeline,
        INIT_LIST_HEAD(&timeline->requests);
 
        i915_syncmap_init(&timeline->sync);
+       i915_active_init(&timeline->active, __timeline_active, __timeline_retire);
 
        return 0;
 }
@@ -279,23 +107,18 @@ void intel_gt_init_timelines(struct intel_gt *gt)
 
        spin_lock_init(&timelines->lock);
        INIT_LIST_HEAD(&timelines->active_list);
-
-       spin_lock_init(&timelines->hwsp_lock);
-       INIT_LIST_HEAD(&timelines->hwsp_free_list);
 }
 
-static void intel_timeline_fini(struct intel_timeline *timeline)
+static void intel_timeline_fini(struct rcu_head *rcu)
 {
-       GEM_BUG_ON(atomic_read(&timeline->pin_count));
-       GEM_BUG_ON(!list_empty(&timeline->requests));
-       GEM_BUG_ON(timeline->retire);
+       struct intel_timeline *timeline =
+               container_of(rcu, struct intel_timeline, rcu);
 
-       if (timeline->hwsp_cacheline)
-               cacheline_free(timeline->hwsp_cacheline);
-       else
-               i915_gem_object_unpin_map(timeline->hwsp_ggtt->obj);
+       i915_gem_object_unpin_map(timeline->hwsp_ggtt->obj);
 
        i915_vma_put(timeline->hwsp_ggtt);
+       i915_active_fini(&timeline->active);
+       kfree(timeline);
 }
 
 struct intel_timeline *
@@ -361,9 +184,9 @@ int intel_timeline_pin(struct intel_timeline *tl, struct i915_gem_ww_ctx *ww)
        GT_TRACE(tl->gt, "timeline:%llx using HWSP offset:%x\n",
                 tl->fence_context, tl->hwsp_offset);
 
-       cacheline_acquire(tl->hwsp_cacheline, tl->hwsp_offset);
+       i915_active_acquire(&tl->active);
        if (atomic_fetch_inc(&tl->pin_count)) {
-               cacheline_release(tl->hwsp_cacheline);
+               i915_active_release(&tl->active);
                __i915_vma_unpin(tl->hwsp_ggtt);
        }
 
@@ -372,9 +195,13 @@ int intel_timeline_pin(struct intel_timeline *tl, struct i915_gem_ww_ctx *ww)
 
 void intel_timeline_reset_seqno(const struct intel_timeline *tl)
 {
+       u32 *hwsp_seqno = (u32 *)tl->hwsp_seqno;
        /* Must be pinned to be writable, and no requests in flight. */
        GEM_BUG_ON(!atomic_read(&tl->pin_count));
-       WRITE_ONCE(*(u32 *)tl->hwsp_seqno, tl->seqno);
+
+       memset(hwsp_seqno + 1, 0, TIMELINE_SEQNO_BYTES - sizeof(*hwsp_seqno));
+       WRITE_ONCE(*hwsp_seqno, tl->seqno);
+       clflush(hwsp_seqno);
 }
 
 void intel_timeline_enter(struct intel_timeline *tl)
@@ -450,106 +277,23 @@ static u32 timeline_advance(struct intel_timeline *tl)
        return tl->seqno += 1 + tl->has_initial_breadcrumb;
 }
 
-static void timeline_rollback(struct intel_timeline *tl)
-{
-       tl->seqno -= 1 + tl->has_initial_breadcrumb;
-}
-
 static noinline int
 __intel_timeline_get_seqno(struct intel_timeline *tl,
-                          struct i915_request *rq,
                           u32 *seqno)
 {
-       struct intel_timeline_cacheline *cl;
-       unsigned int cacheline;
-       struct i915_vma *vma;
-       void *vaddr;
-       int err;
-
-       might_lock(&tl->gt->ggtt->vm.mutex);
-       GT_TRACE(tl->gt, "timeline:%llx wrapped\n", tl->fence_context);
-
-       /*
-        * If there is an outstanding GPU reference to this cacheline,
-        * such as it being sampled by a HW semaphore on another timeline,
-        * we cannot wraparound our seqno value (the HW semaphore does
-        * a strict greater-than-or-equals compare, not i915_seqno_passed).
-        * So if the cacheline is still busy, we must detach ourselves
-        * from it and leave it inflight alongside its users.
-        *
-        * However, if nobody is watching and we can guarantee that nobody
-        * will, we could simply reuse the same cacheline.
-        *
-        * if (i915_active_request_is_signaled(&tl->last_request) &&
-        *     i915_active_is_signaled(&tl->hwsp_cacheline->active))
-        *      return 0;
-        *
-        * That seems unlikely for a busy timeline that needed to wrap in
-        * the first place, so just replace the cacheline.
-        */
-
-       vma = hwsp_alloc(tl, &cacheline);
-       if (IS_ERR(vma)) {
-               err = PTR_ERR(vma);
-               goto err_rollback;
-       }
-
-       err = i915_ggtt_pin(vma, NULL, 0, PIN_HIGH);
-       if (err) {
-               __idle_hwsp_free(vma->private, cacheline);
-               goto err_rollback;
-       }
+       u32 next_ofs = offset_in_page(tl->hwsp_offset + TIMELINE_SEQNO_BYTES);
 
-       cl = cacheline_alloc(vma->private, cacheline);
-       if (IS_ERR(cl)) {
-               err = PTR_ERR(cl);
-               __idle_hwsp_free(vma->private, cacheline);
-               goto err_unpin;
-       }
-       GEM_BUG_ON(cl->hwsp->vma != vma);
-
-       /*
-        * Attach the old cacheline to the current request, so that we only
-        * free it after the current request is retired, which ensures that
-        * all writes into the cacheline from previous requests are complete.
-        */
-       err = i915_active_ref(&tl->hwsp_cacheline->active,
-                             tl->fence_context,
-                             &rq->fence);
-       if (err)
-               goto err_cacheline;
+       /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
+       if (TIMELINE_SEQNO_BYTES <= BIT(5) && (next_ofs & BIT(5)))
+               next_ofs = offset_in_page(next_ofs + BIT(5));
 
-       cacheline_release(tl->hwsp_cacheline); /* ownership now xfered to rq */
-       cacheline_free(tl->hwsp_cacheline);
-
-       i915_vma_unpin(tl->hwsp_ggtt); /* binding kept alive by old cacheline */
-       i915_vma_put(tl->hwsp_ggtt);
-
-       tl->hwsp_ggtt = i915_vma_get(vma);
-
-       vaddr = page_mask_bits(cl->vaddr);
-       tl->hwsp_offset = cacheline * CACHELINE_BYTES;
-       tl->hwsp_seqno =
-               memset(vaddr + tl->hwsp_offset, 0, CACHELINE_BYTES);
-
-       tl->hwsp_offset += i915_ggtt_offset(vma);
-       GT_TRACE(tl->gt, "timeline:%llx using HWSP offset:%x\n",
-                tl->fence_context, tl->hwsp_offset);
-
-       cacheline_acquire(cl, tl->hwsp_offset);
-       tl->hwsp_cacheline = cl;
+       tl->hwsp_offset = i915_ggtt_offset(tl->hwsp_ggtt) + next_ofs;
+       tl->hwsp_seqno = tl->hwsp_map + next_ofs;
+       intel_timeline_reset_seqno(tl);
 
        *seqno = timeline_advance(tl);
        GEM_BUG_ON(i915_seqno_passed(*tl->hwsp_seqno, *seqno));
        return 0;
-
-err_cacheline:
-       cacheline_free(cl);
-err_unpin:
-       i915_vma_unpin(vma);
-err_rollback:
-       timeline_rollback(tl);
-       return err;
 }
 
 int intel_timeline_get_seqno(struct intel_timeline *tl,
@@ -559,51 +303,52 @@ int intel_timeline_get_seqno(struct intel_timeline *tl,
        *seqno = timeline_advance(tl);
 
        /* Replace the HWSP on wraparound for HW semaphores */
-       if (unlikely(!*seqno && tl->hwsp_cacheline))
-               return __intel_timeline_get_seqno(tl, rq, seqno);
+       if (unlikely(!*seqno && tl->has_initial_breadcrumb))
+               return __intel_timeline_get_seqno(tl, seqno);
 
        return 0;
 }
 
-static int cacheline_ref(struct intel_timeline_cacheline *cl,
-                        struct i915_request *rq)
-{
-       return i915_active_add_request(&cl->active, rq);
-}
-
 int intel_timeline_read_hwsp(struct i915_request *from,
                             struct i915_request *to,
                             u32 *hwsp)
 {
-       struct intel_timeline_cacheline *cl;
+       struct intel_timeline *tl;
        int err;
 
-       GEM_BUG_ON(!rcu_access_pointer(from->hwsp_cacheline));
-
        rcu_read_lock();
-       cl = rcu_dereference(from->hwsp_cacheline);
-       if (i915_request_signaled(from)) /* confirm cacheline is valid */
-               goto unlock;
-       if (unlikely(!i915_active_acquire_if_busy(&cl->active)))
-               goto unlock; /* seqno wrapped and completed! */
-       if (unlikely(__i915_request_is_complete(from)))
-               goto release;
+       tl = rcu_dereference(from->timeline);
+       if (i915_request_signaled(from) ||
+           !i915_active_acquire_if_busy(&tl->active))
+               tl = NULL;
+
+       if (tl) {
+               /* hwsp_offset may wraparound, so use from->hwsp_seqno */
+               *hwsp = i915_ggtt_offset(tl->hwsp_ggtt) +
+                       offset_in_page(from->hwsp_seqno);
+       }
+
+       /* ensure we wait on the right request, if not, we completed */
+       if (tl && __i915_request_is_complete(from)) {
+               i915_active_release(&tl->active);
+               tl = NULL;
+       }
        rcu_read_unlock();
 
-       err = cacheline_ref(cl, to);
-       if (err)
+       if (!tl)
+               return 1;
+
+       /* Can't do semaphore waits on kernel context */
+       if (!tl->has_initial_breadcrumb) {
+               err = -EINVAL;
                goto out;
+       }
+
+       err = i915_active_add_request(&tl->active, to);
 
-       *hwsp = cl->ggtt_offset;
 out:
-       i915_active_release(&cl->active);
+       i915_active_release(&tl->active);
        return err;
-
-release:
-       i915_active_release(&cl->active);
-unlock:
-       rcu_read_unlock();
-       return 1;
 }
 
 void intel_timeline_unpin(struct intel_timeline *tl)
@@ -612,8 +357,7 @@ void intel_timeline_unpin(struct intel_timeline *tl)
        if (!atomic_dec_and_test(&tl->pin_count))
                return;
 
-       cacheline_release(tl->hwsp_cacheline);
-
+       i915_active_release(&tl->active);
        __i915_vma_unpin(tl->hwsp_ggtt);
 }
 
@@ -622,8 +366,11 @@ void __intel_timeline_free(struct kref *kref)
        struct intel_timeline *timeline =
                container_of(kref, typeof(*timeline), kref);
 
-       intel_timeline_fini(timeline);
-       kfree_rcu(timeline, rcu);
+       GEM_BUG_ON(atomic_read(&timeline->pin_count));
+       GEM_BUG_ON(!list_empty(&timeline->requests));
+       GEM_BUG_ON(timeline->retire);
+
+       call_rcu(&timeline->rcu, intel_timeline_fini);
 }
 
 void intel_gt_fini_timelines(struct intel_gt *gt)
@@ -631,7 +378,6 @@ void intel_gt_fini_timelines(struct intel_gt *gt)
        struct intel_gt_timelines *timelines = &gt->timelines;
 
        GEM_BUG_ON(!list_empty(&timelines->active_list));
-       GEM_BUG_ON(!list_empty(&timelines->hwsp_free_list));
 }
 
 void intel_gt_show_timelines(struct intel_gt *gt,
index e360f50..9c1d767 100644 (file)
@@ -18,7 +18,6 @@
 struct i915_vma;
 struct i915_syncmap;
 struct intel_gt;
-struct intel_timeline_hwsp;
 
 struct intel_timeline {
        u64 fence_context;
@@ -45,12 +44,11 @@ struct intel_timeline {
        atomic_t pin_count;
        atomic_t active_count;
 
+       void *hwsp_map;
        const u32 *hwsp_seqno;
        struct i915_vma *hwsp_ggtt;
        u32 hwsp_offset;
 
-       struct intel_timeline_cacheline *hwsp_cacheline;
-
        bool has_initial_breadcrumb;
 
        /**
@@ -67,6 +65,8 @@ struct intel_timeline {
         */
        struct i915_active_fence last_request;
 
+       struct i915_active active;
+
        /** A chain of completed timelines ready for early retirement. */
        struct intel_timeline *retire;
 
@@ -90,15 +90,4 @@ struct intel_timeline {
        struct rcu_head rcu;
 };
 
-struct intel_timeline_cacheline {
-       struct i915_active active;
-
-       struct intel_timeline_hwsp *hwsp;
-       void *vaddr;
-
-       u32 ggtt_offset;
-
-       struct rcu_head rcu;
-};
-
 #endif /* __I915_TIMELINE_TYPES_H__ */
index 439c898..fd9414b 100644 (file)
@@ -42,6 +42,9 @@ static int perf_end(struct intel_gt *gt)
 
 static int write_timestamp(struct i915_request *rq, int slot)
 {
+       struct intel_timeline *tl =
+               rcu_dereference_protected(rq->timeline,
+                                         !i915_request_signaled(rq));
        u32 cmd;
        u32 *cs;
 
@@ -54,7 +57,7 @@ static int write_timestamp(struct i915_request *rq, int slot)
                cmd++;
        *cs++ = cmd;
        *cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(rq->engine->mmio_base));
-       *cs++ = i915_request_timeline(rq)->hwsp_offset + slot * sizeof(u32);
+       *cs++ = tl->hwsp_offset + slot * sizeof(u32);
        *cs++ = 0;
 
        intel_ring_advance(rq, cs);
index 6f3a368..860f008 100644 (file)
@@ -35,10 +35,11 @@ static unsigned long hwsp_cacheline(struct intel_timeline *tl)
 {
        unsigned long address = (unsigned long)page_address(hwsp_page(tl));
 
-       return (address + tl->hwsp_offset) / CACHELINE_BYTES;
+       return (address + offset_in_page(tl->hwsp_offset)) / TIMELINE_SEQNO_BYTES;
 }
 
-#define CACHELINES_PER_PAGE (PAGE_SIZE / CACHELINE_BYTES)
+/* Only half of seqno's are usable, see __intel_timeline_get_seqno() */
+#define CACHELINES_PER_PAGE (PAGE_SIZE / TIMELINE_SEQNO_BYTES / 2)
 
 struct mock_hwsp_freelist {
        struct intel_gt *gt;
@@ -666,7 +667,7 @@ static int live_hwsp_wrap(void *arg)
        if (IS_ERR(tl))
                return PTR_ERR(tl);
 
-       if (!tl->has_initial_breadcrumb || !tl->hwsp_cacheline)
+       if (!tl->has_initial_breadcrumb)
                goto out_free;
 
        err = intel_timeline_pin(tl, NULL);
@@ -833,12 +834,26 @@ static int setup_watcher(struct hwsp_watcher *w, struct intel_gt *gt)
        return 0;
 }
 
+static void switch_tl_lock(struct i915_request *from, struct i915_request *to)
+{
+       /* some light mutex juggling required; think co-routines */
+
+       if (from) {
+               lockdep_unpin_lock(&from->context->timeline->mutex, from->cookie);
+               mutex_unlock(&from->context->timeline->mutex);
+       }
+
+       if (to) {
+               mutex_lock(&to->context->timeline->mutex);
+               to->cookie = lockdep_pin_lock(&to->context->timeline->mutex);
+       }
+}
+
 static int create_watcher(struct hwsp_watcher *w,
                          struct intel_engine_cs *engine,
                          int ringsz)
 {
        struct intel_context *ce;
-       struct intel_timeline *tl;
 
        ce = intel_context_create(engine);
        if (IS_ERR(ce))
@@ -851,11 +866,8 @@ static int create_watcher(struct hwsp_watcher *w,
                return PTR_ERR(w->rq);
 
        w->addr = i915_ggtt_offset(w->vma);
-       tl = w->rq->context->timeline;
 
-       /* some light mutex juggling required; think co-routines */
-       lockdep_unpin_lock(&tl->mutex, w->rq->cookie);
-       mutex_unlock(&tl->mutex);
+       switch_tl_lock(w->rq, NULL);
 
        return 0;
 }
@@ -864,15 +876,13 @@ static int check_watcher(struct hwsp_watcher *w, const char *name,
                         bool (*op)(u32 hwsp, u32 seqno))
 {
        struct i915_request *rq = fetch_and_zero(&w->rq);
-       struct intel_timeline *tl = rq->context->timeline;
        u32 offset, end;
        int err;
 
        GEM_BUG_ON(w->addr - i915_ggtt_offset(w->vma) > w->vma->size);
 
        i915_request_get(rq);
-       mutex_lock(&tl->mutex);
-       rq->cookie = lockdep_pin_lock(&tl->mutex);
+       switch_tl_lock(NULL, rq);
        i915_request_add(rq);
 
        if (i915_request_wait(rq, 0, HZ) < 0) {
@@ -901,10 +911,7 @@ out:
 static void cleanup_watcher(struct hwsp_watcher *w)
 {
        if (w->rq) {
-               struct intel_timeline *tl = w->rq->context->timeline;
-
-               mutex_lock(&tl->mutex);
-               w->rq->cookie = lockdep_pin_lock(&tl->mutex);
+               switch_tl_lock(NULL, w->rq);
 
                i915_request_add(w->rq);
        }
@@ -942,7 +949,7 @@ static struct i915_request *wrap_timeline(struct i915_request *rq)
        }
 
        i915_request_put(rq);
-       rq = intel_context_create_request(ce);
+       rq = i915_request_create(ce);
        if (IS_ERR(rq))
                return rq;
 
@@ -977,7 +984,7 @@ static int live_hwsp_read(void *arg)
        if (IS_ERR(tl))
                return PTR_ERR(tl);
 
-       if (!tl->hwsp_cacheline)
+       if (!tl->has_initial_breadcrumb)
                goto out_free;
 
        for (i = 0; i < ARRAY_SIZE(watcher); i++) {
@@ -999,7 +1006,7 @@ static int live_hwsp_read(void *arg)
                do {
                        struct i915_sw_fence *submit;
                        struct i915_request *rq;
-                       u32 hwsp;
+                       u32 hwsp, dummy;
 
                        submit = heap_fence_create(GFP_KERNEL);
                        if (!submit) {
@@ -1017,14 +1024,26 @@ static int live_hwsp_read(void *arg)
                                goto out;
                        }
 
-                       /* Skip to the end, saving 30 minutes of nops */
-                       tl->seqno = -10u + 2 * (count & 3);
-                       WRITE_ONCE(*(u32 *)tl->hwsp_seqno, tl->seqno);
                        ce->timeline = intel_timeline_get(tl);
 
-                       rq = intel_context_create_request(ce);
+                       /* Ensure timeline is mapped, done during first pin */
+                       err = intel_context_pin(ce);
+                       if (err) {
+                               intel_context_put(ce);
+                               goto out;
+                       }
+
+                       /*
+                        * Start at a new wrap, and set seqno right before another wrap,
+                        * saving 30 minutes of nops
+                        */
+                       tl->seqno = -12u + 2 * (count & 3);
+                       __intel_timeline_get_seqno(tl, &dummy);
+
+                       rq = i915_request_create(ce);
                        if (IS_ERR(rq)) {
                                err = PTR_ERR(rq);
+                               intel_context_unpin(ce);
                                intel_context_put(ce);
                                goto out;
                        }
@@ -1034,32 +1053,35 @@ static int live_hwsp_read(void *arg)
                                                            GFP_KERNEL);
                        if (err < 0) {
                                i915_request_add(rq);
+                               intel_context_unpin(ce);
                                intel_context_put(ce);
                                goto out;
                        }
 
-                       mutex_lock(&watcher[0].rq->context->timeline->mutex);
+                       switch_tl_lock(rq, watcher[0].rq);
                        err = intel_timeline_read_hwsp(rq, watcher[0].rq, &hwsp);
                        if (err == 0)
                                err = emit_read_hwsp(watcher[0].rq, /* before */
                                                     rq->fence.seqno, hwsp,
                                                     &watcher[0].addr);
-                       mutex_unlock(&watcher[0].rq->context->timeline->mutex);
+                       switch_tl_lock(watcher[0].rq, rq);
                        if (err) {
                                i915_request_add(rq);
+                               intel_context_unpin(ce);
                                intel_context_put(ce);
                                goto out;
                        }
 
-                       mutex_lock(&watcher[1].rq->context->timeline->mutex);
+                       switch_tl_lock(rq, watcher[1].rq);
                        err = intel_timeline_read_hwsp(rq, watcher[1].rq, &hwsp);
                        if (err == 0)
                                err = emit_read_hwsp(watcher[1].rq, /* after */
                                                     rq->fence.seqno, hwsp,
                                                     &watcher[1].addr);
-                       mutex_unlock(&watcher[1].rq->context->timeline->mutex);
+                       switch_tl_lock(watcher[1].rq, rq);
                        if (err) {
                                i915_request_add(rq);
+                               intel_context_unpin(ce);
                                intel_context_put(ce);
                                goto out;
                        }
@@ -1068,6 +1090,7 @@ static int live_hwsp_read(void *arg)
                        i915_request_add(rq);
 
                        rq = wrap_timeline(rq);
+                       intel_context_unpin(ce);
                        intel_context_put(ce);
                        if (IS_ERR(rq)) {
                                err = PTR_ERR(rq);
@@ -1107,8 +1130,8 @@ static int live_hwsp_read(void *arg)
                            3 * watcher[1].rq->ring->size)
                                break;
 
-               } while (!__igt_timeout(end_time, NULL));
-               WRITE_ONCE(*(u32 *)tl->hwsp_seqno, 0xdeadbeef);
+               } while (!__igt_timeout(end_time, NULL) &&
+                        count < (PAGE_SIZE / TIMELINE_SEQNO_BYTES - 1) / 2);
 
                pr_info("%s: simulated %lu wraps\n", engine->name, count);
                err = check_watcher(&watcher[1], "after", cmp_gte);
@@ -1153,9 +1176,7 @@ static int live_hwsp_rollover_kernel(void *arg)
                }
 
                GEM_BUG_ON(i915_active_fence_isset(&tl->last_request));
-               tl->seqno = 0;
-               timeline_rollback(tl);
-               timeline_rollback(tl);
+               tl->seqno = -2u;
                WRITE_ONCE(*(u32 *)tl->hwsp_seqno, tl->seqno);
 
                for (i = 0; i < ARRAY_SIZE(rq); i++) {
@@ -1235,11 +1256,10 @@ static int live_hwsp_rollover_user(void *arg)
                        goto out;
 
                tl = ce->timeline;
-               if (!tl->has_initial_breadcrumb || !tl->hwsp_cacheline)
+               if (!tl->has_initial_breadcrumb)
                        goto out;
 
-               timeline_rollback(tl);
-               timeline_rollback(tl);
+               tl->seqno = -4u;
                WRITE_ONCE(*(u32 *)tl->hwsp_seqno, tl->seqno);
 
                for (i = 0; i < ARRAY_SIZE(rq); i++) {
index 22e39d9..021535f 100644 (file)
@@ -863,7 +863,6 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
        rq->fence.seqno = seqno;
 
        RCU_INIT_POINTER(rq->timeline, tl);
-       RCU_INIT_POINTER(rq->hwsp_cacheline, tl->hwsp_cacheline);
        rq->hwsp_seqno = tl->hwsp_seqno;
        GEM_BUG_ON(__i915_request_is_complete(rq));
 
@@ -1108,9 +1107,6 @@ emit_semaphore_wait(struct i915_request *to,
        if (i915_request_has_initial_breadcrumb(to))
                goto await_fence;
 
-       if (!rcu_access_pointer(from->hwsp_cacheline))
-               goto await_fence;
-
        /*
         * If this or its dependents are waiting on an external fence
         * that may fail catastrophically, then we want to avoid using
index 1bfe214..ce773c0 100644 (file)
@@ -237,16 +237,6 @@ struct i915_request {
         */
        const u32 *hwsp_seqno;
 
-       /*
-        * If we need to access the timeline's seqno for this request in
-        * another request, we need to keep a read reference to this associated
-        * cacheline, so that we do not free and recycle it before the foreign
-        * observers have completed. Hence, we keep a pointer to the cacheline
-        * inside the timeline's HWSP vma, but it is only valid while this
-        * request has not completed and guarded by the timeline mutex.
-        */
-       struct intel_timeline_cacheline __rcu *hwsp_cacheline;
-
        /** Position in the ring of the start of the request */
        u32 head;
 
@@ -616,4 +606,25 @@ i915_request_active_timeline(const struct i915_request *rq)
                                         lockdep_is_held(&rq->engine->active.lock));
 }
 
+static inline u32
+i915_request_active_seqno(const struct i915_request *rq)
+{
+       u32 hwsp_phys_base =
+               page_mask_bits(i915_request_active_timeline(rq)->hwsp_offset);
+       u32 hwsp_relative_offset = offset_in_page(rq->hwsp_seqno);
+
+       /*
+        * Because of wraparound, we cannot simply take tl->hwsp_offset,
+        * but instead use the fact that the relative for vaddr is the
+        * offset as for hwsp_offset. Take the top bits from tl->hwsp_offset
+        * and combine them with the relative offset in rq->hwsp_seqno.
+        *
+        * As rw->hwsp_seqno is rewritten when signaled, this only works
+        * when the request isn't signaled yet, but at that point you
+        * no longer need the offset.
+        */
+
+       return hwsp_phys_base + hwsp_relative_offset;
+}
+
 #endif /* I915_REQUEST_H */