Message ID | 20200611160529.9558-1-chris@chris-wilson.co.uk (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | drm/i915/gt: Flush gen3 relocs harder, again | expand |
Chris Wilson <chris@chris-wilson.co.uk> writes: > gen3 does not fully flush MI stores to memory on MI_FLUSH, such that a > subsequent read from e.g. the sampler can bypass the store and read the > stale value from memory. This is a serious issue when we are using MI > stores to rewrite the batches for relocation, as it means that the batch > is reading from random user/kernel memory. While it is particularly > sensitive [and detectable] for relocations, reading stale data at any > time is a worry. > > Having started with a small number of delaying stores and doubling until > no more incoherency was seen over a few hours (with and without > background memory pressure), 32 was the magic number. > > Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/2018 > References: a889580c087a ("drm/i915: Flush GPU relocs harder for gen3") > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> > Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> > --- > So gen3 requires a delay after to flush the previous stores, gen5 is > assuming it requires a delay between the seqno and the > MI_USER_INTERRUPT. Here I've made gen5 reuse the gen3 approach, but I > need to verify that it still holds. > --- > drivers/gpu/drm/i915/gt/gen2_engine_cs.c | 39 +++++++++--------------- > 1 file changed, 15 insertions(+), 24 deletions(-) > > diff --git a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > index 3fb0dc1fb910..342c476ec872 100644 > --- a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > +++ b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > @@ -142,19 +142,26 @@ int gen4_emit_flush_vcs(struct i915_request *rq, u32 mode) > return 0; > } > > -u32 *gen3_emit_breadcrumb(struct i915_request *rq, u32 *cs) > +static u32 *__gen2_emit_breadcrumb(struct i915_request *rq, u32 *cs, int count) > { > 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); > > *cs++ = MI_FLUSH; > + *cs++ = MI_NOOP; > + > + while (count--) { > + *cs++ = MI_STORE_DWORD_INDEX; > + *cs++ = I915_GEM_HWS_SCRATCH * sizeof(u32); > + *cs++ = rq->fence.seqno; > + *cs++ = MI_FLUSH | MI_NO_WRITE_FLUSH; Why would you avoid write flush here? -Mika > + } > > *cs++ = MI_STORE_DWORD_INDEX; > *cs++ = I915_GEM_HWS_SEQNO_ADDR; > *cs++ = rq->fence.seqno; > > *cs++ = MI_USER_INTERRUPT; > - *cs++ = MI_NOOP; > > rq->tail = intel_ring_offset(rq, cs); > assert_ring_tail_valid(rq->ring, rq->tail); > @@ -162,31 +169,15 @@ u32 *gen3_emit_breadcrumb(struct i915_request *rq, u32 *cs) > return cs; > } > > -#define GEN5_WA_STORES 8 /* must be at least 1! */ > -u32 *gen5_emit_breadcrumb(struct i915_request *rq, u32 *cs) > +u32 *gen3_emit_breadcrumb(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); > - > - *cs++ = MI_FLUSH; > - > - BUILD_BUG_ON(GEN5_WA_STORES < 1); > - for (i = 0; i < GEN5_WA_STORES; i++) { > - *cs++ = MI_STORE_DWORD_INDEX; > - *cs++ = I915_GEM_HWS_SEQNO_ADDR; > - *cs++ = rq->fence.seqno; > - } > - > - *cs++ = MI_USER_INTERRUPT; > - > - rq->tail = intel_ring_offset(rq, cs); > - assert_ring_tail_valid(rq->ring, rq->tail); > + return __gen2_emit_breadcrumb(rq, cs, 32); > +} > > - return cs; > +u32 *gen5_emit_breadcrumb(struct i915_request *rq, u32 *cs) > +{ > + return __gen2_emit_breadcrumb(rq, cs, 8); > } > -#undef GEN5_WA_STORES > > /* Just userspace ABI convention to limit the wa batch bo to a resonable size */ > #define I830_BATCH_LIMIT SZ_256K > -- > 2.20.1
Quoting Mika Kuoppala (2020-06-12 10:14:55) > Chris Wilson <chris@chris-wilson.co.uk> writes: > > > gen3 does not fully flush MI stores to memory on MI_FLUSH, such that a > > subsequent read from e.g. the sampler can bypass the store and read the > > stale value from memory. This is a serious issue when we are using MI > > stores to rewrite the batches for relocation, as it means that the batch > > is reading from random user/kernel memory. While it is particularly > > sensitive [and detectable] for relocations, reading stale data at any > > time is a worry. > > > > Having started with a small number of delaying stores and doubling until > > no more incoherency was seen over a few hours (with and without > > background memory pressure), 32 was the magic number. > > > > Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/2018 > > References: a889580c087a ("drm/i915: Flush GPU relocs harder for gen3") > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > > Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> > > Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> > > --- > > So gen3 requires a delay after to flush the previous stores, gen5 is > > assuming it requires a delay between the seqno and the > > MI_USER_INTERRUPT. Here I've made gen5 reuse the gen3 approach, but I > > need to verify that it still holds. > > --- > > drivers/gpu/drm/i915/gt/gen2_engine_cs.c | 39 +++++++++--------------- > > 1 file changed, 15 insertions(+), 24 deletions(-) > > > > diff --git a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > > index 3fb0dc1fb910..342c476ec872 100644 > > --- a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > > +++ b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > > @@ -142,19 +142,26 @@ int gen4_emit_flush_vcs(struct i915_request *rq, u32 mode) > > return 0; > > } > > > > -u32 *gen3_emit_breadcrumb(struct i915_request *rq, u32 *cs) > > +static u32 *__gen2_emit_breadcrumb(struct i915_request *rq, u32 *cs, int count) > > { > > 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); > > > > *cs++ = MI_FLUSH; > > + *cs++ = MI_NOOP; > > + > > + while (count--) { > > + *cs++ = MI_STORE_DWORD_INDEX; > > + *cs++ = I915_GEM_HWS_SCRATCH * sizeof(u32); > > + *cs++ = rq->fence.seqno; > > + *cs++ = MI_FLUSH | MI_NO_WRITE_FLUSH; > > Why would you avoid write flush here? It's a flush of the render caches; all I'm using it for here is a delay. As evidenced, MI_FLUSH does not flush the stores by itself. 32 is an awful lot of papering. I should note that for gen5 not only did we have the delay in the breadcrumb but also in the invalidation. Maybe that would help for gen3 -Chris
Quoting Chris Wilson (2020-06-12 10:23:30) > Quoting Mika Kuoppala (2020-06-12 10:14:55) > > Chris Wilson <chris@chris-wilson.co.uk> writes: > > > > > gen3 does not fully flush MI stores to memory on MI_FLUSH, such that a > > > subsequent read from e.g. the sampler can bypass the store and read the > > > stale value from memory. This is a serious issue when we are using MI > > > stores to rewrite the batches for relocation, as it means that the batch > > > is reading from random user/kernel memory. While it is particularly > > > sensitive [and detectable] for relocations, reading stale data at any > > > time is a worry. > > > > > > Having started with a small number of delaying stores and doubling until > > > no more incoherency was seen over a few hours (with and without > > > background memory pressure), 32 was the magic number. > > > > > > Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/2018 > > > References: a889580c087a ("drm/i915: Flush GPU relocs harder for gen3") > > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > > > Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> > > > Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> > > > --- > > > So gen3 requires a delay after to flush the previous stores, gen5 is > > > assuming it requires a delay between the seqno and the > > > MI_USER_INTERRUPT. Here I've made gen5 reuse the gen3 approach, but I > > > need to verify that it still holds. > > > --- > > > drivers/gpu/drm/i915/gt/gen2_engine_cs.c | 39 +++++++++--------------- > > > 1 file changed, 15 insertions(+), 24 deletions(-) > > > > > > diff --git a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > > > index 3fb0dc1fb910..342c476ec872 100644 > > > --- a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > > > +++ b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c > > > @@ -142,19 +142,26 @@ int gen4_emit_flush_vcs(struct i915_request *rq, u32 mode) > > > return 0; > > > } > > > > > > -u32 *gen3_emit_breadcrumb(struct i915_request *rq, u32 *cs) > > > +static u32 *__gen2_emit_breadcrumb(struct i915_request *rq, u32 *cs, int count) > > > { > > > 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); > > > > > > *cs++ = MI_FLUSH; > > > + *cs++ = MI_NOOP; > > > + > > > + while (count--) { > > > + *cs++ = MI_STORE_DWORD_INDEX; > > > + *cs++ = I915_GEM_HWS_SCRATCH * sizeof(u32); > > > + *cs++ = rq->fence.seqno; > > > + *cs++ = MI_FLUSH | MI_NO_WRITE_FLUSH; > > > > Why would you avoid write flush here? > > It's a flush of the render caches; all I'm using it for here is a delay. > As evidenced, MI_FLUSH does not flush the stores by itself. > > 32 is an awful lot of papering. I should note that for gen5 not only did > we have the delay in the breadcrumb but also in the invalidation. Maybe > that would help for gen3 Well that was easy. Splitting the w/a between the breadcrumb and the invalidate does not help to reduce the burden [the number of stores required to make the incoherency go away] of the w/a. -Chris
diff --git a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c index 3fb0dc1fb910..342c476ec872 100644 --- a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c +++ b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c @@ -142,19 +142,26 @@ int gen4_emit_flush_vcs(struct i915_request *rq, u32 mode) return 0; } -u32 *gen3_emit_breadcrumb(struct i915_request *rq, u32 *cs) +static u32 *__gen2_emit_breadcrumb(struct i915_request *rq, u32 *cs, int count) { 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); *cs++ = MI_FLUSH; + *cs++ = MI_NOOP; + + while (count--) { + *cs++ = MI_STORE_DWORD_INDEX; + *cs++ = I915_GEM_HWS_SCRATCH * sizeof(u32); + *cs++ = rq->fence.seqno; + *cs++ = MI_FLUSH | MI_NO_WRITE_FLUSH; + } *cs++ = MI_STORE_DWORD_INDEX; *cs++ = I915_GEM_HWS_SEQNO_ADDR; *cs++ = rq->fence.seqno; *cs++ = MI_USER_INTERRUPT; - *cs++ = MI_NOOP; rq->tail = intel_ring_offset(rq, cs); assert_ring_tail_valid(rq->ring, rq->tail); @@ -162,31 +169,15 @@ u32 *gen3_emit_breadcrumb(struct i915_request *rq, u32 *cs) return cs; } -#define GEN5_WA_STORES 8 /* must be at least 1! */ -u32 *gen5_emit_breadcrumb(struct i915_request *rq, u32 *cs) +u32 *gen3_emit_breadcrumb(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); - - *cs++ = MI_FLUSH; - - BUILD_BUG_ON(GEN5_WA_STORES < 1); - for (i = 0; i < GEN5_WA_STORES; i++) { - *cs++ = MI_STORE_DWORD_INDEX; - *cs++ = I915_GEM_HWS_SEQNO_ADDR; - *cs++ = rq->fence.seqno; - } - - *cs++ = MI_USER_INTERRUPT; - - rq->tail = intel_ring_offset(rq, cs); - assert_ring_tail_valid(rq->ring, rq->tail); + return __gen2_emit_breadcrumb(rq, cs, 32); +} - return cs; +u32 *gen5_emit_breadcrumb(struct i915_request *rq, u32 *cs) +{ + return __gen2_emit_breadcrumb(rq, cs, 8); } -#undef GEN5_WA_STORES /* Just userspace ABI convention to limit the wa batch bo to a resonable size */ #define I830_BATCH_LIMIT SZ_256K
gen3 does not fully flush MI stores to memory on MI_FLUSH, such that a subsequent read from e.g. the sampler can bypass the store and read the stale value from memory. This is a serious issue when we are using MI stores to rewrite the batches for relocation, as it means that the batch is reading from random user/kernel memory. While it is particularly sensitive [and detectable] for relocations, reading stale data at any time is a worry. Having started with a small number of delaying stores and doubling until no more incoherency was seen over a few hours (with and without background memory pressure), 32 was the magic number. Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/2018 References: a889580c087a ("drm/i915: Flush GPU relocs harder for gen3") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> --- So gen3 requires a delay after to flush the previous stores, gen5 is assuming it requires a delay between the seqno and the MI_USER_INTERRUPT. Here I've made gen5 reuse the gen3 approach, but I need to verify that it still holds. --- drivers/gpu/drm/i915/gt/gen2_engine_cs.c | 39 +++++++++--------------- 1 file changed, 15 insertions(+), 24 deletions(-)