[Intel-gfx] [PATCH 3/3] drm/i915/blt: bump the size restriction
Tang, CQ
cq.tang at intel.com
Fri Aug 9 22:25:53 UTC 2019
> -----Original Message-----
> From: Intel-gfx [mailto:intel-gfx-bounces at lists.freedesktop.org] On Behalf
> Of Tang, CQ
> Sent: Friday, August 9, 2019 3:21 PM
> To: Chris Wilson <chris at chris-wilson.co.uk>; Auld, Matthew
> <matthew.auld at intel.com>; intel-gfx at lists.freedesktop.org
> Subject: Re: [Intel-gfx] [PATCH 3/3] drm/i915/blt: bump the size restriction
>
>
>
> > -----Original Message-----
> > From: Intel-gfx [mailto:intel-gfx-bounces at lists.freedesktop.org] On
> > Behalf Of Chris Wilson
> > Sent: Friday, August 9, 2019 2:18 PM
> > To: Auld, Matthew <matthew.auld at intel.com>; intel-
> > gfx at lists.freedesktop.org
> > Subject: Re: [Intel-gfx] [PATCH 3/3] drm/i915/blt: bump the size
> > restriction
> >
> > Quoting Matthew Auld (2019-08-09 21:29:26)
> > > As pointed out by Chris, with our current approach we are actually
> > > limited to S16_MAX * PAGE_SIZE for our size when using the blt to
> > > clear pages. Keeping things simple try to fix this by reducing the
> > > copy to a sequence of S16_MAX * PAGE_SIZE blocks.
>
> Just a general question. If the address space is limited to 1G, for example,
> and we want to copy between two 1G-size objects.
> Do we do fragmentation inside the blitter copying routine?
>
> This could happen during swapping. We pre-allocate 1G PPGTT, and reserve
> the rest of address space, so only 1G space can be used.
Or, can we add offset and size into the copying object, like the pread/pwrite to specify offset/size to read/write.
Then we can do loop in caller. But prefer to do fragmentation inside.
--CQ
>
> --CQ
>
>
> > >
> > > Reported-by: Chris Wilson <chris at chris-wilson.co.uk>
> > > Signed-off-by: Matthew Auld <matthew.auld at intel.com>
> > > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > > ---
> > > .../gpu/drm/i915/gem/i915_gem_client_blt.c | 31 +++-
> > > .../gpu/drm/i915/gem/i915_gem_object_blt.c | 139 ++++++++++++++-
> --
> > -
> > > .../gpu/drm/i915/gem/i915_gem_object_blt.h | 9 +-
> > > .../i915/gem/selftests/i915_gem_client_blt.c | 16 +-
> > > .../i915/gem/selftests/i915_gem_object_blt.c | 22 ++-
> > > 5 files changed, 170 insertions(+), 47 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_client_blt.c
> > > b/drivers/gpu/drm/i915/gem/i915_gem_client_blt.c
> > > index 08a84c940d8d..4b096309a97e 100644
> > > --- a/drivers/gpu/drm/i915/gem/i915_gem_client_blt.c
> > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_client_blt.c
> > > @@ -5,6 +5,8 @@
> > >
> > > #include "i915_drv.h"
> > > #include "gt/intel_context.h"
> > > +#include "gt/intel_engine_pm.h"
> > > +#include "gt/intel_engine_pool.h"
> > > #include "i915_gem_client_blt.h"
> > > #include "i915_gem_object_blt.h"
> > >
> > > @@ -156,7 +158,9 @@ static void clear_pages_worker(struct
> > > work_struct
> > *work)
> > > struct drm_i915_private *i915 = w->ce->engine->i915;
> > > struct drm_i915_gem_object *obj = w->sleeve->vma->obj;
> > > struct i915_vma *vma = w->sleeve->vma;
> > > + struct intel_engine_pool_node *pool;
> > > struct i915_request *rq;
> > > + struct i915_vma *batch;
> > > int err = w->dma.error;
> > >
> > > if (unlikely(err))
> > > @@ -176,10 +180,17 @@ static void clear_pages_worker(struct
> > work_struct *work)
> > > if (unlikely(err))
> > > goto out_unlock;
> > >
> > > + intel_engine_pm_get(w->ce->engine);
> > > + batch = intel_emit_vma_fill_blt(&pool, w->ce, vma,
> > > + w->value);
> >
> > I had to search for where pool was being set!
> >
> > Hmm, batch is from pool right? So we are the owner of the batch, and
> > we could set batch->private = pool.
> >
> > > + if (IS_ERR(batch)) {
> > > + err = PTR_ERR(batch);
> > > + goto out_unpin;
> > > + }
> > > +
> > > rq = intel_context_create_request(w->ce);
> > > if (IS_ERR(rq)) {
> > > err = PTR_ERR(rq);
> > > - goto out_unpin;
> > > + goto out_batch;
> > > }
> > >
> > > /* There's no way the fence has signalled */ @@ -187,6
> > > +198,16 @@ static void clear_pages_worker(struct work_struct *work)
> > > clear_pages_dma_fence_cb))
> > > GEM_BUG_ON(1);
> > >
> > > + i915_vma_lock(batch);
> > > + err = i915_vma_move_to_active(batch, rq, 0);
> > > + i915_vma_unlock(batch);
> > > + if (unlikely(err))
> > > + goto out_request;
> > > +
> > > + err = intel_engine_pool_mark_active(pool, rq);
> > > + if (unlikely(err))
> > > + goto out_request;
> > > +
> > > if (w->ce->engine->emit_init_breadcrumb) {
> > > err = w->ce->engine->emit_init_breadcrumb(rq);
> > > if (unlikely(err))
> > > @@ -202,7 +223,9 @@ static void clear_pages_worker(struct
> > > work_struct
> > *work)
> > > if (err)
> > > goto out_request;
> > >
> > > - err = intel_emit_vma_fill_blt(rq, vma, w->value);
> > > + err = w->ce->engine->emit_bb_start(rq,
> > > + batch->node.start, batch->node.size,
> > > + 0);
> > > out_request:
> > > if (unlikely(err)) {
> > > i915_request_skip(rq, err); @@ -210,7 +233,11 @@
> > > static void clear_pages_worker(struct work_struct *work)
> > > }
> > >
> > > i915_request_add(rq);
> > > +out_batch:
> > > + i915_vma_unpin(batch);
> > > + intel_engine_pool_put(pool);
> > > out_unpin:
> > > + intel_engine_pm_put(w->ce->engine);
> > > i915_vma_unpin(vma);
> > > out_unlock:
> > > mutex_unlock(&i915->drm.struct_mutex);
> > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_blt.c
> > > b/drivers/gpu/drm/i915/gem/i915_gem_object_blt.c
> > > index fa90c38c8b07..c1e5edd1e359 100644
> > > --- a/drivers/gpu/drm/i915/gem/i915_gem_object_blt.c
> > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_object_blt.c
> > > @@ -5,49 +5,103 @@
> > >
> > > #include "i915_drv.h"
> > > #include "gt/intel_context.h"
> > > +#include "gt/intel_engine_pm.h"
> > > +#include "gt/intel_engine_pool.h"
> > > +#include "gt/intel_gt.h"
> > > #include "i915_gem_clflush.h"
> > > #include "i915_gem_object_blt.h"
> > >
> > > -int intel_emit_vma_fill_blt(struct i915_request *rq,
> > > - struct i915_vma *vma,
> > > - u32 value)
> > > +struct i915_vma *intel_emit_vma_fill_blt(struct
> > > +intel_engine_pool_node
> > **p,
> > > + struct intel_context *ce,
> > > + struct i915_vma *vma,
> > > + u32 value)
> > > {
> > > - u32 *cs;
> > > -
> > > - cs = intel_ring_begin(rq, 8);
> > > - if (IS_ERR(cs))
> > > - return PTR_ERR(cs);
> > > -
> > > - if (INTEL_GEN(rq->i915) >= 8) {
> > > - *cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (7 - 2);
> > > - *cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
> > > - *cs++ = 0;
> > > - *cs++ = vma->size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
> > > - *cs++ = lower_32_bits(vma->node.start);
> > > - *cs++ = upper_32_bits(vma->node.start);
> > > - *cs++ = value;
> > > - *cs++ = MI_NOOP;
> > > - } else {
> > > - *cs++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
> > > - *cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | PAGE_SIZE;
> > > - *cs++ = 0;
> > > - *cs++ = vma->size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
> > > - *cs++ = vma->node.start;
> > > - *cs++ = value;
> > > - *cs++ = MI_NOOP;
> > > - *cs++ = MI_NOOP;
> > > + struct drm_i915_private *i915 = ce->vm->i915;
> > > + const u32 block_size = S16_MAX * PAGE_SIZE;
> > > + struct intel_engine_pool_node *pool;
> > > + struct i915_vma *batch;
> > > + u64 offset;
> > > + u64 count;
> > > + u64 rem;
> > > + u32 size;
> > > + u32 *cmd;
> > > + int err;
> > > +
> > > + count = div_u64(vma->size, block_size);
> > > + size = (1 + 8 * count) * sizeof(u32);
> > > + size = round_up(size, PAGE_SIZE);
> > > + pool = intel_engine_pool_get(&ce->engine->pool, size);
> > > + if (IS_ERR(pool))
> > > + return ERR_CAST(pool);
> > > +
> > > + cmd = i915_gem_object_pin_map(pool->obj, I915_MAP_WC);
> > > + if (IS_ERR(cmd)) {
> > > + err = PTR_ERR(cmd);
> > > + goto out_put;
> > > + }
> > > +
> > > + rem = vma->size;
> > > + offset = vma->node.start;
> > > +
> > > + do {
> > > + u32 size = min_t(u64, rem, block_size);
> > > +
> > > + GEM_BUG_ON(size >> PAGE_SHIFT > S16_MAX);
> > > +
> > > + if (INTEL_GEN(i915) >= 8) {
> > > + *cmd++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (7 - 2);
> > > + *cmd++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY |
> > PAGE_SIZE;
> > > + *cmd++ = 0;
> > > + *cmd++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
> > > + *cmd++ = lower_32_bits(offset);
> > > + *cmd++ = upper_32_bits(offset);
> > > + *cmd++ = value;
> > > + } else {
> > > + *cmd++ = XY_COLOR_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
> > > + *cmd++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY |
> > PAGE_SIZE;
> > > + *cmd++ = 0;
> > > + *cmd++ = size >> PAGE_SHIFT << 16 | PAGE_SIZE / 4;
> > > + *cmd++ = offset;
> > > + *cmd++ = value;
> > > + }
> > > +
> > > + /* Allow ourselves to be preempted in between blocks. */
> > > + *cmd++ = MI_ARB_CHECK;
> >
> > That I thought was a nice touch.
> >
> > > +
> > > + offset += size;
> > > + rem -= size;
> > > + } while (rem);
> > > +
> > > + *cmd = MI_BATCH_BUFFER_END;
> > > + intel_gt_chipset_flush(ce->vm->gt);
> > > +
> > > + i915_gem_object_unpin_map(pool->obj);
> > > +
> > > + batch = i915_vma_instance(pool->obj, ce->vm, NULL);
> > > + if (IS_ERR(batch)) {
> > > + err = PTR_ERR(batch);
> > > + goto out_put;
> > > }
> > >
> > > - intel_ring_advance(rq, cs);
> > > + err = i915_vma_pin(batch, 0, 0, PIN_USER);
> > > + if (unlikely(err))
> > > + goto out_put;
> > > +
> > > + *p = pool;
> > > + return batch;
> > >
> > > - return 0;
> > > +out_put:
> > > + intel_engine_pool_put(pool);
> > > + return ERR_PTR(err);
> > > }
> > >
> > > int i915_gem_object_fill_blt(struct drm_i915_gem_object *obj,
> > > struct intel_context *ce,
> > > u32 value) {
> > > + struct intel_engine_pool_node *pool;
> > > struct i915_request *rq;
> > > + struct i915_vma *batch;
> > > struct i915_vma *vma;
> > > int err;
> > >
> > > @@ -65,12 +119,29 @@ int i915_gem_object_fill_blt(struct
> > drm_i915_gem_object *obj,
> > > i915_gem_object_unlock(obj);
> > > }
> > >
> > > + intel_engine_pm_get(ce->engine);
> > > + batch = intel_emit_vma_fill_blt(&pool, ce, vma, value);
> > > + if (IS_ERR(batch)) {
> > > + err = PTR_ERR(batch);
> > > + goto out_unpin;
> > > + }
> > > +
> > > rq = intel_context_create_request(ce);
> > > if (IS_ERR(rq)) {
> > > err = PTR_ERR(rq);
> > > - goto out_unpin;
> > > + goto out_batch;
> > > }
> > >
> > > + i915_vma_lock(batch);
> > > + err = i915_vma_move_to_active(batch, rq, 0);
> > > + i915_vma_unlock(batch);
> > > + if (unlikely(err))
> > > + goto out_request;
> > > +
> > > + err = intel_engine_pool_mark_active(pool, rq);
> > > + if (unlikely(err))
> > > + goto out_request;
> > > +
> > > err = i915_request_await_object(rq, obj, true);
> > > if (unlikely(err))
> > > goto out_request;
> > > @@ -87,13 +158,19 @@ int i915_gem_object_fill_blt(struct
> > drm_i915_gem_object *obj,
> > > if (unlikely(err))
> > > goto out_request;
> > >
> > > - err = intel_emit_vma_fill_blt(rq, vma, value);
> > > + err = ce->engine->emit_bb_start(rq,
> > > + batch->node.start, batch->node.size,
> > > + 0);
> > > out_request:
> > > if (unlikely(err))
> > > i915_request_skip(rq, err);
> > >
> > > i915_request_add(rq);
> > > +out_batch:
> > > + i915_vma_unpin(batch);
> > > + intel_engine_pool_put(pool);
> > > out_unpin:
> > > + intel_engine_pm_put(ce->engine);
> > > i915_vma_unpin(vma);
> > > return err;
> > > }
> > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_blt.h
> > > b/drivers/gpu/drm/i915/gem/i915_gem_object_blt.h
> > > index 7ec7de6ac0c0..a7425c234d50 100644
> > > --- a/drivers/gpu/drm/i915/gem/i915_gem_object_blt.h
> > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_object_blt.h
> > > @@ -9,13 +9,14 @@
> > > #include <linux/types.h>
> > >
> > > struct drm_i915_gem_object;
> > > +struct intel_engine_pool_node;
> > > struct intel_context;
> > > -struct i915_request;
> > > struct i915_vma;
> > >
> > > -int intel_emit_vma_fill_blt(struct i915_request *rq,
> > > - struct i915_vma *vma,
> > > - u32 value);
> > > +struct i915_vma *intel_emit_vma_fill_blt(struct
> > > +intel_engine_pool_node
> > **p,
> > > + struct intel_context *ce,
> > > + struct i915_vma *vma,
> > > + u32 value);
> > >
> > > int i915_gem_object_fill_blt(struct drm_i915_gem_object *obj,
> > > struct intel_context *ce, diff --git
> > > a/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c
> > > b/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c
> > > index 275c28926067..d8804a847945 100644
> > > --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c
> > > +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c
> > > @@ -9,6 +9,7 @@
> > >
> > > #include "selftests/igt_flush_test.h"
> > > #include "selftests/mock_drm.h"
> > > +#include "huge_gem_object.h"
> > > #include "mock_context.h"
> > >
> > > static int igt_client_fill(void *arg) @@ -24,15 +25,19 @@ static
> > > int igt_client_fill(void *arg)
> > > prandom_seed_state(&prng, i915_selftest.random_seed);
> > >
> > > do {
> > > - u32 sz = prandom_u32_state(&prng) % SZ_32M;
> > > + const u32 max_block_size = S16_MAX * PAGE_SIZE;
> > > + u32 sz = min_t(u64, ce->vm->total >> 4,
> > prandom_u32_state(&prng));
> > > + u32 phys_sz = sz % (max_block_size + 1);
> > > u32 val = prandom_u32_state(&prng);
> > > u32 i;
> > >
> > > sz = round_up(sz, PAGE_SIZE);
> > > + phys_sz = round_up(phys_sz, PAGE_SIZE);
> > >
> > > - pr_debug("%s with sz=%x, val=%x\n", __func__, sz, val);
> > > + pr_debug("%s with phys_sz= %x, sz=%x, val=%x\n", __func__,
> > > + phys_sz, sz, val);
> > >
> > > - obj = i915_gem_object_create_internal(i915, sz);
> > > + obj = huge_gem_object(i915, phys_sz, sz);
> > > if (IS_ERR(obj)) {
> > > err = PTR_ERR(obj);
> > > goto err_flush; @@ -54,7 +59,8 @@ static int
> > > igt_client_fill(void *arg)
> > > * values after we do the set_to_cpu_domain and pick it up as a
> > > * test failure.
> > > */
> > > - memset32(vaddr, val ^ 0xdeadbeaf, obj->base.size / sizeof(u32));
> > > + memset32(vaddr, val ^ 0xdeadbeaf,
> > > + huge_gem_object_phys_size(obj) /
> > > + sizeof(u32));
> > >
> > > if (!(obj->cache_coherent &
> > I915_BO_CACHE_COHERENT_FOR_WRITE))
> > > obj->cache_dirty = true; @@ -71,7 +77,7 @@
> > > static int igt_client_fill(void *arg)
> > > if (err)
> > > goto err_unpin;
> > >
> > > - for (i = 0; i < obj->base.size / sizeof(u32); ++i) {
> > > + for (i = 0; i < huge_gem_object_phys_size(obj) /
> > > + sizeof(u32); ++i) {
> > > if (vaddr[i] != val) {
> > > pr_err("vaddr[%u]=%x, expected=%x\n", i,
> > > vaddr[i], val); diff --git
> > > a/drivers/gpu/drm/i915/gem/selftests/i915_gem_object_blt.c
> > > b/drivers/gpu/drm/i915/gem/selftests/i915_gem_object_blt.c
> > > index 19843acc84d3..c6e1eebe53f5 100644
> > > --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_object_blt.c
> > > +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_object_blt.c
> > > @@ -9,6 +9,7 @@
> > >
> > > #include "selftests/igt_flush_test.h"
> > > #include "selftests/mock_drm.h"
> > > +#include "huge_gem_object.h"
> > > #include "mock_context.h"
> > >
> > > static int igt_fill_blt(void *arg)
> > > @@ -23,16 +24,26 @@ static int igt_fill_blt(void *arg)
> > >
> > > prandom_seed_state(&prng, i915_selftest.random_seed);
> > >
> > > + /*
> > > + * XXX: needs some threads to scale all these tests, also maybe
> throw
> > > + * in submission from higher priority context to see if we are
> > > + * preempted for very large objects...
> > > + */
> >
> > A challenge!
> >
> > > +
> > > do {
> > > - u32 sz = prandom_u32_state(&prng) % SZ_32M;
> > > + const u32 max_block_size = S16_MAX * PAGE_SIZE;
> > > + u32 sz = min_t(u64, ce->vm->total >> 4,
> > prandom_u32_state(&prng));
> > > + u32 phys_sz = sz % (max_block_size + 1);
> > > u32 val = prandom_u32_state(&prng);
> > > u32 i;
> > >
> > > sz = round_up(sz, PAGE_SIZE);
> > > + phys_sz = round_up(phys_sz, PAGE_SIZE);
> > >
> > > - pr_debug("%s with sz=%x, val=%x\n", __func__, sz, val);
> > > + pr_debug("%s with phys_sz= %x, sz=%x, val=%x\n", __func__,
> > > + phys_sz, sz, val);
> > >
> > > - obj = i915_gem_object_create_internal(i915, sz);
> > > + obj = huge_gem_object(i915, phys_sz, sz);
> > > if (IS_ERR(obj)) {
> > > err = PTR_ERR(obj);
> > > goto err_flush; @@ -48,7 +59,8 @@ static int
> > > igt_fill_blt(void *arg)
> > > * Make sure the potentially async clflush does its job, if
> > > * required.
> > > */
> > > - memset32(vaddr, val ^ 0xdeadbeaf, obj->base.size / sizeof(u32));
> > > + memset32(vaddr, val ^ 0xdeadbeaf,
> > > + huge_gem_object_phys_size(obj) /
> > > + sizeof(u32));
> >
> > Ok, using huge_object is a little interesting for a fill as we will
> > write the same pages over and over again. Doesn't break the test, but
> > it will be important when considering doing a huge copy.
> >
> > > if (!(obj->cache_coherent &
> > I915_BO_CACHE_COHERENT_FOR_WRITE))
> > > obj->cache_dirty = true; @@ -65,7 +77,7 @@
> > > static int igt_fill_blt(void *arg)
> > > if (err)
> > > goto err_unpin;
> > >
> > > - for (i = 0; i < obj->base.size / sizeof(u32); ++i) {
> > > + for (i = 0; i < huge_gem_object_phys_size(obj) /
> > > + sizeof(u32); ++i) {
> > > if (vaddr[i] != val) {
> > > pr_err("vaddr[%u]=%x, expected=%x\n", i,
> > > vaddr[i], val);
> >
> > I thinking passing pool via batch->private isn't too bad of an idea (I
> > had a double take when I couldn't see an assignment to pool), but
> > nevertheless
> > Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk> -Chris
> > _______________________________________________
> > Intel-gfx mailing list
> > Intel-gfx at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/intel-gfx
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx
More information about the Intel-gfx
mailing list