From 54313503f9a3c34a9e45aad7654976b7f50b9ae7 Mon Sep 17 00:00:00 2001 From: Dan Carpenter Date: Fri, 25 Sep 2015 14:36:55 +0300 Subject: drm/amdgpu: signedness bug in amdgpu_cs_parser_init() The "i" variable should be signed or it leads to a crash in the error handling code. Fixes: 1d263474c441 ('drm/amdgpu: unwind properly in amdgpu_cs_parser_init()') Signed-off-by: Dan Carpenter Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 749420f1ea6f..cb3c274edb0a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -156,7 +156,8 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) uint64_t *chunk_array_user; uint64_t *chunk_array; struct amdgpu_fpriv *fpriv = p->filp->driver_priv; - unsigned size, i; + unsigned size; + int i; int ret; if (cs->in.num_chunks == 0) -- cgit From 028423b0d8fa27c30a46da4af2a5cab230369de6 Mon Sep 17 00:00:00 2001 From: Arnd Bergmann Date: Wed, 7 Oct 2015 09:41:27 +0200 Subject: drm/amdgpu: fix 32-bit compiler warning MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The new amdgpu driver passes a user space pointer in a 64-bit structure member, which is the correct way to do it, but it attempts to directly cast it to a __user pointer in the kernel, which causes a warning in three places: drm/amd/amdgpu/amdgpu_cs.c: In function 'amdgpu_cs_parser_init': drm/amd/amdgpu/amdgpu_cs.c:180:21: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast] chunk_array_user = (uint64_t __user *)(cs->in.chunks); This changes all three to add an intermediate cast to 'unsigned long' as other drivers do. This avoids the warning and works correctly on both 32-bit and 64-bit architectures. Fixes: e60b344f6c0eff ("drm/amdgpu: optimize amdgpu_parser_init") Reviewed-by: Christian König Signed-off-by: Arnd Bergmann Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index cb3c274edb0a..fd16652aa277 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -177,7 +177,7 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) /* get chunks */ INIT_LIST_HEAD(&p->validated); - chunk_array_user = (uint64_t __user *)(cs->in.chunks); + chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks); if (copy_from_user(chunk_array, chunk_array_user, sizeof(uint64_t)*cs->in.num_chunks)) { ret = -EFAULT; @@ -197,7 +197,7 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) struct drm_amdgpu_cs_chunk user_chunk; uint32_t __user *cdata; - chunk_ptr = (void __user *)chunk_array[i]; + chunk_ptr = (void __user *)(unsigned long)chunk_array[i]; if (copy_from_user(&user_chunk, chunk_ptr, sizeof(struct drm_amdgpu_cs_chunk))) { ret = -EFAULT; @@ -208,7 +208,7 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) p->chunks[i].length_dw = user_chunk.length_dw; size = p->chunks[i].length_dw; - cdata = (void __user *)user_chunk.chunk_data; + cdata = (void __user *)(unsigned long)user_chunk.chunk_data; p->chunks[i].user_ptr = cdata; p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t)); -- cgit From c113ea1c4f4a7592f17e53d658873b83af0dbb4b Mon Sep 17 00:00:00 2001 From: Alex Deucher Date: Thu, 8 Oct 2015 16:30:37 -0400 Subject: drm/amdgpu: rework sdma structures MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Rework the sdma structures in the driver to consolidate all of the sdma info into a single structure and allow for asics that may have different numbers of sdma instances. Reviewed-by: Christian König Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 749420f1ea6f..29fc45ce64dd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -104,10 +104,11 @@ int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type, } break; case AMDGPU_HW_IP_DMA: - if (ring < 2) { - *out_ring = &adev->sdma[ring].ring; + if (ring < adev->sdma.num_instances) { + *out_ring = &adev->sdma.instance[ring].ring; } else { - DRM_ERROR("only two SDMA rings are supported\n"); + DRM_ERROR("only %d SDMA rings are supported\n", + adev->sdma.num_instances); return -EINVAL; } break; -- cgit From b495bd3a54e732e846bfde49ba2ea93b62b1de91 Mon Sep 17 00:00:00 2001 From: Christian König Date: Thu, 10 Sep 2015 14:00:35 +0200 Subject: drm/amdgpu: add option to clear VM page tables after every submit MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This makes it much easier to find when userspace misses to send some buffers. Signed-off-by: Christian König Reviewed-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 29fc45ce64dd..25012c790f8f 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -567,9 +567,24 @@ static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p, if (r) return r; } + + } + + r = amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync); + + if (amdgpu_vm_debug && p->bo_list) { + /* Invalidate all BOs to test for userspace bugs */ + for (i = 0; i < p->bo_list->num_entries; i++) { + /* ignore duplicates */ + bo = p->bo_list->array[i].robj; + if (!bo) + continue; + + amdgpu_vm_bo_invalidate(adev, bo); + } } - return amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync); + return r; } static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev, -- cgit From 0c418f10104d4aa1d6b83698790898dc9ef1c12d Mon Sep 17 00:00:00 2001 From: Christian König Date: Tue, 1 Sep 2015 15:13:53 +0200 Subject: drm/amdgpu: remove the exclusive lock MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Finally getting rid of it. Signed-off-by: Christian König --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 10 ++-------- 1 file changed, 2 insertions(+), 8 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 25012c790f8f..5fdc0394561e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -831,11 +831,8 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) bool reserved_buffers = false; int i, r; - down_read(&adev->exclusive_lock); - if (!adev->accel_working) { - up_read(&adev->exclusive_lock); + if (!adev->accel_working) return -EBUSY; - } parser = amdgpu_cs_parser_create(adev, filp, NULL, NULL, 0); if (!parser) @@ -843,8 +840,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) r = amdgpu_cs_parser_init(parser, data); if (r) { DRM_ERROR("Failed to initialize parser !\n"); - kfree(parser); - up_read(&adev->exclusive_lock); + amdgpu_cs_parser_fini(parser, r, false); r = amdgpu_cs_handle_lockup(adev, r); return r; } @@ -915,14 +911,12 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) mutex_unlock(&job->job_lock); amdgpu_cs_parser_fini_late(parser); - up_read(&adev->exclusive_lock); return 0; } cs->out.handle = parser->ibs[parser->num_ibs - 1].sequence; out: amdgpu_cs_parser_fini(parser, r, reserved_buffers); - up_read(&adev->exclusive_lock); r = amdgpu_cs_handle_lockup(adev, r); return r; } -- cgit From f48b2659f521301753f9b3b67e308a79c6110346 Mon Sep 17 00:00:00 2001 From: Chunming Zhou Date: Fri, 16 Oct 2015 14:06:19 +0800 Subject: drm/amdgpu: fix the broken vm->mutex V2 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit fix the vm->mutex and ww_mutex confilcts. vm->mutex is always token first, then ww_mutex. V2: remove unneccessary checking for pt bo. Change-Id: Iea56e183752c02831126d06d2f5b7a474a6e4743 Signed-off-by: Chunming Zhou Reviewed-by: Christian König --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 5fdc0394561e..8727c3099554 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -608,7 +608,6 @@ static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev, } } - mutex_lock(&vm->mutex); r = amdgpu_bo_vm_update_pte(parser, vm); if (r) { goto out; @@ -619,7 +618,6 @@ static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev, parser->filp); out: - mutex_unlock(&vm->mutex); return r; } @@ -827,6 +825,8 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) { struct amdgpu_device *adev = dev->dev_private; union drm_amdgpu_cs *cs = data; + struct amdgpu_fpriv *fpriv = filp->driver_priv; + struct amdgpu_vm *vm = &fpriv->vm; struct amdgpu_cs_parser *parser; bool reserved_buffers = false; int i, r; @@ -844,7 +844,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) r = amdgpu_cs_handle_lockup(adev, r); return r; } - + mutex_lock(&vm->mutex); r = amdgpu_cs_parser_relocs(parser); if (r == -ENOMEM) DRM_ERROR("Not enough memory for command submission!\n"); @@ -911,12 +911,14 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) mutex_unlock(&job->job_lock); amdgpu_cs_parser_fini_late(parser); + mutex_unlock(&vm->mutex); return 0; } cs->out.handle = parser->ibs[parser->num_ibs - 1].sequence; out: amdgpu_cs_parser_fini(parser, r, reserved_buffers); + mutex_unlock(&vm->mutex); r = amdgpu_cs_handle_lockup(adev, r); return r; } -- cgit From 4cfdcd9cecc05646b08805ba449811a645b3f4a8 Mon Sep 17 00:00:00 2001 From: Dan Carpenter Date: Wed, 4 Nov 2015 16:25:09 +0300 Subject: drm/amdgpu: cleanup on error in amdgpu_cs_ioctl() MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit We recently changed the locking in this function and now there is a missing unlock on error. Also there are some other resources that we should probably release as well... Fixes: f48b2659f521 ('drm/amdgpu: fix the broken vm->mutex V2') Reviewed-by: Christian König Signed-off-by: Dan Carpenter --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 8727c3099554..233ac9c64eb3 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -875,8 +875,10 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) struct amdgpu_job *job; struct amdgpu_ring * ring = parser->ibs->ring; job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL); - if (!job) - return -ENOMEM; + if (!job) { + r = -ENOMEM; + goto out; + } job->base.sched = &ring->sched; job->base.s_entity = &parser->ctx->rings[ring->idx].entity; job->adev = parser->adev; -- cgit From e4a58a28b50f30e72292b6659d94410cbf7355ad Mon Sep 17 00:00:00 2001 From: Christian König Date: Thu, 5 Nov 2015 17:00:25 +0100 Subject: drm/amdgpu: fix leaking the IBs on error MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Fixing a memory leak when the scheduler is enabled. Signed-off-by: Christian König Reviewed-by: Junwei Zhang Reviewed-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 25 ++++++++++++------------- 1 file changed, 12 insertions(+), 13 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index dfc4d02c7a38..ecc82dfe83f8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -499,16 +499,12 @@ static void amdgpu_cs_parser_fini_late(struct amdgpu_cs_parser *parser) for (i = 0; i < parser->nchunks; i++) drm_free_large(parser->chunks[i].kdata); kfree(parser->chunks); - if (!amdgpu_enable_scheduler) - { - if (parser->ibs) - for (i = 0; i < parser->num_ibs; i++) - amdgpu_ib_free(parser->adev, &parser->ibs[i]); - kfree(parser->ibs); - if (parser->uf.bo) - drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base); - } - + if (parser->ibs) + for (i = 0; i < parser->num_ibs; i++) + amdgpu_ib_free(parser->adev, &parser->ibs[i]); + kfree(parser->ibs); + if (parser->uf.bo) + drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base); kfree(parser); } @@ -888,11 +884,14 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) job->base.owner = parser->filp; mutex_init(&job->job_lock); if (job->ibs[job->num_ibs - 1].user) { - memcpy(&job->uf, &parser->uf, - sizeof(struct amdgpu_user_fence)); + job->uf = parser->uf; job->ibs[job->num_ibs - 1].user = &job->uf; + parser->uf.bo = NULL; } + parser->ibs = NULL; + parser->num_ibs = 0; + job->free_job = amdgpu_cs_free_job; mutex_lock(&job->job_lock); r = amd_sched_entity_push_job(&job->base); @@ -905,7 +904,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) cs->out.handle = amdgpu_ctx_add_fence(parser->ctx, ring, &job->base.s_fence->base); - parser->ibs[parser->num_ibs - 1].sequence = cs->out.handle; + job->ibs[job->num_ibs - 1].sequence = cs->out.handle; list_sort(NULL, &parser->validated, cmp_size_smaller_first); ttm_eu_fence_buffer_objects(&parser->ticket, -- cgit From 7e52a81c2f0326a85d3ebc005829bcd604731c6d Mon Sep 17 00:00:00 2001 From: Christian König Date: Wed, 4 Nov 2015 15:44:39 +0100 Subject: drm/amdgpu: cleanup amdgpu_cs_parser handling MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit No need any more to allocate that structure dynamically, just put it on the stack. This is a start to cleanup some of the scheduler fallouts. Signed-off-by: Christian König Reviewed-by: Junwei Zhang Reviewed-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 88 +++++++++++++--------------------- 1 file changed, 33 insertions(+), 55 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index ecc82dfe83f8..bf32096b8eb7 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -127,30 +127,6 @@ int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type, return 0; } -struct amdgpu_cs_parser *amdgpu_cs_parser_create(struct amdgpu_device *adev, - struct drm_file *filp, - struct amdgpu_ctx *ctx, - struct amdgpu_ib *ibs, - uint32_t num_ibs) -{ - struct amdgpu_cs_parser *parser; - int i; - - parser = kzalloc(sizeof(struct amdgpu_cs_parser), GFP_KERNEL); - if (!parser) - return NULL; - - parser->adev = adev; - parser->filp = filp; - parser->ctx = ctx; - parser->ibs = ibs; - parser->num_ibs = num_ibs; - for (i = 0; i < num_ibs; i++) - ibs[i].ctx = ctx; - - return parser; -} - int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data) { union drm_amdgpu_cs *cs = data; @@ -490,6 +466,7 @@ static void amdgpu_cs_parser_fini_early(struct amdgpu_cs_parser *parser, int err static void amdgpu_cs_parser_fini_late(struct amdgpu_cs_parser *parser) { unsigned i; + if (parser->ctx) amdgpu_ctx_put(parser->ctx); if (parser->bo_list) @@ -505,7 +482,6 @@ static void amdgpu_cs_parser_fini_late(struct amdgpu_cs_parser *parser) kfree(parser->ibs); if (parser->uf.bo) drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base); - kfree(parser); } /** @@ -824,36 +800,36 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) union drm_amdgpu_cs *cs = data; struct amdgpu_fpriv *fpriv = filp->driver_priv; struct amdgpu_vm *vm = &fpriv->vm; - struct amdgpu_cs_parser *parser; + struct amdgpu_cs_parser parser = {}; bool reserved_buffers = false; int i, r; if (!adev->accel_working) return -EBUSY; - parser = amdgpu_cs_parser_create(adev, filp, NULL, NULL, 0); - if (!parser) - return -ENOMEM; - r = amdgpu_cs_parser_init(parser, data); + parser.adev = adev; + parser.filp = filp; + + r = amdgpu_cs_parser_init(&parser, data); if (r) { DRM_ERROR("Failed to initialize parser !\n"); - amdgpu_cs_parser_fini(parser, r, false); + amdgpu_cs_parser_fini(&parser, r, false); r = amdgpu_cs_handle_lockup(adev, r); return r; } mutex_lock(&vm->mutex); - r = amdgpu_cs_parser_relocs(parser); + r = amdgpu_cs_parser_relocs(&parser); if (r == -ENOMEM) DRM_ERROR("Not enough memory for command submission!\n"); else if (r && r != -ERESTARTSYS) DRM_ERROR("Failed to process the buffer list %d!\n", r); else if (!r) { reserved_buffers = true; - r = amdgpu_cs_ib_fill(adev, parser); + r = amdgpu_cs_ib_fill(adev, &parser); } if (!r) { - r = amdgpu_cs_dependencies(adev, parser); + r = amdgpu_cs_dependencies(adev, &parser); if (r) DRM_ERROR("Failed in the dependencies handling %d!\n", r); } @@ -861,36 +837,38 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) if (r) goto out; - for (i = 0; i < parser->num_ibs; i++) - trace_amdgpu_cs(parser, i); + for (i = 0; i < parser.num_ibs; i++) + trace_amdgpu_cs(&parser, i); - r = amdgpu_cs_ib_vm_chunk(adev, parser); + r = amdgpu_cs_ib_vm_chunk(adev, &parser); if (r) goto out; - if (amdgpu_enable_scheduler && parser->num_ibs) { + if (amdgpu_enable_scheduler && parser.num_ibs) { struct amdgpu_job *job; - struct amdgpu_ring * ring = parser->ibs->ring; + struct amdgpu_ring * ring = parser.ibs->ring; + job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL); if (!job) { r = -ENOMEM; goto out; } + job->base.sched = &ring->sched; - job->base.s_entity = &parser->ctx->rings[ring->idx].entity; - job->adev = parser->adev; - job->ibs = parser->ibs; - job->num_ibs = parser->num_ibs; - job->base.owner = parser->filp; + job->base.s_entity = &parser.ctx->rings[ring->idx].entity; + job->adev = parser.adev; + job->ibs = parser.ibs; + job->num_ibs = parser.num_ibs; + job->base.owner = parser.filp; mutex_init(&job->job_lock); if (job->ibs[job->num_ibs - 1].user) { - job->uf = parser->uf; + job->uf = parser.uf; job->ibs[job->num_ibs - 1].user = &job->uf; - parser->uf.bo = NULL; + parser.uf.bo = NULL; } - parser->ibs = NULL; - parser->num_ibs = 0; + parser.ibs = NULL; + parser.num_ibs = 0; job->free_job = amdgpu_cs_free_job; mutex_lock(&job->job_lock); @@ -902,24 +880,24 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) goto out; } cs->out.handle = - amdgpu_ctx_add_fence(parser->ctx, ring, + amdgpu_ctx_add_fence(parser.ctx, ring, &job->base.s_fence->base); job->ibs[job->num_ibs - 1].sequence = cs->out.handle; - list_sort(NULL, &parser->validated, cmp_size_smaller_first); - ttm_eu_fence_buffer_objects(&parser->ticket, - &parser->validated, + list_sort(NULL, &parser.validated, cmp_size_smaller_first); + ttm_eu_fence_buffer_objects(&parser.ticket, + &parser.validated, &job->base.s_fence->base); mutex_unlock(&job->job_lock); - amdgpu_cs_parser_fini_late(parser); + amdgpu_cs_parser_fini_late(&parser); mutex_unlock(&vm->mutex); return 0; } - cs->out.handle = parser->ibs[parser->num_ibs - 1].sequence; + cs->out.handle = parser.ibs[parser.num_ibs - 1].sequence; out: - amdgpu_cs_parser_fini(parser, r, reserved_buffers); + amdgpu_cs_parser_fini(&parser, r, reserved_buffers); mutex_unlock(&vm->mutex); r = amdgpu_cs_handle_lockup(adev, r); return r; -- cgit From 7034decf6a5b1ff778d83ff9d7ce1f0b404804e4 Mon Sep 17 00:00:00 2001 From: Chunming Zhou Date: Wed, 11 Nov 2015 14:56:00 +0800 Subject: drm/amdgpu: add command submission workflow tracepoint OGL needs these tracepoints to investigate performance issue. Change-Id: I5e58187d061253f7d665dfce8e4e163ba91d3e2b Signed-off-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index bf32096b8eb7..2ae73d5232dd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -888,7 +888,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) ttm_eu_fence_buffer_objects(&parser.ticket, &parser.validated, &job->base.s_fence->base); - + trace_amdgpu_cs_ioctl(job); mutex_unlock(&job->job_lock); amdgpu_cs_parser_fini_late(&parser); mutex_unlock(&vm->mutex); -- cgit From e284022163716ecf11c37fd1057c35d689ef2c11 Mon Sep 17 00:00:00 2001 From: Christian König Date: Thu, 5 Nov 2015 19:49:48 +0100 Subject: drm/amdgpu: fix incorrect mutex usage v3 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Before this patch the scheduler fence was created when we push the job into the queue, so we could only get the fence after pushing it. The mutex now was necessary to prevent the thread pushing the jobs to the hardware from running faster than the thread pushing the jobs into the queue. Otherwise the thread pushing jobs into the queue would have accessed possible freed up memory when it tries to get a reference to the fence. So what you get in the end is thread A: mutex_lock(&job->lock); ... Kick of thread B. ... mutex_unlock(&job->lock); And thread B: mutex_lock(&job->lock); .... mutex_unlock(&job->lock); kfree(job); I'm actually not sure if I'm still up to date on this, but this usage pattern used to be not allowed with mutexes. See here as well https://lwn.net/Articles/575460/. v2: remove unrelated changes, fix missing owner v3: rebased, add more commit message Signed-off-by: Christian König Reviewed-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 43 +++++++++++++++++++--------------- 1 file changed, 24 insertions(+), 19 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 2ae73d5232dd..44cf977ae4f6 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -845,8 +845,9 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) goto out; if (amdgpu_enable_scheduler && parser.num_ibs) { - struct amdgpu_job *job; struct amdgpu_ring * ring = parser.ibs->ring; + struct amd_sched_fence *fence; + struct amdgpu_job *job; job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL); if (!job) { @@ -859,37 +860,41 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) job->adev = parser.adev; job->ibs = parser.ibs; job->num_ibs = parser.num_ibs; - job->base.owner = parser.filp; - mutex_init(&job->job_lock); + job->owner = parser.filp; + job->free_job = amdgpu_cs_free_job; + if (job->ibs[job->num_ibs - 1].user) { job->uf = parser.uf; job->ibs[job->num_ibs - 1].user = &job->uf; parser.uf.bo = NULL; } - parser.ibs = NULL; - parser.num_ibs = 0; - - job->free_job = amdgpu_cs_free_job; - mutex_lock(&job->job_lock); - r = amd_sched_entity_push_job(&job->base); - if (r) { - mutex_unlock(&job->job_lock); + fence = amd_sched_fence_create(job->base.s_entity, + parser.filp); + if (!fence) { + r = -ENOMEM; amdgpu_cs_free_job(job); kfree(job); goto out; } - cs->out.handle = - amdgpu_ctx_add_fence(parser.ctx, ring, - &job->base.s_fence->base); + job->base.s_fence = fence; + fence_get(&fence->base); + + cs->out.handle = amdgpu_ctx_add_fence(parser.ctx, ring, + &fence->base); job->ibs[job->num_ibs - 1].sequence = cs->out.handle; - list_sort(NULL, &parser.validated, cmp_size_smaller_first); - ttm_eu_fence_buffer_objects(&parser.ticket, - &parser.validated, - &job->base.s_fence->base); + parser.ibs = NULL; + parser.num_ibs = 0; + trace_amdgpu_cs_ioctl(job); - mutex_unlock(&job->job_lock); + amd_sched_entity_push_job(&job->base); + + list_sort(NULL, &parser.validated, cmp_size_smaller_first); + ttm_eu_fence_buffer_objects(&parser.ticket, &parser.validated, + &fence->base); + fence_put(&fence->base); + amdgpu_cs_parser_fini_late(&parser); mutex_unlock(&vm->mutex); return 0; -- cgit From 5d82730af746abca2aa74e00de6370d338df7e95 Mon Sep 17 00:00:00 2001 From: Christian König Date: Fri, 13 Nov 2015 13:04:50 +0100 Subject: drm/amdgpu: fix handling order in scheduler CS MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit We need to clear parser.ibs and num_ibs before amd_sched_fence_create, otherwise the IB could be freed twice if fence creates fails. Signed-off-by: Christian König Reviewed-by: Chunming Zhou --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 44cf977ae4f6..6096effd6a56 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -858,11 +858,14 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) job->base.sched = &ring->sched; job->base.s_entity = &parser.ctx->rings[ring->idx].entity; job->adev = parser.adev; - job->ibs = parser.ibs; - job->num_ibs = parser.num_ibs; job->owner = parser.filp; job->free_job = amdgpu_cs_free_job; + job->ibs = parser.ibs; + job->num_ibs = parser.num_ibs; + parser.ibs = NULL; + parser.num_ibs = 0; + if (job->ibs[job->num_ibs - 1].user) { job->uf = parser.uf; job->ibs[job->num_ibs - 1].user = &job->uf; @@ -884,9 +887,6 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) &fence->base); job->ibs[job->num_ibs - 1].sequence = cs->out.handle; - parser.ibs = NULL; - parser.num_ibs = 0; - trace_amdgpu_cs_ioctl(job); amd_sched_entity_push_job(&job->base); -- cgit From 984810fc45389c545719fbb4219e8a12b27032a4 Mon Sep 17 00:00:00 2001 From: Christian König Date: Sat, 14 Nov 2015 21:05:35 +0100 Subject: drm/amdgpu: cleanup scheduler command submission MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Unify the two code path again, cause they do pretty much the same thing. Signed-off-by: Christian König Reviewed-by: Chunming Zhou Reviewed-by: Junwei Zhang --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 63 +++++++++++++--------------------- 1 file changed, 24 insertions(+), 39 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index 6096effd6a56..3afcf0237c25 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -439,8 +439,18 @@ static int cmp_size_smaller_first(void *priv, struct list_head *a, return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages; } -static void amdgpu_cs_parser_fini_early(struct amdgpu_cs_parser *parser, int error, bool backoff) +/** + * cs_parser_fini() - clean parser states + * @parser: parser structure holding parsing context. + * @error: error number + * + * If error is set than unvalidate buffer, otherwise just free memory + * used by parsing context. + **/ +static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff) { + unsigned i; + if (!error) { /* Sort the buffer list from the smallest to largest buffer, * which affects the order of buffers in the LRU list. @@ -455,17 +465,13 @@ static void amdgpu_cs_parser_fini_early(struct amdgpu_cs_parser *parser, int err list_sort(NULL, &parser->validated, cmp_size_smaller_first); ttm_eu_fence_buffer_objects(&parser->ticket, - &parser->validated, - &parser->ibs[parser->num_ibs-1].fence->base); + &parser->validated, + parser->fence); } else if (backoff) { ttm_eu_backoff_reservation(&parser->ticket, &parser->validated); } -} - -static void amdgpu_cs_parser_fini_late(struct amdgpu_cs_parser *parser) -{ - unsigned i; + fence_put(parser->fence); if (parser->ctx) amdgpu_ctx_put(parser->ctx); @@ -484,20 +490,6 @@ static void amdgpu_cs_parser_fini_late(struct amdgpu_cs_parser *parser) drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base); } -/** - * cs_parser_fini() - clean parser states - * @parser: parser structure holding parsing context. - * @error: error number - * - * If error is set than unvalidate buffer, otherwise just free memory - * used by parsing context. - **/ -static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff) -{ - amdgpu_cs_parser_fini_early(parser, error, backoff); - amdgpu_cs_parser_fini_late(parser); -} - static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p, struct amdgpu_vm *vm) { @@ -582,15 +574,9 @@ static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev, } r = amdgpu_bo_vm_update_pte(parser, vm); - if (r) { - goto out; - } - amdgpu_cs_sync_rings(parser); - if (!amdgpu_enable_scheduler) - r = amdgpu_ib_schedule(adev, parser->num_ibs, parser->ibs, - parser->filp); + if (!r) + amdgpu_cs_sync_rings(parser); -out: return r; } @@ -881,7 +867,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) goto out; } job->base.s_fence = fence; - fence_get(&fence->base); + parser.fence = fence_get(&fence->base); cs->out.handle = amdgpu_ctx_add_fence(parser.ctx, ring, &fence->base); @@ -890,17 +876,16 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) trace_amdgpu_cs_ioctl(job); amd_sched_entity_push_job(&job->base); - list_sort(NULL, &parser.validated, cmp_size_smaller_first); - ttm_eu_fence_buffer_objects(&parser.ticket, &parser.validated, - &fence->base); - fence_put(&fence->base); + } else { + struct amdgpu_fence *fence; - amdgpu_cs_parser_fini_late(&parser); - mutex_unlock(&vm->mutex); - return 0; + r = amdgpu_ib_schedule(adev, parser.num_ibs, parser.ibs, + parser.filp); + fence = parser.ibs[parser.num_ibs - 1].fence; + parser.fence = fence_get(&fence->base); + cs->out.handle = parser.ibs[parser.num_ibs - 1].sequence; } - cs->out.handle = parser.ibs[parser.num_ibs - 1].sequence; out: amdgpu_cs_parser_fini(&parser, r, reserved_buffers); mutex_unlock(&vm->mutex); -- cgit