| Message ID | 20200728041736.20689-1-airlied@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm/ttm/nouveau: don't call tt destroy callback on alloc failure. | expand |
Am 28.07.20 um 06:17 schrieb Dave Airlie: > From: Dave Airlie <airlied@redhat.com> > > This is confusing, and from my reading of all the drivers only > nouveau got this right. > > Just make the API act under driver control of it's own allocation > failing, and don't call destroy, if the page table fails to > create there is nothing to cleanup here. > > (I'm willing to believe I've missed something here, so please > review deeply). > > Signed-off-by: Dave Airlie <airlied@redhat.com> That looks right to me as well, Reviewed-by: Christian König <christian.koenig@amd.com> > --- > drivers/gpu/drm/nouveau/nouveau_sgdma.c | 9 +++------ > drivers/gpu/drm/ttm/ttm_tt.c | 3 --- > 2 files changed, 3 insertions(+), 9 deletions(-) > > diff --git a/drivers/gpu/drm/nouveau/nouveau_sgdma.c b/drivers/gpu/drm/nouveau/nouveau_sgdma.c > index 20b6d0b3de5c..c3ccf661b7a6 100644 > --- a/drivers/gpu/drm/nouveau/nouveau_sgdma.c > +++ b/drivers/gpu/drm/nouveau/nouveau_sgdma.c > @@ -95,12 +95,9 @@ nouveau_sgdma_create_ttm(struct ttm_buffer_object *bo, uint32_t page_flags) > else > nvbe->ttm.ttm.func = &nv50_sgdma_backend; > > - if (ttm_dma_tt_init(&nvbe->ttm, bo, page_flags)) > - /* > - * A failing ttm_dma_tt_init() will call ttm_tt_destroy() > - * and thus our nouveau_sgdma_destroy() hook, so we don't need > - * to free nvbe here. > - */ > + if (ttm_dma_tt_init(&nvbe->ttm, bo, page_flags)) { > + kfree(nvbe); > return NULL; > + } > return &nvbe->ttm.ttm; > } > diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c > index bab67873cfd4..9d1c7177384c 100644 > --- a/drivers/gpu/drm/ttm/ttm_tt.c > +++ b/drivers/gpu/drm/ttm/ttm_tt.c > @@ -244,7 +244,6 @@ int ttm_tt_init(struct ttm_tt *ttm, struct ttm_buffer_object *bo, > ttm_tt_init_fields(ttm, bo, page_flags); > > if (ttm_tt_alloc_page_directory(ttm)) { > - ttm_tt_destroy(ttm); > pr_err("Failed allocating page table\n"); > return -ENOMEM; > } > @@ -268,7 +267,6 @@ int ttm_dma_tt_init(struct ttm_dma_tt *ttm_dma, struct ttm_buffer_object *bo, > > INIT_LIST_HEAD(&ttm_dma->pages_list); > if (ttm_dma_tt_alloc_page_directory(ttm_dma)) { > - ttm_tt_destroy(ttm); > pr_err("Failed allocating page table\n"); > return -ENOMEM; > } > @@ -290,7 +288,6 @@ int ttm_sg_tt_init(struct ttm_dma_tt *ttm_dma, struct ttm_buffer_object *bo, > else > ret = ttm_dma_tt_alloc_page_directory(ttm_dma); > if (ret) { > - ttm_tt_destroy(ttm); > pr_err("Failed allocating page table\n"); > return -ENOMEM; > }
On Tue, 28 Jul 2020 at 17:49, Christian König <christian.koenig@amd.com> wrote: > > Am 28.07.20 um 06:17 schrieb Dave Airlie: > > From: Dave Airlie <airlied@redhat.com> > > > > This is confusing, and from my reading of all the drivers only > > nouveau got this right. > > > > Just make the API act under driver control of it's own allocation > > failing, and don't call destroy, if the page table fails to > > create there is nothing to cleanup here. > > > > (I'm willing to believe I've missed something here, so please > > review deeply). > > > > Signed-off-by: Dave Airlie <airlied@redhat.com> > > That looks right to me as well, Reviewed-by: Christian König > <christian.koenig@amd.com> Landed this in drm-next as well. Dave.
Am 29.07.20 um 02:08 schrieb Dave Airlie: > [SNIP] > Landed this in drm-next as well. By the way there is another design bug in TTM which only affects Nouveau. See ttm_mem_io_reserve_vm() and ttm_mem_io_free_vm(). What happens is that as soon as you have an userspace VM mapping the BO ends up on the LRU and a flag noting that so that we don't free the VM reference twice. Additional to that we have a reference count to avoid calling the io_mem_reserve and io_mem_free callbacks unbalanced if there are both userspace VM mappings and kernel mappings for CPU copies of BOs or kmap() etc... But this design is complete nonsense. What we need instead is some handling which prevents adding the io_mem to the LRU as soon as there are any kernel mappings, because we can't restore those using a fault. I already tried to move the whole handling into Nouveau itself, so that we can properly fix this up there. But Ben reported that my only compile tested patches don't work and without hardware I can't finish them up. Regards, Christian.
diff --git a/drivers/gpu/drm/nouveau/nouveau_sgdma.c b/drivers/gpu/drm/nouveau/nouveau_sgdma.c index 20b6d0b3de5c..c3ccf661b7a6 100644 --- a/drivers/gpu/drm/nouveau/nouveau_sgdma.c +++ b/drivers/gpu/drm/nouveau/nouveau_sgdma.c @@ -95,12 +95,9 @@ nouveau_sgdma_create_ttm(struct ttm_buffer_object *bo, uint32_t page_flags) else nvbe->ttm.ttm.func = &nv50_sgdma_backend; - if (ttm_dma_tt_init(&nvbe->ttm, bo, page_flags)) - /* - * A failing ttm_dma_tt_init() will call ttm_tt_destroy() - * and thus our nouveau_sgdma_destroy() hook, so we don't need - * to free nvbe here. - */ + if (ttm_dma_tt_init(&nvbe->ttm, bo, page_flags)) { + kfree(nvbe); return NULL; + } return &nvbe->ttm.ttm; } diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c index bab67873cfd4..9d1c7177384c 100644 --- a/drivers/gpu/drm/ttm/ttm_tt.c +++ b/drivers/gpu/drm/ttm/ttm_tt.c @@ -244,7 +244,6 @@ int ttm_tt_init(struct ttm_tt *ttm, struct ttm_buffer_object *bo, ttm_tt_init_fields(ttm, bo, page_flags); if (ttm_tt_alloc_page_directory(ttm)) { - ttm_tt_destroy(ttm); pr_err("Failed allocating page table\n"); return -ENOMEM; } @@ -268,7 +267,6 @@ int ttm_dma_tt_init(struct ttm_dma_tt *ttm_dma, struct ttm_buffer_object *bo, INIT_LIST_HEAD(&ttm_dma->pages_list); if (ttm_dma_tt_alloc_page_directory(ttm_dma)) { - ttm_tt_destroy(ttm); pr_err("Failed allocating page table\n"); return -ENOMEM; } @@ -290,7 +288,6 @@ int ttm_sg_tt_init(struct ttm_dma_tt *ttm_dma, struct ttm_buffer_object *bo, else ret = ttm_dma_tt_alloc_page_directory(ttm_dma); if (ret) { - ttm_tt_destroy(ttm); pr_err("Failed allocating page table\n"); return -ENOMEM; }