| Message ID | 20191118103536.17675-2-daniel.vetter@ffwll.ch |
|---|---|
| State | Not Applicable |
| Headers | show |
| Series | [01/15] drm/tegra: Map cmdbuf once for reloc processing | expand |
On Mon, Nov 18, 2019 at 11:35:22AM +0100, Daniel Vetter wrote: > A few reasons to drop kmap: > > - For native objects all we do is look at obj->vaddr anyway, so might > as well not call functions for every page. > > - Reloc-processing on dma-buf is ... questionable. > > - Plus most dma-buf that bother kernel cpu mmaps give you at least > vmap, much less kmaps. And all the ones relevant for arm-soc are > again doing a obj->vaddr game anyway, there's no real kmap going on > on arm it seems. > > Plus this seems to be the only real in-tree user of dma_buf_kmap, and > I'd like to get rid of that. > > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> > Cc: Thierry Reding <thierry.reding@gmail.com> > Cc: linux-tegra@vger.kernel.org Ping for testing/review on these first 2 tegra patches. They're holding up the entire series, and I got acks for all the other bits surprisingly fast. So would like to land this rather sooner than later. I'm also working on a lot more dma-buf patches ... Thanks, Daniel > --- > drivers/gpu/host1x/job.c | 21 +++++++-------------- > 1 file changed, 7 insertions(+), 14 deletions(-) > > diff --git a/drivers/gpu/host1x/job.c b/drivers/gpu/host1x/job.c > index 25ca54de8fc5..60b2fedd0061 100644 > --- a/drivers/gpu/host1x/job.c > +++ b/drivers/gpu/host1x/job.c > @@ -244,8 +244,7 @@ static unsigned int pin_job(struct host1x *host, struct host1x_job *job) > > static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) > { > - u32 last_page = ~0; > - void *cmdbuf_page_addr = NULL; > + void *cmdbuf_addr = NULL; > struct host1x_bo *cmdbuf = g->bo; > unsigned int i; > > @@ -267,28 +266,22 @@ static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) > goto patch_reloc; > } > > - if (last_page != reloc->cmdbuf.offset >> PAGE_SHIFT) { > - if (cmdbuf_page_addr) > - host1x_bo_kunmap(cmdbuf, last_page, > - cmdbuf_page_addr); > + if (!cmdbuf_addr) { > + cmdbuf_addr = host1x_bo_mmap(cmdbuf); > > - cmdbuf_page_addr = host1x_bo_kmap(cmdbuf, > - reloc->cmdbuf.offset >> PAGE_SHIFT); > - last_page = reloc->cmdbuf.offset >> PAGE_SHIFT; > - > - if (unlikely(!cmdbuf_page_addr)) { > + if (unlikely(!cmdbuf_addr)) { > pr_err("Could not map cmdbuf for relocation\n"); > return -ENOMEM; > } > } > > - target = cmdbuf_page_addr + (reloc->cmdbuf.offset & ~PAGE_MASK); > + target = cmdbuf_addr + reloc->cmdbuf.offset; > patch_reloc: > *target = reloc_addr; > } > > - if (cmdbuf_page_addr) > - host1x_bo_kunmap(cmdbuf, last_page, cmdbuf_page_addr); > + if (cmdbuf_addr) > + host1x_bo_munmap(cmdbuf, cmdbuf_addr); > > return 0; > } > -- > 2.24.0 >
On Mon, Nov 25, 2019 at 10:58:56AM +0100, Daniel Vetter wrote: > On Mon, Nov 18, 2019 at 11:35:22AM +0100, Daniel Vetter wrote: > > A few reasons to drop kmap: > > > > - For native objects all we do is look at obj->vaddr anyway, so might > > as well not call functions for every page. > > > > - Reloc-processing on dma-buf is ... questionable. > > > > - Plus most dma-buf that bother kernel cpu mmaps give you at least > > vmap, much less kmaps. And all the ones relevant for arm-soc are > > again doing a obj->vaddr game anyway, there's no real kmap going on > > on arm it seems. > > > > Plus this seems to be the only real in-tree user of dma_buf_kmap, and > > I'd like to get rid of that. > > > > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> > > Cc: Thierry Reding <thierry.reding@gmail.com> > > Cc: linux-tegra@vger.kernel.org > > Ping for testing/review on these first 2 tegra patches. They're holding up > the entire series, and I got acks for all the other bits surprisingly > fast. So would like to land this rather sooner than later. I'm also > working on a lot more dma-buf patches ... Right, I had forgotten about this series. Let me go test it right away. Thierry > > --- > > drivers/gpu/host1x/job.c | 21 +++++++-------------- > > 1 file changed, 7 insertions(+), 14 deletions(-) > > > > diff --git a/drivers/gpu/host1x/job.c b/drivers/gpu/host1x/job.c > > index 25ca54de8fc5..60b2fedd0061 100644 > > --- a/drivers/gpu/host1x/job.c > > +++ b/drivers/gpu/host1x/job.c > > @@ -244,8 +244,7 @@ static unsigned int pin_job(struct host1x *host, struct host1x_job *job) > > > > static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) > > { > > - u32 last_page = ~0; > > - void *cmdbuf_page_addr = NULL; > > + void *cmdbuf_addr = NULL; > > struct host1x_bo *cmdbuf = g->bo; > > unsigned int i; > > > > @@ -267,28 +266,22 @@ static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) > > goto patch_reloc; > > } > > > > - if (last_page != reloc->cmdbuf.offset >> PAGE_SHIFT) { > > - if (cmdbuf_page_addr) > > - host1x_bo_kunmap(cmdbuf, last_page, > > - cmdbuf_page_addr); > > + if (!cmdbuf_addr) { > > + cmdbuf_addr = host1x_bo_mmap(cmdbuf); > > > > - cmdbuf_page_addr = host1x_bo_kmap(cmdbuf, > > - reloc->cmdbuf.offset >> PAGE_SHIFT); > > - last_page = reloc->cmdbuf.offset >> PAGE_SHIFT; > > - > > - if (unlikely(!cmdbuf_page_addr)) { > > + if (unlikely(!cmdbuf_addr)) { > > pr_err("Could not map cmdbuf for relocation\n"); > > return -ENOMEM; > > } > > } > > > > - target = cmdbuf_page_addr + (reloc->cmdbuf.offset & ~PAGE_MASK); > > + target = cmdbuf_addr + reloc->cmdbuf.offset; > > patch_reloc: > > *target = reloc_addr; > > } > > > > - if (cmdbuf_page_addr) > > - host1x_bo_kunmap(cmdbuf, last_page, cmdbuf_page_addr); > > + if (cmdbuf_addr) > > + host1x_bo_munmap(cmdbuf, cmdbuf_addr); > > > > return 0; > > } > > -- > > 2.24.0 > > > > -- > Daniel Vetter > Software Engineer, Intel Corporation > http://blog.ffwll.ch
On Mon, Nov 18, 2019 at 11:35:22AM +0100, Daniel Vetter wrote: > A few reasons to drop kmap: > > - For native objects all we do is look at obj->vaddr anyway, so might > as well not call functions for every page. > > - Reloc-processing on dma-buf is ... questionable. > > - Plus most dma-buf that bother kernel cpu mmaps give you at least > vmap, much less kmaps. And all the ones relevant for arm-soc are > again doing a obj->vaddr game anyway, there's no real kmap going on > on arm it seems. > > Plus this seems to be the only real in-tree user of dma_buf_kmap, and > I'd like to get rid of that. > > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> > Cc: Thierry Reding <thierry.reding@gmail.com> > Cc: linux-tegra@vger.kernel.org > --- > drivers/gpu/host1x/job.c | 21 +++++++-------------- > 1 file changed, 7 insertions(+), 14 deletions(-) This looks correct to me, and running some of the grate project's tests against this works just fine, so: Reviewed-by: Thierry Reding <treding@nvidia.com> Tested-by: Thierry Reding <treding@nvidia.com> > diff --git a/drivers/gpu/host1x/job.c b/drivers/gpu/host1x/job.c > index 25ca54de8fc5..60b2fedd0061 100644 > --- a/drivers/gpu/host1x/job.c > +++ b/drivers/gpu/host1x/job.c > @@ -244,8 +244,7 @@ static unsigned int pin_job(struct host1x *host, struct host1x_job *job) > > static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) > { > - u32 last_page = ~0; > - void *cmdbuf_page_addr = NULL; > + void *cmdbuf_addr = NULL; > struct host1x_bo *cmdbuf = g->bo; > unsigned int i; > > @@ -267,28 +266,22 @@ static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) > goto patch_reloc; > } > > - if (last_page != reloc->cmdbuf.offset >> PAGE_SHIFT) { > - if (cmdbuf_page_addr) > - host1x_bo_kunmap(cmdbuf, last_page, > - cmdbuf_page_addr); > + if (!cmdbuf_addr) { > + cmdbuf_addr = host1x_bo_mmap(cmdbuf); > > - cmdbuf_page_addr = host1x_bo_kmap(cmdbuf, > - reloc->cmdbuf.offset >> PAGE_SHIFT); > - last_page = reloc->cmdbuf.offset >> PAGE_SHIFT; > - > - if (unlikely(!cmdbuf_page_addr)) { > + if (unlikely(!cmdbuf_addr)) { > pr_err("Could not map cmdbuf for relocation\n"); > return -ENOMEM; > } > } > > - target = cmdbuf_page_addr + (reloc->cmdbuf.offset & ~PAGE_MASK); > + target = cmdbuf_addr + reloc->cmdbuf.offset; > patch_reloc: > *target = reloc_addr; > } > > - if (cmdbuf_page_addr) > - host1x_bo_kunmap(cmdbuf, last_page, cmdbuf_page_addr); > + if (cmdbuf_addr) > + host1x_bo_munmap(cmdbuf, cmdbuf_addr); > > return 0; > } > -- > 2.24.0 >
diff --git a/drivers/gpu/host1x/job.c b/drivers/gpu/host1x/job.c index 25ca54de8fc5..60b2fedd0061 100644 --- a/drivers/gpu/host1x/job.c +++ b/drivers/gpu/host1x/job.c @@ -244,8 +244,7 @@ static unsigned int pin_job(struct host1x *host, struct host1x_job *job) static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) { - u32 last_page = ~0; - void *cmdbuf_page_addr = NULL; + void *cmdbuf_addr = NULL; struct host1x_bo *cmdbuf = g->bo; unsigned int i; @@ -267,28 +266,22 @@ static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g) goto patch_reloc; } - if (last_page != reloc->cmdbuf.offset >> PAGE_SHIFT) { - if (cmdbuf_page_addr) - host1x_bo_kunmap(cmdbuf, last_page, - cmdbuf_page_addr); + if (!cmdbuf_addr) { + cmdbuf_addr = host1x_bo_mmap(cmdbuf); - cmdbuf_page_addr = host1x_bo_kmap(cmdbuf, - reloc->cmdbuf.offset >> PAGE_SHIFT); - last_page = reloc->cmdbuf.offset >> PAGE_SHIFT; - - if (unlikely(!cmdbuf_page_addr)) { + if (unlikely(!cmdbuf_addr)) { pr_err("Could not map cmdbuf for relocation\n"); return -ENOMEM; } } - target = cmdbuf_page_addr + (reloc->cmdbuf.offset & ~PAGE_MASK); + target = cmdbuf_addr + reloc->cmdbuf.offset; patch_reloc: *target = reloc_addr; } - if (cmdbuf_page_addr) - host1x_bo_kunmap(cmdbuf, last_page, cmdbuf_page_addr); + if (cmdbuf_addr) + host1x_bo_munmap(cmdbuf, cmdbuf_addr); return 0; }
A few reasons to drop kmap: - For native objects all we do is look at obj->vaddr anyway, so might as well not call functions for every page. - Reloc-processing on dma-buf is ... questionable. - Plus most dma-buf that bother kernel cpu mmaps give you at least vmap, much less kmaps. And all the ones relevant for arm-soc are again doing a obj->vaddr game anyway, there's no real kmap going on on arm it seems. Plus this seems to be the only real in-tree user of dma_buf_kmap, and I'd like to get rid of that. Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> Cc: Thierry Reding <thierry.reding@gmail.com> Cc: linux-tegra@vger.kernel.org --- drivers/gpu/host1x/job.c | 21 +++++++-------------- 1 file changed, 7 insertions(+), 14 deletions(-)