| Message ID | 20180209172750.15675-1-tom.stdenis@amd.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 09/02/18 12:27 PM, Tom St Denis wrote: > From: Christian König <ckoenig.leichtzumerken@gmail.com> Oops, I'll remove this from the commit message before pushing :-) I did give you credit below though. Tom > > This allows access to pages allocated through the driver with optional > IOMMU mapping. > > v2: Fix number of bytes copied and add write method > > Original-by: Christian König <christian.koenig@amd.com> > Signed-off-by: Tom St Denis <tom.stdenis@amd.com> > --- > drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 110 ++++++++++++++++++++++++++------ > 1 file changed, 89 insertions(+), 21 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > index b372d8d650a5..d6c56b001a2c 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > @@ -1929,38 +1929,106 @@ static const struct file_operations amdgpu_ttm_gtt_fops = { > > #endif > > -static ssize_t amdgpu_iova_to_phys_read(struct file *f, char __user *buf, > - size_t size, loff_t *pos) > +static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf, > + size_t size, loff_t *pos) > { > struct amdgpu_device *adev = file_inode(f)->i_private; > - int r; > - uint64_t phys; > struct iommu_domain *dom; > + ssize_t result = 0; > + int r; > > - // always return 8 bytes > - if (size != 8) > - return -EINVAL; > + dom = iommu_get_domain_for_dev(adev->dev); > > - // only accept page addresses > - if (*pos & 0xFFF) > - return -EINVAL; > + while (size) { > + phys_addr_t addr = *pos & PAGE_MASK; > + loff_t off = *pos & ~PAGE_MASK; > + size_t bytes = PAGE_SIZE - off; > + unsigned long pfn; > + struct page *p; > + void *ptr; > + > + bytes = bytes < size ? bytes : size; > + > + addr = dom ? iommu_iova_to_phys(dom, addr) : addr; > + > + pfn = addr >> PAGE_SHIFT; > + if (!pfn_valid(pfn)) > + return -EPERM; > + > + p = pfn_to_page(pfn); > + if (p->mapping != adev->mman.bdev.dev_mapping) > + return -EPERM; > + > + ptr = kmap(p); > + if (ptr) { > + r = copy_to_user(buf, ptr, bytes); > + kunmap(p); > + if (r) > + return -EFAULT; > + } else { > + return -EFAULT; > + } > + > + size -= bytes; > + *pos += bytes; > + result += bytes; > + } > + > + return result; > +} > + > +static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf, > + size_t size, loff_t *pos) > +{ > + struct amdgpu_device *adev = file_inode(f)->i_private; > + struct iommu_domain *dom; > + ssize_t result = 0; > + int r; > > dom = iommu_get_domain_for_dev(adev->dev); > - if (dom) > - phys = iommu_iova_to_phys(dom, *pos); > - else > - phys = *pos; > > - r = copy_to_user(buf, &phys, 8); > - if (r) > - return -EFAULT; > + while (size) { > + phys_addr_t addr = *pos & PAGE_MASK; > + loff_t off = *pos & ~PAGE_MASK; > + size_t bytes = PAGE_SIZE - off; > + unsigned long pfn; > + struct page *p; > + void *ptr; > + > + bytes = bytes < size ? bytes : size; > + > + addr = dom ? iommu_iova_to_phys(dom, addr) : addr; > > - return 8; > + pfn = addr >> PAGE_SHIFT; > + if (!pfn_valid(pfn)) > + return -EPERM; > + > + p = pfn_to_page(pfn); > + if (p->mapping != adev->mman.bdev.dev_mapping) > + return -EPERM; > + > + ptr = kmap(p); > + if (ptr) { > + r = copy_from_user(ptr, buf, bytes); > + kunmap(p); > + if (r) > + return -EFAULT; > + } else { > + return -EFAULT; > + } > + > + size -= bytes; > + *pos += bytes; > + result += bytes; > + } > + > + return result; > } > > -static const struct file_operations amdgpu_ttm_iova_fops = { > +static const struct file_operations amdgpu_ttm_iomem_fops = { > .owner = THIS_MODULE, > - .read = amdgpu_iova_to_phys_read, > + .read = amdgpu_iomem_read, > + .write = amdgpu_iomem_write, > .llseek = default_llseek > }; > > @@ -1973,7 +2041,7 @@ static const struct { > #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS > { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT }, > #endif > - { "amdgpu_iova", &amdgpu_ttm_iova_fops, TTM_PL_SYSTEM }, > + { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM }, > }; > > #endif >
Am 09.02.2018 um 18:28 schrieb Tom St Denis: > On 09/02/18 12:27 PM, Tom St Denis wrote: >> From: Christian König <ckoenig.leichtzumerken@gmail.com> > > Oops, I'll remove this from the commit message before pushing :-) > > I did give you credit below though. The patch before this one isn't merged yet because I'm still not sure if that works under all circumstances or not. Could you give it some extended testing? Especially if it work with eviction. Christian. > > Tom > >> >> This allows access to pages allocated through the driver with optional >> IOMMU mapping. >> >> v2: Fix number of bytes copied and add write method >> >> Original-by: Christian König <christian.koenig@amd.com> >> Signed-off-by: Tom St Denis <tom.stdenis@amd.com> >> --- >> drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 110 >> ++++++++++++++++++++++++++------ >> 1 file changed, 89 insertions(+), 21 deletions(-) >> >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c >> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c >> index b372d8d650a5..d6c56b001a2c 100644 >> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c >> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c >> @@ -1929,38 +1929,106 @@ static const struct file_operations >> amdgpu_ttm_gtt_fops = { >> #endif >> -static ssize_t amdgpu_iova_to_phys_read(struct file *f, char >> __user *buf, >> - size_t size, loff_t *pos) >> +static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf, >> + size_t size, loff_t *pos) >> { >> struct amdgpu_device *adev = file_inode(f)->i_private; >> - int r; >> - uint64_t phys; >> struct iommu_domain *dom; >> + ssize_t result = 0; >> + int r; >> - // always return 8 bytes >> - if (size != 8) >> - return -EINVAL; >> + dom = iommu_get_domain_for_dev(adev->dev); >> - // only accept page addresses >> - if (*pos & 0xFFF) >> - return -EINVAL; >> + while (size) { >> + phys_addr_t addr = *pos & PAGE_MASK; >> + loff_t off = *pos & ~PAGE_MASK; >> + size_t bytes = PAGE_SIZE - off; >> + unsigned long pfn; >> + struct page *p; >> + void *ptr; >> + >> + bytes = bytes < size ? bytes : size; >> + >> + addr = dom ? iommu_iova_to_phys(dom, addr) : addr; >> + >> + pfn = addr >> PAGE_SHIFT; >> + if (!pfn_valid(pfn)) >> + return -EPERM; >> + >> + p = pfn_to_page(pfn); >> + if (p->mapping != adev->mman.bdev.dev_mapping) >> + return -EPERM; >> + >> + ptr = kmap(p); >> + if (ptr) { >> + r = copy_to_user(buf, ptr, bytes); >> + kunmap(p); >> + if (r) >> + return -EFAULT; >> + } else { >> + return -EFAULT; >> + } >> + >> + size -= bytes; >> + *pos += bytes; >> + result += bytes; >> + } >> + >> + return result; >> +} >> + >> +static ssize_t amdgpu_iomem_write(struct file *f, const char __user >> *buf, >> + size_t size, loff_t *pos) >> +{ >> + struct amdgpu_device *adev = file_inode(f)->i_private; >> + struct iommu_domain *dom; >> + ssize_t result = 0; >> + int r; >> dom = iommu_get_domain_for_dev(adev->dev); >> - if (dom) >> - phys = iommu_iova_to_phys(dom, *pos); >> - else >> - phys = *pos; >> - r = copy_to_user(buf, &phys, 8); >> - if (r) >> - return -EFAULT; >> + while (size) { >> + phys_addr_t addr = *pos & PAGE_MASK; >> + loff_t off = *pos & ~PAGE_MASK; >> + size_t bytes = PAGE_SIZE - off; >> + unsigned long pfn; >> + struct page *p; >> + void *ptr; >> + >> + bytes = bytes < size ? bytes : size; >> + >> + addr = dom ? iommu_iova_to_phys(dom, addr) : addr; >> - return 8; >> + pfn = addr >> PAGE_SHIFT; >> + if (!pfn_valid(pfn)) >> + return -EPERM; >> + >> + p = pfn_to_page(pfn); >> + if (p->mapping != adev->mman.bdev.dev_mapping) >> + return -EPERM; >> + >> + ptr = kmap(p); >> + if (ptr) { >> + r = copy_from_user(ptr, buf, bytes); >> + kunmap(p); >> + if (r) >> + return -EFAULT; >> + } else { >> + return -EFAULT; >> + } >> + >> + size -= bytes; >> + *pos += bytes; >> + result += bytes; >> + } >> + >> + return result; >> } >> -static const struct file_operations amdgpu_ttm_iova_fops = { >> +static const struct file_operations amdgpu_ttm_iomem_fops = { >> .owner = THIS_MODULE, >> - .read = amdgpu_iova_to_phys_read, >> + .read = amdgpu_iomem_read, >> + .write = amdgpu_iomem_write, >> .llseek = default_llseek >> }; >> @@ -1973,7 +2041,7 @@ static const struct { >> #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS >> { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT }, >> #endif >> - { "amdgpu_iova", &amdgpu_ttm_iova_fops, TTM_PL_SYSTEM }, >> + { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM }, >> }; >> #endif >> >
On 09/02/18 01:17 PM, Christian König wrote: > Am 09.02.2018 um 18:28 schrieb Tom St Denis: >> On 09/02/18 12:27 PM, Tom St Denis wrote: >>> From: Christian König <ckoenig.leichtzumerken@gmail.com> >> >> Oops, I'll remove this from the commit message before pushing :-) >> >> I did give you credit below though. > > The patch before this one isn't merged yet because I'm still not sure if > that works under all circumstances or not. > > Could you give it some extended testing? Especially if it work with > eviction. I supposed there is a race on the kmap'ed memory which is why I added a ptr check. Not perfect but since it's a debugfs entry probably better than it needs to be. I'll give it a try on some live traffic next. Cheers, Tom
Am 09.02.2018 um 19:19 schrieb Tom St Denis: > On 09/02/18 01:17 PM, Christian König wrote: >> Am 09.02.2018 um 18:28 schrieb Tom St Denis: >>> On 09/02/18 12:27 PM, Tom St Denis wrote: >>>> From: Christian König <ckoenig.leichtzumerken@gmail.com> >>> >>> Oops, I'll remove this from the commit message before pushing :-) >>> >>> I did give you credit below though. >> >> The patch before this one isn't merged yet because I'm still not sure >> if that works under all circumstances or not. >> >> Could you give it some extended testing? Especially if it work with >> eviction. > > I supposed there is a race on the kmap'ed memory which is why I added > a ptr check. Not perfect but since it's a debugfs entry probably > better than it needs to be. I think you can drop that, kmap can only return NULL on 32bit systems when we ran out of vmap area and then you can basically call panic() as well. The question is if setting page->mapping during allocation has some undesired side effect. Christian. > > I'll give it a try on some live traffic next. > > Cheers, > Tom > _______________________________________________ > amd-gfx mailing list > amd-gfx@lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/amd-gfx
On 09/02/18 01:44 PM, Christian König wrote: > Am 09.02.2018 um 19:19 schrieb Tom St Denis: >> On 09/02/18 01:17 PM, Christian König wrote: >>> Am 09.02.2018 um 18:28 schrieb Tom St Denis: >>>> On 09/02/18 12:27 PM, Tom St Denis wrote: >>>>> From: Christian König <ckoenig.leichtzumerken@gmail.com> >>>> >>>> Oops, I'll remove this from the commit message before pushing :-) >>>> >>>> I did give you credit below though. >>> >>> The patch before this one isn't merged yet because I'm still not sure >>> if that works under all circumstances or not. >>> >>> Could you give it some extended testing? Especially if it work with >>> eviction. >> >> I supposed there is a race on the kmap'ed memory which is why I added >> a ptr check. Not perfect but since it's a debugfs entry probably >> better than it needs to be. > > I think you can drop that, kmap can only return NULL on 32bit systems > when we ran out of vmap area and then you can basically call panic() as > well. > > The question is if setting page->mapping during allocation has some > undesired side effect. Haven't noticed anything with piglit or other FOSS GL games. Is there a specific test you'd like me to consider? Tom
Am 12.02.2018 um 17:40 schrieb Tom St Denis: > On 09/02/18 01:44 PM, Christian König wrote: >> Am 09.02.2018 um 19:19 schrieb Tom St Denis: >>> On 09/02/18 01:17 PM, Christian König wrote: >>>> Am 09.02.2018 um 18:28 schrieb Tom St Denis: >>>>> On 09/02/18 12:27 PM, Tom St Denis wrote: >>>>>> From: Christian König <ckoenig.leichtzumerken@gmail.com> >>>>> >>>>> Oops, I'll remove this from the commit message before pushing :-) >>>>> >>>>> I did give you credit below though. >>>> >>>> The patch before this one isn't merged yet because I'm still not >>>> sure if that works under all circumstances or not. >>>> >>>> Could you give it some extended testing? Especially if it work with >>>> eviction. >>> >>> I supposed there is a race on the kmap'ed memory which is why I >>> added a ptr check. Not perfect but since it's a debugfs entry >>> probably better than it needs to be. >> >> I think you can drop that, kmap can only return NULL on 32bit systems >> when we ran out of vmap area and then you can basically call panic() >> as well. >> >> The question is if setting page->mapping during allocation has some >> undesired side effect. > > > Haven't noticed anything with piglit or other FOSS GL games. > > Is there a specific test you'd like me to consider? Try to restrict VRAM/GTT/memory in general and run something which needs a lot of CPU access to BOs. When it affects anything then the CPU page table unmap of BOs during eviction. Christian. > > Tom
On 12/02/18 12:16 PM, Christian König wrote: > Am 12.02.2018 um 17:40 schrieb Tom St Denis: >> On 09/02/18 01:44 PM, Christian König wrote: >>> Am 09.02.2018 um 19:19 schrieb Tom St Denis: >>>> On 09/02/18 01:17 PM, Christian König wrote: >>>>> Am 09.02.2018 um 18:28 schrieb Tom St Denis: >>>>>> On 09/02/18 12:27 PM, Tom St Denis wrote: >>>>>>> From: Christian König <ckoenig.leichtzumerken@gmail.com> >>>>>> >>>>>> Oops, I'll remove this from the commit message before pushing :-) >>>>>> >>>>>> I did give you credit below though. >>>>> >>>>> The patch before this one isn't merged yet because I'm still not >>>>> sure if that works under all circumstances or not. >>>>> >>>>> Could you give it some extended testing? Especially if it work with >>>>> eviction. >>>> >>>> I supposed there is a race on the kmap'ed memory which is why I >>>> added a ptr check. Not perfect but since it's a debugfs entry >>>> probably better than it needs to be. >>> >>> I think you can drop that, kmap can only return NULL on 32bit systems >>> when we ran out of vmap area and then you can basically call panic() >>> as well. >>> >>> The question is if setting page->mapping during allocation has some >>> undesired side effect. >> >> >> Haven't noticed anything with piglit or other FOSS GL games. >> >> Is there a specific test you'd like me to consider? > > Try to restrict VRAM/GTT/memory in general and run something which needs > a lot of CPU access to BOs. When it affects anything then the CPU page > table unmap of BOs during eviction. Limiting vram to 256M seems to be fine for both my iGPU and dGPU (thought it drops heaven from 40fps to about 3fps hehehe). Trying to play with gttsize=256M seems to have broken things I get: [ 165.080976] [drm:amdgpu_cs_ioctl [amdgpu]] *ERROR* amdgpu_cs_list_validate(validated) failed. [ 165.081163] [drm:amdgpu_cs_ioctl [amdgpu]] *ERROR* Not enough memory for command submission! and it seems heaven is blocked. There are fences being submitted (and signalled) but nothing is being drawn. Tom
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index b372d8d650a5..d6c56b001a2c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -1929,38 +1929,106 @@ static const struct file_operations amdgpu_ttm_gtt_fops = { #endif -static ssize_t amdgpu_iova_to_phys_read(struct file *f, char __user *buf, - size_t size, loff_t *pos) +static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf, + size_t size, loff_t *pos) { struct amdgpu_device *adev = file_inode(f)->i_private; - int r; - uint64_t phys; struct iommu_domain *dom; + ssize_t result = 0; + int r; - // always return 8 bytes - if (size != 8) - return -EINVAL; + dom = iommu_get_domain_for_dev(adev->dev); - // only accept page addresses - if (*pos & 0xFFF) - return -EINVAL; + while (size) { + phys_addr_t addr = *pos & PAGE_MASK; + loff_t off = *pos & ~PAGE_MASK; + size_t bytes = PAGE_SIZE - off; + unsigned long pfn; + struct page *p; + void *ptr; + + bytes = bytes < size ? bytes : size; + + addr = dom ? iommu_iova_to_phys(dom, addr) : addr; + + pfn = addr >> PAGE_SHIFT; + if (!pfn_valid(pfn)) + return -EPERM; + + p = pfn_to_page(pfn); + if (p->mapping != adev->mman.bdev.dev_mapping) + return -EPERM; + + ptr = kmap(p); + if (ptr) { + r = copy_to_user(buf, ptr, bytes); + kunmap(p); + if (r) + return -EFAULT; + } else { + return -EFAULT; + } + + size -= bytes; + *pos += bytes; + result += bytes; + } + + return result; +} + +static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf, + size_t size, loff_t *pos) +{ + struct amdgpu_device *adev = file_inode(f)->i_private; + struct iommu_domain *dom; + ssize_t result = 0; + int r; dom = iommu_get_domain_for_dev(adev->dev); - if (dom) - phys = iommu_iova_to_phys(dom, *pos); - else - phys = *pos; - r = copy_to_user(buf, &phys, 8); - if (r) - return -EFAULT; + while (size) { + phys_addr_t addr = *pos & PAGE_MASK; + loff_t off = *pos & ~PAGE_MASK; + size_t bytes = PAGE_SIZE - off; + unsigned long pfn; + struct page *p; + void *ptr; + + bytes = bytes < size ? bytes : size; + + addr = dom ? iommu_iova_to_phys(dom, addr) : addr; - return 8; + pfn = addr >> PAGE_SHIFT; + if (!pfn_valid(pfn)) + return -EPERM; + + p = pfn_to_page(pfn); + if (p->mapping != adev->mman.bdev.dev_mapping) + return -EPERM; + + ptr = kmap(p); + if (ptr) { + r = copy_from_user(ptr, buf, bytes); + kunmap(p); + if (r) + return -EFAULT; + } else { + return -EFAULT; + } + + size -= bytes; + *pos += bytes; + result += bytes; + } + + return result; } -static const struct file_operations amdgpu_ttm_iova_fops = { +static const struct file_operations amdgpu_ttm_iomem_fops = { .owner = THIS_MODULE, - .read = amdgpu_iova_to_phys_read, + .read = amdgpu_iomem_read, + .write = amdgpu_iomem_write, .llseek = default_llseek }; @@ -1973,7 +2041,7 @@ static const struct { #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS { "amdgpu_gtt", &amdgpu_ttm_gtt_fops, TTM_PL_TT }, #endif - { "amdgpu_iova", &amdgpu_ttm_iova_fops, TTM_PL_SYSTEM }, + { "amdgpu_iomem", &amdgpu_ttm_iomem_fops, TTM_PL_SYSTEM }, }; #endif