From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
This commit fixes this by using phys_to_page() since we get a physical address from dma_alloc_coherent(). Note that this is not a proper fix because if an IOMMU is set up to translate addresses for the GPU this address will be an I/O virtual address rather than a physical one. The proper fix probably involves not getting a pointer to the struct page in the first place, but that would be a much more intrusive change, if at all possible.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Signed-off-by: Thierry Reding treding@nvidia.com --- Arnd, I realize that this isn't a proper fix according to what we discussed on IRC yesterday, but I can't see a way to remove access to the pages array that would be as simple as this. I've marked this as RFC in the hope that it will trigger some discussion that will lead to a proper solution.
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c index c96db433f8af..d7993985752c 100644 --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -343,7 +343,11 @@ static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) &d_page->dma, pool->gfp_flags); if (d_page->vaddr) +#if defined(CONFIG_ARM) || defined(CONFIG_ARM64) + d_page->p = phys_to_page(d_page->dma); +#else d_page->p = virt_to_page(d_page->vaddr); +#endif else { kfree(d_page); d_page = NULL;
On Wed, Nov 12, 2014 at 01:39:05PM +0100, Thierry Reding wrote:
From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
Oh! That is not good!
This commit fixes this by using phys_to_page() since we get a physical address from dma_alloc_coherent(). Note that this is not a proper fix because if an IOMMU is set up to translate addresses for the GPU this address will be an I/O virtual address rather than a physical one. The proper fix probably involves not getting a pointer to the struct page in the first place, but that would be a much more intrusive change, if at all possible.
What type of caching types are there on ARM? We use the 'struct page' on the set_pages_to_[wc|uc|wb] but all of those are X86 specfic.
But I think you could by passing the 'struct dma_page' instead of 'struct page' (and the array uses) around. That should solve the touching of 'struct page' and we can treat it as an opaque type.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Is there a way to query the 'struct device' to see if the IOMMU translation is enabled/disabled for said device?
Now your patch looks to get the 'struct page' by doing some form of translation. Could you explain to me which type of memory have a 'struct page' and which ones do not ?
It is OK if you explain this in nauseating details :-)
Signed-off-by: Thierry Reding treding@nvidia.com
Arnd, I realize that this isn't a proper fix according to what we discussed on IRC yesterday, but I can't see a way to remove access to the pages array that would be as simple as this. I've marked this as RFC in the hope that it will trigger some discussion that will lead to a proper solution.
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c index c96db433f8af..d7993985752c 100644 --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -343,7 +343,11 @@ static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) &d_page->dma, pool->gfp_flags); if (d_page->vaddr) +#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);+#else d_page->p = virt_to_page(d_page->vaddr); +#endif else { kfree(d_page); d_page = NULL; -- 2.1.3
On Wednesday 12 November 2014 09:18:59 Konrad Rzeszutek Wilk wrote:
On Wed, Nov 12, 2014 at 01:39:05PM +0100, Thierry Reding wrote:
From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
Oh! That is not good!
I think what Thierry meant is that the returned pointer is /not/ in the linear range.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Is there a way to query the 'struct device' to see if the IOMMU translation is enabled/disabled for said device?
Now your patch looks to get the 'struct page' by doing some form of translation. Could you explain to me which type of memory have a 'struct page' and which ones do not ?
It is OK if you explain this in nauseating details
Basically there are two types of memory that have a struct page:
- directly mapped cacheable memory, i.e. anything that can be accessed through a kernel pointer without having to go though ioremap/vmalloc/...
- highmem pages on 32-bit system.
On noncoherent ARM systems, dma_alloc_coherent will return memory that is was unmapped from the linear range to avoid having both cacheable and noncachable mappings for the same page.
Arnd
On Wed, Nov 12, 2014 at 06:03:49PM +0100, Arnd Bergmann wrote:
On Wednesday 12 November 2014 09:18:59 Konrad Rzeszutek Wilk wrote:
On Wed, Nov 12, 2014 at 01:39:05PM +0100, Thierry Reding wrote:
From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
Oh! That is not good!
I think what Thierry meant is that the returned pointer is /not/ in the linear range.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Is there a way to query the 'struct device' to see if the IOMMU translation is enabled/disabled for said device?
?
Now your patch looks to get the 'struct page' by doing some form of translation. Could you explain to me which type of memory have a 'struct page' and which ones do not ?
It is OK if you explain this in nauseating details
Basically there are two types of memory that have a struct page:
directly mapped cacheable memory, i.e. anything that can be accessed through a kernel pointer without having to go though ioremap/vmalloc/...
highmem pages on 32-bit system.
On noncoherent ARM systems, dma_alloc_coherent will return memory that is was unmapped from the linear range to avoid having both cacheable and noncachable mappings for the same page.
Arnd
On 11/12/2014 09:39 PM, Thierry Reding wrote:
From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
This commit fixes this by using phys_to_page() since we get a physical address from dma_alloc_coherent(). Note that this is not a proper fix because if an IOMMU is set up to translate addresses for the GPU this address will be an I/O virtual address rather than a physical one. The proper fix probably involves not getting a pointer to the struct page in the first place, but that would be a much more intrusive change, if at all possible.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Signed-off-by: Thierry Reding treding@nvidia.com
Arnd, I realize that this isn't a proper fix according to what we discussed on IRC yesterday, but I can't see a way to remove access to the pages array that would be as simple as this. I've marked this as RFC in the hope that it will trigger some discussion that will lead to a proper solution.
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c index c96db433f8af..d7993985752c 100644 --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -343,7 +343,11 @@ static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) &d_page->dma, pool->gfp_flags); if (d_page->vaddr) +#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);+#else d_page->p = virt_to_page(d_page->vaddr); +#endif
Since I am messing with the IOMMU I just happened to have hit the issue you are mentioning. Wouldn't the following work:
- if (d_page->vaddr) -#if defined(CONFIG_ARM) || defined(CONFIG_ARM64) - d_page->p = phys_to_page(d_page->dma); -#else - d_page->p = virt_to_page(d_page->vaddr); -#endif - else { + if (d_page->vaddr) { + if (is_vmalloc_addr(d_page->vaddr)) { + d_page->p = vmalloc_to_page(d_page->vaddr); + } else { + d_page->p = virt_to_page(d_page->vaddr); + } + } else {
A remapped page will end up in the vmalloc range of the address space, and in this case we can use vmalloc_to_page() to get the right page. Pages outside of this range are part of the linear mapping and can be resolved using virt_to_page().
Jetson seems to be mostly happy with this, although I sometimes get the following trace:
[ 13.174763] kernel BUG at ../mm/slab.c:2593! [ 13.174767] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP ARM [ 13.174790] Modules linked in: nouveau_platform(O+) nouveau(O) cfbfillrect cfbimgblt cfbcopyarea ttm ... [ 13.175234] [<c00de238>] (cache_alloc_refill) from [<c00de528>] (__kmalloc+0x100/0x13c) [ 13.175247] [<c00de528>] (__kmalloc) from [<c001d564>] (arm_iommu_alloc_attrs+0x94/0x3a8) [ 13.175269] [<c001d564>] (arm_iommu_alloc_attrs) from [<bf008f4c>] (ttm_dma_populate+0x498/0x76c [ttm]) [ 13.175294] [<bf008f4c>] (ttm_dma_populate [ttm]) from [<bf000bb8>] (ttm_tt_bind+0x38/0x68 [ttm]) [ 13.175315] [<bf000bb8>] (ttm_tt_bind [ttm]) from [<bf00298c>] (ttm_bo_handle_move_mem+0x408/0x47c [ttm]) [ 13.175337] [<bf00298c>] (ttm_bo_handle_move_mem [ttm]) from [<bf003758>] (ttm_bo_validate+0x220/0x22c [ttm]) [ 13.175359] [<bf003758>] (ttm_bo_validate [ttm]) from [<bf003984>] (ttm_bo_init+0x220/0x338 [ttm]) [ 13.175480] [<bf003984>] (ttm_bo_init [ttm]) from [<bf0c70a0>] (nouveau_bo_new+0x1c0/0x294 [nouveau]) [ 13.175688] [<bf0c70a0>] (nouveau_bo_new [nouveau]) from [<bf0ce88c>] (nv84_fence_create+0x1cc/0x240 [nouveau]) [ 13.175891] [<bf0ce88c>] (nv84_fence_create [nouveau]) from [<bf0cec90>] (nvc0_fence_create+0xc/0x24 [nouveau]) [ 13.176094] [<bf0cec90>] (nvc0_fence_create [nouveau]) from [<bf0c1480>] (nouveau_accel_init+0xec/0x450 [nouveau])
I suspect this is related to this change, but it might also be the side-effect of another bug in my code.
On 12/08/2014 04:14 PM, Alexandre Courbot wrote:
On 11/12/2014 09:39 PM, Thierry Reding wrote:
From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
This commit fixes this by using phys_to_page() since we get a physical address from dma_alloc_coherent(). Note that this is not a proper fix because if an IOMMU is set up to translate addresses for the GPU this address will be an I/O virtual address rather than a physical one. The proper fix probably involves not getting a pointer to the struct page in the first place, but that would be a much more intrusive change, if at all possible.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Signed-off-by: Thierry Reding treding@nvidia.com
Arnd, I realize that this isn't a proper fix according to what we discussed on IRC yesterday, but I can't see a way to remove access to the pages array that would be as simple as this. I've marked this as RFC in the hope that it will trigger some discussion that will lead to a proper solution.
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c index c96db433f8af..d7993985752c 100644 --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -343,7 +343,11 @@ static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) &d_page->dma, pool->gfp_flags); if (d_page->vaddr) +#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);+#else d_page->p = virt_to_page(d_page->vaddr); +#endif
Since I am messing with the IOMMU I just happened to have hit the issue you are mentioning. Wouldn't the following work:
if (d_page->vaddr)-#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);-#else
d_page->p = virt_to_page(d_page->vaddr);-#endif
else {
if (d_page->vaddr) {if (is_vmalloc_addr(d_page->vaddr)) {d_page->p = vmalloc_to_page(d_page->vaddr);} else {d_page->p = virt_to_page(d_page->vaddr);}} else {A remapped page will end up in the vmalloc range of the address space, and in this case we can use vmalloc_to_page() to get the right page. Pages outside of this range are part of the linear mapping and can be resolved using virt_to_page().
Jetson seems to be mostly happy with this, although I sometimes get the following trace:
[ 13.174763] kernel BUG at ../mm/slab.c:2593! [ 13.174767] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP ARM [ 13.174790] Modules linked in: nouveau_platform(O+) nouveau(O) cfbfillrect cfbimgblt cfbcopyarea ttm ... [ 13.175234] [<c00de238>] (cache_alloc_refill) from [<c00de528>] (__kmalloc+0x100/0x13c) [ 13.175247] [<c00de528>] (__kmalloc) from [<c001d564>] (arm_iommu_alloc_attrs+0x94/0x3a8) [ 13.175269] [<c001d564>] (arm_iommu_alloc_attrs) from [<bf008f4c>] (ttm_dma_populate+0x498/0x76c [ttm]) [ 13.175294] [<bf008f4c>] (ttm_dma_populate [ttm]) from [<bf000bb8>] (ttm_tt_bind+0x38/0x68 [ttm]) [ 13.175315] [<bf000bb8>] (ttm_tt_bind [ttm]) from [<bf00298c>] (ttm_bo_handle_move_mem+0x408/0x47c [ttm]) [ 13.175337] [<bf00298c>] (ttm_bo_handle_move_mem [ttm]) from [<bf003758>] (ttm_bo_validate+0x220/0x22c [ttm]) [ 13.175359] [<bf003758>] (ttm_bo_validate [ttm]) from [<bf003984>] (ttm_bo_init+0x220/0x338 [ttm]) [ 13.175480] [<bf003984>] (ttm_bo_init [ttm]) from [<bf0c70a0>] (nouveau_bo_new+0x1c0/0x294 [nouveau]) [ 13.175688] [<bf0c70a0>] (nouveau_bo_new [nouveau]) from [<bf0ce88c>] (nv84_fence_create+0x1cc/0x240 [nouveau]) [ 13.175891] [<bf0ce88c>] (nv84_fence_create [nouveau]) from [<bf0cec90>] (nvc0_fence_create+0xc/0x24 [nouveau]) [ 13.176094] [<bf0cec90>] (nvc0_fence_create [nouveau]) from [<bf0c1480>] (nouveau_accel_init+0xec/0x450 [nouveau])
I suspect this is related to this change, but it might also be the side-effect of another bug in my code.
FWIW and after some more testing, I noticed that without IOMMU vmalloc_to_page() and phys_to_page() both return the same valid page. With the IOMMU enabled vmalloc_to_page() still returns what seems to be valid pages (phys_to_page() of course doesn't make sense anymore).
So AFAICT the change I proposed seems valid. I am not sure what causes the BUG() in the slab allocator.
On 12/08/2014 04:14 PM, Alexandre Courbot wrote:
On 11/12/2014 09:39 PM, Thierry Reding wrote:
From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
This commit fixes this by using phys_to_page() since we get a physical address from dma_alloc_coherent(). Note that this is not a proper fix because if an IOMMU is set up to translate addresses for the GPU this address will be an I/O virtual address rather than a physical one. The proper fix probably involves not getting a pointer to the struct page in the first place, but that would be a much more intrusive change, if at all possible.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Signed-off-by: Thierry Reding treding@nvidia.com
Arnd, I realize that this isn't a proper fix according to what we discussed on IRC yesterday, but I can't see a way to remove access to the pages array that would be as simple as this. I've marked this as RFC in the hope that it will trigger some discussion that will lead to a proper solution.
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c index c96db433f8af..d7993985752c 100644 --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -343,7 +343,11 @@ static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) &d_page->dma, pool->gfp_flags); if (d_page->vaddr) +#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);+#else d_page->p = virt_to_page(d_page->vaddr); +#endif
Since I am messing with the IOMMU I just happened to have hit the issue you are mentioning. Wouldn't the following work:
if (d_page->vaddr)-#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);-#else
d_page->p = virt_to_page(d_page->vaddr);-#endif
else {
if (d_page->vaddr) {if (is_vmalloc_addr(d_page->vaddr)) {d_page->p = vmalloc_to_page(d_page->vaddr);} else {d_page->p = virt_to_page(d_page->vaddr);}} else {A remapped page will end up in the vmalloc range of the address space, and in this case we can use vmalloc_to_page() to get the right page. Pages outside of this range are part of the linear mapping and can be resolved using virt_to_page().
Jetson seems to be mostly happy with this, although I sometimes get the following trace:
[ 13.174763] kernel BUG at ../mm/slab.c:2593! [ 13.174767] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP ARM [ 13.174790] Modules linked in: nouveau_platform(O+) nouveau(O) cfbfillrect cfbimgblt cfbcopyarea ttm ... [ 13.175234] [<c00de238>] (cache_alloc_refill) from [<c00de528>] (__kmalloc+0x100/0x13c) [ 13.175247] [<c00de528>] (__kmalloc) from [<c001d564>] (arm_iommu_alloc_attrs+0x94/0x3a8) [ 13.175269] [<c001d564>] (arm_iommu_alloc_attrs) from [<bf008f4c>] (ttm_dma_populate+0x498/0x76c [ttm]) [ 13.175294] [<bf008f4c>] (ttm_dma_populate [ttm]) from [<bf000bb8>] (ttm_tt_bind+0x38/0x68 [ttm]) [ 13.175315] [<bf000bb8>] (ttm_tt_bind [ttm]) from [<bf00298c>] (ttm_bo_handle_move_mem+0x408/0x47c [ttm]) [ 13.175337] [<bf00298c>] (ttm_bo_handle_move_mem [ttm]) from [<bf003758>] (ttm_bo_validate+0x220/0x22c [ttm]) [ 13.175359] [<bf003758>] (ttm_bo_validate [ttm]) from [<bf003984>] (ttm_bo_init+0x220/0x338 [ttm]) [ 13.175480] [<bf003984>] (ttm_bo_init [ttm]) from [<bf0c70a0>] (nouveau_bo_new+0x1c0/0x294 [nouveau]) [ 13.175688] [<bf0c70a0>] (nouveau_bo_new [nouveau]) from [<bf0ce88c>] (nv84_fence_create+0x1cc/0x240 [nouveau]) [ 13.175891] [<bf0ce88c>] (nv84_fence_create [nouveau]) from [<bf0cec90>] (nvc0_fence_create+0xc/0x24 [nouveau]) [ 13.176094] [<bf0cec90>] (nvc0_fence_create [nouveau]) from [<bf0c1480>] (nouveau_accel_init+0xec/0x450 [nouveau])
I suspect this is related to this change, but it might also be the side-effect of another bug in my code.
Ok, I finally understand where this trace comes from.
slab.c:2593: BUG_ON(flags & GFP_SLAB_BUG_MASK);
gfp.h: #define GFP_SLAB_BUG_MASK (__GFP_DMA32|__GFP_HIGHMEM|~__GFP_BITS_MASK)
and ttm_dma_populate() calls ttm_dma_pool_init() with gfp_flags == GFP_HIGHUSER, which includes __GFP_HIGHMEM. Somehow these flags are propagated through arm_iommu_alloc_attrs() up to the slab allocator, which meets this BUG_ON() when it needs to grow its cache.
Note that ttm_dma_pool_init() can also be called with GFP_DMA32, which will trigger the same error.
So although I am still not sure how this should be fixed, it seems like this has nothing to do with the change I am proposing.
Temporary workaround:
--- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -882,9 +882,9 @@ int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
type = ttm_to_type(ttm->page_flags, ttm->caching_state); if (ttm->page_flags & TTM_PAGE_FLAG_DMA32) - gfp_flags = GFP_USER | GFP_DMA32; + gfp_flags = GFP_USER; else - gfp_flags = GFP_HIGHUSER; + gfp_flags = GFP_USER; if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC) gfp_flags |= __GFP_ZERO;
... which obviously doesn't look right.
On 12/08/2014 05:11 PM, Alexandre Courbot wrote:
On 12/08/2014 04:14 PM, Alexandre Courbot wrote:
On 11/12/2014 09:39 PM, Thierry Reding wrote:
From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
This commit fixes this by using phys_to_page() since we get a physical address from dma_alloc_coherent(). Note that this is not a proper fix because if an IOMMU is set up to translate addresses for the GPU this address will be an I/O virtual address rather than a physical one. The proper fix probably involves not getting a pointer to the struct page in the first place, but that would be a much more intrusive change, if at all possible.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Signed-off-by: Thierry Reding treding@nvidia.com
Arnd, I realize that this isn't a proper fix according to what we discussed on IRC yesterday, but I can't see a way to remove access to the pages array that would be as simple as this. I've marked this as RFC in the hope that it will trigger some discussion that will lead to a proper solution.
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c index c96db433f8af..d7993985752c 100644 --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -343,7 +343,11 @@ static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) &d_page->dma, pool->gfp_flags); if (d_page->vaddr) +#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);+#else d_page->p = virt_to_page(d_page->vaddr); +#endif
Since I am messing with the IOMMU I just happened to have hit the issue you are mentioning. Wouldn't the following work:
if (d_page->vaddr)-#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);-#else
d_page->p = virt_to_page(d_page->vaddr);-#endif
else {
if (d_page->vaddr) {if (is_vmalloc_addr(d_page->vaddr)) {d_page->p = vmalloc_to_page(d_page->vaddr);} else {d_page->p = virt_to_page(d_page->vaddr);}} else {A remapped page will end up in the vmalloc range of the address space, and in this case we can use vmalloc_to_page() to get the right page. Pages outside of this range are part of the linear mapping and can be resolved using virt_to_page().
Jetson seems to be mostly happy with this, although I sometimes get the following trace:
[ 13.174763] kernel BUG at ../mm/slab.c:2593! [ 13.174767] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP ARM [ 13.174790] Modules linked in: nouveau_platform(O+) nouveau(O) cfbfillrect cfbimgblt cfbcopyarea ttm ... [ 13.175234] [<c00de238>] (cache_alloc_refill) from [<c00de528>] (__kmalloc+0x100/0x13c) [ 13.175247] [<c00de528>] (__kmalloc) from [<c001d564>] (arm_iommu_alloc_attrs+0x94/0x3a8) [ 13.175269] [<c001d564>] (arm_iommu_alloc_attrs) from [<bf008f4c>] (ttm_dma_populate+0x498/0x76c [ttm]) [ 13.175294] [<bf008f4c>] (ttm_dma_populate [ttm]) from [<bf000bb8>] (ttm_tt_bind+0x38/0x68 [ttm]) [ 13.175315] [<bf000bb8>] (ttm_tt_bind [ttm]) from [<bf00298c>] (ttm_bo_handle_move_mem+0x408/0x47c [ttm]) [ 13.175337] [<bf00298c>] (ttm_bo_handle_move_mem [ttm]) from [<bf003758>] (ttm_bo_validate+0x220/0x22c [ttm]) [ 13.175359] [<bf003758>] (ttm_bo_validate [ttm]) from [<bf003984>] (ttm_bo_init+0x220/0x338 [ttm]) [ 13.175480] [<bf003984>] (ttm_bo_init [ttm]) from [<bf0c70a0>] (nouveau_bo_new+0x1c0/0x294 [nouveau]) [ 13.175688] [<bf0c70a0>] (nouveau_bo_new [nouveau]) from [<bf0ce88c>] (nv84_fence_create+0x1cc/0x240 [nouveau]) [ 13.175891] [<bf0ce88c>] (nv84_fence_create [nouveau]) from [<bf0cec90>] (nvc0_fence_create+0xc/0x24 [nouveau]) [ 13.176094] [<bf0cec90>] (nvc0_fence_create [nouveau]) from [<bf0c1480>] (nouveau_accel_init+0xec/0x450 [nouveau])
I suspect this is related to this change, but it might also be the side-effect of another bug in my code.
Ok, I finally understand where this trace comes from.
slab.c:2593: BUG_ON(flags & GFP_SLAB_BUG_MASK);
gfp.h: #define GFP_SLAB_BUG_MASK (__GFP_DMA32|__GFP_HIGHMEM|~__GFP_BITS_MASK)
and ttm_dma_populate() calls ttm_dma_pool_init() with gfp_flags == GFP_HIGHUSER, which includes __GFP_HIGHMEM. Somehow these flags are propagated through arm_iommu_alloc_attrs() up to the slab allocator, which meets this BUG_ON() when it needs to grow its cache.
Note that ttm_dma_pool_init() can also be called with GFP_DMA32, which will trigger the same error.
So although I am still not sure how this should be fixed, it seems like this has nothing to do with the change I am proposing.
Temporary workaround:
--- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -882,9 +882,9 @@ int ttm_dma_populate(struct ttm_dma_tt *ttm_dma, struct device *dev)
type = ttm_to_type(ttm->page_flags, ttm->caching_state); if (ttm->page_flags & TTM_PAGE_FLAG_DMA32)
gfp_flags = GFP_USER | GFP_DMA32;
gfp_flags = GFP_USER; else
gfp_flags = GFP_HIGHUSER;
gfp_flags = GFP_USER; if (ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC) gfp_flags |= __GFP_ZERO;... which obviously doesn't look right.
This, however, looks better:
diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c index e8907117861e..bc495354c802 100644 --- a/arch/arm/mm/dma-mapping.c +++ b/arch/arm/mm/dma-mapping.c @@ -1106,7 +1106,7 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size, int i = 0;
if (array_size <= PAGE_SIZE) - pages = kzalloc(array_size, gfp); + pages = kzalloc(array_size, GFP_KERNEL); else pages = vzalloc(array_size); if (!pages)
I don't see any reason why the array of pages should be allocated with the same properties as the buffer itself. Looks like this is a DMA allocator bug. Sorry about the noise.
Alex.
On Mon, Dec 8, 2014 at 4:14 PM, Alexandre Courbot acourbot@nvidia.com wrote:
On 11/12/2014 09:39 PM, Thierry Reding wrote:
From: Thierry Reding treding@nvidia.com
dma_alloc_coherent() returns a kernel virtual address that is part of the linear range. Passing such an address to virt_to_page() is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained from virt_to_page() points to unmapped memory.
This commit fixes this by using phys_to_page() since we get a physical address from dma_alloc_coherent(). Note that this is not a proper fix because if an IOMMU is set up to translate addresses for the GPU this address will be an I/O virtual address rather than a physical one. The proper fix probably involves not getting a pointer to the struct page in the first place, but that would be a much more intrusive change, if at all possible.
Until that time, this temporary fix will allow TTM to work on 32-bit and 64-bit ARM as well, provided that no IOMMU translations are enabled for the GPU.
Signed-off-by: Thierry Reding treding@nvidia.com
Arnd, I realize that this isn't a proper fix according to what we discussed on IRC yesterday, but I can't see a way to remove access to the pages array that would be as simple as this. I've marked this as RFC in the hope that it will trigger some discussion that will lead to a proper solution.
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c index c96db433f8af..d7993985752c 100644 --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c @@ -343,7 +343,11 @@ static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) &d_page->dma, pool->gfp_flags); if (d_page->vaddr) +#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);+#else d_page->p = virt_to_page(d_page->vaddr); +#endif
Since I am messing with the IOMMU I just happened to have hit the issue you are mentioning. Wouldn't the following work:
if (d_page->vaddr)-#if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
d_page->p = phys_to_page(d_page->dma);-#else
d_page->p = virt_to_page(d_page->vaddr);-#endif
else {
if (d_page->vaddr) {if (is_vmalloc_addr(d_page->vaddr)) {d_page->p = vmalloc_to_page(d_page->vaddr);} else {d_page->p = virt_to_page(d_page->vaddr);}} else {A remapped page will end up in the vmalloc range of the address space, and in this case we can use vmalloc_to_page() to get the right page. Pages outside of this range are part of the linear mapping and can be resolved using virt_to_page().
Thierry, have you had a chance to try this? If not, do you want to to try to push this patch? It seems to solve the issue AFAICT, but needs more testing.
dri-devel@lists.freedesktop.org
-
Alexandre Courbot -
Alexandre Courbot -
Arnd Bergmann -
Konrad Rzeszutek Wilk -
Thierry Reding