drm/ttm: dma: Fixes for 32-bit and 64-bit ARM
Subsystems (4)
RFC 1/1 drm/ttm: dma: Fixes for 32-bit and 64-bit ARM
From: Thierry Reding <[email protected]>
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.
---
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(+)
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +4 -0
@@ -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 ;
Signature · 6 lines
2.1.3 _______________________________________________ dri-devel mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 7 lines · Thierry Reding
From: Thierry Reding <[email protected]>
dma_alloc_coherent()returns a kernel virtual address that is part of the linear range. Passing such an address tovirt_to_page()is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained fromvirt_to_page()points to unmapped memory.
Oh! That is not good!
Quoted 8 lines
This commit fixes this by using
phys_to_page()since we get a physical address fromdma_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.
Quoted 4 lines
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 :-)
Quoted 30 lines
Signed-off-by Thierry Reding <[email protected]>---
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(+)drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +4
-0@@ -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 ;Signature · 1 line
2.1.3
_______________________________________________
dri-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 11 lines · On Wednesday 12 November 2014 09:18:59 Konrad Rzeszutek Wilk wrote:
Quoted 7 lines · Thierry Reding
From: Thierry Reding <[email protected]>
dma_alloc_coherent()returns a kernel virtual address that is part of the linear range. Passing such an address tovirt_to_page()is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained fromvirt_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.
Quoted 4 lines
Quoted 3 lines
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.
Quoted 8 lines
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
_______________________________________________
dri-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 10 lines · Arnd Bergmann
Quoted 9 lines · On Wednesday 12 November 2014 09:18:59 Konrad Rzeszutek Wilk wrote:
Quoted 7 lines · Thierry Reding
From: Thierry Reding <[email protected]>
dma_alloc_coherent()returns a kernel virtual address that is part of the linear range. Passing such an address tovirt_to_page()is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained fromvirt_to_page()points to unmapped memory.
Quoted 11 lines
Quoted 2 lines
Oh! That is not good!
I think what Thierry meant is that the returned pointer is /not/ in the linear range.
Quoted 6 lines
Quoted 3 lines
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?
?
Quoted 12 lines
Quoted 6 lines
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/...
Quoted 7 lines
- highmem pages on 32-bit system.
On noncoherent ARM systems,
dma_alloc_coherentwill return memory that is was unmapped from the linear range to avoid having both cacheable and noncachable mappings for the same page.Arnd
_______________________________________________
dri-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 35 lines · Thierry Reding
From: Thierry Reding <[email protected]>
dma_alloc_coherent()returns a kernel virtual address that is part of the linear range. Passing such an address tovirt_to_page()is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained fromvirt_to_page()points to unmapped memory.This commit fixes this by using
phys_to_page()since we get a physical address fromdma_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 <[email protected]>---
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(+)drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +0
-0
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool) · line 350
Quoted 8 lines
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +4
-0@@ -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.
_______________________________________________
dri-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 102 lines · Alexandre Courbot
Quoted 39 lines · Thierry Reding
From: Thierry Reding <[email protected]>
dma_alloc_coherent()returns a kernel virtual address that is part of the linear range. Passing such an address tovirt_to_page()is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained fromvirt_to_page()points to unmapped memory.This commit fixes this by using
phys_to_page()since we get a physical address fromdma_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 <[email protected]>---
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(+)drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +0
-0drivers/gpu/drm/ttm/ttm_page_alloc_dma.c static struct dma_page · line 351
Quoted 9 lines
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +4
-0@@ -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:
(quoted diff) +7
-7- 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 usingvirt_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
Quoted 4 lines
[<
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.
_______________________________________________
dri-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 102 lines · Alexandre Courbot
Quoted 39 lines · Thierry Reding
From: Thierry Reding <[email protected]>
dma_alloc_coherent()returns a kernel virtual address that is part of the linear range. Passing such an address tovirt_to_page()is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained fromvirt_to_page()points to unmapped memory.This commit fixes this by using
phys_to_page()since we get a physical address fromdma_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 <[email protected]>---
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(+)drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +0
-0drivers/gpu/drm/ttm/ttm_page_alloc_dma.c static struct dma_page · line 351
Quoted 9 lines
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +4
-0@@ -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:
(quoted diff) +7
-7- 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 usingvirt_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
Quoted 4 lines
[<
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:
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +2 -2
@@ -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.
_______________________________________________
dri-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 144 lines · Alexandre Courbot
Quoted 102 lines · Alexandre Courbot
Quoted 39 lines · Thierry Reding
From: Thierry Reding <[email protected]>
dma_alloc_coherent()returns a kernel virtual address that is part of the linear range. Passing such an address tovirt_to_page()is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained fromvirt_to_page()points to unmapped memory.This commit fixes this by using
phys_to_page()since we get a physical address fromdma_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 <[email protected]>---
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(+)drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +0
-0drivers/gpu/drm/ttm/ttm_page_alloc_dma.c static struct dma_page · line 351
Quoted 9 lines
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +4
-0@@ -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:
(quoted diff) +7
-7- 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 usingvirt_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]) fromQuoted 4 lines
[<
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:
#defineGFP_SLAB_BUG_MASK(__GFP_DMA32|__GFP_HIGHMEM|~__GFP_BITS_MASK)and
ttm_dma_populate()callsttm_dma_pool_init()withgfp_flags==GFP_HIGHUSER, which includes__GFP_HIGHMEM. Somehow these flags are propagated througharm_iommu_alloc_attrs()up to the slab allocator, which meets thisBUG_ON()when it needs to grow its cache.Note that
ttm_dma_pool_init()can also be called withGFP_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:
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +2
-2@@ -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;
Quoted 1 line
... which obviously doesn't look right.
This, however, looks better:
arch/arm/mm/dma-mapping.c +1 -1
@@ -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.
_______________________________________________
dri-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 70 lines · Alexandre Courbot
Quoted 40 lines · Thierry Reding
From: Thierry Reding <[email protected]>
dma_alloc_coherent()returns a kernel virtual address that is part of the linear range. Passing such an address tovirt_to_page()is illegal on non-coherent architectures. This causes the kernel to oops on 64-bit ARM because the struct page * obtained fromvirt_to_page()points to unmapped memory.This commit fixes this by using
phys_to_page()since we get a physical address fromdma_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 <[email protected]>---
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(+)drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +0
-0drivers/gpu/drm/ttm/ttm_page_alloc_dma.c static struct dma_page *__ttm_dma_alloc_page(struct · line 351
Quoted 9 lines
drivers/gpu/drm/ttm/ttm_page_alloc_dma.c +4
-0@@ -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:
(quoted diff) +7
-7- 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 {
Quoted 4 lines
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 usingvirt_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 mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/dri-devel