AMDGPU / RADEON / DRM Fix mapping of user pages
Subsystems (3)
v4 cover AMDGPU / RADEON / DRM Fix mapping of user pages
This patch set is to fix a bug in amdgpu / radeon drm that results in a crash when dma_map_sg combines elemnets within a scatterlist table.
There are 2 shortfalls in the current kernel.
1) AMDGPU / RADEON assumes that the requested and created scatterlist table lengths using from dma_map_sg are equal. This may not be the case using the newer dma-iommu implementation
2) drm_prime does not fetch the length of the scatterlist via the correct dma macro, this can use the incorrect length being used (>0) in places where dma_map_sg has updated the table elements.
The sg_dma_len macro is representative of the length of the sg item after dma_map_sg
Example Crash :
Quoted 1 line
[drm:amdgpu_ttm_backend_bind [amdgpu]] *ERROR* failed to pin userptr
This happens in OpenCL applications, causing them to crash or hang, on either amdgpu-pro or ROCm OpenCL implementations
I have verified this fixes the above on kernel 5.5 and 5.5rc using an AMD Vega 64 GPU
Shane Francis (3):
drm/prime: use dma length macro when mapping sg to arrays drm/amdgpu: fix scatter-gather mapping with user pages drm/radeon: fix scatter-gather mapping with user pages
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 2 +- drivers/gpu/drm/drm_prime.c | 2 +- drivers/gpu/drm/radeon/radeon_ttm.c | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-)
Signature · 6 lines
2.26.0 _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
PATCH v4 1/3 drm/prime: use dma length macro when mapping sg
As dma_map_sg can reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use the sg_dma_len macro to fetch the actual length.
This could now be 0 and not need to be mapped to a page or address array
--- drivers/gpu/drm/drm_prime.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/drm_prime.c +1 -1
@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index =0 ;for_each_sg (sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len =sg_dma_len (sg); page =sg_page (sg); addr =sg_dma_address (sg);
Signature · 6 lines
2.26.0 _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
RE: [PATCH v4 1/3] drm/prime: use dma length macro when mapping sg
Quoted 26 lines
-----Original Message-----
From: dri-devel <[email protected]> On Behalf Of Shane Francis
Sent: Wednesday, March 25, 2020 5:08 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Subject: [PATCH v4 1/3] drm/prime: use dma length macro when mapping sgAs
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>---
drivers/gpu/drm/drm_prime.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/drm_prime.c +0
-0
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct · line 973
Quoted 8 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index =0 ;for_each_sg (sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len =sg_dma_len (sg); page =sg_page (sg); addr =sg_dma_address (sg);
This looks correct to me.
M
Quoted 8 lines
--
2.26.0_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
In reply to Shane FrancisHi All,
Quoted 9 lines · Shane Francis
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>
This patch landed in linux-next 20200326 and it causes a kernel panic on various Exynos SoC based boards.
Quoted 9 lines
--- drivers/gpu/drm/drm_prime.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/drm_prime.c +0
-0
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, · line 973
Quoted 8 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index =0 ;for_each_sg (sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len =sg_dma_len (sg); page =sg_page (sg); addr =sg_dma_address (sg);
Sorry, but this code is wrong :(
The scatterlist elements (sg) describes memory chunks in physical memory and in the DMA (IO virtual) space. However in general, you cannot assume 1:1 mapping between them. If you access sg_page(sg) (basically sg->page), you must match it with sg->length. When you access sg_dma_address(sg) (again, in most cases it is sg->dma_address), then you must match it with sg_dma_len(sg). The sg->dma_address might not be the dma address of the sg->page.
In some cases (when IOMMU is available, it performs aggregation of the scatterlist chunks and a few other, minor requirements), the whole scatterlist might be mapped into contiguous DMA address space and filled only to the first sg element.
The proper way to iterate over a scatterlists to get both the pages and the DMA addresses assigned to them is:
int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages,
dma_addr_t *addrs, int max_entries)
{
unsigned count;
struct scatterlist *sg;
struct page *page;
u32 page_len, page_index;
dma_addr_t addr;
u32 dma_len, dma_index;
page_index = 0;
dma_index = 0;
for_each_sg(sgt->sgl, sg, sgt->nents, count) { page_len = sg->length;
page = sg_page(sg);
dma_len = sg_dma_len(sg); addr = sg_dma_address(sg);
while (pages && page_len > 0) {
if (WARN_ON(page_index >= max_entries)) return -1; pages[page_index] = page; page++;
page_len -= PAGE_SIZE; page_index++;
}
while (addrs && dma_len > 0) {
if (WARN_ON(dma_index >= max_entries)) return -1; addrs[dma_index] = addr; addr += PAGE_SIZE;
dma_len -= PAGE_SIZE; dma_index++;
}
}
return 0;
}
I will send a patch in a few minutes with the above fixed code.
Best regards
Signature · 7 lines
Marek Szyprowski, PhD Samsung R&D Institute Poland _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
Christian König
maintainerRe: [v4,1/3] drm/prime: use dma length macro when mapping sg
Am 27.03.20 um 08:54 schrieb Marek Szyprowski:
Quoted 24 lines
Hi All,
Quoted 9 lines · Shane Francis
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>This patch landed in linux-next 20200326 and it causes a kernel panic on various Exynos SoC based boards.
Quoted 10 lines
--- drivers/gpu/drm/drm_prime.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/drm_prime.c +0
-0@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages,
Quoted 8 lines
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, · line 973
Quoted 7 lines
drivers/gpu/drm/drm_prime.c +1
-1index =0 ;for_each_sg (sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len =sg_dma_len (sg); page =sg_page (sg); addr =sg_dma_address (sg);Sorry, but this code is wrong :(
Well it is at least better than before because it makes most drivers work correctly again.
See we only fill the pages array because some drivers (like Exynos) are still buggy and require this.
Accessing the pages of an DMA-buf imported sg_table is illegal and should be fixed in the drivers.
Quoted 3 lines
[SNIP]
I will send a patch in a few minutes with the above fixed code.
That is certainly a good idea for now, but could we also put on somebodies todo list an item to fix Exynos?
Thanks in advance,
Christian.
Quoted 2 lines
Best regards
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
Hi Christian,
Quoted 31 lines · Christian König
Am 27.03.20 um 08:54 schrieb Marek Szyprowski:
Quoted 16 lines · Shane Francis
Quoted 9 lines
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>This patch landed in linux-next 20200326 and it causes a kernel panic on various Exynos SoC based boards.
Quoted 4 lines
---
drivers/gpu/drm/drm_prime.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct · line 972
Quoted 12 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index = 0; for_each_sg(sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len = sg_dma_len(sg); page = sg_page(sg); addr = sg_dma_address(sg);Sorry, but this code is wrong :(
Quoted 2 lines
Well it is at least better than before because it makes most drivers work correctly again.
Well, I'm not sure that a half-broken fix should be considered as a fix ;)
Anyway, I just got the comment from Shane, that my patch is fixing the issues with amdgpu and radeon, while still working fine for exynos, so it is indeed a proper fix.
Quoted 2 lines
See we only fill the pages array because some drivers (like Exynos) are still buggy and require this.
Exynos driver use this pages array internally.
Quoted 3 lines
Accessing the pages of an DMA-buf imported
sg_tableis illegal and should be fixed in the drivers.
True, but in meantime we should avoid breaking stuff which worked fine for ages.
Quoted 7 lines
Quoted 4 lines
[SNIP]
I will send a patch in a few minutes with the above fixed code.
That is certainly a good idea for now, but could we also put on somebodies todo list an item to fix Exynos?
Yes, I can take a look into removing the use of the internal pages array. It is used mainly for implementing vmap for fbdev emulation, but this can be handled in a different way.
Best regards
Signature · 7 lines
Marek Szyprowski, PhD Samsung R&D Institute Poland _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
Hi
Quoted 38 lines · Marek Szyprowski
Hi Christian,
Quoted 32 lines · Christian König
Am 27.03.20 um 08:54 schrieb Marek Szyprowski:
Quoted 17 lines · Shane Francis
Quoted 9 lines
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>This patch landed in linux-next 20200326 and it causes a kernel panic on
various Exynos SoC based boards.Quoted 4 lines
---
drivers/gpu/drm/drm_prime.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct · line 972
Quoted 12 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index = 0; for_each_sg(sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len = sg_dma_len(sg); page = sg_page(sg); addr = sg_dma_address(sg);Sorry, but this code is wrong :(
Quoted 2 lines
Well it is at least better than before because it makes most drivers work correctly again.
Quoted 6 lines
Well, I'm not sure that a half-broken fix should be considered as a fix ;)
Anyway, I just got the comment from Shane, that my patch is fixing the issues with amdgpu and radeon, while still working fine for exynos, so it is indeed a proper fix.
Today I've noticed that this patch went to final v5.6 without even a day of testing in linux-next, so v5.6 is broken on Exynos and probably a few other ARM archs, which rely on the drm_prime_sg_to_page_addr_arrays function.
Best regards
Signature · 7 lines
Marek Szyprowski, PhD Samsung R&D Institute Poland _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
Quoted 10 lines · Marek Szyprowski
Today I've noticed that this patch went to final v5.6 without even a day of testing in linux-next, so v5.6 is broken on Exynos and probably a few other ARM archs, which rely on the
drm_prime_sg_to_page_addr_arraysfunction.Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
Not sure what the full merge pipeline is here, but my original patch was not sent to the stable mailing list. So I would assume that it went through the normal release gates / checks ? If there was a fault on the way I uploaded the patches I do apologise however I did follow the normal guidelines as far as I understood them.
Personally I did validate this patch on systems with Intel and AMD GFX, unfortunately I do not have any ARM based dev kits that are able to run mainline (was never able to get an Nvidia TK1 to boot outside of L4T)
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
In reply to Marek SzyprowskiQuoted 10 lines · Marek Szyprowski
Today I've noticed that this patch went to final v5.6 without even a day of testing in linux-next, so v5.6 is broken on Exynos and probably a few other ARM archs, which rely on the
drm_prime_sg_to_page_addr_arraysfunction.Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
In reply to Marek SzyprowskiQuoted 10 lines · Marek Szyprowski
Today I've noticed that this patch went to final v5.6 without even a day of testing in linux-next, so v5.6 is broken on Exynos and probably a few other ARM archs, which rely on the
drm_prime_sg_to_page_addr_arraysfunction.Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
FYI These changes are now in the 5.5-stable queue.
_______________________________________________
dri-devel mailing list [email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
In reply to Marek SzyprowskiQuoted 45 lines · Marek Szyprowski
Hi
Quoted 41 lines · Marek Szyprowski
Hi Christian,
Quoted 32 lines · Christian König
Am 27.03.20 um 08:54 schrieb Marek Szyprowski:
Quoted 17 lines · Shane Francis
Quoted 9 lines
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>This patch landed in linux-next 20200326 and it causes a kernel panic on
various Exynos SoC based boards.Quoted 4 lines
--- drivers/gpu/drm/drm_prime.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct · line 972
Quoted 12 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index =0 ;for_each_sg (sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len =sg_dma_len (sg); page =sg_page (sg); addr =sg_dma_address (sg);Sorry, but this code is wrong :(
Quoted 2 lines
Well it is at least better than before because it makes most drivers work correctly again.
Well, I'm not sure that a half-broken fix should be considered as a fix ;)
Quoted 8 lines
Quoted 3 lines
Anyway, I just got the comment from Shane, that my patch is fixing the issues with amdgpu and radeon, while still working fine for exynos, so it is indeed a proper fix.
Today I've noticed that this patch went to final v5.6 without even a day of testing in linux-next, so v5.6 is broken on Exynos and probably a few other ARM archs, which rely on the
drm_prime_sg_to_page_addr_arraysfunction.
Please commit your patch and cc stable.
Alex
Quoted 10 lines
Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
Hi Alex,
Quoted 44 lines · Alex Deucher
Quoted 42 lines · Marek Szyprowski
Hi
Quoted 39 lines · Marek Szyprowski
Hi Christian,
Quoted 31 lines · Christian König
Am 27.03.20 um 08:54 schrieb Marek Szyprowski:
Quoted 17 lines · Shane Francis
Quoted 9 lines
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>This patch landed in linux-next 20200326 and it causes a kernel panic on
various Exynos SoC based boards.Quoted 4 lines
--- drivers/gpu/drm/drm_prime.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct · line 972
Quoted 12 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index =0 ;for_each_sg (sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len =sg_dma_len (sg); page =sg_page (sg); addr =sg_dma_address (sg);Sorry, but this code is wrong :(
Quoted 2 lines
Well it is at least better than before because it makes most drivers work correctly again.
Well, I'm not sure that a half-broken fix should be considered as a fix ;)
Quoted 8 lines
Quoted 7 lines
Quoted 3 lines
Anyway, I just got the comment from Shane, that my patch is fixing the issues with amdgpu and radeon, while still working fine for exynos, so it is indeed a proper fix.
Today I've noticed that this patch went to final v5.6 without even a day of testing in linux-next, so v5.6 is broken on Exynos and probably a few other ARM archs, which rely on the
drm_prime_sg_to_page_addr_arraysfunction.Please commit your patch and cc stable.
I've already did that: https://lkml.org/lkml/2020/3/27/555
Best regards
Signature · 7 lines
Marek Szyprowski, PhD Samsung R&D Institute Poland _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
Quoted 51 lines · Marek Szyprowski
Hi Alex,
Quoted 47 lines · Alex Deucher
Quoted 45 lines · Marek Szyprowski
Hi
Quoted 42 lines · Marek Szyprowski
Hi Christian,
Quoted 31 lines · Christian König
Am 27.03.20 um 08:54 schrieb Marek Szyprowski:
Quoted 17 lines · Shane Francis
Quoted 9 lines
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>This patch landed in linux-next 20200326 and it causes a kernel panic on
various Exynos SoC based boards.Quoted 4 lines
--- drivers/gpu/drm/drm_prime.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct · line 972
Quoted 12 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index =0 ;for_each_sg (sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len =sg_dma_len (sg); page =sg_page (sg); addr =sg_dma_address (sg);Sorry, but this code is wrong :(
Quoted 2 lines
Well it is at least better than before because it makes most drivers work correctly again.
Well, I'm not sure that a half-broken fix should be considered as a fix ;)
Anyway, I just got the comment from Shane, that my patch is fixing the issues with amdgpu and radeon, while still working fine for exynos, so it is indeed a proper fix.
Quoted 7 lines
Quoted 5 lines
Quoted 4 lines
Today I've noticed that this patch went to final v5.6 without even a day of testing in linux-next, so v5.6 is broken on Exynos and probably a few other ARM archs, which rely on the
drm_prime_sg_to_page_addr_arraysfunction.Please commit your patch and cc stable.
I've already did that: https://lkml.org/lkml/2020/3/27/555
Do you have drm-misc commit rights or do you need someone to commit this for you?
Alex
Quoted 6 lines
Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
Hi Alex,
Quoted 52 lines · Alex Deucher
Quoted 50 lines · Marek Szyprowski
Hi Alex,
Quoted 47 lines · Alex Deucher
Quoted 45 lines · Marek Szyprowski
Hi
Quoted 42 lines · Marek Szyprowski
Hi Christian,
Quoted 31 lines · Christian König
Am 27.03.20 um 08:54 schrieb Marek Szyprowski:
Quoted 17 lines · Shane Francis
Quoted 9 lines
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>This patch landed in linux-next 20200326 and it causes a kernel panic on
various Exynos SoC based boards.Quoted 4 lines
--- drivers/gpu/drm/drm_prime.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct · line 972
Quoted 12 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index =0 ;for_each_sg (sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len =sg_dma_len (sg); page =sg_page (sg); addr =sg_dma_address (sg);Sorry, but this code is wrong :(
Quoted 2 lines
Well it is at least better than before because it makes most drivers work correctly again.
Well, I'm not sure that a half-broken fix should be considered as a fix ;)
Anyway, I just got the comment from Shane, that my patch is fixing the issues with amdgpu and radeon, while still working fine for exynos, so it is indeed a proper fix.
Quoted 8 lines
Quoted 6 lines
Quoted 5 lines
Quoted 4 lines
Today I've noticed that this patch went to final v5.6 without even a day of testing in linux-next, so v5.6 is broken on Exynos and probably a few other ARM archs, which rely on the
drm_prime_sg_to_page_addr_arraysfunction.Please commit your patch and cc stable.
I've already did that: https%3A%2F%2Flkml.org%2Flkml%2F2020%2F3%2F27%2F555
Do you have drm-misc commit rights or do you need someone to commit this for you?
I have no access to drm-misc.
Best regards
Signature · 7 lines
Marek Szyprowski, PhD Samsung R&D Institute Poland _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [v4,1/3] drm/prime: use dma length macro when mapping sg
In reply to Marek SzyprowskiQuoted 85 lines · Marek Szyprowski
Quoted 9 lines · Shane Francis
As
dma_map_sgcan reorganize scatter-gather lists in a way that can cause some later segments to be empty we should always use thesg_dma_lenmacro to fetch the actual length.This could now be 0 and not need to be mapped to a page or address array
Signed-off-by Shane Francis <[email protected]>Reviewed-by Michael J. Ruhl <[email protected]>This patch landed in linux-next 20200326 and it causes a kernel panic on various Exynos SoC based boards.
Quoted 9 lines
---
drivers/gpu/drm/drm_prime.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/drm_prime.c +0
-0drivers/gpu/drm/drm_prime.c int drm_prime_sg_to_page_addr_arrays(struct · line 972
Quoted 7 lines
drivers/gpu/drm/drm_prime.c +1
-1@@ -967,7 +967,7 @@ int drm_prime_sg_to_page_addr_arrays(struct sg_table *sgt, struct page **pages, index = 0; for_each_sg(sgt->sgl, sg, sgt->nents, count) { - len = sg->length; + len = sg_dma_len(sg); page = sg_page(sg); addr = sg_dma_address(sg);Sorry, but this code is wrong :(
The scatterlist elements (sg) describes memory chunks in physical memory and in the DMA (IO virtual) space. However in general, you cannot assume 1:1 mapping between them. If you access
sg_page(sg)(basicallysg->page), you must match it withsg->length. When you accesssg_dma_address(sg)(again, in most cases it issg->dma_address), then you must match it withsg_dma_len(sg). Thesg->dma_addressmight not be the dma address of thesg->page.In some cases (when IOMMU is available, it performs aggregation of the scatterlist chunks and a few other, minor requirements), the whole scatterlist might be mapped into contiguous DMA address space and filled only to the first sg element.
The proper way to iterate over a scatterlists to get both the pages and the DMA addresses assigned to them is:
int drm_prime_sg_to_page_addr_arrays(struct
sg_table*sgt, struct page **pages,
dma_addr_t*addrs, intmax_entries) {
unsigned count;
struct scatterlist *sg;
struct page *page;
u32page_len,page_index;
dma_addr_taddr;
u32dma_len,dma_index;
page_index= 0;
dma_index= 0;
for_each_sg(sgt->sgl, sg, sgt->nents, count){page_len=sg->length;
page =sg_page(sg);
dma_len=sg_dma_len(sg); addr =sg_dma_address(sg);while (pages &&
page_len> 0) {
if (WARN_ON(page_index >= max_entries)) return -1; pages[page_index] = page; page++;
page_len-=PAGE_SIZE; page_index++;
}while (addrs &&
dma_len> 0) {
if (WARN_ON(dma_index >= max_entries)) return -1; addrs[dma_index] = addr; addr +=PAGE_SIZE;
dma_len-=PAGE_SIZE; dma_index++;
}
}
Quoted 4 lines
return 0;
}I will send a patch in a few minutes with the above fixed code.
Here is the fix: https://patchwork.freedesktop.org/patch/359081/
Best regards
Signature · 7 lines
Marek Szyprowski, PhD Samsung R&D Institute Poland _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
PATCH v4 2/3 drm/amdgpu: fix scatter-gather mapping with user pages
In reply to Shane FrancisCalls to dma_map_sg may return segments / entries than requested if they fall on page bounderies. The old implementation did not support this use case.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1 -1
@@ -974,7 +974,7 @@ static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)/* Map SG to device */ r = -ENOMEM ; nents =dma_map_sg (adev->dev, ttm->sg->sgl, ttm->sg->nents, direction); -if (nents != ttm->sg->nents) +if (nents ==0 ) goto release_sg;/* convert SG to linear array of pages and dma addresses */
Signature · 1 line
2.26.0
RE: [PATCH v4 2/3] drm/amdgpu: fix scatter-gather mapping with user pages
Quoted 11 lines
-----Original Message-----
From: dri-devel <[email protected]> On Behalf Of Shane Francis
Sent: Wednesday, March 25, 2020 5:08 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Subject: [PATCH v4 2/3] drm/amdgpu: fix scatter-gather mapping with user pages
Quoted 1 line
Calls to
dma_map_sgmay return segments / entries than requested
"may return less segments/entries" ?
^^^
Quoted 14 lines
if they fall on page bounderies. The old implementation did not support this use case.
Signed-off-by Shane Francis <[email protected]>---
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +0
-0
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt · line 980
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +1
-1@@ -974,7 +974,7 @@ static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)/* Map SG to device */ r = -ENOMEM ; nents =dma_map_sg (adev->dev, ttm->sg->sgl, ttm->sg->nents, direction); -if (nents != ttm->sg->nents) +if (nents ==0 ) goto release_sg;
this looks correct to me.
Quoted 8 lines
/* convert SG to linear array of pages and dma addresses */--
2.26.0_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
PATCH v4 3/3 drm/radeon: fix scatter-gather mapping with user pages
In reply to Shane FrancisCalls to dma_map_sg may return segments / entries than requested if they fall on page bounderies. The old implementation did not support this use case.
--- drivers/gpu/drm/radeon/radeon_ttm.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/radeon/radeon_ttm.c +1 -1
@@ -528,7 +528,7 @@ static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm) r = -ENOMEM ; nents =dma_map_sg (rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction); -if (nents != ttm->sg->nents) +if (nents ==0 ) goto release_sg; drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
Signature · 6 lines
2.26.0 _______________________________________________ dri-devel mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/dri-devel
RE: [PATCH v4 3/3] drm/radeon: fix scatter-gather mapping with user pages
Quoted 11 lines
-----Original Message-----
From: dri-devel <[email protected]> On Behalf Of Shane Francis
Sent: Wednesday, March 25, 2020 5:08 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Subject: [PATCH v4 3/3] drm/radeon: fix scatter-gather mapping with user pages
Quoted 1 line
Calls to
dma_map_sgmay return segments / entries than requested
"may return less segment..." ?
^^^
Quoted 14 lines
if they fall on page bounderies. The old implementation did not support this use case.
Signed-off-by Shane Francis <[email protected]>---
drivers/gpu/drm/radeon/radeon_ttm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/radeon/radeon_ttm.c +0
-0
drivers/gpu/drm/radeon/radeon_ttm.c static int radeon_ttm_tt_pin_userptr(struct ttm_tt · line 534
Quoted 8 lines
drivers/gpu/drm/radeon/radeon_ttm.c +1
-1@@ -528,7 +528,7 @@ static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm) r = -ENOMEM ; nents =dma_map_sg (rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction); -if (nents != ttm->sg->nents) +if (nents ==0 ) goto release_sg;
This looks correct to me.
M
Quoted 8 lines
drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,--
2.26.0_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH v4 3/3] drm/radeon: fix scatter-gather mapping with user pages
Quoted 11 lines
Quoted 11 lines
-----Original Message-----
From: dri-devel <[email protected]> On Behalf Of Shane Francis
Sent: Wednesday, March 25, 2020 5:08 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Subject: [PATCH v4 3/3] drm/radeon: fix scatter-gather mapping with user pages
Quoted 4 lines
Quoted 1 line
Calls to
dma_map_sgmay return segments / entries than requested"may return less segment..." ?
^^^
I will reword / fix the highlighted issues with the text and send a updated patch set later today.
Quoted 26 lines
Quoted 15 lines
if they fall on page bounderies. The old implementation did not support this use case.
Signed-off-by Shane Francis <[email protected]>---
drivers/gpu/drm/radeon/radeon_ttm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/radeon/radeon_ttm.c +0
-0drivers/gpu/drm/radeon/radeon_ttm.c static int radeon_ttm_tt_pin_userptr(struct ttm_tt · line 534
Quoted 8 lines
drivers/gpu/drm/radeon/radeon_ttm.c +1
-1@@ -528,7 +528,7 @@ static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm) r = -ENOMEM ; nents =dma_map_sg (rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction); -if (nents != ttm->sg->nents) +if (nents ==0 ) goto release_sg;This looks correct to me.
Quoted 11 lines
Reviewed-by Michael J. Ruhl <[email protected]>M
Quoted 8 lines
drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,--
2.26.0_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH v4 3/3] drm/radeon: fix scatter-gather mapping with user pages
Quoted 12 lines · Shane Francis
Quoted 12 lines
Quoted 12 lines
-----Original Message-----
From: dri-devel <[email protected]> On Behalf Of Shane Francis
Sent: Wednesday, March 25, 2020 5:08 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Subject: [PATCH v4 3/3] drm/radeon: fix scatter-gather mapping with user pages
Quoted 7 lines
Quoted 4 lines
Quoted 1 line
Calls to
dma_map_sgmay return segments / entries than requested"may return less segment..." ?
^^^I will reword / fix the highlighted issues with the text and send a updated patch set later today.
I'll fix it up locally when I apply it. Thanks!
Alex
Quoted 34 lines
Quoted 34 lines
Quoted 16 lines
if they fall on page bounderies. The old implementation did not support this use case.
Signed-off-by Shane Francis <[email protected]>---
drivers/gpu/drm/radeon/radeon_ttm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/radeon/radeon_ttm.c +0
-0drivers/gpu/drm/radeon/radeon_ttm.c static int radeon_ttm_tt_pin_userptr(struct ttm_tt · line 534
Quoted 8 lines
drivers/gpu/drm/radeon/radeon_ttm.c +1
-1@@ -528,7 +528,7 @@ static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm) r = -ENOMEM ; nents =dma_map_sg (rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction); -if (nents != ttm->sg->nents) +if (nents ==0 ) goto release_sg;This looks correct to me.
Reviewed-by Michael J. Ruhl <[email protected]>M
Quoted 4 lines
drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,--
2.26.0
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Quoted 10 lines
-----Original Message-----
From: dri-devel <[email protected]> On Behalf Of Shane Francis
Sent: Wednesday, March 25, 2020 5:08 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Subject: [PATCH v4 0/3] AMDGPU / RADEON / DRM Fix mapping of user pages
Quoted 2 lines
This patch set is to fix a bug in amdgpu / radeon drm that results in a crash when
dma_map_sgcombines elemnets within a scatterlist table.
s/elemnets/elements
Quoted 40 lines
There are 2 shortfalls in the current kernel.
1) AMDGPU / RADEON assumes that the requested and created scatterlist table lengths using from
dma_map_sgare equal. This may not be the case using the newer dma-iommu implementation2)
drm_primedoes not fetch the length of the scatterlist via the correct dma macro, this can use the incorrect length being used (>0) in places wheredma_map_sghas updated the table elements.The
sg_dma_lenmacro is representative of the length of the sg item afterdma_map_sgExample Crash :
Quoted 1 line
[drm:amdgpu_ttm_backend_bind [amdgpu]] *ERROR* failed to pin userptr
This happens in OpenCL applications, causing them to crash or hang, on either amdgpu-pro or ROCm OpenCL implementations
I have verified this fixes the above on kernel 5.5 and 5.5rc using an AMD Vega 64 GPU
Shane Francis (3):
drm/prime: use dma length macro when mapping sg to arrays drm/amdgpu: fix scatter-gather mapping with user pages drm/radeon: fix scatter-gather mapping with user pagesdrivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 2 +- drivers/gpu/drm/drm_prime.c | 2 +- drivers/gpu/drm/radeon/radeon_ttm.c | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-)--
2.26.0_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel