mm/gup: Use put_user_page*() instead of put_page*()
Subsystems (12)
PATCH 1/1 mm/gup: Use put_user_page*() instead of put_page*()
This patch converts all call sites of get_user_pages to use put_user_page*() instead of put_page*() functions to release reference to gup pinned pages.
This is a bunch of trivial conversions which is a part of an effort by John Hubbard to solve issues with gup pinned pages and filesystem writeback.
The issue is more clearly described in John Hubbard's patch[1] where put_user_page*() functions are introduced.
Currently put_user_page*() simply does put_page but future implementations look to change that once treewide change of put_page callsites to put_user_page*() is finished.
The lwn article describing the issue with gup pinned pages and filesystem writeback [2].
This patch has been tested by building and booting the kernel as I don't have the required hardware to test the device drivers.
I did not modify gpu/drm drivers which use release_pages instead of put_page() to release reference of gup pinned pages as I am not clear whether release_pages and put_page are interchangable.
[1] https://lkml.org/lkml/2019/3/26/1396
[2] https://lwn.net/Articles/784574/
--- drivers/media/v4l2-core/videobuf-dma-sg.c | 3 +-- drivers/misc/sgi-gru/grufault.c | 2 +- drivers/staging/kpc2000/kpc_dma/fileops.c | 4 +--- drivers/vfio/vfio_iommu_type1.c | 2 +- fs/io_uring.c | 7 +++---- mm/gup_benchmark.c | 6 +----- net/xdp/xdp_umem.c | 7 +------ 7 files changed, 9 insertions(+), 22 deletions(-)
drivers/media/v4l2-core/videobuf-dma-sg.c +1 -2
@@ -349,8 +349,7 @@ int videobuf_dma_free(struct videobuf_dmabuf *dma)BUG_ON (dma->sglen);if (dma->pages) { -for (i =0 ; i < dma->nr_pages; i++) -put_page (dma->pages[i]); +put_user_pages (dma->pages, dma->nr_pages);kfree (dma->pages); dma->pages =NULL ; }
drivers/misc/sgi-gru/grufault.c +1 -1
@@ -188,7 +188,7 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,if (get_user_pages (vaddr,1 , write ?FOLL_WRITE :0 , &page,NULL ) <=0 )return -EFAULT ; *paddr =page_to_phys (page); -put_page (page); +put_user_page (page);return 0 ; }
drivers/staging/kpc2000/kpc_dma/fileops.c +1 -3
@@ -198,9 +198,7 @@ int kpc_dma_transfer(struct dev_private_data *priv, struct kiocb *kcb, unsignedsg_free_table (&acd->sgt); err_dma_map_sg: err_alloc_sg_table: -for (i =0 ; i < acd->page_count ; i++){ -put_page (acd->user_pages[i]); - } +put_user_pages (acd->user_pages, acd->page_count); err_get_user_pages:kfree (acd->user_pages); err_alloc_userpages:
drivers/vfio/vfio_iommu_type1.c +1 -1
@@ -369,7 +369,7 @@ static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,*/ if (ret >0 &&vma_is_fsdax (vmas[0 ])) { ret = -EOPNOTSUPP ; -put_page (page[0 ]); +put_user_page (page[0 ]); } }up_read (&mm->mmap_sem);
fs/io_uring.c +3 -4
@@ -2694,10 +2694,9 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,* if we did partial map, or found file backed vmas, * release any pages we did get */ -if (pret >0 ) { -for (j =0 ; j < pret; j++) -put_page (pages[j]); - } +if (pret >0 ) +put_user_pages (pages, pret); +if (ctx->account_mem)io_unaccount_mem (ctx->user, nr_pages);kvfree (imu->bvec);
mm/gup_benchmark.c +1 -5
@@ -76,11 +76,7 @@ static int __gup_benchmark_ioctl(unsigned int cmd, gup->size = addr - gup->addr; start_time =ktime_get (); -for (i =0 ; i < nr_pages; i++) { -if (!pages[i]) -break ; -put_page (pages[i]); - } +put_user_pages (pages, nr_pages); end_time =ktime_get (); gup->put_delta_usec =ktime_us_delta (end_time, start_time);
net/xdp/xdp_umem.c +1 -6
@@ -173,12 +173,7 @@ static void xdp_umem_unpin_pages(struct xdp_umem *umem) {unsigned int i; -for (i =0 ; i < umem->npgs; i++) { -struct page *page = umem->pgs[i]; - -set_page_dirty_lock (page); -put_page (page); - } +put_user_pages_dirty_lock (umem->pgs, umem->npgs);kfree (umem->pgs); umem->pgs =NULL ;
Signature · 1 line
1.8.3.1
Quoted 3 lines · Bharath Vedartham
This patch converts all call sites of
get_user_pagesto use put_user_page*() instead of put_page*() functions to release reference to gup pinned pages.
Hi Bharath,
Thanks for jumping in to help, and welcome to the party!
You've caught everyone in the middle of a merge window, btw. As a result, I'm busy rebasing and reworking the get_user_pages call sites, and gup tracking, in the wake of some semi-traumatic changes to bio and gup and such. I plan to re-post right after 5.3-rc1 shows up, from here:
https://github.com/johnhubbard/linux/commits/gup_dma_core
...which you'll find already covers the changes you've posted, except for:
drivers/misc/sgi-gru/grufault.c
drivers/staging/kpc2000/kpc_dma/fileops.c
...and this one, which is undergoing to larger local changes, due to bvec, so let's leave it out of the choices:
fs/io_uring.c
Therefore, until -rc1, if you'd like to help, I'd recommend one or more of the following ideas:
1. Pull down https://github.com/johnhubbard/linux/commits/gup_dma_core
and find missing conversions: look for any additional missing get_user_pages/put_page conversions. You've already found a couple missing ones. I haven't re-run a search in a long time, so there's probably even more.
a) And find more, after I rebase to 5.3-rc1: people probably are adding get_user_pages() calls as we speak. :)
2. Patches: Focus on just one subsystem at a time, and perfect the patch for it. For example, I think this the staging driver would be perfect to start with:
drivers/staging/kpc2000/kpc_dma/fileops.c
a) verify that you've really, corrected converted the whole driver. (Hint: I think you might be overlooking a put_page call.)
b) Attempt to test it if you can (I'm being hypocritical in the extreme here, but one of my problems is that testing has been light, so any help is very valuable). qemu...? OTOH, maybe even qemu cannot easily test a kpc2000, but perhaps git blame and talking to the authors would help figure out a way to validate the changes.
Thinking about whether you can run a test that would prove or disprove my claim in (a), above, could be useful in coming up with tests to run.
In other words, a few very high quality conversions (even just one) that we can really put our faith in, is what I value most here. Tested patches are awesome.
3. Once I re-post, turn on the new CONFIG_DEBUG_GET_USER_PAGES_REFERENCES and run things such as xfstest/fstest. (Again, doing so would be going further than I have yet--very helpful). Help clarify what conversions have actually been tested and work, and which ones remain unvalidated.
Other: Please note that this:
https://github.com/johnhubbard/linux/commits/gup_dma_core
a) gets rebased often, and
b) has a bunch of commits (iov_iter and related) that conflict with the latest linux.git,
c) has some bugs in the bio area, that I'm fixing, so I don't trust that's it's safely runnable, for a few more days.
One note below, for the future:
Quoted 15 lines
This is a bunch of trivial conversions which is a part of an effort by John Hubbard to solve issues with gup pinned pages and filesystem writeback.
The issue is more clearly described in John Hubbard's patch[1] where put_user_page*() functions are introduced.
Currently put_user_page*() simply does
put_pagebut future implementations look to change that once treewide change ofput_pagecallsites to put_user_page*() is finished.The lwn article describing the issue with gup pinned pages and filesystem writeback [2].
Quoted 8 lines
This patch has been tested by building and booting the kernel as I don't have the required hardware to test the device drivers.
I did not modify gpu/drm drivers which use
release_pagesinstead ofput_page()to release reference of gup pinned pages as I am not clear whetherrelease_pagesandput_pageare interchangable.
When referring to patches in a commit description, please use the commit hash, not an external link. See Submitting Patches [1] for details.
Also, once you figure out the right maintainers and other involved people, putting Cc: in the commit description is common practice, too.
[1] https://www.kernel.org/doc/html/latest/process/submitting-patches.html
thanks,
Signature · 2 lines
John Hubbard NVIDIA
Quoted 123 lines
[2] https://lwn.net/Articles/784574/
Signed-off-by Bharath Vedartham <[email protected]>--- drivers/media/v4l2-core/videobuf-dma-sg.c | 3 +-- drivers/misc/sgi-gru/grufault.c | 2 +- drivers/staging/kpc2000/kpc_dma/fileops.c | 4 +--- drivers/vfio/vfio_iommu_type1.c | 2 +- fs/io_uring.c | 7 +++---- mm/gup_benchmark.c | 6 +----- net/xdp/xdp_umem.c | 7 +------ 7 files changed, 9 insertions(+), 22 deletions(-)drivers/media/v4l2-core/videobuf-dma-sg.c +1
-2@@ -349,8 +349,7 @@ int videobuf_dma_free(struct videobuf_dmabuf *dma)BUG_ON (dma->sglen);if (dma->pages) { -for (i =0 ; i < dma->nr_pages; i++) -put_page (dma->pages[i]); +put_user_pages (dma->pages, dma->nr_pages);kfree (dma->pages); dma->pages =NULL ; }drivers/misc/sgi-gru/grufault.c +1
-1@@ -188,7 +188,7 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,if (get_user_pages (vaddr,1 , write ?FOLL_WRITE :0 , &page,NULL ) <=0 )return -EFAULT ; *paddr =page_to_phys (page); -put_page (page); +put_user_page (page);return 0 ; }drivers/staging/kpc2000/kpc_dma/fileops.c +1
-3@@ -198,9 +198,7 @@ int kpc_dma_transfer(struct dev_private_data *priv, struct kiocb *kcb, unsignedsg_free_table (&acd->sgt); err_dma_map_sg: err_alloc_sg_table: -for (i =0 ; i < acd->page_count ; i++){ -put_page (acd->user_pages[i]); - } +put_user_pages (acd->user_pages, acd->page_count); err_get_user_pages:kfree (acd->user_pages); err_alloc_userpages:drivers/vfio/vfio_iommu_type1.c +1
-1@@ -369,7 +369,7 @@ static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,*/ if (ret >0 &&vma_is_fsdax (vmas[0 ])) { ret = -EOPNOTSUPP ; -put_page (page[0 ]); +put_user_page (page[0 ]); } }up_read (&mm->mmap_sem);fs/io_uring.c +3
-4@@ -2694,10 +2694,9 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,* if we did partial map, or found file backed vmas, * release any pages we did get */ -if (pret >0 ) { -for (j =0 ; j < pret; j++) -put_page (pages[j]); - } +if (pret >0 ) +put_user_pages (pages, pret); +if (ctx->account_mem)io_unaccount_mem (ctx->user, nr_pages);kvfree (imu->bvec);mm/gup_benchmark.c +1
-5@@ -76,11 +76,7 @@ static int __gup_benchmark_ioctl(unsigned int cmd, gup->size = addr - gup->addr; start_time =ktime_get (); -for (i =0 ; i < nr_pages; i++) { -if (!pages[i]) -break ; -put_page (pages[i]); - } +put_user_pages (pages, nr_pages); end_time =ktime_get (); gup->put_delta_usec =ktime_us_delta (end_time, start_time);net/xdp/xdp_umem.c +1
-6@@ -173,12 +173,7 @@ static void xdp_umem_unpin_pages(struct xdp_umem *umem) {unsigned int i; -for (i =0 ; i < umem->npgs; i++) { -struct page *page = umem->pgs[i]; - -set_page_dirty_lock (page); -put_page (page); - } +put_user_pages_dirty_lock (umem->pgs, umem->npgs);kfree (umem->pgs); umem->pgs =NULL ;
Quoted 4 lines · John Hubbard
Quoted 3 lines · Bharath Vedartham
This patch converts all call sites of
get_user_pagesto use put_user_page*() instead of put_page*() functions to release reference to gup pinned pages.
Hi John,
Quoted 25 lines
Hi Bharath,
Thanks for jumping in to help, and welcome to the party!
You've caught everyone in the middle of a merge window, btw. As a result, I'm busy rebasing and reworking the
get_user_pagescall sites, and gup tracking, in the wake of some semi-traumatic changes to bio and gup and such. I plan to re-post right after 5.3-rc1 shows up, from here:https://github.com/johnhubbard/linux/commits/gup_dma_core
...which you'll find already covers the changes you've posted, except for:
drivers/misc/sgi-gru/grufault.c
drivers/staging/kpc2000/kpc_dma/fileops.c...and this one, which is undergoing to larger local changes, due to bvec, so let's leave it out of the choices:
fs/io_uring.cTherefore, until -rc1, if you'd like to help, I'd recommend one or more of the following ideas:
Quoted 6 lines
1. Pull down https://github.com/johnhubbard/linux/commits/gup_dma_core
and find missing conversions: look for any additional missing get_user_pages/put_page conversions. You've already found a couple missing ones. I haven't re-run a search in a long time, so there's probably even more.
a) And find more, after I rebase to 5.3-rc1: people probably are addingget_user_pages()calls as we speak. :)
Shouldn't this be documented then? I don't see any docs for using put_user_page*() in v5.2.1 in the memory management API section?
Quoted 7 lines
2. Patches: Focus on just one subsystem at a time, and perfect the patch for it. For example, I think this the staging driver would be perfect to start with:
drivers/staging/kpc2000/kpc_dma/fileops.ca) verify that you've really, corrected converted the whole driver. (Hint: I think you might be overlooking a
put_pagecall.)
Yup. I did see that! Will fix it!
Quoted 6 lines
b) Attempt to test it if you can (I'm being hypocritical in the extreme here, but one of my problems is that testing has been light, so any help is very valuable). qemu...? OTOH, maybe even qemu cannot easily test a kpc2000, but perhaps
git blameand talking to the authors would help figure out a way to validate the changes.
Great! I ll do that, I ll mail the patch authors and ask them for help in testing.
Quoted 7 lines
Thinking about whether you can run a test that would prove or disprove my claim in (a), above, could be useful in coming up with tests to run.
In other words, a few very high quality conversions (even just one) that we can really put our faith in, is what I value most here. Tested patches are awesome.
I understand that!
Quoted 5 lines
3. Once I re-post, turn on the new
CONFIG_DEBUG_GET_USER_PAGES_REFERENCESand run things such as xfstest/fstest. (Again, doing so would be going further than I have yet--very helpful). Help clarify what conversions have actually been tested and work, and which ones remain unvalidated.
Other: Please note that this:
Yup will do that.
Quoted 9 lines
https://github.com/johnhubbard/linux/commits/gup_dma_core
a) gets rebased often, and
b) has a bunch of commits (
iov_iterand related) that conflict with the latest linux.git,c) has some bugs in the bio area, that I'm fixing, so I don't trust that's it's safely runnable, for a few more days.
I assume your repo contains only work related to fixing gup issues and not the main repo for gup development? i.e where gup changes are merged? Also are release_pages and put_user_pages interchangable?
Quoted 26 lines
One note below, for the future:
Quoted 15 lines
This is a bunch of trivial conversions which is a part of an effort by John Hubbard to solve issues with gup pinned pages and filesystem writeback.
The issue is more clearly described in John Hubbard's patch[1] where put_user_page*() functions are introduced.
Currently put_user_page*() simply does
put_pagebut future implementations look to change that once treewide change ofput_pagecallsites to put_user_page*() is finished.The lwn article describing the issue with gup pinned pages and filesystem writeback [2].
Quoted 8 lines
This patch has been tested by building and booting the kernel as I don't have the required hardware to test the device drivers.
I did not modify gpu/drm drivers which use
release_pagesinstead ofput_page()to release reference of gup pinned pages as I am not clear whetherrelease_pagesandput_pageare interchangable.
Quoted 7 lines
When referring to patches in a commit description, please use the commit hash, not an external link. See Submitting Patches [1] for details.
Also, once you figure out the right maintainers and other involved people, putting Cc: in the commit description is common practice, too.
[1] https://www.kernel.org/doc/html/latest/process/submitting-patches.html
Will work on that! Thanks!
Quoted 128 lines
thanks,
Signature · 2 lines
John Hubbard NVIDIAQuoted 123 lines
[2] https://lwn.net/Articles/784574/
Signed-off-by Bharath Vedartham <[email protected]>--- drivers/media/v4l2-core/videobuf-dma-sg.c | 3 +-- drivers/misc/sgi-gru/grufault.c | 2 +- drivers/staging/kpc2000/kpc_dma/fileops.c | 4 +--- drivers/vfio/vfio_iommu_type1.c | 2 +- fs/io_uring.c | 7 +++---- mm/gup_benchmark.c | 6 +----- net/xdp/xdp_umem.c | 7 +------ 7 files changed, 9 insertions(+), 22 deletions(-)drivers/media/v4l2-core/videobuf-dma-sg.c +1
-2@@ -349,8 +349,7 @@ int videobuf_dma_free(struct videobuf_dmabuf *dma)BUG_ON (dma->sglen);if (dma->pages) { -for (i =0 ; i < dma->nr_pages; i++) -put_page (dma->pages[i]); +put_user_pages (dma->pages, dma->nr_pages);kfree (dma->pages); dma->pages =NULL ; }drivers/misc/sgi-gru/grufault.c +1
-1@@ -188,7 +188,7 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,if (get_user_pages (vaddr,1 , write ?FOLL_WRITE :0 , &page,NULL ) <=0 )return -EFAULT ; *paddr =page_to_phys (page); -put_page (page); +put_user_page (page);return 0 ; }drivers/staging/kpc2000/kpc_dma/fileops.c +1
-3@@ -198,9 +198,7 @@ int kpc_dma_transfer(struct dev_private_data *priv, struct kiocb *kcb, unsignedsg_free_table (&acd->sgt); err_dma_map_sg: err_alloc_sg_table: -for (i =0 ; i < acd->page_count ; i++){ -put_page (acd->user_pages[i]); - } +put_user_pages (acd->user_pages, acd->page_count); err_get_user_pages:kfree (acd->user_pages); err_alloc_userpages:drivers/vfio/vfio_iommu_type1.c +1
-1@@ -369,7 +369,7 @@ static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,*/ if (ret >0 &&vma_is_fsdax (vmas[0 ])) { ret = -EOPNOTSUPP ; -put_page (page[0 ]); +put_user_page (page[0 ]); } }up_read (&mm->mmap_sem);fs/io_uring.c +3
-4@@ -2694,10 +2694,9 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,* if we did partial map, or found file backed vmas, * release any pages we did get */ -if (pret >0 ) { -for (j =0 ; j < pret; j++) -put_page (pages[j]); - } +if (pret >0 ) +put_user_pages (pages, pret); +if (ctx->account_mem)io_unaccount_mem (ctx->user, nr_pages);kvfree (imu->bvec);mm/gup_benchmark.c +1
-5@@ -76,11 +76,7 @@ static int __gup_benchmark_ioctl(unsigned int cmd, gup->size = addr - gup->addr; start_time =ktime_get (); -for (i =0 ; i < nr_pages; i++) { -if (!pages[i]) -break ; -put_page (pages[i]); - } +put_user_pages (pages, nr_pages); end_time =ktime_get (); gup->put_delta_usec =ktime_us_delta (end_time, start_time);net/xdp/xdp_umem.c +1
-6@@ -173,12 +173,7 @@ static void xdp_umem_unpin_pages(struct xdp_umem *umem) {unsigned int i; -for (i =0 ; i < umem->npgs; i++) { -struct page *page = umem->pgs[i]; - -set_page_dirty_lock (page); -put_page (page); - } +put_user_pages_dirty_lock (umem->pgs, umem->npgs);kfree (umem->pgs); umem->pgs =NULL ;
Quoted 73 lines · Bharath Vedartham
Quoted 4 lines · John Hubbard
Quoted 3 lines · Bharath Vedartham
This patch converts all call sites of
get_user_pagesto use put_user_page*() instead of put_page*() functions to release reference to gup pinned pages.Hi John,
Quoted 25 lines
Hi Bharath,
Thanks for jumping in to help, and welcome to the party!
You've caught everyone in the middle of a merge window, btw. As a result, I'm busy rebasing and reworking the
get_user_pagescall sites, and gup tracking, in the wake of some semi-traumatic changes to bio and gup and such. I plan to re-post right after 5.3-rc1 shows up, from here:https://github.com/johnhubbard/linux/commits/gup_dma_core
...which you'll find already covers the changes you've posted, except for:
drivers/misc/sgi-gru/grufault.c
drivers/staging/kpc2000/kpc_dma/fileops.c...and this one, which is undergoing to larger local changes, due to bvec, so let's leave it out of the choices:
fs/io_uring.cTherefore, until -rc1, if you'd like to help, I'd recommend one or more of the following ideas:
Quoted 6 lines
1. Pull down https://github.com/johnhubbard/linux/commits/gup_dma_core
and find missing conversions: look for any additional missing get_user_pages/put_page conversions. You've already found a couple missing ones. I haven't re-run a search in a long time, so there's probably even more.
a) And find more, after I rebase to 5.3-rc1: people probably are addingget_user_pages()calls as we speak. :)Shouldn't this be documented then? I don't see any docs for using put_user_page*() in v5.2.1 in the memory management API section?
Quoted 7 lines
2. Patches: Focus on just one subsystem at a time, and perfect the patch for it. For example, I think this the staging driver would be perfect to start with:
drivers/staging/kpc2000/kpc_dma/fileops.ca) verify that you've really, corrected converted the whole driver. (Hint: I think you might be overlooking a
put_pagecall.)Yup. I did see that! Will fix it!
Quoted 6 lines
b) Attempt to test it if you can (I'm being hypocritical in the extreme here, but one of my problems is that testing has been light, so any help is very valuable). qemu...? OTOH, maybe even qemu cannot easily test a kpc2000, but perhaps
git blameand talking to the authors would help figure out a way to validate the changes.Great! I ll do that, I ll mail the patch authors and ask them for help in testing.
Quoted 7 lines
Thinking about whether you can run a test that would prove or disprove my claim in (a), above, could be useful in coming up with tests to run.
In other words, a few very high quality conversions (even just one) that we can really put our faith in, is what I value most here. Tested patches are awesome.
I understand that!
Quoted 5 lines
3. Once I re-post, turn on the new
CONFIG_DEBUG_GET_USER_PAGES_REFERENCESand run things such as xfstest/fstest. (Again, doing so would be going further than I have yet--very helpful). Help clarify what conversions have actually been tested and work, and which ones remain unvalidated.
Other: Please note that this:Yup will do that.
Quoted 4 lines
https://github.com/johnhubbard/linux/commits/gup_dma_core
a) gets rebased often, and
Quoted 7 lines
Quoted 5 lines
b) has a bunch of commits (
iov_iterand related) that conflict with the latest linux.git,c) has some bugs in the bio area, that I'm fixing, so I don't trust that's it's safely runnable, for a few more days.
I assume your repo contains only work related to fixing gup issues and not the main repo for gup development? i.e where gup changes are merged?
We have been using Andrews tree for merging.
Quoted 1 line
Also are
release_pagesandput_user_pagesinterchangable?
Conceptually yes. But release_pages is more efficient. There was some discussion around this starting here:
https://lore.kernel.org/lkml/[email protected]/
And a resulting bug fix.
https://lkml.org/lkml/2019/6/21/95
Ira
Quoted 162 lines
Quoted 26 lines
One note below, for the future:
Quoted 15 lines
This is a bunch of trivial conversions which is a part of an effort by John Hubbard to solve issues with gup pinned pages and filesystem writeback.
The issue is more clearly described in John Hubbard's patch[1] where put_user_page*() functions are introduced.
Currently put_user_page*() simply does
put_pagebut future implementations look to change that once treewide change ofput_pagecallsites to put_user_page*() is finished.The lwn article describing the issue with gup pinned pages and filesystem writeback [2].
Quoted 8 lines
This patch has been tested by building and booting the kernel as I don't have the required hardware to test the device drivers.
I did not modify gpu/drm drivers which use
release_pagesinstead ofput_page()to release reference of gup pinned pages as I am not clear whetherrelease_pagesandput_pageare interchangable.Quoted 7 lines
When referring to patches in a commit description, please use the commit hash, not an external link. See Submitting Patches [1] for details.
Also, once you figure out the right maintainers and other involved people, putting Cc: in the commit description is common practice, too.
[1] https://www.kernel.org/doc/html/latest/process/submitting-patches.html
Will work on that! Thanks!
Quoted 128 lines
thanks,
Signature · 2 lines
John Hubbard NVIDIAQuoted 123 lines
[2] https://lwn.net/Articles/784574/
Signed-off-by Bharath Vedartham <[email protected]>--- drivers/media/v4l2-core/videobuf-dma-sg.c | 3 +-- drivers/misc/sgi-gru/grufault.c | 2 +- drivers/staging/kpc2000/kpc_dma/fileops.c | 4 +--- drivers/vfio/vfio_iommu_type1.c | 2 +- fs/io_uring.c | 7 +++---- mm/gup_benchmark.c | 6 +----- net/xdp/xdp_umem.c | 7 +------ 7 files changed, 9 insertions(+), 22 deletions(-)drivers/media/v4l2-core/videobuf-dma-sg.c +1
-2@@ -349,8 +349,7 @@ int videobuf_dma_free(struct videobuf_dmabuf *dma)BUG_ON (dma->sglen);if (dma->pages) { -for (i =0 ; i < dma->nr_pages; i++) -put_page (dma->pages[i]); +put_user_pages (dma->pages, dma->nr_pages);kfree (dma->pages); dma->pages =NULL ; }drivers/misc/sgi-gru/grufault.c +1
-1@@ -188,7 +188,7 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,if (get_user_pages (vaddr,1 , write ?FOLL_WRITE :0 , &page,NULL ) <=0 )return -EFAULT ; *paddr =page_to_phys (page); -put_page (page); +put_user_page (page);return 0 ; }drivers/staging/kpc2000/kpc_dma/fileops.c +1
-3@@ -198,9 +198,7 @@ int kpc_dma_transfer(struct dev_private_data *priv, struct kiocb *kcb, unsignedsg_free_table (&acd->sgt); err_dma_map_sg: err_alloc_sg_table: -for (i =0 ; i < acd->page_count ; i++){ -put_page (acd->user_pages[i]); - } +put_user_pages (acd->user_pages, acd->page_count); err_get_user_pages:kfree (acd->user_pages); err_alloc_userpages:drivers/vfio/vfio_iommu_type1.c +1
-1@@ -369,7 +369,7 @@ static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,*/ if (ret >0 &&vma_is_fsdax (vmas[0 ])) { ret = -EOPNOTSUPP ; -put_page (page[0 ]); +put_user_page (page[0 ]); } }up_read (&mm->mmap_sem);fs/io_uring.c +3
-4@@ -2694,10 +2694,9 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,* if we did partial map, or found file backed vmas, * release any pages we did get */ -if (pret >0 ) { -for (j =0 ; j < pret; j++) -put_page (pages[j]); - } +if (pret >0 ) +put_user_pages (pages, pret); +if (ctx->account_mem)io_unaccount_mem (ctx->user, nr_pages);kvfree (imu->bvec);mm/gup_benchmark.c +1
-5@@ -76,11 +76,7 @@ static int __gup_benchmark_ioctl(unsigned int cmd, gup->size = addr - gup->addr; start_time =ktime_get (); -for (i =0 ; i < nr_pages; i++) { -if (!pages[i]) -break ; -put_page (pages[i]); - } +put_user_pages (pages, nr_pages); end_time =ktime_get (); gup->put_delta_usec =ktime_us_delta (end_time, start_time);net/xdp/xdp_umem.c +1
-6@@ -173,12 +173,7 @@ static void xdp_umem_unpin_pages(struct xdp_umem *umem) {unsigned int i; -for (i =0 ; i < umem->npgs; i++) { -struct page *page = umem->pgs[i]; - -set_page_dirty_lock (page); -put_page (page); - } +put_user_pages_dirty_lock (umem->pgs, umem->npgs);kfree (umem->pgs); umem->pgs =NULL ;
Quoted 89 lines · Ira Weiny
Quoted 73 lines · Bharath Vedartham
Quoted 4 lines · John Hubbard
Quoted 3 lines · Bharath Vedartham
This patch converts all call sites of
get_user_pagesto use put_user_page*() instead of put_page*() functions to release reference to gup pinned pages.Hi John,
Quoted 25 lines
Hi Bharath,
Thanks for jumping in to help, and welcome to the party!
You've caught everyone in the middle of a merge window, btw. As a result, I'm busy rebasing and reworking the
get_user_pagescall sites, and gup tracking, in the wake of some semi-traumatic changes to bio and gup and such. I plan to re-post right after 5.3-rc1 shows up, from here:https://github.com/johnhubbard/linux/commits/gup_dma_core
...which you'll find already covers the changes you've posted, except for:
drivers/misc/sgi-gru/grufault.c
drivers/staging/kpc2000/kpc_dma/fileops.c...and this one, which is undergoing to larger local changes, due to bvec, so let's leave it out of the choices:
fs/io_uring.cTherefore, until -rc1, if you'd like to help, I'd recommend one or more of the following ideas:
Quoted 6 lines
1. Pull down https://github.com/johnhubbard/linux/commits/gup_dma_core
and find missing conversions: look for any additional missing get_user_pages/put_page conversions. You've already found a couple missing ones. I haven't re-run a search in a long time, so there's probably even more.
a) And find more, after I rebase to 5.3-rc1: people probably are addingget_user_pages()calls as we speak. :)Shouldn't this be documented then? I don't see any docs for using put_user_page*() in v5.2.1 in the memory management API section?
Quoted 7 lines
2. Patches: Focus on just one subsystem at a time, and perfect the patch for it. For example, I think this the staging driver would be perfect to start with:
drivers/staging/kpc2000/kpc_dma/fileops.ca) verify that you've really, corrected converted the whole driver. (Hint: I think you might be overlooking a
put_pagecall.)Yup. I did see that! Will fix it!
Quoted 6 lines
b) Attempt to test it if you can (I'm being hypocritical in the extreme here, but one of my problems is that testing has been light, so any help is very valuable). qemu...? OTOH, maybe even qemu cannot easily test a kpc2000, but perhaps
git blameand talking to the authors would help figure out a way to validate the changes.Great! I ll do that, I ll mail the patch authors and ask them for help in testing.
Quoted 7 lines
Thinking about whether you can run a test that would prove or disprove my claim in (a), above, could be useful in coming up with tests to run.
In other words, a few very high quality conversions (even just one) that we can really put our faith in, is what I value most here. Tested patches are awesome.
I understand that!
Quoted 5 lines
3. Once I re-post, turn on the new
CONFIG_DEBUG_GET_USER_PAGES_REFERENCESand run things such as xfstest/fstest. (Again, doing so would be going further than I have yet--very helpful). Help clarify what conversions have actually been tested and work, and which ones remain unvalidated.
Other: Please note that this:Yup will do that.
Quoted 4 lines
https://github.com/johnhubbard/linux/commits/gup_dma_core
a) gets rebased often, and
Quoted 7 lines
Quoted 5 lines
b) has a bunch of commits (
iov_iterand related) that conflict with the latest linux.git,c) has some bugs in the bio area, that I'm fixing, so I don't trust that's it's safely runnable, for a few more days.
I assume your repo contains only work related to fixing gup issues and not the main repo for gup development? i.e where gup changes are merged?
We have been using Andrews tree for merging.
Quoted 1 line
Also are
release_pagesandput_user_pagesinterchangable?Conceptually yes. But
release_pagesis more efficient. There was some discussion around this starting here:
Quoted 7 lines
https://lore.kernel.org/lkml/[email protected]/
And a resulting bug fix.
https://lkml.org/lkml/2019/6/21/95
Ira
Thanks! Will have a look at these links!
Quoted 162 lines
Quoted 162 lines
Quoted 26 lines
One note below, for the future:
Quoted 15 lines
This is a bunch of trivial conversions which is a part of an effort by John Hubbard to solve issues with gup pinned pages and filesystem writeback.
The issue is more clearly described in John Hubbard's patch[1] where put_user_page*() functions are introduced.
Currently put_user_page*() simply does
put_pagebut future implementations look to change that once treewide change ofput_pagecallsites to put_user_page*() is finished.The lwn article describing the issue with gup pinned pages and filesystem writeback [2].
Quoted 8 lines
This patch has been tested by building and booting the kernel as I don't have the required hardware to test the device drivers.
I did not modify gpu/drm drivers which use
release_pagesinstead ofput_page()to release reference of gup pinned pages as I am not clear whetherrelease_pagesandput_pageare interchangable.Quoted 7 lines
When referring to patches in a commit description, please use the commit hash, not an external link. See Submitting Patches [1] for details.
Also, once you figure out the right maintainers and other involved people, putting Cc: in the commit description is common practice, too.
[1] https://www.kernel.org/doc/html/latest/process/submitting-patches.html
Will work on that! Thanks!
Quoted 128 lines
thanks,
Signature · 2 lines
John Hubbard NVIDIAQuoted 123 lines
[2] https://lwn.net/Articles/784574/
Signed-off-by Bharath Vedartham <[email protected]>--- drivers/media/v4l2-core/videobuf-dma-sg.c | 3 +-- drivers/misc/sgi-gru/grufault.c | 2 +- drivers/staging/kpc2000/kpc_dma/fileops.c | 4 +--- drivers/vfio/vfio_iommu_type1.c | 2 +- fs/io_uring.c | 7 +++---- mm/gup_benchmark.c | 6 +----- net/xdp/xdp_umem.c | 7 +------ 7 files changed, 9 insertions(+), 22 deletions(-)drivers/media/v4l2-core/videobuf-dma-sg.c +1
-2@@ -349,8 +349,7 @@ int videobuf_dma_free(struct videobuf_dmabuf *dma)BUG_ON (dma->sglen);if (dma->pages) { -for (i =0 ; i < dma->nr_pages; i++) -put_page (dma->pages[i]); +put_user_pages (dma->pages, dma->nr_pages);kfree (dma->pages); dma->pages =NULL ; }drivers/misc/sgi-gru/grufault.c +1
-1@@ -188,7 +188,7 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,if (get_user_pages (vaddr,1 , write ?FOLL_WRITE :0 , &page,NULL ) <=0 )return -EFAULT ; *paddr =page_to_phys (page); -put_page (page); +put_user_page (page);return 0 ; }drivers/staging/kpc2000/kpc_dma/fileops.c +1
-3@@ -198,9 +198,7 @@ int kpc_dma_transfer(struct dev_private_data *priv, struct kiocb *kcb, unsignedsg_free_table (&acd->sgt); err_dma_map_sg: err_alloc_sg_table: -for (i =0 ; i < acd->page_count ; i++){ -put_page (acd->user_pages[i]); - } +put_user_pages (acd->user_pages, acd->page_count); err_get_user_pages:kfree (acd->user_pages); err_alloc_userpages:drivers/vfio/vfio_iommu_type1.c +1
-1@@ -369,7 +369,7 @@ static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,*/ if (ret >0 &&vma_is_fsdax (vmas[0 ])) { ret = -EOPNOTSUPP ; -put_page (page[0 ]); +put_user_page (page[0 ]); } }up_read (&mm->mmap_sem);fs/io_uring.c +3
-4@@ -2694,10 +2694,9 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,* if we did partial map, or found file backed vmas, * release any pages we did get */ -if (pret >0 ) { -for (j =0 ; j < pret; j++) -put_page (pages[j]); - } +if (pret >0 ) +put_user_pages (pages, pret); +if (ctx->account_mem)io_unaccount_mem (ctx->user, nr_pages);kvfree (imu->bvec);mm/gup_benchmark.c +1
-5@@ -76,11 +76,7 @@ static int __gup_benchmark_ioctl(unsigned int cmd, gup->size = addr - gup->addr; start_time =ktime_get (); -for (i =0 ; i < nr_pages; i++) { -if (!pages[i]) -break ; -put_page (pages[i]); - } +put_user_pages (pages, nr_pages); end_time =ktime_get (); gup->put_delta_usec =ktime_us_delta (end_time, start_time);net/xdp/xdp_umem.c +1
-6@@ -173,12 +173,7 @@ static void xdp_umem_unpin_pages(struct xdp_umem *umem) {unsigned int i; -for (i =0 ; i < umem->npgs; i++) { -struct page *page = umem->pgs[i]; - -set_page_dirty_lock (page); -put_page (page); - } +put_user_pages_dirty_lock (umem->pgs, umem->npgs);kfree (umem->pgs); umem->pgs =NULL ;
Quoted 2 lines · Bharath Vedartham
Quoted 1 line · John Hubbard
On 7/14/19 12:08 PM, Bharath Vedartham wrote:
[...]
Quoted 8 lines
Quoted 6 lines
1. Pull down https://github.com/johnhubbard/linux/commits/gup_dma_core
and find missing conversions: look for any additional missing get_user_pages/put_page conversions. You've already found a couple missing ones. I haven't re-run a search in a long time, so there's probably even more.
a) And find more, after I rebase to 5.3-rc1: people probably are addingget_user_pages()calls as we speak. :)Shouldn't this be documented then? I don't see any docs for using put_user_page*() in v5.2.1 in the memory management API section?
Yes, it needs documentation. My first try (which is still in the above git repo) was reviewed and found badly wanting, so I'm going to rewrite it. Meanwhile, I agree that an interim note would be helpful, let me put something together.
[...]
Quoted 11 lines
Quoted 9 lines
https://github.com/johnhubbard/linux/commits/gup_dma_core
a) gets rebased often, and
b) has a bunch of commits (
iov_iterand related) that conflict with the latest linux.git,c) has some bugs in the bio area, that I'm fixing, so I don't trust that's it's safely runnable, for a few more days.
I assume your repo contains only work related to fixing gup issues and not the main repo for gup development? i.e where gup changes are merged?
Correct, this is just a private tree, not a maintainer tree. But I'll try to keep the gup_dma_core branch something that is usable by others, during the transition over to put_user_page(), because the page-tracking patches are the main way to test any put_user_page() conversions.
As Ira said, we're using linux-mm as the real (maintainer) tree.
thanks,
Signature · 2 lines
John Hubbard NVIDIA
Quoted 30 lines · John Hubbard
Quoted 2 lines · Bharath Vedartham
Quoted 1 line · John Hubbard
On 7/14/19 12:08 PM, Bharath Vedartham wrote:
[...]
Quoted 8 lines
Quoted 6 lines
1. Pull down https://github.com/johnhubbard/linux/commits/gup_dma_core
and find missing conversions: look for any additional missing get_user_pages/put_page conversions. You've already found a couple missing ones. I haven't re-run a search in a long time, so there's probably even more.
a) And find more, after I rebase to 5.3-rc1: people probably are addingget_user_pages()calls as we speak. :)Shouldn't this be documented then? I don't see any docs for using put_user_page*() in v5.2.1 in the memory management API section?
Yes, it needs documentation. My first try (which is still in the above git repo) was reviewed and found badly wanting, so I'm going to rewrite it. Meanwhile, I agree that an interim note would be helpful, let me put something together.
[...]
Quoted 11 lines
Quoted 9 lines
https://github.com/johnhubbard/linux/commits/gup_dma_core
a) gets rebased often, and
b) has a bunch of commits (
iov_iterand related) that conflict with the latest linux.git,c) has some bugs in the bio area, that I'm fixing, so I don't trust that's it's safely runnable, for a few more days.
I assume your repo contains only work related to fixing gup issues and not the main repo for gup development? i.e where gup changes are merged?
Quoted 6 lines
Correct, this is just a private tree, not a maintainer tree. But I'll try to keep the
gup_dma_corebranch something that is usable by others, during the transition over toput_user_page(), because the page-tracking patches are the main way to test anyput_user_page()conversions.As Ira said, we're using linux-mm as the real (maintainer) tree.
Thanks for the info!
Quoted 5 lines
thanks,
Signature · 2 lines
John Hubbard NVIDIA
Jens Axboe
maintainerQuoted 8 lines · Bharath Vedartham
fs/io_uring.c +0
-0@@ -2694,10 +2694,9 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,* if we did partial map, or found file backed vmas, * release any pages we did get */
fs/io_uring.c static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, · line 2702
Quoted 10 lines
fs/io_uring.c +3
-4-if (pret >0 ) { -for (j =0 ; j < pret; j++) -put_page (pages[j]); - } +if (pret >0 ) +put_user_pages (pages, pret); +if (ctx->account_mem)io_unaccount_mem (ctx->user, nr_pages);kvfree (imu->bvec);
You handled just the failure case of the buffer registration, but not the actual free in io_sqe_buffer_unregister().
Signature · 1 line
Jens Axboe
Quoted 14 lines · Jens Axboe
Quoted 13 lines · Bharath Vedartham
fs/io_uring.c +1
-4@@ -2694,10 +2694,9 @@ static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,* if we did partial map, or found file backed vmas, * release any pages we did get */ -if (pret >0 ) { -for (j =0 ; j < pret; j++) -put_page (pages[j]); - } +if (pret >0 )
Quoted 11 lines
fs/io_uring.c static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, · line 2702
Quoted 5 lines
fs/io_uring.c +2
-0+put_user_pages (pages, pret); +if (ctx->account_mem)io_unaccount_mem (ctx->user, nr_pages);kvfree (imu->bvec);You handled just the failure case of the buffer registration, but not the actual free in
io_sqe_buffer_unregister().Signature · 1 line
Jens Axboe
Yup got it! Thanks! I won't be sending a patch again as fs/io_uring.c may have larger local changes for put_user_pages.
Thanks