Confused by commit 56a76f [fs: fix page_mkwrite error cases in core code and btrfs]
Nick,
While visiting some issues in ocfs2_page_mkwrite(), I realized
that we're returning 0 to mean "Please try again Mr VM", but the caller of page_mkwrite() now expects VM_FAULT_NOPAGE. This is all due to 56a76f [fs: fix page_mkwrite error cases in core code and btrfs].
In the comment for 56a76f, you state that btrfs is the example
but all other filesystems need to be fixed...yet ocfs2, ext4, and gfs2 continue to return 0 or VM_FAULT_SIGBUS.
I guess this continues to work because we return the page
unlocked (thus triggering the "if (unlikely(!(tmp & VM_FAULT_LOCKED)))" logic). Was this expected to continue to work, or do you consider these filesystems broken until they are updated with the new return codes?
Back to the ocfs2 issue. Am I correctly reading the current
code that we can safely throw away the page passed in to page_mkwrite() if a pagecache flush has dropped it? Currently, we presume that the page passed in must be the one we make writable. We make a quick check of page->mapping == inode->i_mapping, returning 0 (for "try again") immediately if that's false. But once we get into the meat of our locking and finally lock the page for good, we assume mapping==i_mapping still holds. That obviously breaks when the pagecache gets truncated. At this late stage, we -EINVAL (clearly wrong).
It looks hard to lock the page for good high up at our first
check of the mapping. Wengang has proposed to simply ignore the page passed into page_mkwrite() and just find_or_create_page() the sucker at the place we're ready to consider it stable. As I see it, the two places that call page_mkwrite() are going to revalidate the pte anyway, and they'll find the newly created page if find_or_create_page() gets a different. Is that safe behavior?
Joel
Signature · 9 lines
Life's Little Instruction Book #3 "Watch a sunrise at least once a year." Joel Becker Senior Development Manager Oracle E-mail: [email protected] Phone: (650) 506-8127
Hi Joel,
Quoted 3 lines · Joel Becker
Nick,
While visiting some issues inocfs2_page_mkwrite(), I realized that we're returning 0 to mean "Please try again Mr VM", but the caller
0 has always meant minor fault, which means the filesystem satisfied the request without doing IO. In the change to bit mask return values, I kept it compatible by having major fault be presence of a bit, and minor fault indicate absence.
Quoted 2 lines
of
page_mkwrite()now expectsVM_FAULT_NOPAGE. This is all due to 56a76f [fs: fixpage_mkwriteerror cases in core code and btrfs].
NOPAGE basically means no further action on part of the VM is required. It was used to consolidate pfn based fault handler with page based fault handler. And then, later, it was further used in this case to allow the filesystem to have the VM try again in rare / difficult to handle error cases exactly like truncate races.
Quoted 7 lines
In the comment for 56a76f, you state that btrfs is the example but all other filesystems need to be fixed...yet ocfs2, ext4, and gfs2 continue to return 0 or
VM_FAULT_SIGBUS.
I guess this continues to work because we return the page unlocked (thus triggering the "if (unlikely(!(tmp & VM_FAULT_LOCKED)))" logic). Was this expected to continue to work, or do you consider these filesystems broken until they are updated with the new return codes?
No it was all back compatible. That is to say, the ones that return SIGBUS in these cases were already broken, but the patch didn't break them further. Actaully it closed up races a bit, if I recall. But yes they should have all been converted to the new calling convention and returning with page locked.
If the filesystem returns 0, it means minor fault, and the VM will actually install the page (unless the hack to check page->mapping catches it).
Quoted 3 lines
Back to the ocfs2 issue. Am I correctly reading the current code that we can safely throw away the page passed in to
page_mkwrite()if a pagecache flush has dropped it?
Well you just return NOPAGE and the VM throws the page away.
Quoted 7 lines
Currently, we presume that the
page passed in must be the one we make writable. We make a quick check ofpage->mapping==inode->i_mapping, returning 0 (for "try again") immediately if that's false. But once we get into the meat of our locking and finally lock the page for good, we assume mapping==i_mapping still holds. That obviously breaks when the pagecache gets truncated. At this late stage, we-EINVAL(clearly wrong).
The better way to do this would be to just return VM_FAULT_NOPAGE in any case you need the VM to retry the fault. When you reach the business end of your handler, you want to hold the page locked, after you verify it is correct, and return that to the fault handler.
Quoted 7 lines
It looks hard to lock the page for good high up at our first check of the mapping. Wengang has proposed to simply ignore the page passed into
page_mkwrite()and justfind_or_create_page()the sucker at the place we're ready to consider it stable. As I see it, the two places that callpage_mkwrite()are going to revalidate the pte anyway, and they'll find the newly created page iffind_or_create_page()gets a different. Is that safe behavior?
I can't see the point. If you can find_or_create_page, then you can lock_page, by definition. Then check the mapping and VM_FAULT_SIGBUS if it is wrong.
Messing with the wrong page will only see the result ignored by the VM, and push an incorrect page through parts of your fault handler, which is potentially confusing at best, I would have thought.
Thanks,
Nick
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Joel Becker
maintainerQuoted 9 lines · Nick Piggin
Quoted 3 lines · Joel Becker
Nick,
While visiting some issues inocfs2_page_mkwrite(), I realized that we're returning 0 to mean "Please try again Mr VM", but the caller0 has always meant minor fault, which means the filesystem satisfied the request without doing IO. In the change to bit mask return values, I kept it compatible by having major fault be presence of a bit, and minor fault indicate absence.
Ok, good, the 0 is intentionally somewhat-working.
Quoted 5 lines
NOPAGE basically means no further action on part of the VM is required. It was used to consolidate pfn based fault handler with page based fault handler. And then, later, it was further used in this case to allow the filesystem to have the VM try again in rare / difficult to handle error cases exactly like truncate races.
Ok, so it's not "no further action" anymore, it is "I don't have
a page, check again to see if I'm supposed to give you one." That's how I read the code. Cool.
Quoted 8 lines
No it was all back compatible. That is to say, the ones that return SIGBUS in these cases were already broken, but the patch didn't break them further. Actaully it closed up races a bit, if I recall. But yes they should have all been converted to the new calling convention and returning with page locked.
If the filesystem returns 0, it means minor fault, and the VM will actually install the page (unless the hack to check
page->mappingcatches it).
Right, but does it install the page passed into page_mkwrite()?
The way I read the code, it actually rechecks the pte and installs the page it now finds.
Quoted 5 lines
Quoted 3 lines
Back to the ocfs2 issue. Am I correctly reading the current code that we can safely throw away the page passed in to
page_mkwrite()if a pagecache flush has dropped it?Well you just return NOPAGE and the VM throws the page away.
I mean, as we discuss below, that I ignore the page passed
to page_mkwrite() and rediscover the mapping ourselves.
Quoted 8 lines
Quoted 7 lines
Currently, we presume that the
page passed in must be the one we make writable. We make a quick check ofpage->mapping==inode->i_mapping, returning 0 (for "try again") immediately if that's false. But once we get into the meat of our locking and finally lock the page for good, we assume mapping==i_mapping still holds. That obviously breaks when the pagecache gets truncated. At this late stage, we-EINVAL(clearly wrong).
Quoted 4 lines
The better way to do this would be to just return
VM_FAULT_NOPAGEin any case you need the VM to retry the fault. When you reach the business end of your handler, you want to hold the page locked, after you verify it is correct, and return that to the fault handler.
This is going to be hard. Our write_end() assumes it must
unlock the pages (which is normal behavior for write(2)), but in the page_mkwrite() case we need to avoid the unlock to follow your recommendation (we use our write_begin/write_end pair to trigger any allocation or zeroing needed before the page is writable).
Quoted 11 lines
Quoted 7 lines
It looks hard to lock the page for good high up at our first check of the mapping. Wengang has proposed to simply ignore the page passed into
page_mkwrite()and justfind_or_create_page()the sucker at the place we're ready to consider it stable. As I see it, the two places that callpage_mkwrite()are going to revalidate the pte anyway, and they'll find the newly created page iffind_or_create_page()gets a different. Is that safe behavior?I can't see the point. If you can
find_or_create_page, then you canlock_page, by definition. Then check the mapping andVM_FAULT_SIGBUSif it is wrong.
The find_or_create_page() is deep at the meat of the function,
not the cursory check at the top. The idea is that at this point, find_or_create_page() will return a locked page that must, by definition, be part of the correct mapping.
Quoted 3 lines
Messing with the wrong page will only see the result ignored by the VM, and push an incorrect page through parts of your fault handler, which is potentially confusing at best, I would have thought.
If the VM is rechecking the pte after we return frompage_mkwrite(), won't it see any new page created?
Joel
Signature · 9 lines
"Egotist: a person more interested in himself than in me."
- Ambrose Bierce
http://www.jlbec.org/
[email protected]
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.htmlQuoted 9 lines · Joel Becker
Quoted 4 lines · Nick Piggin
On Fri, Dec 31, 2010 at 10:29 AM, Joel Becker <[email protected]> wrote:
If the filesystem returns 0, it means minor fault, and the VM will actually install the page (unless the hack to check
page->mappingcatches it).Right, but does it install the page passed into
page_mkwrite()? The way I read the code, it actually rechecks the pte and installs the page it now finds.
No, it doesn't install anything, it just bails out and the fault will be restarted if required.
Quoted 10 lines
Quoted 4 lines
The better way to do this would be to just return
VM_FAULT_NOPAGEin any case you need the VM to retry the fault. When you reach the business end of your handler, you want to hold the page locked, after you verify it is correct, and return that to the fault handler.This is going to be hard. Our
write_end()assumes it must unlock the pages (which is normal behavior for write(2)), but in thepage_mkwrite()case we need to avoid the unlock to follow your recommendation (we use our write_begin/write_end pair to trigger any allocation or zeroing needed before the page is writable).
Yes that would be the best option. It is possible to use the unlocked return, but that still gives possibility of races in some cases. Consider if the filesystem has to allocate some persistent metadata while the page is dirty, then it must do so in page_mkwrite (set_page_dirty may not sleep), but if the page is returned with no locks, it might be possible that metadata is freed before the page can be installed in the dirty pte.
Basically it would be nice for all filesystems to move to this convention so we can remove the old cruft.
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
It looks hard to lock the page for good high up at our first check of the mapping. Wengang has proposed to simply ignore the page passed into
page_mkwrite()and justfind_or_create_page()the sucker at the place we're ready to consider it stable. As I see it, the two places that callpage_mkwrite()are going to revalidate the pte anyway, and they'll find the newly created page iffind_or_create_page()gets a different. Is that safe behavior?
Quoted 8 lines
Quoted 3 lines
I can't see the point. If you can
find_or_create_page, then you canlock_page, by definition. Then check the mapping andVM_FAULT_SIGBUSif it is wrong.The
find_or_create_page()is deep at the meat of the function, not the cursory check at the top. The idea is that at this point,find_or_create_page()will return a locked page that must, by definition, be part of the correct mapping.
But you must still handle failures there, because find_or_create_page may return -ENOMEM. So just lock the page, recheck the mapping there, and then do exactly the same error handling.
Quoted 6 lines
Quoted 3 lines
Messing with the wrong page will only see the result ignored by the VM, and push an incorrect page through parts of your fault handler, which is potentially confusing at best, I would have thought.
If the VM is rechecking the pte after we return from
page_mkwrite(), won't it see any new page created?
But the point of page_mkwrite is a dirty notifier for the fs. If this new different page was installed due to say truncate then a read-only fault, the filesystem would miss the notification. So it just does the simple thing and retries the whole sequence (if needed).
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 9 lines · Nick Piggin
Quoted 8 lines · Joel Becker
Quoted 6 lines · Nick Piggin
On Fri, Dec 31, 2010 at 10:29 AM, Joel Becker <[email protected]> wrote:
The better way to do this would be to just return
VM_FAULT_NOPAGEin any case you need the VM to retry the fault. When you reach the business end of your handler, you want to hold the page locked, after you verify it is correct, and return that to the fault handler.
Quoted 8 lines
Quoted 5 lines
This is going to be hard. Our
write_end()assumes it must unlock the pages (which is normal behavior for write(2)), but in thepage_mkwrite()case we need to avoid the unlock to follow your recommendation (we use our write_begin/write_end pair to trigger any allocation or zeroing needed before the page is writable).Yes that would be the best option. It is possible to use the unlocked return, but that still gives possibility of races in some cases. Consider
Yes, I am aware of that.
Quoted 2 lines
Basically it would be nice for all filesystems to move to this convention so we can remove the old cruft.
I can appreciate that. And you've just answered the "Do you
want us to get there, or are minor faults in the old-0-style OK?" question.
Quoted 8 lines
Quoted 4 lines
The
find_or_create_page()is deep at the meat of the function, not the cursory check at the top. The idea is that at this point,find_or_create_page()will return a locked page that must, by definition, be part of the correct mapping.But you must still handle failures there, because
find_or_create_pagemay return-ENOMEM. So just lock the page, recheck the mapping there, and then do exactly the same error handling.
Of course it can error, but error is different than clean
restart. Though cleanup should be the same; I'm looking at our code trying to convince myself that this is so ;-) Btw, -ENOMEM is that OOM fault error, right? ;-)
Quoted 7 lines
Quoted 2 lines
If the VM is rechecking the pte after we return from
page_mkwrite(), won't it see any new page created?But the point of
page_mkwriteis a dirty notifier for the fs. If this new different page was installed due to say truncate then a read-only fault, the filesystem would miss the notification. So it just does the simple thing and retries the whole sequence (if needed).
Fair enough, even if we caused the new page and thus are already
notified ;-)
Essentially, you would really like us (and ext4, and gfs2, etc)
to start returning VM_FAULT_NOPAGE for any place we aren't sure what happened to our page, VM_FAULT_OOM when we get an -ENOMEM, and VM_FAULT_LOCKED for all successful operations. That the long and short of it? Thank you for helping me understand your plan for this code.
Btw, what not use VM_FAULT_RETRY for "I'd like you to retry"?
Especially since the comment for NOPAGE doesn't read anything like its actual behavior.
Joel
Signature · 12 lines
"Friends may come and go, but enemies accumulate."
- Thomas Jones
Joel Becker
Senior Development Manager
Oracle
E-mail: [email protected]
Phone: (650) 506-8127
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.htmlQuoted 19 lines · Joel Becker
Quoted 9 lines · Nick Piggin
Quoted 8 lines · Joel Becker
Quoted 6 lines · Nick Piggin
On Fri, Dec 31, 2010 at 10:29 AM, Joel Becker <[email protected]> wrote:
The better way to do this would be to just return
VM_FAULT_NOPAGEin any case you need the VM to retry the fault. When you reach the business end of your handler, you want to hold the page locked, after you verify it is correct, and return that to the fault handler.Quoted 8 lines
Quoted 5 lines
This is going to be hard. Our
write_end()assumes it must unlock the pages (which is normal behavior for write(2)), but in thepage_mkwrite()case we need to avoid the unlock to follow your recommendation (we use our write_begin/write_end pair to trigger any allocation or zeroing needed before the page is writable).Yes that would be the best option. It is possible to use the unlocked return, but that still gives possibility of races in some cases. Consider
Quoted 8 lines
Yes, I am aware of that.
Quoted 2 lines
Basically it would be nice for all filesystems to move to this convention so we can remove the old cruft.
I can appreciate that. And you've just answered the "Do you want us to get there, or are minor faults in the old-0-style OK?" question.
It is OK in that it will work. It is not preferred, if you are reworking pleaase use the new style.
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines
The
find_or_create_page()is deep at the meat of the function, not the cursory check at the top. The idea is that at this point,find_or_create_page()will return a locked page that must, by definition, be part of the correct mapping.
Quoted 7 lines
Quoted 3 lines
But you must still handle failures there, because
find_or_create_pagemay return-ENOMEM. So just lock the page, recheck the mapping there, and then do exactly the same error handling.Of course it can error, but error is different than clean restart. Though cleanup should be the same; I'm looking at our code trying to convince myself that this is so ;-)
Cleanup is exactly the same, yes, because the only difference in different error codes is what the VM does with them.
Quoted 2 lines
Btw,
-ENOMEMis that OOM
fault error, right? ;-)
Right.
Quoted 10 lines
Quoted 7 lines
Quoted 2 lines
If the VM is rechecking the pte after we return from
page_mkwrite(), won't it see any new page created?But the point of
page_mkwriteis a dirty notifier for the fs. If this new different page was installed due to say truncate then a read-only fault, the filesystem would miss the notification. So it just does the simple thing and retries the whole sequence (if needed).Fair enough, even if we caused the new page and thus are already notified ;-)
Well, you would be notified via ->fault, but not not writable fault and no page_mkwrite.
Quoted 5 lines
Essentially, you would really like us (and ext4, and gfs2, etc) to start returning
VM_FAULT_NOPAGEfor any place we aren't sure what happened to our page,VM_FAULT_OOMwhen we get an-ENOMEM, andVM_FAULT_LOCKEDfor all successful operations. That the long and short of it? Thank you for helping me understand your plan for this code.
Exactly.
Quoted 3 lines
Btw, what not use
VM_FAULT_RETRYfor "I'd like you to retry"? Especially since the comment for NOPAGE doesn't read anything like its actual behavior.
It was designed first to replace that old ->nopfn handler, so it does need comments updating. RETRY doesn't cover the ->nopfn usage, but NOPAGE does cover the retry usage, I think, so names are OK.
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html