Message ID | 20191020155935.12297-6-ira.weiny@intel.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | Enable per-file/directory DAX operations | expand |
On Sun, Oct 20, 2019 at 08:59:35AM -0700, ira.weiny@intel.com wrote: > From: Ira Weiny <ira.weiny@intel.com> > > Switching between DAX and non-DAX on a file is racy with respect to data > operations. However, if no data is involved the flag is safe to switch. > > Allow toggling the physical flag if a file is empty. The file length > check is not racy with respect to other operations as it is performed > under XFS_MMAPLOCK_EXCL and XFS_IOLOCK_EXCL locks. > > Signed-off-by: Ira Weiny <ira.weiny@intel.com> > --- > fs/xfs/xfs_ioctl.c | 9 +++++++-- > 1 file changed, 7 insertions(+), 2 deletions(-) > > diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c > index d3a7340d3209..3839684c249b 100644 > --- a/fs/xfs/xfs_ioctl.c > +++ b/fs/xfs/xfs_ioctl.c > @@ -33,6 +33,7 @@ > #include "xfs_sb.h" > #include "xfs_ag.h" > #include "xfs_health.h" > +#include "libxfs/xfs_dir2.h" > > #include <linux/mount.h> > #include <linux/namei.h> > @@ -1232,12 +1233,10 @@ xfs_diflags_to_linux( > inode->i_flags |= S_NOATIME; > else > inode->i_flags &= ~S_NOATIME; > -#if 0 /* disabled until the flag switching races are sorted out */ > if (xflags & FS_XFLAG_DAX) > inode->i_flags |= S_DAX; > else > inode->i_flags &= ~S_DAX; > -#endif This code has bit-rotted. See xfs_setup_iops(), where we now have a different inode->i_mapping->a_ops for DAX inodes. That, fundamentally, is the issue here - it's not setting/clearing the DAX flag that is the issue, it's doing a swap of the mapping->a_ops while there may be other code using that ops structure. IOWs, if there is any code anywhere in the kernel that calls an address space op without holding one of the three locks we hold here (i_rwsem, MMAPLOCK, ILOCK) then it can race with the swap of the address space operations. By limiting the address space swap to file sizes of zero, we rule out the page fault path (mmap of a zero length file segv's with an access beyond EOF on the first read/write page fault, right?). However, other aops callers that might run unlocked and do the wrong thing if the aops pointer is swapped between check of the aop method existing and actually calling it even if the file size is zero? A quick look shows that FIBMAP (ioctl_fibmap())) looks susceptible to such a race condition with the current definitions of the XFS DAX aops. I'm guessing there will be others, but I haven't looked further than this... > /* lock, flush and invalidate mapping in preparation for flag change */ > xfs_ilock(ip, XFS_MMAPLOCK_EXCL | XFS_IOLOCK_EXCL); > + > + if (i_size_read(inode) != 0) { > + error = -EOPNOTSUPP; > + goto out_unlock; > + } Wrong error. Should be the same as whatever is returned when we try to change the extent size hint and can't because the file is non-zero in length (-EINVAL, I think). Also needs a comment explainging why this check exists, and probably better written as i_size_read() > 0 .... Cheers, Dave.
On Mon, Oct 21, 2019 at 11:45:36AM +1100, Dave Chinner wrote: > On Sun, Oct 20, 2019 at 08:59:35AM -0700, ira.weiny@intel.com wrote: > > @@ -1232,12 +1233,10 @@ xfs_diflags_to_linux( > > inode->i_flags |= S_NOATIME; > > else > > inode->i_flags &= ~S_NOATIME; > > -#if 0 /* disabled until the flag switching races are sorted out */ > > if (xflags & FS_XFLAG_DAX) > > inode->i_flags |= S_DAX; > > else > > inode->i_flags &= ~S_DAX; > > -#endif > > This code has bit-rotted. See xfs_setup_iops(), where we now have a > different inode->i_mapping->a_ops for DAX inodes. :-( > > That, fundamentally, is the issue here - it's not setting/clearing > the DAX flag that is the issue, it's doing a swap of the > mapping->a_ops while there may be other code using that ops > structure. > > IOWs, if there is any code anywhere in the kernel that > calls an address space op without holding one of the three locks we > hold here (i_rwsem, MMAPLOCK, ILOCK) then it can race with the swap > of the address space operations. > > By limiting the address space swap to file sizes of zero, we rule > out the page fault path (mmap of a zero length file segv's with an > access beyond EOF on the first read/write page fault, right?). Yes I checked that and thought we were safe here... > However, other aops callers that might run unlocked and do the wrong > thing if the aops pointer is swapped between check of the aop method > existing and actually calling it even if the file size is zero? > > A quick look shows that FIBMAP (ioctl_fibmap())) looks susceptible > to such a race condition with the current definitions of the XFS DAX > aops. I'm guessing there will be others, but I haven't looked > further than this... I'll check for others and think on what to do about this. ext4 will have the same problem I think. :-( I don't suppose using a single a_ops for both DAX and non-DAX is palatable? > > > /* lock, flush and invalidate mapping in preparation for flag change */ > > xfs_ilock(ip, XFS_MMAPLOCK_EXCL | XFS_IOLOCK_EXCL); > > + > > + if (i_size_read(inode) != 0) { > > + error = -EOPNOTSUPP; > > + goto out_unlock; > > + } > > Wrong error. Should be the same as whatever is returned when we try > to change the extent size hint and can't because the file is > non-zero in length (-EINVAL, I think). Also needs a comment > explainging why this check exists, and probably better written as > i_size_read() > 0 .... Done, done, and done. Thank you, Ira > > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com
On Mon, Oct 21, 2019 at 03:49:31PM -0700, Ira Weiny wrote: > On Mon, Oct 21, 2019 at 11:45:36AM +1100, Dave Chinner wrote: > > On Sun, Oct 20, 2019 at 08:59:35AM -0700, ira.weiny@intel.com wrote: > > > @@ -1232,12 +1233,10 @@ xfs_diflags_to_linux( > > > inode->i_flags |= S_NOATIME; > > > else > > > inode->i_flags &= ~S_NOATIME; > > > -#if 0 /* disabled until the flag switching races are sorted out */ > > > if (xflags & FS_XFLAG_DAX) > > > inode->i_flags |= S_DAX; > > > else > > > inode->i_flags &= ~S_DAX; > > > -#endif > > > > This code has bit-rotted. See xfs_setup_iops(), where we now have a > > different inode->i_mapping->a_ops for DAX inodes. > > :-( > > > > > That, fundamentally, is the issue here - it's not setting/clearing > > the DAX flag that is the issue, it's doing a swap of the > > mapping->a_ops while there may be other code using that ops > > structure. > > > > IOWs, if there is any code anywhere in the kernel that > > calls an address space op without holding one of the three locks we > > hold here (i_rwsem, MMAPLOCK, ILOCK) then it can race with the swap > > of the address space operations. > > > > By limiting the address space swap to file sizes of zero, we rule > > out the page fault path (mmap of a zero length file segv's with an > > access beyond EOF on the first read/write page fault, right?). > > Yes I checked that and thought we were safe here... > > > However, other aops callers that might run unlocked and do the wrong > > thing if the aops pointer is swapped between check of the aop method > > existing and actually calling it even if the file size is zero? > > > > A quick look shows that FIBMAP (ioctl_fibmap())) looks susceptible > > to such a race condition with the current definitions of the XFS DAX > > aops. I'm guessing there will be others, but I haven't looked > > further than this... > > I'll check for others and think on what to do about this. ext4 will have the > same problem I think. :-( > > I don't suppose using a single a_ops for both DAX and non-DAX is palatable? IMO, no. It means we have to check IS_DAX() in every aops, and replicate a bunch of callouts to generic code. i.e this sort of thing: if (aops->method) return aops->method(...) /* do something else */ results in us having to replicate that logic as something like: if (!IS_DAX) return filesystem_aops_method() /* do something else */ Indeed, the calling code may well do the wrong thing if we have methods defined just to add IS_DAX() checks to avoid using that functionality because the caller now thinks that functionality is supported when in fact it isn't. So it seems to me like an even bigger can of worms to try to use a single aops structure for vastly different functionality.... Cheers, Dave.
On Mon 21-10-19 15:49:31, Ira Weiny wrote: > On Mon, Oct 21, 2019 at 11:45:36AM +1100, Dave Chinner wrote: > > On Sun, Oct 20, 2019 at 08:59:35AM -0700, ira.weiny@intel.com wrote: > > That, fundamentally, is the issue here - it's not setting/clearing > > the DAX flag that is the issue, it's doing a swap of the > > mapping->a_ops while there may be other code using that ops > > structure. > > > > IOWs, if there is any code anywhere in the kernel that > > calls an address space op without holding one of the three locks we > > hold here (i_rwsem, MMAPLOCK, ILOCK) then it can race with the swap > > of the address space operations. > > > > By limiting the address space swap to file sizes of zero, we rule > > out the page fault path (mmap of a zero length file segv's with an > > access beyond EOF on the first read/write page fault, right?). > > Yes I checked that and thought we were safe here... > > > However, other aops callers that might run unlocked and do the wrong > > thing if the aops pointer is swapped between check of the aop method > > existing and actually calling it even if the file size is zero? > > > > A quick look shows that FIBMAP (ioctl_fibmap())) looks susceptible > > to such a race condition with the current definitions of the XFS DAX > > aops. I'm guessing there will be others, but I haven't looked > > further than this... > > I'll check for others and think on what to do about this. ext4 will have the > same problem I think. :-( Just as a datapoint, ext4 is bold and sets inode->i_mapping->a_ops on existing inodes when switching journal data flag and so far it has not blown up. What we did to deal with issues Dave describes is that we introduced percpu rw-semaphore guarding switching of aops and then inside problematic functions redirect callbacks in the right direction under this semaphore. Somewhat ugly but it seems to work. Honza
On Fri 08-11-19 14:12:38, Jan Kara wrote: > On Mon 21-10-19 15:49:31, Ira Weiny wrote: > > On Mon, Oct 21, 2019 at 11:45:36AM +1100, Dave Chinner wrote: > > > On Sun, Oct 20, 2019 at 08:59:35AM -0700, ira.weiny@intel.com wrote: > > > That, fundamentally, is the issue here - it's not setting/clearing > > > the DAX flag that is the issue, it's doing a swap of the > > > mapping->a_ops while there may be other code using that ops > > > structure. > > > > > > IOWs, if there is any code anywhere in the kernel that > > > calls an address space op without holding one of the three locks we > > > hold here (i_rwsem, MMAPLOCK, ILOCK) then it can race with the swap > > > of the address space operations. > > > > > > By limiting the address space swap to file sizes of zero, we rule > > > out the page fault path (mmap of a zero length file segv's with an > > > access beyond EOF on the first read/write page fault, right?). > > > > Yes I checked that and thought we were safe here... > > > > > However, other aops callers that might run unlocked and do the wrong > > > thing if the aops pointer is swapped between check of the aop method > > > existing and actually calling it even if the file size is zero? > > > > > > A quick look shows that FIBMAP (ioctl_fibmap())) looks susceptible > > > to such a race condition with the current definitions of the XFS DAX > > > aops. I'm guessing there will be others, but I haven't looked > > > further than this... > > > > I'll check for others and think on what to do about this. ext4 will have the > > same problem I think. :-( > > Just as a datapoint, ext4 is bold and sets inode->i_mapping->a_ops on > existing inodes when switching journal data flag and so far it has not > blown up. What we did to deal with issues Dave describes is that we > introduced percpu rw-semaphore guarding switching of aops and then inside > problematic functions redirect callbacks in the right direction under this > semaphore. Somewhat ugly but it seems to work. Thinking about this some more, perhaps this scheme could be actually transformed in something workable. We could have a global (or maybe per-sb but I'm not sure it's worth it) percpu rwsem and we could transform aops calls into: percpu_down_read(aops_rwsem); do_call(); percpu_up_read(aops_rwsem); With some macro magic it needn't be even that ugly. Honza
On Fri, Nov 08, 2019 at 02:46:06PM +0100, Jan Kara wrote: > On Fri 08-11-19 14:12:38, Jan Kara wrote: > > On Mon 21-10-19 15:49:31, Ira Weiny wrote: > > > On Mon, Oct 21, 2019 at 11:45:36AM +1100, Dave Chinner wrote: > > > > On Sun, Oct 20, 2019 at 08:59:35AM -0700, ira.weiny@intel.com wrote: > > > > That, fundamentally, is the issue here - it's not setting/clearing > > > > the DAX flag that is the issue, it's doing a swap of the > > > > mapping->a_ops while there may be other code using that ops > > > > structure. > > > > > > > > IOWs, if there is any code anywhere in the kernel that > > > > calls an address space op without holding one of the three locks we > > > > hold here (i_rwsem, MMAPLOCK, ILOCK) then it can race with the swap > > > > of the address space operations. > > > > > > > > By limiting the address space swap to file sizes of zero, we rule > > > > out the page fault path (mmap of a zero length file segv's with an > > > > access beyond EOF on the first read/write page fault, right?). > > > > > > Yes I checked that and thought we were safe here... > > > > > > > However, other aops callers that might run unlocked and do the wrong > > > > thing if the aops pointer is swapped between check of the aop method > > > > existing and actually calling it even if the file size is zero? > > > > > > > > A quick look shows that FIBMAP (ioctl_fibmap())) looks susceptible > > > > to such a race condition with the current definitions of the XFS DAX > > > > aops. I'm guessing there will be others, but I haven't looked > > > > further than this... > > > > > > I'll check for others and think on what to do about this. ext4 will have the > > > same problem I think. :-( > > > > Just as a datapoint, ext4 is bold and sets inode->i_mapping->a_ops on > > existing inodes when switching journal data flag and so far it has not > > blown up. What we did to deal with issues Dave describes is that we > > introduced percpu rw-semaphore guarding switching of aops and then inside > > problematic functions redirect callbacks in the right direction under this > > semaphore. Somewhat ugly but it seems to work. Ah I am glad you brought this up. I had not seen this before. Is that s_journal_flag_rwsem? In the general case I don't think that correctly protects against: if (a_ops->call) a_ops->call(); Because not all operations are defined in both ext4_aops and ext4_journalled_aops. Specifically migratepage. move_to_new_page() specifically follows the pattern above with migratepage. So is there a bug here? > > Thinking about this some more, perhaps this scheme could be actually > transformed in something workable. We could have a global (or maybe per-sb > but I'm not sure it's worth it) percpu rwsem and we could transform aops > calls into: > > percpu_down_read(aops_rwsem); > do_call(); > percpu_up_read(aops_rwsem); > > With some macro magic it needn't be even that ugly. I think this is safer. And what I have been investigating/coding up. Because that also would protect against the above with: percpu_down_read(aops_rwsem); if (a_ops->call) a_ops->call(); percpu_up_read(aops_rwsem); However I have been looking at SRCU because we also have patterns like: generic_file_buffered_read if (a_ops->is_partially_uptodate) a_ops->is_partially_uptodate() page_cache_sync_readahead force_page_cache_readahead if (!a_ops->readpage && !a_ops->readpages) return; __do_page_cache_readahead read_pages if (a_ops->readpages) a_ops->readpages() a_ops->readpage So we would have to pass the a_ops through to use a rwsem. Where SRCU I think would be fine to just take the SRCU read lock multiple times. Am I wrong? We also have a 3rd (2nd?) issue. There are callers who check for the presence of an operation to be used later. For example do_dentry_open(): do_dentry_open() { ... if (<flags> & O_DIRECT) if (!<a_ops> || !<a_ops>->direct_IO) return -EINVAL; ... } After this open direct_IO better be there AFAICT so changing the a_ops later would not be good. For ext4 direct_IO is defined for all the a_ops... so I guess that is not a big deal. However, is the user really getting the behavior they expect in this case? I'm afraid of requiring FSs to have to follow rules in defining their a_ops. Because I'm afraid maintaining those rules would be hard and would eventually lead to crashes when someone did it wrong. :-/ So for this 3rd (2nd) case I think we should simply take a reference to the a_ops and fail changing the mode. For the DAX case that means the user is best served by taking a write lease on the file to ensure there are no other opens which could cause issues. Would that work for changing the journaling mode? And I _think_ this is the only issue we have with this right now. But if other callers of a_ops needed the pattern of using the a_ops at a time across context changes they would need to ensure this reference was taken. What I have come up with thus far is an interface like: /* * as_get_a_ops() -- safely get the a_ops from the address_space specified * * @as: address space to get a_ops from * @ref: used to indicate if a reference is required on this a_ops * @tok: srcu token to be returned in as_put_a_ops() * * The a_ops returned is protected from changing until as_put_a_ops(). * * If ref is specified then ref must also be specified in as_put_a_ops() to * release this reference. In this case a reference is taken on the a_ops * which will prevent it from changing until the reference is released. * * References should _ONLY_ be taken when the a_ops needs to be constant * across a user context switch because doing so will block changing the a_ops * until that reference is released. * * Examples of using a reference are checks for specific a_ops pointers which * are expected to support functionality at a later date (example direct_IO) */ static inline const struct address_space_operations * as_get_a_ops(struct address_space *as, bool ref, int *tok) { ... } static inline void as_assign_a_ops(struct address_space *as, const struct address_space_operations *a_ops) { ... } static inline void as_put_a_ops(struct address_space *as, int tok, bool ref) { ... } I'm still working out the details of using SRCU and a ref count. I have made at least 1 complete pass of all the a_ops users and I think this would cover them all. Thoughts? Ira > > Honza > -- > Jan Kara <jack@suse.com> > SUSE Labs, CR
On Fri 08-11-19 11:36:13, Ira Weiny wrote: > On Fri, Nov 08, 2019 at 02:46:06PM +0100, Jan Kara wrote: > > On Fri 08-11-19 14:12:38, Jan Kara wrote: > > > On Mon 21-10-19 15:49:31, Ira Weiny wrote: > > > > On Mon, Oct 21, 2019 at 11:45:36AM +1100, Dave Chinner wrote: > > > > > On Sun, Oct 20, 2019 at 08:59:35AM -0700, ira.weiny@intel.com wrote: > > > > > That, fundamentally, is the issue here - it's not setting/clearing > > > > > the DAX flag that is the issue, it's doing a swap of the > > > > > mapping->a_ops while there may be other code using that ops > > > > > structure. > > > > > > > > > > IOWs, if there is any code anywhere in the kernel that > > > > > calls an address space op without holding one of the three locks we > > > > > hold here (i_rwsem, MMAPLOCK, ILOCK) then it can race with the swap > > > > > of the address space operations. > > > > > > > > > > By limiting the address space swap to file sizes of zero, we rule > > > > > out the page fault path (mmap of a zero length file segv's with an > > > > > access beyond EOF on the first read/write page fault, right?). > > > > > > > > Yes I checked that and thought we were safe here... > > > > > > > > > However, other aops callers that might run unlocked and do the wrong > > > > > thing if the aops pointer is swapped between check of the aop method > > > > > existing and actually calling it even if the file size is zero? > > > > > > > > > > A quick look shows that FIBMAP (ioctl_fibmap())) looks susceptible > > > > > to such a race condition with the current definitions of the XFS DAX > > > > > aops. I'm guessing there will be others, but I haven't looked > > > > > further than this... > > > > > > > > I'll check for others and think on what to do about this. ext4 will have the > > > > same problem I think. :-( > > > > > > Just as a datapoint, ext4 is bold and sets inode->i_mapping->a_ops on > > > existing inodes when switching journal data flag and so far it has not > > > blown up. What we did to deal with issues Dave describes is that we > > > introduced percpu rw-semaphore guarding switching of aops and then inside > > > problematic functions redirect callbacks in the right direction under this > > > semaphore. Somewhat ugly but it seems to work. > > Ah I am glad you brought this up. I had not seen this before. > > Is that s_journal_flag_rwsem? Yes. > In the general case I don't think that correctly protects against: > > if (a_ops->call) > a_ops->call(); > > Because not all operations are defined in both ext4_aops and > ext4_journalled_aops. Specifically migratepage. > > move_to_new_page() specifically follows the pattern above with migratepage. So > is there a bug here? Looks like there could be. > > Thinking about this some more, perhaps this scheme could be actually > > transformed in something workable. We could have a global (or maybe per-sb > > but I'm not sure it's worth it) percpu rwsem and we could transform aops > > calls into: > > > > percpu_down_read(aops_rwsem); > > do_call(); > > percpu_up_read(aops_rwsem); > > > > With some macro magic it needn't be even that ugly. > > I think this is safer. And what I have been investigating/coding up. > Because that also would protect against the above with: > > percpu_down_read(aops_rwsem); > if (a_ops->call) > a_ops->call(); > percpu_up_read(aops_rwsem); > > However I have been looking at SRCU because we also have patterns like: > > > generic_file_buffered_read > if (a_ops->is_partially_uptodate) > a_ops->is_partially_uptodate() > page_cache_sync_readahead > force_page_cache_readahead > if (!a_ops->readpage && !a_ops->readpages) > return; > __do_page_cache_readahead > read_pages > if (a_ops->readpages) > a_ops->readpages() > a_ops->readpage > > > So we would have to pass the a_ops through to use a rwsem. Where SRCU I > think would be fine to just take the SRCU read lock multiple times. Am I > wrong? So the idea I had would not solve this issue because we'd release the rwsem once we return from ->is_partially_uptodate(). This example shows that we actually expect consistency among different aops as they are called in sequence and that's much more difficult to achieve than just a consistency within single aop call. > We also have a 3rd (2nd?) issue. There are callers who check for the > presence of an operation to be used later. For example do_dentry_open(): > > do_dentry_open() > { > ... > if (<flags> & O_DIRECT) > if (!<a_ops> || !<a_ops>->direct_IO) > return -EINVAL; > ... > } > > After this open direct_IO better be there AFAICT so changing the a_ops > later would not be good. For ext4 direct_IO is defined for all the > a_ops... so I guess that is not a big deal. However, is the user really > getting the behavior they expect in this case? In this particular case I don't think there's any practical harm for any filesystem but in general this is another instance where consistency of aops over time is assumed. > I'm afraid of requiring FSs to have to follow rules in defining their a_ops. > Because I'm afraid maintaining those rules would be hard and would eventually > lead to crashes when someone did it wrong. I guess this very much depends on the rules. But yes, anything non-obvious or hard to check would quickly lead to bugs, I agree. But IMHO fully general solution to above problems would clutter the generic code in rather ugly way as well because usage of aops is pretty widespread in mm and fs code. It isn't just a few places that call them... But I think we could significantly reduce the problem by looking at what's in aops. We have lots of operations there that operate on pages. If we mandate that before and during switching of aops, you must make sure there's nothing in page cache for the inode, you've already dealt with 90% of the problems. Beside these we have: * write_begin - that creates page in page cache so above rule should stop it as well * bmap - honestly I'd be inclined to just move this to inode_operations just like fiemap. There's nothing about address_space in its functionality. * swap_activate / swap_deactivate - Either I'd move these to file_operations (what's there about address_space, right), or since all instances of this only care about the inode, we can as well just pass only inode to the function and move it to inode_operations. And then the really problematic ones: * direct_IO - Logically with how the IO path is structured, it belongs in aops so I wouldn't move it. With the advance of iomap it is on its way to being removed altogether but that will take a long time to happen completely. So for now I'd mandate that direct_IO path must be locked out while switching aops. * readpages - these should be locked out by the rule that page creation is forbidden. * writepages - these need to be locked out when switching aops. And that should be it. So I don't think there's a need for reference-counting of aops in the generic code, especially since I don't think it can be done in an elegant way (but feel free to correct me). I think that just providing a way to lock-out above three calls would be enough. > So for this 3rd (2nd) case I think we should simply take a reference to the > a_ops and fail changing the mode. For the DAX case that means the user is best > served by taking a write lease on the file to ensure there are no other opens > which could cause issues. > > Would that work for changing the journaling mode? > > And I _think_ this is the only issue we have with this right now. But if other > callers of a_ops needed the pattern of using the a_ops at a time across context > changes they would need to ensure this reference was taken. > > What I have come up with thus far is an interface like: > > /* > * as_get_a_ops() -- safely get the a_ops from the address_space specified > * > * @as: address space to get a_ops from > * @ref: used to indicate if a reference is required on this a_ops > * @tok: srcu token to be returned in as_put_a_ops() > * > * The a_ops returned is protected from changing until as_put_a_ops(). > * > * If ref is specified then ref must also be specified in as_put_a_ops() to > * release this reference. In this case a reference is taken on the a_ops > * which will prevent it from changing until the reference is released. > * > * References should _ONLY_ be taken when the a_ops needs to be constant > * across a user context switch because doing so will block changing the a_ops > * until that reference is released. > * > * Examples of using a reference are checks for specific a_ops pointers which > * are expected to support functionality at a later date (example direct_IO) > */ > static inline const struct address_space_operations * > as_get_a_ops(struct address_space *as, bool ref, int *tok) > { > ... > } > > static inline void > as_assign_a_ops(struct address_space *as, > const struct address_space_operations *a_ops) > { > ... > } > > static inline void as_put_a_ops(struct address_space *as, int tok, bool ref) > { > ... > } > > > I'm still working out the details of using SRCU and a ref count. I have made > at least 1 complete pass of all the a_ops users and I think this would cover > them all. Well, my concern with the use of interface like this is: a) The clutter in the generic code b) It's difficult to make this work with SRCU because presumably you want to use synchronize_srcu() while switching aops. But then you have three operations to do: 1) switch aops 2) set inode flag 3) synchronize_srcu and depending on the order in which you do these either "old aops" operations will see inode with a flag or "new aops" will see the inode without a flag and either can confuse those functions... Honza
> > > In the general case I don't think that correctly protects against: > > > > if (a_ops->call) > > a_ops->call(); > > > > Because not all operations are defined in both ext4_aops and > > ext4_journalled_aops. Specifically migratepage. > > > > move_to_new_page() specifically follows the pattern above with migratepage. So > > is there a bug here? > > Looks like there could be. Ok I'm going to leave this alone and whatever I come up with try and make sure that the ext4 journal a_ops fits. [snip] > > > > However I have been looking at SRCU because we also have patterns like: > > > > > > generic_file_buffered_read > > if (a_ops->is_partially_uptodate) > > a_ops->is_partially_uptodate() > > page_cache_sync_readahead > > force_page_cache_readahead > > if (!a_ops->readpage && !a_ops->readpages) > > return; > > __do_page_cache_readahead > > read_pages > > if (a_ops->readpages) > > a_ops->readpages() > > a_ops->readpage > > > > > > So we would have to pass the a_ops through to use a rwsem. Where SRCU I > > think would be fine to just take the SRCU read lock multiple times. Am I > > wrong? > > So the idea I had would not solve this issue because we'd release the rwsem > once we return from ->is_partially_uptodate(). This example shows that we > actually expect consistency among different aops as they are called in > sequence and that's much more difficult to achieve than just a consistency > within single aop call. I can't be sure but is seems like consistency for an operation is somewhat important. OR we have to develop rules around when filesystems can, or can not change a_ops so that the consistency does not matter. I think what scares me the most is that I'm not really sure what those rules are... > > > We also have a 3rd (2nd?) issue. There are callers who check for the > > presence of an operation to be used later. For example do_dentry_open(): > > > > do_dentry_open() > > { > > ... > > if (<flags> & O_DIRECT) > > if (!<a_ops> || !<a_ops>->direct_IO) > > return -EINVAL; > > ... > > } > > > > After this open direct_IO better be there AFAICT so changing the a_ops > > later would not be good. For ext4 direct_IO is defined for all the > > a_ops... so I guess that is not a big deal. However, is the user really > > getting the behavior they expect in this case? > > In this particular case I don't think there's any practical harm for any > filesystem but in general this is another instance where consistency of > aops over time is assumed. Yes... But this is just a more complex situation to maintain consistency. Again I don't have any idea if consistency is required. But the current definition/use of a_ops is very static (with the exception of the ext4 case you gave), so I'm thinking that much of the code is written with the assumption that the vectors do not change. > > > I'm afraid of requiring FSs to have to follow rules in defining their a_ops. > > Because I'm afraid maintaining those rules would be hard and would eventually > > lead to crashes when someone did it wrong. > > I guess this very much depends on the rules. But yes, anything non-obvious > or hard to check would quickly lead to bugs, I agree. But IMHO fully > general solution to above problems would clutter the generic code in rather > ugly way as well because usage of aops is pretty widespread in mm and fs > code. It isn't just a few places that call them... I agree it does clutter the code. and the patches I have are not pretty and I'm not even sure they are not broken... So yea I'm open to ideas! > > But I think we could significantly reduce the problem by looking at what's > in aops. We have lots of operations there that operate on pages. If we > mandate that before and during switching of aops, you must make sure > there's nothing in page cache for the inode, you've already dealt with 90% > of the problems. Sounds promising... But digging into this it looks like we need a similar rule for the DAX side to have no mappings outstanding. Perhaps you meant that when you said page cache? > > Beside these we have: > * write_begin - that creates page in page cache so above rule should stop > it as well Just to make sure I understand, do you propose that we put a check in pagecache_write_[begin|end]() to protect from these calls while changing? I just want to be sure because I've wondered if we can get away with minimal checks, or checks on individual functions, in the generic code. But that seemed kind of ugly as well. > * bmap - honestly I'd be inclined to just move this to inode_operations > just like fiemap. There's nothing about address_space in its functionality. Hmmm... What about these call paths? ext4_bmap() filemap_write_and_wait() filemap_fdatawrite() __filemap_fdatawrite() __filemap_fdatawrite_range() do_writepages() a_ops->writepages() or xfs_vm_bmap() iomap_bmap() filemap_write_and_wait() ... :-/ maybe we should leave it and have it covered under the page cache rule you propose? > * swap_activate / swap_deactivate - Either I'd move these to > file_operations (what's there about address_space, right), or since all > instances of this only care about the inode, we can as well just pass > only inode to the function and move it to inode_operations. XFS calls iomap_swapfile_activate() which calls vfs_fsync() which needs the file. Seems like file_operations would be better. I like the idea of cleaning this up a lot. I've gone ahead with a couple of patches to do this. At least this simplifies things a little bit... > > And then the really problematic ones: > * direct_IO - Logically with how the IO path is structured, it belongs in > aops so I wouldn't move it. With the advance of iomap it is on its way to > being removed altogether but that will take a long time to happen > completely. So for now I'd mandate that direct_IO path must be locked out > while switching aops. How do we lock this out between checking for this support on open and using it later? I think if we go down this path we have to make a rule that says that direct_IO must be defined for both a_ops. Right now for our 2 use case we are lucky that direct_IO is also defined as the same function. So there is little danger of odd behavior. Let me explore more. > * readpages - these should be locked out by the rule that page creation is > forbidden. Agreed. Same question applies here as for pagecache_write_[begin|end](). Should I special case a check for this? > * writepages - these need to be locked out when switching aops. If nothing is in the pagecache could we ignore this as well? > > And that should be it. So I don't think there's a need for reference-counting > of aops in the generic code, especially since I don't think it can be done > in an elegant way (but feel free to correct me). I think that just > providing a way to lock-out above three calls would be enough. Ok, I've been thinking about this whole email more today. And what if we add a couple of FS specific lockout callbacks in struct address_space itself. If they are defined the FS has dynamic a_ops capability and these 3 functions will need to be locked out by a rw_sem controlled by the FS. We can then document the "rules" for dynamic a_ops better for FS's to support them by referencing the special cases and the fact that dynamic a_ops requires these callbacks to be defined. It clutters the generic code a bit, but not as much as my idea. At the same time it helps to self document the special cases in both the FS's and the core code. [snip] > > > > > > I'm still working out the details of using SRCU and a ref count. I have made > > at least 1 complete pass of all the a_ops users and I think this would cover > > them all. > > Well, my concern with the use of interface like this is: > > a) The clutter in the generic code There is a bit of clutter. What I'm most concerned about is the amount of special casing. I still think it would be cleaner in the long run... And force some structure on the use of a_ops. But after looking at it more there may be a middle ground. > b) It's difficult to make this work with SRCU because presumably you want > to use synchronize_srcu() while switching aops. But then you have three > operations to do: > 1) switch aops > 2) set inode flag > 3) synchronize_srcu > > and depending on the order in which you do these either "old aops" > operations will see inode with a flag or "new aops" will see the inode > without a flag and either can confuse those functions... Yes that might be a challenge. Ira
diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c index d3a7340d3209..3839684c249b 100644 --- a/fs/xfs/xfs_ioctl.c +++ b/fs/xfs/xfs_ioctl.c @@ -33,6 +33,7 @@ #include "xfs_sb.h" #include "xfs_ag.h" #include "xfs_health.h" +#include "libxfs/xfs_dir2.h" #include <linux/mount.h> #include <linux/namei.h> @@ -1232,12 +1233,10 @@ xfs_diflags_to_linux( inode->i_flags |= S_NOATIME; else inode->i_flags &= ~S_NOATIME; -#if 0 /* disabled until the flag switching races are sorted out */ if (xflags & FS_XFLAG_DAX) inode->i_flags |= S_DAX; else inode->i_flags &= ~S_DAX; -#endif } static int @@ -1320,6 +1319,12 @@ xfs_ioctl_setattr_dax_invalidate( /* lock, flush and invalidate mapping in preparation for flag change */ xfs_ilock(ip, XFS_MMAPLOCK_EXCL | XFS_IOLOCK_EXCL); + + if (i_size_read(inode) != 0) { + error = -EOPNOTSUPP; + goto out_unlock; + } + error = filemap_write_and_wait(inode->i_mapping); if (error) goto out_unlock;