On 12/8/23 21:46, Amir Goldstein wrote:
On Fri, Dec 8, 2023 at 9:50 PM Bernd Schubert
<bernd.schubert@xxxxxxxxxxx> wrote:
On 12/8/23 09:39, Amir Goldstein wrote:
On Thu, Dec 7, 2023 at 8:38 PM Bernd Schubert
<bernd.schubert@xxxxxxxxxxx> wrote:
On 12/7/23 08:39, Amir Goldstein wrote:
On Thu, Dec 7, 2023 at 1:28 AM Bernd Schubert
<bernd.schubert@xxxxxxxxxxx> wrote:
On 12/6/23 09:25, Amir Goldstein wrote:
Is it actually important for FUSE_DIRECT_IO_ALLOW_MMAP fs
(e.g. virtiofsd) to support FOPEN_PARALLEL_DIRECT_WRITES?
I guess not otherwise, the combination would have been tested.
I'm not sure how many people are aware of these different flags/features.
I had just finalized the backport of the related patches to RHEL8 on
Friday, as we (or our customers) need both for different jobs.
FOPEN_PARALLEL_DIRECT_WRITES is typically important for
network fs and FUSE_DIRECT_IO_ALLOW_MMAP is typically not
for network fs. Right?
We kind of have these use cases for our network file systems
FOPEN_PARALLEL_DIRECT_WRITES:
- Traditional HPC, large files, parallel IO
- Large file used on local node as container for many small files
FUSE_DIRECT_IO_ALLOW_MMAP:
- compilation through gcc (not so important, just not nice when it
does not work)
- rather recent: python libraries using mmap _reads_. As it is read
only no issue of consistency.
These jobs do not intermix - no issue as in generic/095. If such
applications really exist, I have no issue with a serialization penalty.
Just disabling FOPEN_PARALLEL_DIRECT_WRITES because other
nodes/applications need FUSE_DIRECT_IO_ALLOW_MMAP is not so nice.
Final goal is also to have FOPEN_PARALLEL_DIRECT_WRITES to work on plain
O_DIRECT and not only for FUSE_DIRECT_IO - I need to update this branch
and post the next version
https://github.com/bsbernd/linux/commits/fuse-dio-v4
In the mean time I have another idea how to solve
FOPEN_PARALLEL_DIRECT_WRITES + FUSE_DIRECT_IO_ALLOW_MMAP
Please find attached what I had in my mind. With that generic/095 is not
crashing for me anymore. I just finished the initial coding - it still
needs a bit cleanup and maybe a few comments.
Nice. I like the FUSE_I_CACHE_WRITES state.
For FUSE_PASSTHROUGH I will need to track if inode is open/mapped
in caching mode, so FUSE_I_CACHE_WRITES can be cleared on release
of the last open file of the inode.
I did not understand some of the complexity here:
/* The inode ever got page writes and we do not know for sure
* in the DIO path if these are pending - shared lock not possible */
spin_lock(&fi->lock);
if (!test_bit(FUSE_I_CACHE_WRITES, &fi->state)) {
if (!(*cnt_increased)) {
How can *cnt_increased be true here?
I think you missed the 2nd entry into this function, when the shared
lock was already taken?
Yeh, I did.
I have changed the code now to have all
complexity in this function (test, lock, retest with lock, release,
wakeup). I hope that will make it easier to see the intention of the
code. Will post the new patches in the morning.
Sounds good. Current version was a bit hard to follow.
fi->shared_lock_direct_io_ctr++;
*cnt_increased = true;
}
excl_lock = false;
Seems like in every outcome of this function
*cnt_increased = !excl_lock
so there is not need for out arg cnt_increased
If excl_lock would be used as input - yeah, would have worked as well.
Or a parameter like "retest-under-lock". Code is changed now to avoid
going in and out.
}
spin_unlock(&fi->lock);
out:
if (excl_lock && *cnt_increased) {
bool wake = false;
spin_lock(&fi->lock);
if (--fi->shared_lock_direct_io_ctr == 0)
wake = true;
spin_unlock(&fi->lock);
if (wake)
wake_up(&fi->direct_io_waitq);
}
I don't see how this wake_up code is reachable.
TBH, I don't fully understand the expected result.
Surely, the behavior of dio mixed with mmap is undefined. Right?
IIUC, your patch does not prevent dirtying page cache while dio is in
flight. It only prevents writeback while dio is in flight, which is the same
behavior as with exclusive inode lock. Right?
Yeah, thanks. I will add it in the patch description.
And there was actually an issue with the patch, as cache flushing needs
to be initiated before doing the lock decision, fixed now.
I thought there was, because of the wait in fuse_send_writepage()
but wasn't sure if I was following the flow correctly.
Maybe this interaction is spelled out somewhere else, but if not
better spell it out for people like me that are new to this code.
Sure, thanks a lot for your helpful comments!
Just to be clear, this patch looks like a good improvement and
is mostly independent of the "inode caching mode" and
FOPEN_CACHE_MMAP idea that I suggested.
The only thing that my idea changes is replacing the
FUSE_I_CACHE_WRITES state with a FUSE_I_CACHE_IO_MODE
state, which is set earlier than FUSE_I_CACHE_WRITES
on caching file open or first direct_io mmap and unlike
FUSE_I_CACHE_WRITES, it is cleared on the last file close.
FUSE_I_CACHE_WRITES means that caching writes happened.
FUSE_I_CACHE_IO_MODE means the caching writes and reads
may happen.
FOPEN_PARALLEL_DIRECT_WRITES obviously shouldn't care
about "caching reads may happen", but IMO that is a small trade off
to make for maintaining the same state for
"do not allow parallel dio" and "do not allow passthrough open".
I think the attached patches should do, it now also unsets
IMO, your patch is still more complicated than it should be.
There is no need for the complicated retest state machine.
If you split the helpers to:
bool exclusive_lock fuse_dio_wr_needs_exclusive_lock();
...
fuse_dio_lock_inode(iocb, &exclusive);
...
fuse_dio_unlock_inode(iocb, &exclusive);
Then you only need to test FUSE_I_CACHE_IO_MODE in
fuse_dio_wr_needs_exclusive_lock()
and you only need to increment shared_lock_direct_io_ctr
after taking shared lock and re-testing FUSE_I_CACHE_IO_MODE.
Hmm, I'm not sure.
I changed fuse_file_mmap() to call this function
/*
* direct-io with shared locks cannot handle page cache io - set an inode
* flag to disable shared locks and wait until remaining threads are done
*/
static void fuse_file_mmap_handle_dio_writers(struct inode *inode)
{
struct fuse_inode *fi = get_fuse_inode(inode);
spin_lock(&fi->lock);
set_bit(FUSE_I_CACHE_IO_MODE, &fi->state);
while (fi->shared_lock_direct_io_ctr > 0) {
spin_unlock(&fi->lock);
wait_event_interruptible(fi->direct_io_waitq,
fi->shared_lock_direct_io_ctr == 0);
spin_lock(&fi->lock);
}
spin_unlock(&fi->lock);
}
Before we had indeed a race. Idea for fuse_file_mmap_handle_dio_writers()
and fuse_dio_lock_inode() is to either have FUSE_I_CACHE_IO_MODE set,
or fi->shared_lock_direct_io_ctr is greater 0, but that requires that
FUSE_I_CACHE_IO_MODE is checked for when fi->lock is taken.
I'm going to think about over the weekend if your suggestion
to increase fi->shared_lock_direct_io_ctr only after taking the shared
lock is possible. Right now I don't see how to do that.
FUSE_I_CACHE_IO_MODE. Setting the flag actually has to be done from
fuse_file_mmap (and not from fuse_send_writepage) to avoid a dead stall,
but that aligns with passthrough anyway?
Yes.
I see that shared_lock_direct_io_ctr is checked without lock or barriers
in and the wait_event() should be interruptible.
Thanks, fixed with the function above.
I am also not sure if it breaks any locking order for mmap because
the task that is going to wake it up is holding the shared inode lock...
The waitq has its own lock. We have
fuse_file_mmap - called under some mmap lock, waitq lock
fuse_dio_lock_inode: no lock taken before calling wakeup
fuse_direct_write_iter: wakeup after release of all locks
So I don't think we have a locker issue (lockdep also doesn't annotate
anything).
I don't think that lockdep can understand this dependency.
What we definitely cannot do it to take the inode i_rwsem lock in fuse_file_mmap
It's complicated. I need to look at the whole thing again.
While looking at this code, the invalidate_inode_pages2() looks suspicious.
If inode is already in FUSE_I_CACHE_IO_MODE when performing
another mmap, doesn't that have potential for data loss?
(even before your patch I mean)
Amir, right now it only sets
FUSE_I_CACHE_IO_MODE for VM_MAYWRITE. Maybe you could add a condition
for passthrough there?
We could add a condition, but I don't think that we should.
I think we should refrain from different behavior when it is not justified.
I think it is not justified to allow parallel dio if any file is open in
caching mode on the inode and any mmap (private or shared)
exists on the inode.
That means that FUSE_I_CACHE_IO_MODE should be set on
any mmap, and already on open for non direct_io files.
Ok, I can change and add that. Doing it in open is definitely needed
for O_DIRECT (in my other dio branch).
Good, the more common code the better.
Mixing caching and direct io on the same inode is hard as it is
already and there is no need to add complexity by allowing
parallel dio in that case. IMO it wins us nothing.
So the slight issue I see are people like me, who check the content
of a file during a long running computation. Like an HPC application
is doing some long term runs. Then in the middle of
the run the user wants to see the current content of the file and
reads it - if that is done through mmap (and from a node that runs
the application), parallel DIO is disabled with the current patch
until the file is closed - I see the use case to check for writes.
That's what I thought.
The FUSE_I_CACHE_IO_MODE could be cleared on last file
close (as your patch did) but it could be cleared earlier if
instead of tracking refcount of open file, we track refcount of
files open in caching mode or mmaped, which is what the
FOPEN_MMAP_CACHE flag I suggested is for.
But how does open() know that a file/fd is used for mmap?
Because what I tried to suggest is a trick/hack:
first mmap on direct_io file sets FOPEN_MMAP_CACHE on the file
and bumps the cached_opens on the inode as if file was
opened in caching mode or in FOPEN_MMAP_CACHE mode.
When the file that was used for mmap is closed and all the rest
of the open files have only ever been used for direct_io, then
inode exists the caching io mode.
Using an FOPEN flag for that is kind of a hack.
We could add an internal file state bits for that as well,
but my thinking was that FOPEN_MMAP_CACHE could really
be set by the server to mean per-file ALLOW_MMAP instead of
the per-filesystem ALLOW_MMAP. Not sure if that will be useful.
Ok, I will try to add that in a different patch to have better
visibility. Will also put these patch here in front of my dio branch and
rebase these patches. There comes in a bit additional complexity to
handle O_DIRECT, but it also consolidates direct-IO writes code paths.
At least I hope this is still possible with the new changes.
Sorry for the hand waving. I was trying to send out a demo
patch that explains it better, but got caught up with other things.
No problem at all, I think I know what you mean and I can try add this
myself.
Thanks,
Bernd
