Re: [PATCH 2/2] NFS: Improve nfsiod workqueue detection for allocation flags

[Trond Myklebust <trondmy@kernel.org>]

On Mon, 2025-07-07 at 16:12 -0400, Benjamin Coddington wrote:
> On 7 Jul 2025, at 15:25, Trond Myklebust wrote:
> 
> > On Mon, 2025-07-07 at 14:46 -0400, Benjamin Coddington wrote:
> > > The NFS client writeback paths change which flags are passed to
> > > their
> > > memory allocation calls based on whether the current task is
> > > running
> > > from
> > > within a workqueue or not.  More specifically, it appears that
> > > during
> > > writeback allocations with PF_WQ_WORKER set on current->flags
> > > will
> > > add
> > > __GFP_NORETRY | __GFP_NOWARN.  Presumably this is because nfsiod
> > > can
> > > simply fail quickly and later retry to write back that specific
> > > page
> > > should
> > > the allocation fail.
> > > 
> > > However, the check for PF_WQ_WORKER is too general because tasks
> > > can
> > > enter NFS
> > > writeback paths from other workqueues.  Specifically, the
> > > loopback
> > > driver
> > > tends to perform writeback into backing files on NFS with
> > > PF_WQ_WORKER set,
> > > and additionally sets PF_MEMALLOC_NOIO.  The combination of
> > > PF_MEMALLOC_NOIO with __GFP_NORETRY can easily result in
> > > allocation
> > > failures and the loopback driver has no retry functionality.  As
> > > a
> > > result,
> > > after commit 0bae835b63c5 ("NFS: Avoid writeback threads getting
> > > stuck in
> > > mempool_alloc()") users are seeing corrupted loop-mounted
> > > filesystems
> > > backed
> > > by image files on NFS.
> > > 
> > > In a preceding patch, we introduced a function to allow NFS to
> > > detect
> > > if
> > > the task is executing within a specific workqueue.  Here we use
> > > that
> > > helper
> > > to set __GFP_NORETRY | __GFP_NOWARN only if the workqueue is
> > > nfsiod.
> > > 
> > > Fixes: 0bae835b63c5 ("NFS: Avoid writeback threads getting stuck
> > > in
> > > mempool_alloc()")
> > > Signed-off-by: Benjamin Coddington <bcodding@redhat.com>
> > > ---
> > >  fs/nfs/internal.h | 12 +++++++++++-
> > >  1 file changed, 11 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/fs/nfs/internal.h b/fs/nfs/internal.h
> > > index 69c2c10ee658..173172afa3f5 100644
> > > --- a/fs/nfs/internal.h
> > > +++ b/fs/nfs/internal.h
> > > @@ -12,6 +12,7 @@
> > >  #include <linux/nfs_page.h>
> > >  #include <linux/nfslocalio.h>
> > >  #include <linux/wait_bit.h>
> > > +#include <linux/workqueue.h>
> > >  
> > >  #define NFS_SB_MASK
> > > (SB_NOSUID|SB_NODEV|SB_NOEXEC|SB_SYNCHRONOUS)
> > >  
> > > @@ -669,9 +670,18 @@ nfs_write_match_verf(const struct
> > > nfs_writeverf
> > > *verf,
> > >  		!nfs_write_verifier_cmp(&req->wb_verf, &verf-
> > > > verifier);
> > >  }
> > >  
> > > +static inline bool is_nfsiod(void)
> > > +{
> > > +	struct workqueue_struct *current_wq =
> > > current_workqueue();
> > > +
> > > +	if (current_wq)
> > > +		return current_wq == nfsiod_workqueue;
> > > +	return false;
> > > +}
> > > +
> > >  static inline gfp_t nfs_io_gfp_mask(void)
> > >  {
> > > -	if (current->flags & PF_WQ_WORKER)
> > > +	if (is_nfsiod())
> > >  		return GFP_KERNEL | __GFP_NORETRY |
> > > __GFP_NOWARN;
> > >  	return GFP_KERNEL;
> > >  }
> > 
> > 
> > Instead of trying to identify the nfsiod_workqueue, why not apply
> > current_gfp_context() in order to weed out callers that set
> > PF_MEMALLOC_NOIO and PF_MEMALLOC_NOFS?
> > 
> > i.e.
> > 
> > 
> > static inline gfp_t nfs_io_gfp_mask(void)
> > {
> > 	gfp_t ret = current_gfp_context(GFP_KERNEL);
> > 
> > 	if ((current->flags & PF_WQ_WORKER) && ret == GFP_KERNEL)
> > 		ret |= __GFP_NORETRY | __GFP_NOWARN;
> > 	return ret;
> > }
> 
> This would fix the problem we see, but we'll also end up carrying the
> flags
> from the layer above NFS into the client's current allocation
> strategy.
> That seems fragile to part of the original intent - we have static
> known
> flags for NFS' allocation in either context.

Yes, but if the PF_MEMALLOC_NOIO or PF_MEMALLOC_NOFS flags are set, the
memory manager will in any case water those flags down using the same
call to current_gfp_context().

This is really just making sure that we don't set the __GFP_NORETRY
flag in that case, because in a low memory situation that could end up
deadlocking due to being unable to kick off I/O in order to free up
memory.

> On the other hand, perhaps we want to honor those flags if the upper
> layer
> is setting them, because it should have a good reason -- to avoid
> deadlocks.
> 
> We originally considered your suggested flag-checking approach, but
> went the
> "is_nfsiod" way because that seems like the actual intent of checking
> for
> PF_WQ_WORKER.  The code then clarifies what's actually wanted, and we
> don't
> end up with future problems (what if nfsiod changes its PF_ flags in
> the
> future but the author doesn't know to update this function?)

If that were ever to happen, then we'd be well up the creek and without
a paddle.

Firstly, there is so much VFS work going on in nfsiod (dput(),
path_put(), iput(),....), that we really do not want to encumber it
with any PF_MEMALLOC restrictions.

Secondly, if we were to do so anyway, we definitely would want to
revisit this function, in addition to all those RPC callbacks that
would be affected.

> Do you prefer this approach?  I can send it with your Suggested-by or
> authorship.
> 
> The other way to fix this might be to create a mempool for nfs_page -
> which
> is the one place that uses nfs_io_gfp_mask() that doesn't fall back
> to a
> mempool.  We haven't tested that.

I think I prefer an approach that is aware of the limitations imposed
by the memory manager rather than one that just worries about which
workqueue we're on.

Note that one of the main differences between rpciod and nfsiod is
precisely the PF_MEMALLOC settings.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trondmy@kernel.org, trond.myklebust@hammerspace.com