[PATCH 01/11] nfsd: don't hold inode_lock while waiting for a delegation in nfsd_setattr()
From: Jeff Layton
Date: Fri Nov 20 2026 - 11:41:24 EST
When notify_change() breaks a delegation, nfsd_setattr() waits up to
NFSD_DELEGRETURN_TIMEOUT for the holder to return it and then retries
once. It waits while still holding inode_lock, but what the holder
sends before DELEGRETURN can need that lock: WRITEs that flush its
dirty data take i_rwsem exclusively on ext4 and xfs, and a SETATTR
with delegated timestamps in the DELEGRETURN compound goes through
nfsd_setattr() itself. The wait then always times out and SETATTR
returns NFS4ERR_DELAY, after holding up the return it was waiting for.
Drop inode_lock across the wait, as callers of break_deleg_wait() do.
The file can change while the lock is dropped, so start over: check
that the truncate is allowed, take fresh pre-op attributes, and check
the guard again, so that all of them stay atomic with the attempt that
changes the inode. fh_fill_pre_attrs() does nothing once pre-op
attributes are saved, so restore fh_pre_saved to its value on entry.
If the wait times out, still try once more rather than finishing with
the attributes from before the wait. A lease that is already being
broken isn't broken again, so the retry just gets -EAGAIN.
Don't start over once the size change is done, though, as when a
filesystem returns -EAGAIN for the other attributes: the retry would
truncate again, over whatever was written while the lock was dropped,
and its new pre-op attributes would hide the truncate. Return the error
instead, as if the wait had failed.
Fixes: 34b91dda7124 ("NFSD: Make nfsd4_setattr() wait before returning NFS4ERR_DELAY")
---
fs/nfsd/vfs.c | 42 +++++++++++++++++++++++++++---------------
1 file changed, 27 insertions(+), 15 deletions(-)
diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
index 6905d3285ce1..4f3e480a3981 100644
--- a/fs/nfsd/vfs.c
+++ b/fs/nfsd/vfs.c
@@ -550,7 +550,9 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp,
int host_err = 0;
bool get_write_count;
bool size_change = (iap->ia_valid & ATTR_SIZE);
- int retries;
+ bool pre_saved = fhp->fh_pre_saved;
+ bool retried = false;
+ struct iattr attrs;
trace_nfsd_vfs_setattr(rqstp, fhp, iap, guardtime);
@@ -612,6 +614,7 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp,
}
}
+retry:
inode_lock(inode);
if (size_change) {
err = nfsd_may_truncate(rqstp, fhp, iap);
@@ -631,20 +634,29 @@ nfsd_setattr(struct svc_rqst *rqstp, struct svc_fh *fhp,
}
}
- for (retries = 1;;) {
- struct iattr attrs;
-
- /*
- * notify_change() can alter its iattr argument, making
- * @iap unsuitable for submission multiple times. Make a
- * copy for every loop iteration.
- */
- attrs = *iap;
- host_err = __nfsd_setattr(dentry, &attrs);
- if (host_err != -EAGAIN || !retries--)
- break;
- if (!nfsd_wait_for_delegreturn(rqstp, inode))
- break;
+ /*
+ * notify_change() can alter its iattr argument, making @iap
+ * unsuitable for submission multiple times. Pass a copy.
+ */
+ attrs = *iap;
+ host_err = __nfsd_setattr(dentry, &attrs);
+ /*
+ * A delegation was broken. Give its holder a chance to return
+ * it, without inode_lock: what it sends first, such as WRITEs
+ * that flush its dirty data, can need the lock. The file can
+ * change meanwhile, so start over, taking fresh pre-op
+ * attributes and checking the guard again. That's only safe if
+ * nothing was changed: a size change that is already done
+ * (__nfsd_setattr() clears ATTR_SIZE) would be repeated over
+ * whatever was written after it, so return the error then.
+ */
+ if (host_err == -EAGAIN && !retried &&
+ !(size_change && !(attrs.ia_valid & ATTR_SIZE))) {
+ inode_unlock(inode);
+ retried = true;
+ nfsd_wait_for_delegreturn(rqstp, inode);
+ fhp->fh_pre_saved = pre_saved;
+ goto retry;
}
if (attr->na_seclabel && attr->na_seclabel->len)
attr->na_labelerr = security_inode_setsecctx(dentry,
--
2.56.1