Skip to contentSkip to conversation

nfsd: filecache cleanups and optimizations

PATCH In review Jeff Layton 2 participants 11 messages Jan 5, 2023 – Jan 18, 2023

Subsystems (2)
lore
Share & tools
lore

Jeff Layton

author maintainer
#0 lore raw

cover nfsd: filecache cleanups and optimizations

This is just a small set of filecache fixes and cleanups that I put together while going over this code. None of these fix critical problems, so these are probably v6.3 material.

Jeff Layton (4):
nfsd: don't open-code clear_and_wake_up_bit
nfsd: NFSD_FILE_KEY_INODE only needs to find GC'ed entries
nfsd: don't kill nfsd_files because of lease break error
nfsd: add some comments to nfsd_file_do_acquire

 fs/nfsd/filecache.c | 52 ++++++++++++++++++++++++++-------------------
 1 file changed, 30 insertions(+), 22 deletions(-)
Signature · 1 line
2.39.0
author maintainer
#1 lore raw

PATCH 1/4 nfsd: don't open-code clear_and_wake_up_bit

---
 fs/nfsd/filecache.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
fs/nfsd/filecache.c +1 -3
index 6674a86e1917..9fff1fa09d08 100644
@@ -1184,9 +1184,7 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 		status = nfserr_jukebox;
 	if (status != nfs_ok)
 		nfsd_file_unhash(nf);
-	clear_bit_unlock(NFSD_FILE_PENDING, &nf->nf_flags);
-	smp_mb__after_atomic();
-	wake_up_bit(&nf->nf_flags, NFSD_FILE_PENDING);
+	clear_and_wake_up_bit(NFSD_FILE_PENDING, &nf->nf_flags);
 	goto out;
 }
 
Signature · 1 line
2.39.0
author maintainer
#2 lore raw

PATCH 2/4 nfsd: NFSD_FILE_KEY_INODE only needs to find GC'ed entries

In reply to Jeff Layton

Since v4 files are expected to be long-lived, there's little value in closing them out of the cache when there is conflicting access.

Rename NFSD_FILE_KEY_INODE to NFSD_FILE_KEY_INODE_GC, and change the comparator to also match the gc value in the key. Change both of the current users of that key to set the gc value in the key to "true".

Also, test_bit returns bool, AFAICT, so I think we're ok to drop the !! from the condition later in the comparator.

---
 fs/nfsd/filecache.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
fs/nfsd/filecache.c +9 -5
index 9fff1fa09d08..a67b22579c6e 100644
@@ -78,7 +78,7 @@ static struct rhashtable		nfsd_file_rhash_tbl
 						____cacheline_aligned_in_smp;
 
 enum nfsd_file_lookup_type {
-	NFSD_FILE_KEY_INODE,
+	NFSD_FILE_KEY_INODE_GC,
 	NFSD_FILE_KEY_FULL,
 };
 
@@ -174,7 +174,9 @@ static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
 	const struct nfsd_file *nf = ptr;
 
 	switch (key->type) {
-	case NFSD_FILE_KEY_INODE:
+	case NFSD_FILE_KEY_INODE_GC:
+		if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
+			return 1;
 		if (nf->nf_inode != key->inode)
 			return 1;
 		break;
@@ -187,7 +189,7 @@ static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
 			return 1;
 		if (!nfsd_match_cred(nf->nf_cred, key->cred))
 			return 1;
-		if (!!test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
+		if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
 			return 1;
 		if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0)
 			return 1;
@@ -681,8 +683,9 @@ static void
 nfsd_file_queue_for_close(struct inode *inode, struct list_head *dispose)
 {
 	struct nfsd_file_lookup_key key = {
-		.type	= NFSD_FILE_KEY_INODE,
+		.type	= NFSD_FILE_KEY_INODE_GC,
 		.inode	= inode,
+		.gc	= true,
 	};
 	struct nfsd_file *nf;
 
@@ -1057,8 +1060,9 @@ bool
 nfsd_file_is_cached(struct inode *inode)
 {
 	struct nfsd_file_lookup_key key = {
-		.type	= NFSD_FILE_KEY_INODE,
+		.type	= NFSD_FILE_KEY_INODE_GC,
 		.inode	= inode,
+		.gc	= true,
 	};
 	bool ret = false;
 
Signature · 1 line
2.39.0
#3 lore raw

Re: [PATCH 2/4] nfsd: NFSD_FILE_KEY_INODE only needs to find GC'ed entries

Hi Jeff-

Quoted 4 lines · Jeff Layton

Since v4 files are expected to be long-lived, there's little value in closing them out of the cache when there is conflicting access.

Seems like targeting long-lived nfsd_file items could actually be a hazardous behavior. Are you sure it's safe to leave it in stable kernels?

Quoted 1 line

Rename NFSD_FILE_KEY_INODE to NFSD_FILE_KEY_INODE_GC,

I'd prefer to keep the name as it is. It's for searching for inodes, and adding the ".gc = true" to the search criteria is enough to show what you are looking for.

Quoted 6 lines

and change the
comparator to also match the gc value in the key. Change both of the current users of that key to set the gc value in the key to "true".

Also, test_bit returns bool, AFAICT, so I think we're ok to drop the !! from the condition later in the comparator.

And I'd prefer that you leave this clean up for another patch.

Quoted 63 lines

---
fs/nfsd/filecache.c | 14 +++++++++-----

1 file changed, 9 insertions(+), 5 deletions(-)
fs/nfsd/filecache.c +9 -5
index 9fff1fa09d08..a67b22579c6e 100644
@@ -78,7 +78,7 @@ static struct rhashtable		nfsd_file_rhash_tbl
 						____cacheline_aligned_in_smp;
 
 enum nfsd_file_lookup_type {
-	NFSD_FILE_KEY_INODE,
+	NFSD_FILE_KEY_INODE_GC,
 	NFSD_FILE_KEY_FULL,
 };
 
@@ -174,7 +174,9 @@ static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
 	const struct nfsd_file *nf = ptr;
 
 	switch (key->type) {
-	case NFSD_FILE_KEY_INODE:
+	case NFSD_FILE_KEY_INODE_GC:
+		if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
+			return 1;
 		if (nf->nf_inode != key->inode)
 			return 1;
 		break;
@@ -187,7 +189,7 @@ static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
 			return 1;
 		if (!nfsd_match_cred(nf->nf_cred, key->cred))
 			return 1;
-		if (!!test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
+		if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
 			return 1;
 		if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0)
 			return 1;
@@ -681,8 +683,9 @@ static void
 nfsd_file_queue_for_close(struct inode *inode, struct list_head *dispose)
 {
 	struct nfsd_file_lookup_key key = {
-		.type	= NFSD_FILE_KEY_INODE,
+		.type	= NFSD_FILE_KEY_INODE_GC,
 		.inode	= inode,
+		.gc	= true,
 	};
 	struct nfsd_file *nf;
 
@@ -1057,8 +1060,9 @@ bool
 nfsd_file_is_cached(struct inode *inode)
 {
 	struct nfsd_file_lookup_key key = {
-		.type	= NFSD_FILE_KEY_INODE,
+		.type	= NFSD_FILE_KEY_INODE_GC,
 		.inode	= inode,
+		.gc	= true,
 	};
 	bool ret = false;
 
Signature · 1 line
2.39.0

--
Chuck Lever

author maintainer
#4 lore raw

Re: [PATCH 2/4] nfsd: NFSD_FILE_KEY_INODE only needs to find GC'ed entries

Quoted 12 lines · Chuck Lever III

Hi Jeff-

Quoted 4 lines · Jeff Layton

Since v4 files are expected to be long-lived, there's little value in closing them out of the cache when there is conflicting access.

Seems like targeting long-lived nfsd_file items could actually be a hazardous behavior. Are you sure it's safe to leave it in stable kernels?

Basically it just means we end up doing more opens than are needed in this situation.

The notifiers just unnecessarily unhash a nfsd_file in this case, even though it doesn't have any chance of freeing it until the client issues a CLOSE, due to the persistent references held by the stateids.

So, this really is just an optimization and not a "real bug".

Quoted 7 lines
Quoted 2 lines

Rename NFSD_FILE_KEY_INODE to NFSD_FILE_KEY_INODE_GC,

I'd prefer to keep the name as it is. It's for searching for inodes, and adding the ".gc = true" to the search criteria is enough to show what you are looking for.

Ok.

Quoted 10 lines
Quoted 7 lines

and change the
comparator to also match the gc value in the key. Change both of the current users of that key to set the gc value in the key to "true".

Also, test_bit returns bool, AFAICT, so I think we're ok to drop the !! from the condition later in the comparator.

And I'd prefer that you leave this clean up for another patch.

Ok, I'll drop that hunk.

Quoted 70 lines
Quoted 64 lines

---
fs/nfsd/filecache.c | 14 +++++++++-----

1 file changed, 9 insertions(+), 5 deletions(-)
fs/nfsd/filecache.c +9 -5
index 9fff1fa09d08..a67b22579c6e 100644
@@ -78,7 +78,7 @@ static struct rhashtable		nfsd_file_rhash_tbl
 						____cacheline_aligned_in_smp;
 
 enum nfsd_file_lookup_type {
-	NFSD_FILE_KEY_INODE,
+	NFSD_FILE_KEY_INODE_GC,
 	NFSD_FILE_KEY_FULL,
 };
 
@@ -174,7 +174,9 @@ static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
 	const struct nfsd_file *nf = ptr;
 
 	switch (key->type) {
-	case NFSD_FILE_KEY_INODE:
+	case NFSD_FILE_KEY_INODE_GC:
+		if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
+			return 1;
 		if (nf->nf_inode != key->inode)
 			return 1;
 		break;
@@ -187,7 +189,7 @@ static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
 			return 1;
 		if (!nfsd_match_cred(nf->nf_cred, key->cred))
 			return 1;
-		if (!!test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
+		if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
 			return 1;
 		if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0)
 			return 1;
@@ -681,8 +683,9 @@ static void
 nfsd_file_queue_for_close(struct inode *inode, struct list_head *dispose)
 {
 	struct nfsd_file_lookup_key key = {
-		.type	= NFSD_FILE_KEY_INODE,
+		.type	= NFSD_FILE_KEY_INODE_GC,
 		.inode	= inode,
+		.gc	= true,
 	};
 	struct nfsd_file *nf;
 
@@ -1057,8 +1060,9 @@ bool
 nfsd_file_is_cached(struct inode *inode)
 {
 	struct nfsd_file_lookup_key key = {
-		.type	= NFSD_FILE_KEY_INODE,
+		.type	= NFSD_FILE_KEY_INODE_GC,
 		.inode	= inode,
+		.gc	= true,
 	};
 	bool ret = false;
 
Signature · 1 line
2.39.0

--
Chuck Lever

Signature · 1 line
Jeff Layton <[email protected]>
#5 lore raw

Re: [PATCH 2/4] nfsd: NFSD_FILE_KEY_INODE only needs to find GC'ed entries

Quoted 18 lines · Jeff Layton
Quoted 12 lines · Chuck Lever III

Hi Jeff-

Quoted 4 lines · Jeff Layton

Since v4 files are expected to be long-lived, there's little value in closing them out of the cache when there is conflicting access.

Seems like targeting long-lived nfsd_file items could actually be a hazardous behavior. Are you sure it's safe to leave it in stable kernels?

Quoted 8 lines

Basically it just means we end up doing more opens than are needed in this situation.

The notifiers just unnecessarily unhash a nfsd_file in this case, even though it doesn't have any chance of freeing it until the client issues a CLOSE, due to the persistent references held by the stateids.

So, this really is just an optimization and not a "real bug".

Good to know. Let's add this to the patch description when you redrive...?

Quoted 83 lines
Quoted 6 lines
Quoted 1 line

Rename NFSD_FILE_KEY_INODE to NFSD_FILE_KEY_INODE_GC,

I'd prefer to keep the name as it is. It's for searching for inodes, and adding the ".gc = true" to the search criteria is enough to show what you are looking for.

Ok.

Quoted 10 lines
Quoted 7 lines

and change the
comparator to also match the gc value in the key. Change both of the current users of that key to set the gc value in the key to "true".

Also, test_bit returns bool, AFAICT, so I think we're ok to drop the !! from the condition later in the comparator.

And I'd prefer that you leave this clean up for another patch.

Ok, I'll drop that hunk.

Quoted 61 lines
Quoted 61 lines

---
fs/nfsd/filecache.c | 14 +++++++++-----

1 file changed, 9 insertions(+), 5 deletions(-)
fs/nfsd/filecache.c +9 -5
index 9fff1fa09d08..a67b22579c6e 100644
@@ -78,7 +78,7 @@ static struct rhashtable		nfsd_file_rhash_tbl
 						____cacheline_aligned_in_smp;
 
 enum nfsd_file_lookup_type {
-	NFSD_FILE_KEY_INODE,
+	NFSD_FILE_KEY_INODE_GC,
 	NFSD_FILE_KEY_FULL,
 };
 
@@ -174,7 +174,9 @@ static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
 	const struct nfsd_file *nf = ptr;
 
 	switch (key->type) {
-	case NFSD_FILE_KEY_INODE:
+	case NFSD_FILE_KEY_INODE_GC:
+		if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
+			return 1;
 		if (nf->nf_inode != key->inode)
 			return 1;
 		break;
@@ -187,7 +189,7 @@ static int nfsd_file_obj_cmpfn(struct rhashtable_compare_arg *arg,
 			return 1;
 		if (!nfsd_match_cred(nf->nf_cred, key->cred))
 			return 1;
-		if (!!test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
+		if (test_bit(NFSD_FILE_GC, &nf->nf_flags) != key->gc)
 			return 1;
 		if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0)
 			return 1;
@@ -681,8 +683,9 @@ static void
 nfsd_file_queue_for_close(struct inode *inode, struct list_head *dispose)
 {
 	struct nfsd_file_lookup_key key = {
-		.type	= NFSD_FILE_KEY_INODE,
+		.type	= NFSD_FILE_KEY_INODE_GC,
 		.inode	= inode,
+		.gc	= true,
 	};
 	struct nfsd_file *nf;
 
@@ -1057,8 +1060,9 @@ bool
 nfsd_file_is_cached(struct inode *inode)
 {
 	struct nfsd_file_lookup_key key = {
-		.type	= NFSD_FILE_KEY_INODE,
+		.type	= NFSD_FILE_KEY_INODE_GC,
 		.inode	= inode,
+		.gc	= true,
 	};
 	bool ret = false;
 
Quoted 12 lines
Quoted 9 lines
Quoted 3 lines
Signature · 1 line
2.39.0

--
Chuck Lever

Signature · 1 line
Jeff Layton <[email protected]>

--
Chuck Lever

author maintainer
#6 lore raw

PATCH 3/4 nfsd: don't kill nfsd_files because of lease break error

In reply to Jeff Layton

An error from break_lease is non-fatal, so we needn't destroy the nfsd_file in that case. Just put the reference like we normally would and return the error.

---
 fs/nfsd/filecache.c | 29 +++++++++++++++--------------
 1 file changed, 15 insertions(+), 14 deletions(-)
fs/nfsd/filecache.c +15 -14
index a67b22579c6e..f0ca9501edb2 100644
@@ -1113,7 +1113,7 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	nf = nfsd_file_alloc(&key, may_flags);
 	if (!nf) {
 		status = nfserr_jukebox;
-		goto out_status;
+		goto out;
 	}
 
 	ret = rhashtable_lookup_insert_key(&nfsd_file_rhash_tbl,
@@ -1122,13 +1122,11 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	if (likely(ret == 0))
 		goto open_file;
 
-	nfsd_file_slab_free(&nf->nf_rcu);
-	nf = NULL;
 	if (ret == -EEXIST)
 		goto retry;
 	trace_nfsd_file_insert_err(rqstp, key.inode, may_flags, ret);
 	status = nfserr_jukebox;
-	goto out_status;
+	goto construction_err;
 
 wait_for_construction:
 	wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
@@ -1138,28 +1136,24 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 		trace_nfsd_file_cons_err(rqstp, key.inode, may_flags, nf);
 		if (!open_retry) {
 			status = nfserr_jukebox;
-			goto out;
+			goto construction_err;
 		}
 		open_retry = false;
-		if (refcount_dec_and_test(&nf->nf_ref))
-			nfsd_file_free(nf);
 		goto retry;
 	}
-
 	this_cpu_inc(nfsd_file_cache_hits);
 
 	status = nfserrno(nfsd_open_break_lease(file_inode(nf->nf_file), may_flags));
+	if (status != nfs_ok) {
+		nfsd_file_put(nf);
+		nf = NULL;
+	}
+
 out:
 	if (status == nfs_ok) {
 		this_cpu_inc(nfsd_file_acquisitions);
 		*pnf = nf;
-	} else {
-		if (refcount_dec_and_test(&nf->nf_ref))
-			nfsd_file_free(nf);
-		nf = NULL;
 	}
-
-out_status:
 	put_cred(key.cred);
 	trace_nfsd_file_acquire(rqstp, key.inode, may_flags, nf, status);
 	return status;
@@ -1189,6 +1183,13 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	if (status != nfs_ok)
 		nfsd_file_unhash(nf);
 	clear_and_wake_up_bit(NFSD_FILE_PENDING, &nf->nf_flags);
+	if (status == nfs_ok)
+		goto out;
+
+construction_err:
+	if (refcount_dec_and_test(&nf->nf_ref))
+		nfsd_file_free(nf);
+	nf = NULL;
 	goto out;
 }
 
Signature · 1 line
2.39.0
author maintainer
#7 lore raw

Re: [PATCH 3/4] nfsd: don't kill nfsd_files because of lease break error

Quoted 76 lines · Jeff Layton

An error from break_lease is non-fatal, so we needn't destroy the nfsd_file in that case. Just put the reference like we normally would and return the error.

---
 fs/nfsd/filecache.c | 29 +++++++++++++++--------------
 1 file changed, 15 insertions(+), 14 deletions(-)
fs/nfsd/filecache.c +8 -14
index a67b22579c6e..f0ca9501edb2 100644
@@ -1113,7 +1113,7 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	nf = nfsd_file_alloc(&key, may_flags);
 	if (!nf) {
 		status = nfserr_jukebox;
-		goto out_status;
+		goto out;
 	}
 
 	ret = rhashtable_lookup_insert_key(&nfsd_file_rhash_tbl,
@@ -1122,13 +1122,11 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	if (likely(ret == 0))
 		goto open_file;
 
-	nfsd_file_slab_free(&nf->nf_rcu);
-	nf = NULL;
 	if (ret == -EEXIST)
 		goto retry;
 	trace_nfsd_file_insert_err(rqstp, key.inode, may_flags, ret);
 	status = nfserr_jukebox;
-	goto out_status;
+	goto construction_err;
 
 wait_for_construction:
 	wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
@@ -1138,28 +1136,24 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 		trace_nfsd_file_cons_err(rqstp, key.inode, may_flags, nf);
 		if (!open_retry) {
 			status = nfserr_jukebox;
-			goto out;
+			goto construction_err;
 		}
 		open_retry = false;
-		if (refcount_dec_and_test(&nf->nf_ref))
-			nfsd_file_free(nf);
 		goto retry;
 	}
-
 	this_cpu_inc(nfsd_file_cache_hits);
 
 	status = nfserrno(nfsd_open_break_lease(file_inode(nf->nf_file), may_flags));
+	if (status != nfs_ok) {
+		nfsd_file_put(nf);
+		nf = NULL;
+	}
+
 out:
 	if (status == nfs_ok) {
 		this_cpu_inc(nfsd_file_acquisitions);
 		*pnf = nf;
-	} else {
-		if (refcount_dec_and_test(&nf->nf_ref))
-			nfsd_file_free(nf);
-		nf = NULL;
 	}
-
-out_status:
 	put_cred(key.cred);
 	trace_nfsd_file_acquire(rqstp, key.inode, may_flags, nf, status);
 	return status;
@@ -1189,6 +1183,13 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	if (status != nfs_ok)
 		nfsd_file_unhash(nf);
 	clear_and_wake_up_bit(NFSD_FILE_PENDING, &nf->nf_flags);

fs/nfsd/filecache.c nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp, · line 1195

Quoted 10 lines
fs/nfsd/filecache.c +7 -0
+	if (status == nfs_ok)
+		goto out;
+
+construction_err:
+	if (refcount_dec_and_test(&nf->nf_ref))
+		nfsd_file_free(nf);
+	nf = NULL;
 	goto out;
 }
 

Chuck, ping?

You never responded to this patch and I don't see it in your tree. Any thoughts?

Signature · 1 line
Jeff Layton <[email protected]>
#8 lore raw

Re: [PATCH 3/4] nfsd: don't kill nfsd_files because of lease break error

Quoted 94 lines · Jeff Layton
Quoted 76 lines · Jeff Layton

An error from break_lease is non-fatal, so we needn't destroy the nfsd_file in that case. Just put the reference like we normally would and return the error.

---
fs/nfsd/filecache.c | 29 +++++++++++++++--------------

1 file changed, 15 insertions(+), 14 deletions(-)
fs/nfsd/filecache.c +2 -4
index a67b22579c6e..f0ca9501edb2 100644
@@ -1113,7 +1113,7 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	nf = nfsd_file_alloc(&key, may_flags);
 	if (!nf) {
 		status = nfserr_jukebox;
-		goto out_status;
+		goto out;
 	}
 
 	ret = rhashtable_lookup_insert_key(&nfsd_file_rhash_tbl,
@@ -1122,13 +1122,11 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	if (likely(ret == 0))
 		goto open_file;
 
-	nfsd_file_slab_free(&nf->nf_rcu);
-	nf = NULL;
 	if (ret == -EEXIST)
 		goto retry;
 	trace_nfsd_file_insert_err(rqstp, key.inode, may_flags, ret);
 	status = nfserr_jukebox;
-	goto out_status;
+	goto construction_err;
 

wait_for_construction:

	wait_on_bit(&nf->nf_flags, NFSD_FILE_PENDING, TASK_UNINTERRUPTIBLE);
fs/nfsd/filecache.c +6 -4
@@ -1138,28 +1136,24 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 		trace_nfsd_file_cons_err(rqstp, key.inode, may_flags, nf);
 		if (!open_retry) {
 			status = nfserr_jukebox;
-			goto out;
+			goto construction_err;
 		}
 		open_retry = false;
-		if (refcount_dec_and_test(&nf->nf_ref))
-			nfsd_file_free(nf);
 		goto retry;
 	}
-
 	this_cpu_inc(nfsd_file_cache_hits);
 
 	status = nfserrno(nfsd_open_break_lease(file_inode(nf->nf_file), may_flags));
+	if (status != nfs_ok) {
+		nfsd_file_put(nf);
+		nf = NULL;
+	}
+

out:

	if (status == nfs_ok) {
		this_cpu_inc(nfsd_file_acquisitions);
		*pnf = nf;
fs/nfsd/filecache.c +0 -4
-	} else {
-		if (refcount_dec_and_test(&nf->nf_ref))
-			nfsd_file_free(nf);
-		nf = NULL;
	}
fs/nfsd/filecache.c +0 -2
-
-out_status:
	put_cred(key.cred);
	trace_nfsd_file_acquire(rqstp, key.inode, may_flags, nf, status);
	return status;
fs/nfsd/filecache.c +0 -0
@@ -1189,6 +1183,13 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	if (status != nfs_ok)
 		nfsd_file_unhash(nf);
 	clear_and_wake_up_bit(NFSD_FILE_PENDING, &nf->nf_flags);

fs/nfsd/filecache.c nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp, · line 1195

Quoted 10 lines
fs/nfsd/filecache.c +7 -0
+	if (status == nfs_ok)
+		goto out;
+
+construction_err:
+	if (refcount_dec_and_test(&nf->nf_ref))
+		nfsd_file_free(nf);
+	nf = NULL;
 	goto out;
 }
 

Chuck, ping?

Quoted 2 lines

You never responded to this patch and I don't see it in your tree. Any thoughts?

It's been in nfsd-next for a bit, but I had to drop a bunch of patches from nfsd-next until yesterday's -rc PR was applied upstream because they depended on one of Dai's crasher fixes.

That's been sorted, and I've restored these to nfsd-next.

--
Chuck Lever

author maintainer
#9 lore raw

PATCH 4/4 nfsd: add some comments to nfsd_file_do_acquire

In reply to Jeff Layton

David Howells mentioned that he found this bit of code confusing, so sprinkle in some comments to clarify.

---
 fs/nfsd/filecache.c | 5 +++++
 1 file changed, 5 insertions(+)
fs/nfsd/filecache.c +5 -0
index f0ca9501edb2..5724baad09ec 100644
@@ -1105,6 +1105,11 @@ nfsd_file_do_acquire(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	rcu_read_unlock();
 
 	if (nf) {
+		/*
+		 * If the nf is on the LRU then it holds an extra reference
+		 * that must be put if it's removed. It had better not be
+		 * the last one however, since we should hold another.
+		 */
 		if (nfsd_file_lru_remove(nf))
 			WARN_ON_ONCE(refcount_dec_and_test(&nf->nf_ref));
 		goto wait_for_construction;
Signature · 1 line
2.39.0
#10 lore raw
In reply to Jeff Layton
Quoted 8 lines · Jeff Layton

This is just a small set of filecache fixes and cleanups that I put together while going over this code. None of these fix critical problems, so these are probably v6.3 material.

Quoted 12 lines

Jeff Layton (4):
nfsd: don't open-code clear_and_wake_up_bit
nfsd: NFSD_FILE_KEY_INODE only needs to find GC'ed entries
nfsd: don't kill nfsd_files because of lease break error
nfsd: add some comments to nfsd_file_do_acquire

fs/nfsd/filecache.c | 52 ++++++++++++++++++++++++++-------------------

1 file changed, 30 insertions(+), 22 deletions(-)
Signature · 1 line
2.39.0

These have been applied to nfsd's for-next. Thanks!

--
Chuck Lever