cifs: Revert "smb: client: Avoid race in open_cached_dir with lease breaks"
Subsystems (2)
PATCH 1/7 cifs: Revert "smb: client: Avoid race in open_cached_dir with lease breaks"
From: Shyam Prasad N <[email protected]>
This reverts commit 3ca02e63edccb78ef3659bebc68579c7224a6ca2. Upcoming patches in this code path introduces a new mutex, which takes care of the same problem which this change fixes, but with the use of mutex for the same section. So the net effect is that there will be no blocking with spinlocks held.
--- fs/smb/client/cached_dir.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)
fs/smb/client/cached_dir.c +8 -2
@@ -29,6 +29,7 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids, {struct cached_fid *cfid; +spin_lock (&cfids->cfid_list_lock);list_for_each_entry (cfid, &cfids->entries, entry) {if (!strcmp (cfid->path, path)) {/*
@@ -37,20 +38,25 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids,* being deleted due to a lease break. */ if (!cfid->time || !cfid->has_lease) { +spin_unlock (&cfids->cfid_list_lock);return NULL ; }kref_get (&cfid->refcount); +spin_unlock (&cfids->cfid_list_lock);return cfid; } }if (lookup_only) { +spin_unlock (&cfids->cfid_list_lock);return NULL ; }if (cfids->num_entries >= max_cached_dirs) { +spin_unlock (&cfids->cfid_list_lock);return NULL ; } cfid =init_cached_dir (path);if (cfid ==NULL ) { +spin_unlock (&cfids->cfid_list_lock);return NULL ; } cfid->cfids = cfids;
@@ -68,6 +74,7 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids,*/ cfid->has_lease = true; +spin_unlock (&cfids->cfid_list_lock);return cfid; }
@@ -181,10 +188,8 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,if (!utf16_path)return -ENOMEM ; -spin_lock (&cfids->cfid_list_lock); cfid =find_or_create_cached_dir (cfids, path, lookup_only, tcon->max_cached_dirs);if (cfid ==NULL ) { -spin_unlock (&cfids->cfid_list_lock);kfree (utf16_path);return -ENOENT ; }
@@ -193,6 +198,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,* Otherwise, it is either a new entry or laundromat worker removed it * from @cfids->entries. Caller will put last reference if the latter. */ +spin_lock (&cfids->cfid_list_lock);if (cfid->has_lease && cfid->time) {spin_unlock (&cfids->cfid_list_lock); *ret_cfid = cfid;
Signature · 1 line
2.43.0
PATCH 2/7 cifs: protect cfid accesses with fid_lock
From: Shyam Prasad N <[email protected]>
There are several accesses to cfid structure today without locking fid_lock. This can lead to race conditions that are hard to debug.
With this change, I'm trying to make sure that accesses to cfid struct members happen with fid_lock held.
--- fs/smb/client/cached_dir.c | 86 ++++++++++++++++++++++---------------- 1 file changed, 49 insertions(+), 37 deletions(-)
fs/smb/client/cached_dir.c +49 -37
@@ -31,6 +31,7 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids,spin_lock (&cfids->cfid_list_lock);list_for_each_entry (cfid, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (!strcmp (cfid->path, path)) {/* * If it doesn't have a lease it is either not yet
@@ -38,13 +39,16 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids,* being deleted due to a lease break. */ if (!cfid->time || !cfid->has_lease) { +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);return NULL ; }kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);return cfid; } +spin_unlock (&cfid->fid_lock); }if (lookup_only) {spin_unlock (&cfids->cfid_list_lock);
@@ -193,19 +197,20 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,kfree (utf16_path);return -ENOENT ; } +/* * Return cached fid if it is valid (has a lease and has a time). * Otherwise, it is either a new entry or laundromat worker removed it * from @cfids->entries. Caller will put last reference if the latter. */ -spin_lock (&cfids->cfid_list_lock); +spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) { -spin_unlock (&cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock); *ret_cfid = cfid;kfree (utf16_path);return 0 ; } -spin_unlock (&cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock);/* * Skip any prefix paths in @path as lookup_noperm_positive_unlocked() ends up
@@ -220,17 +225,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, goto out; } -if (!npath[0 ]) { - dentry =dget (cifs_sb->root); - }else { - dentry =path_to_dentry (cifs_sb, npath); -if (IS_ERR (dentry)) { - rc = -ENOENT ; - goto out; - } - } - cfid->dentry = dentry; - cfid->tcon = tcon;/* * We do not hold the lock for the open because in case
@@ -302,9 +296,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, } goto oshr_free; } - cfid->is_open = true; - -spin_lock (&cfids->cfid_list_lock); o_rsp = (struct smb2_create_rsp *)rsp_iov[0 ].iov_base; oparms.fid->persistent_fid = o_rsp->PersistentFileId;
@@ -315,7 +306,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,if (o_rsp->OplockLevel !=SMB2_OPLOCK_LEVEL_LEASE ) { -spin_unlock (&cfids->cfid_list_lock); rc = -EINVAL ; goto oshr_free; }
@@ -324,21 +314,15 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, &oparms.fid->epoch, oparms.fid->lease_key, &oplock, NULL, NULL); -if (rc) { -spin_unlock (&cfids->cfid_list_lock); +if (rc) goto oshr_free; - } rc = -EINVAL ; -if (!(oplock &SMB2_LEASE_READ_CACHING_HE )) { -spin_unlock (&cfids->cfid_list_lock); +if (!(oplock &SMB2_LEASE_READ_CACHING_HE )) goto oshr_free; - } qi_rsp = (struct smb2_query_info_rsp *)rsp_iov[1 ].iov_base; -if (le32_to_cpu (qi_rsp->OutputBufferLength) <sizeof (struct smb2_file_all_info )) { -spin_unlock (&cfids->cfid_list_lock); +if (le32_to_cpu (qi_rsp->OutputBufferLength) <sizeof (struct smb2_file_all_info )) goto oshr_free; - } if (!smb2_validate_and_copy_iov( le16_to_cpu(qi_rsp->OutputBufferOffset), sizeof(struct smb2_file_all_info),
@@ -346,10 +330,25 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, (char *)&cfid->file_all_info)) cfid->file_all_info_is_valid = true; - cfid->time = jiffies; -spin_unlock (&cfids->cfid_list_lock);/* At this point the directory handle is fully cached */ rc =0 ; +if (!cfid->dentry) { +if (!npath[0 ]) { + dentry =dget (cifs_sb->root); + }else { + dentry =path_to_dentry (cifs_sb, npath); +if (IS_ERR (dentry)) { + rc = -ENOENT ; + goto out; + } + } + } +spin_lock (&cfid->fid_lock); + cfid->dentry = dentry; + cfid->tcon = tcon; + cfid->is_open = true; + cfid->time = jiffies; +spin_unlock (&cfid->fid_lock); oshr_free:SMB2_open_free (&rqst[0 ]);
@@ -364,6 +363,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->on_list = false; cfids->num_entries--; } +spin_lock (&cfid->fid_lock);if (cfid->has_lease) {/* * We are guaranteed to have two references at this
@@ -373,6 +373,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->has_lease = false;kref_put (&cfid->refcount, smb2_close_cached_fid); } +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);kref_put (&cfid->refcount, smb2_close_cached_fid);
@@ -401,13 +402,16 @@ int open_cached_dir_by_dentry(struct cifs_tcon *tcon,spin_lock (&cfids->cfid_list_lock);list_for_each_entry (cfid, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (dentry && cfid->dentry == dentry) {cifs_dbg (FYI ,"found a cached file handle by dentry\n" );kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock); *ret_cfid = cfid;spin_unlock (&cfids->cfid_list_lock);return 0 ; } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);return -ENOENT ;
@@ -428,8 +432,11 @@ smb2_close_cached_fid(struct kref *ref) }spin_unlock (&cfid->cfids->cfid_list_lock); -dput (cfid->dentry); - cfid->dentry =NULL ; +/* no locking necessary as we're the last user of this cfid */ +if (cfid->dentry) { +dput (cfid->dentry); + cfid->dentry =NULL ; + }if (cfid->is_open) { rc = SMB2_close(0, cfid->tcon, cfid->fid.persistent_fid,
@@ -452,12 +459,13 @@ void drop_cached_dir_by_name(const unsigned int xid, struct cifs_tcon *tcon,cifs_dbg (FYI ,"no cached dir found for rmdir(%s)\n" , name);return ; } -spin_lock (&cfid->cfids->cfid_list_lock); +spin_lock (&cfid->fid_lock);if (cfid->has_lease) { +/* mark as invalid */ cfid->has_lease = false;kref_put (&cfid->refcount, smb2_close_cached_fid); } -spin_unlock (&cfid->cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock);close_cached_dir (cfid); }
@@ -539,6 +547,7 @@ void invalidate_all_cached_dirs(struct cifs_tcon *tcon) cfids->num_entries--; cfid->is_open = false; cfid->on_list = false; +spin_lock (&cfid->fid_lock);if (cfid->has_lease) {/* * The lease was never cancelled from the server,
@@ -547,6 +556,7 @@ void invalidate_all_cached_dirs(struct cifs_tcon *tcon) cfid->has_lease = false; }else kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock); }/* * Queue dropping of the dentries once locks have been dropped
@@ -601,6 +611,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])spin_lock (&cfids->cfid_list_lock); list_for_each_entry(cfid, &cfids->entries, entry) { + spin_lock(&cfid->fid_lock); if (cfid->has_lease && !memcmp(lease_key, cfid->fid.lease_key,
@@ -613,6 +624,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])*/ list_del (&cfid->entry); cfid->on_list = false; +spin_unlock (&cfid->fid_lock); cfids->num_entries--; ++tcon->tc_count;
@@ -622,6 +634,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])spin_unlock (&cfids->cfid_list_lock);return true; } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);return false;
@@ -657,9 +670,6 @@ static void free_cached_dir(struct cached_fid *cfid)WARN_ON (work_pending (&cfid->close_work));WARN_ON (work_pending (&cfid->put_work)); -dput (cfid->dentry); - cfid->dentry =NULL ; -/* * Delete all cached dirent names */
@@ -704,6 +714,7 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (cfid->time &&time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { cfid->on_list = false;
@@ -718,6 +729,7 @@ static void cfids_laundromat_worker(struct work_struct *work) }else kref_get (&cfid->refcount); } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);
@@ -728,8 +740,8 @@ static void cfids_laundromat_worker(struct work_struct *work) dentry = cfid->dentry; cfid->dentry =NULL ;spin_unlock (&cfid->fid_lock); -dput (dentry); +if (cfid->is_open) {spin_lock (&cifs_tcp_ses_lock); ++cfid->tcon->tc_count;
Signature · 1 line
2.43.0
Re: [PATCH 2/7] cifs: protect cfid accesses with fid_lock
Quoted 72 lines · On 2025-06-04 15:48:11 +0530, [email protected] wrote:
From: Shyam Prasad N <[email protected]>
There are several accesses to cfid structure today without locking
fid_lock. This can lead to race conditions that are hard to debug.With this change, I'm trying to make sure that accesses to cfid struct members happen with
fid_lockheld.Signed-off-by Shyam Prasad N <[email protected]>---
fs/smb/client/cached_dir.c| 86 ++++++++++++++++++++++----------------1 file changed, 49 insertions(+), 37 deletions(-)fs/smb/client/cached_dir.c +8
-4@@ -31,6 +31,7 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids,spin_lock (&cfids->cfid_list_lock);list_for_each_entry (cfid, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (!strcmp (cfid->path, path)) {/* * If it doesn't have a lease it is either not yet @@ -38,13 +39,16 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids,* being deleted due to a lease break. */ if (!cfid->time || !cfid->has_lease) { +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);return NULL ; }kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);return cfid; } +spin_unlock (&cfid->fid_lock); }if (lookup_only) {spin_unlock (&cfids->cfid_list_lock);@@ -193,19 +197,20 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,kfree (utf16_path);return -ENOENT ; } +/* * Return cached fid if it is valid (has a lease and has a time). * Otherwise, it is either a new entry or laundromat worker removed it * from @cfids->entries. Caller will put last reference if the latter. */ -spin_lock (&cfids->cfid_list_lock); +spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) { -spin_unlock (&cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock); *ret_cfid = cfid;kfree (utf16_path);return 0 ; } -spin_unlock (&cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock);/* * Skip any prefix paths in @path as lookup_noperm_positive_unlocked() ends up @@ -220,17 +225,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, goto out; } -if (!npath[0 ]) {
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 233
Quoted 10 lines
fs/smb/client/cached_dir.c +0
-10- dentry =dget (cifs_sb->root); - }else { - dentry =path_to_dentry (cifs_sb, npath); -if (IS_ERR (dentry)) { - rc = -ENOENT ; - goto out; - } - } - cfid->dentry = dentry; - cfid->tcon = tcon;
I think moving this down below to after the "At this point the directory handle is fully cached" comment is going to regress c353ee4fb119 ("smb: Initialize cfid->tcon before performing network ops").
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 305
Quoted 7 lines
fs/smb/client/cached_dir.c +0
-1/* * We do not hold the lock for the open because in case @@ -302,9 +296,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, } goto oshr_free; } - cfid->is_open = true;
Moving this down below is going to regress part of 66d45ca1350a ("cifs: Check the lease context if we actually got a lease"), I think. If parsing the lease fails, the cfid is disposed, but since is_open is false, the code won't call SMB2_close.
Quoted 207 lines
fs/smb/client/cached_dir.c +20
-14- -spin_lock (&cfids->cfid_list_lock); o_rsp = (struct smb2_create_rsp *)rsp_iov[0 ].iov_base; oparms.fid->persistent_fid = o_rsp->PersistentFileId;@@ -315,7 +306,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,if (o_rsp->OplockLevel !=SMB2_OPLOCK_LEVEL_LEASE ) { -spin_unlock (&cfids->cfid_list_lock); rc = -EINVAL ; goto oshr_free; }@@ -324,21 +314,15 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, &oparms.fid->epoch, oparms.fid->lease_key, &oplock, NULL, NULL); -if (rc) { -spin_unlock (&cfids->cfid_list_lock); +if (rc) goto oshr_free; - } rc = -EINVAL ; -if (!(oplock &SMB2_LEASE_READ_CACHING_HE )) { -spin_unlock (&cfids->cfid_list_lock); +if (!(oplock &SMB2_LEASE_READ_CACHING_HE )) goto oshr_free; - } qi_rsp = (struct smb2_query_info_rsp *)rsp_iov[1 ].iov_base; -if (le32_to_cpu (qi_rsp->OutputBufferLength) <sizeof (struct smb2_file_all_info )) { -spin_unlock (&cfids->cfid_list_lock); +if (le32_to_cpu (qi_rsp->OutputBufferLength) <sizeof (struct smb2_file_all_info )) goto oshr_free; - } if (!smb2_validate_and_copy_iov( le16_to_cpu(qi_rsp->OutputBufferOffset), sizeof(struct smb2_file_all_info),@@ -346,10 +330,25 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, (char *)&cfid->file_all_info)) cfid->file_all_info_is_valid = true; - cfid->time = jiffies; -spin_unlock (&cfids->cfid_list_lock);/* At this point the directory handle is fully cached */ rc =0 ; +if (!cfid->dentry) { +if (!npath[0 ]) { + dentry =dget (cifs_sb->root); + }else { + dentry =path_to_dentry (cifs_sb, npath); +if (IS_ERR (dentry)) { + rc = -ENOENT ; + goto out; + } + } + } +spin_lock (&cfid->fid_lock); + cfid->dentry = dentry; + cfid->tcon = tcon; + cfid->is_open = true; + cfid->time = jiffies; +spin_unlock (&cfid->fid_lock);
oshr_free:SMB2_open_free(&rqst[0]);fs/smb/client/cached_dir.c +21
-8@@ -364,6 +363,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->on_list = false; cfids->num_entries--; } +spin_lock (&cfid->fid_lock);if (cfid->has_lease) {/* * We are guaranteed to have two references at this @@ -373,6 +373,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->has_lease = false;kref_put (&cfid->refcount, smb2_close_cached_fid); } +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);kref_put (&cfid->refcount, smb2_close_cached_fid);@@ -401,13 +402,16 @@ int open_cached_dir_by_dentry(struct cifs_tcon *tcon,spin_lock (&cfids->cfid_list_lock);list_for_each_entry (cfid, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (dentry && cfid->dentry == dentry) {cifs_dbg (FYI ,"found a cached file handle by dentry\n" );kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock); *ret_cfid = cfid;spin_unlock (&cfids->cfid_list_lock);return 0 ; } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);return -ENOENT ;@@ -428,8 +432,11 @@ smb2_close_cached_fid(struct kref *ref) }spin_unlock (&cfid->cfids->cfid_list_lock); -dput (cfid->dentry); - cfid->dentry =NULL ; +/* no locking necessary as we're the last user of this cfid */ +if (cfid->dentry) { +dput (cfid->dentry); + cfid->dentry =NULL ; + }if (cfid->is_open) { rc = SMB2_close(0, cfid->tcon, cfid->fid.persistent_fid,@@ -452,12 +459,13 @@ void drop_cached_dir_by_name(const unsigned int xid, struct cifs_tcon *tcon,cifs_dbg (FYI ,"no cached dir found for rmdir(%s)\n" , name);return ; } -spin_lock (&cfid->cfids->cfid_list_lock); +spin_lock (&cfid->fid_lock);if (cfid->has_lease) { +/* mark as invalid */ cfid->has_lease = false;kref_put (&cfid->refcount, smb2_close_cached_fid); } -spin_unlock (&cfid->cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock);close_cached_dir (cfid); }@@ -539,6 +547,7 @@ void invalidate_all_cached_dirs(struct cifs_tcon *tcon) cfids->num_entries--; cfid->is_open = false; cfid->on_list = false; +spin_lock (&cfid->fid_lock);if (cfid->has_lease) {/* * The lease was never cancelled from the server, @@ -547,6 +556,7 @@ void invalidate_all_cached_dirs(struct cifs_tcon *tcon) cfid->has_lease = false; }else kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock); }/* * Queue dropping of the dentries once locks have been dropped @@ -601,6 +611,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])spin_lock (&cfids->cfid_list_lock); list_for_each_entry(cfid, &cfids->entries, entry) { + spin_lock(&cfid->fid_lock); if (cfid->has_lease && !memcmp(lease_key, cfid->fid.lease_key,@@ -613,6 +624,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])*/ list_del (&cfid->entry); cfid->on_list = false; +spin_unlock (&cfid->fid_lock); cfids->num_entries--; ++tcon->tc_count;@@ -622,6 +634,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])spin_unlock (&cfids->cfid_list_lock);return true; } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);return false;@@ -657,9 +670,6 @@ static void free_cached_dir(struct cached_fid *cfid)WARN_ON (work_pending (&cfid->close_work));WARN_ON (work_pending (&cfid->put_work)); -dput (cfid->dentry); - cfid->dentry =NULL ; -/* * Delete all cached dirent names */ @@ -704,6 +714,7 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (cfid->time &&time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { cfid->on_list = false;@@ -718,6 +729,7 @@ static void cfids_laundromat_worker(struct work_struct *work) }else kref_get (&cfid->refcount); } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);@@ -728,8 +740,8 @@ static void cfids_laundromat_worker(struct work_struct *work) dentry = cfid->dentry; cfid->dentry =NULL ;spin_unlock (&cfid->fid_lock); -dput (dentry); +if (cfid->is_open) {spin_lock (&cifs_tcp_ses_lock); ++cfid->tcon->tc_count;Signature · 1 line
2.43.0
~Paul
Re: [PATCH 2/7] cifs: protect cfid accesses with fid_lock
Quoted 73 lines · Paul Aurich
Quoted 71 lines · On 2025-06-04 15:48:11 +0530, [email protected] wrote:
From: Shyam Prasad N <[email protected]>
There are several accesses to cfid structure today without locking
fid_lock. This can lead to race conditions that are hard to debug.With this change, I'm trying to make sure that accesses to cfid struct members happen with
fid_lockheld.Signed-off-by Shyam Prasad N <[email protected]>---
fs/smb/client/cached_dir.c| 86 ++++++++++++++++++++++----------------1 file changed, 49 insertions(+), 37 deletions(-)fs/smb/client/cached_dir.c +8
-3@@ -31,6 +31,7 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids,spin_lock (&cfids->cfid_list_lock);list_for_each_entry (cfid, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (!strcmp (cfid->path, path)) {/* * If it doesn't have a lease it is either not yet @@ -38,13 +39,16 @@ static struct cached_fid *find_or_create_cached_dir(struct cached_fids *cfids,* being deleted due to a lease break. */ if (!cfid->time || !cfid->has_lease) { +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);return NULL ; }kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);return cfid; } +spin_unlock (&cfid->fid_lock); }if (lookup_only) {spin_unlock (&cfids->cfid_list_lock);@@ -193,19 +197,20 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,kfree (utf16_path);return -ENOENT ; } +/* * Return cached fid if it is valid (has a lease and has a time). * Otherwise, it is either a new entry or laundromat worker removed it * from @cfids->entries. Caller will put last reference if the latter. */ -spin_lock (&cfids->cfid_list_lock); +spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) { -spin_unlock (&cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock); *ret_cfid = cfid;kfree (utf16_path);return 0 ; } -spin_unlock (&cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock);/* * Skip any prefix paths in @path as lookup_noperm_positive_unlocked() ends up @@ -220,17 +225,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, goto out; }
Quoted 15 lines
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 233
Quoted 11 lines
fs/smb/client/cached_dir.c +0
-11-if (!npath[0 ]) { - dentry =dget (cifs_sb->root); - }else { - dentry =path_to_dentry (cifs_sb, npath); -if (IS_ERR (dentry)) { - rc = -ENOENT ; - goto out; - } - } - cfid->dentry = dentry; - cfid->tcon = tcon;I think moving this down below to after the "At this point the directory handle is fully cached" comment is going to regress
c353ee4fb119("smb:
Initializecfid->tconbefore performing network ops").
Hi Paul,
Are you referring to the tcon reference?
If so, I think the way to fix that is to get a reference (tc_count++) on tcon right when we add it to the cfid->tcon, and put the reference in smb2_close_cached_fid.
I think that should take care of any concerns of race. Let me know if you disagree.
Quoted 13 lines
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 305
Quoted 8 lines
fs/smb/client/cached_dir.c +0
-1/* * We do not hold the lock for the open because in case @@ -302,9 +296,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, } goto oshr_free; } - cfid->is_open = true;Moving this down below is going to regress part of
66d45ca1350a("cifs: Check the lease context if we actually got a lease"), I think. If parsing the lease fails, the cfid is disposed, but sinceis_openis false, the code won't callSMB2_close.
That's a good point. I'll submit a patch to fix this.
Quoted 211 lines
Quoted 207 lines
fs/smb/client/cached_dir.c +20
-14- -spin_lock (&cfids->cfid_list_lock); o_rsp = (struct smb2_create_rsp *)rsp_iov[0 ].iov_base; oparms.fid->persistent_fid = o_rsp->PersistentFileId;@@ -315,7 +306,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,if (o_rsp->OplockLevel !=SMB2_OPLOCK_LEVEL_LEASE ) { -spin_unlock (&cfids->cfid_list_lock); rc = -EINVAL ; goto oshr_free; }@@ -324,21 +314,15 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, &oparms.fid->epoch, oparms.fid->lease_key, &oplock, NULL, NULL); -if (rc) { -spin_unlock (&cfids->cfid_list_lock); +if (rc) goto oshr_free; - } rc = -EINVAL ; -if (!(oplock &SMB2_LEASE_READ_CACHING_HE )) { -spin_unlock (&cfids->cfid_list_lock); +if (!(oplock &SMB2_LEASE_READ_CACHING_HE )) goto oshr_free; - } qi_rsp = (struct smb2_query_info_rsp *)rsp_iov[1 ].iov_base; -if (le32_to_cpu (qi_rsp->OutputBufferLength) <sizeof (struct smb2_file_all_info )) { -spin_unlock (&cfids->cfid_list_lock); +if (le32_to_cpu (qi_rsp->OutputBufferLength) <sizeof (struct smb2_file_all_info )) goto oshr_free; - } if (!smb2_validate_and_copy_iov( le16_to_cpu(qi_rsp->OutputBufferOffset), sizeof(struct smb2_file_all_info),@@ -346,10 +330,25 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, (char *)&cfid->file_all_info)) cfid->file_all_info_is_valid = true; - cfid->time = jiffies; -spin_unlock (&cfids->cfid_list_lock);/* At this point the directory handle is fully cached */ rc =0 ; +if (!cfid->dentry) { +if (!npath[0 ]) { + dentry =dget (cifs_sb->root); + }else { + dentry =path_to_dentry (cifs_sb, npath); +if (IS_ERR (dentry)) { + rc = -ENOENT ; + goto out; + } + } + } +spin_lock (&cfid->fid_lock); + cfid->dentry = dentry; + cfid->tcon = tcon; + cfid->is_open = true; + cfid->time = jiffies; +spin_unlock (&cfid->fid_lock);
oshr_free:SMB2_open_free(&rqst[0]);fs/smb/client/cached_dir.c +21
-9@@ -364,6 +363,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->on_list = false; cfids->num_entries--; } +spin_lock (&cfid->fid_lock);if (cfid->has_lease) {/* * We are guaranteed to have two references at this @@ -373,6 +373,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->has_lease = false;kref_put (&cfid->refcount, smb2_close_cached_fid); } +spin_unlock (&cfid->fid_lock);spin_unlock (&cfids->cfid_list_lock);kref_put (&cfid->refcount, smb2_close_cached_fid);@@ -401,13 +402,16 @@ int open_cached_dir_by_dentry(struct cifs_tcon *tcon,spin_lock (&cfids->cfid_list_lock);list_for_each_entry (cfid, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (dentry && cfid->dentry == dentry) {cifs_dbg (FYI ,"found a cached file handle by dentry\n" );kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock); *ret_cfid = cfid;spin_unlock (&cfids->cfid_list_lock);return 0 ; } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);return -ENOENT ;@@ -428,8 +432,11 @@ smb2_close_cached_fid(struct kref *ref) }spin_unlock (&cfid->cfids->cfid_list_lock); -dput (cfid->dentry); - cfid->dentry =NULL ; +/* no locking necessary as we're the last user of this cfid */ +if (cfid->dentry) { +dput (cfid->dentry); + cfid->dentry =NULL ; + }if (cfid->is_open) { rc = SMB2_close(0, cfid->tcon, cfid->fid.persistent_fid,@@ -452,12 +459,13 @@ void drop_cached_dir_by_name(const unsigned int xid, struct cifs_tcon *tcon,cifs_dbg (FYI ,"no cached dir found for rmdir(%s)\n" , name);return ; } -spin_lock (&cfid->cfids->cfid_list_lock); +spin_lock (&cfid->fid_lock);if (cfid->has_lease) { +/* mark as invalid */ cfid->has_lease = false;kref_put (&cfid->refcount, smb2_close_cached_fid); } -spin_unlock (&cfid->cfids->cfid_list_lock); +spin_unlock (&cfid->fid_lock);close_cached_dir (cfid); }@@ -539,6 +547,7 @@ void invalidate_all_cached_dirs(struct cifs_tcon *tcon) cfids->num_entries--; cfid->is_open = false; cfid->on_list = false; +spin_lock (&cfid->fid_lock);if (cfid->has_lease) {/* * The lease was never cancelled from the server, @@ -547,6 +556,7 @@ void invalidate_all_cached_dirs(struct cifs_tcon *tcon) cfid->has_lease = false; }else kref_get (&cfid->refcount); +spin_unlock (&cfid->fid_lock); }/* * Queue dropping of the dentries once locks have been dropped @@ -601,6 +611,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])spin_lock (&cfids->cfid_list_lock); list_for_each_entry(cfid, &cfids->entries, entry) { + spin_lock(&cfid->fid_lock); if (cfid->has_lease && !memcmp(lease_key, cfid->fid.lease_key,@@ -613,6 +624,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])*/ list_del (&cfid->entry); cfid->on_list = false; +spin_unlock (&cfid->fid_lock); cfids->num_entries--; ++tcon->tc_count;@@ -622,6 +634,7 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])spin_unlock (&cfids->cfid_list_lock);return true; } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);return false;@@ -657,9 +670,6 @@ static void free_cached_dir(struct cached_fid *cfid)WARN_ON (work_pending (&cfid->close_work));WARN_ON (work_pending (&cfid->put_work)); -dput (cfid->dentry); - cfid->dentry =NULL ; -/* * Delete all cached dirent names */ @@ -704,6 +714,7 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) { +spin_lock (&cfid->fid_lock);if (cfid->time &&time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { cfid->on_list = false;@@ -718,6 +729,7 @@ static void cfids_laundromat_worker(struct work_struct *work) }else kref_get (&cfid->refcount); } +spin_unlock (&cfid->fid_lock); }spin_unlock (&cfids->cfid_list_lock);@@ -728,8 +740,8 @@ static void cfids_laundromat_worker(struct work_struct *work) dentry = cfid->dentry; cfid->dentry =NULL ;spin_unlock (&cfid->fid_lock); -dput (dentry); +if (cfid->is_open) {spin_lock (&cifs_tcp_ses_lock); ++cfid->tcon->tc_count; --Quoted 1 line
2.43.0
~Paul
Signature · 2 lines
Regards, Shyam
Re: [PATCH 2/7] cifs: protect cfid accesses with fid_lock
Quoted 25 lines · Shyam Prasad N
Quoted 8 lines · Paul Aurich
Quoted 6 lines · On 2025-06-04 15:48:11 +0530, [email protected] wrote:
From: Shyam Prasad N <[email protected]>
fs/smb/client/cached_dir.c +0
-0@@ -220,17 +225,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, goto out; }Quoted 15 lines
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 233
Quoted 11 lines
fs/smb/client/cached_dir.c +0
-11-if (!npath[0 ]) { - dentry =dget (cifs_sb->root); - }else { - dentry =path_to_dentry (cifs_sb, npath); -if (IS_ERR (dentry)) { - rc = -ENOENT ; - goto out; - } - } - cfid->dentry = dentry; - cfid->tcon = tcon;I think moving this down below to after the "At this point the directory handle is fully cached" comment is going to regress
c353ee4fb119("smb:
Initializecfid->tconbefore performing network ops").
Quoted 8 lines
Hi Paul,
Are you referring to the tcon reference?
If so, I think the way to fix that is to get a reference (tc_count++) on tcon right when we add it to thecfid->tcon, and put the reference insmb2_close_cached_fid.
I think that should take care of any concerns of race. Let me know if you disagree.
I don't think that is sufficient. cfid->tcon needs to be initialized prior to the network operations in open_cached_dir. If we receive a lease break from the server and the delayed work involved in that runs *before* cfid->tcon is initialized, it leaks a ref. (cached_dir_lease_break has a direct copy of the tcon passed as an arg and takes a ref from that, but cfid->tcon is still NULL)
T1 T2
open_cached_dir
<open+lease acquisition>
(gets far enough along that the file is open and lease key initialized)
// receives a lease break from the server
cached_dir_lease_break
// takes a ref on tcon (via arg to function, not cfid->tcon)
cached_dir_put_work
cached_dir_offload_close
// cifs_put_tcon on cfid->tcon, which is still NULLopen_cached_dir continues on at this point
Another reason you need to initialize cfid->tcon earlier is that in the error paths, cfid->tcon needs to be valid so that smb2_close_cached_fid can actually close it. Before c353ee4fb119, the tcon was initialized at the same time cfid->is_open was set.
Quoted 20 lines
Quoted 13 lines
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 305
Quoted 8 lines
fs/smb/client/cached_dir.c +0
-1/* * We do not hold the lock for the open because in case @@ -302,9 +296,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, } goto oshr_free; } - cfid->is_open = true;Moving this down below is going to regress part of
66d45ca1350a("cifs: Check the lease context if we actually got a lease"), I think. If parsing the lease fails, the cfid is disposed, but sinceis_openis false, the code won't callSMB2_close.That's a good point. I'll submit a patch to fix this.
Signature · 2 lines
Regards, Shyam
Signature · 1 line
~Paul
Re: [PATCH 2/7] cifs: protect cfid accesses with fid_lock
Quoted 55 lines · Paul Aurich
Quoted 25 lines · Shyam Prasad N
Quoted 8 lines · Paul Aurich
Quoted 6 lines · On 2025-06-04 15:48:11 +0530, [email protected] wrote:
From: Shyam Prasad N <[email protected]>
fs/smb/client/cached_dir.c +0
-0@@ -220,17 +225,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, goto out; }Quoted 15 lines
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 233
Quoted 11 lines
fs/smb/client/cached_dir.c +0
-11-if (!npath[0 ]) { - dentry =dget (cifs_sb->root); - }else { - dentry =path_to_dentry (cifs_sb, npath); -if (IS_ERR (dentry)) { - rc = -ENOENT ; - goto out; - } - } - cfid->dentry = dentry; - cfid->tcon = tcon;I think moving this down below to after the "At this point the directory handle is fully cached" comment is going to regress
c353ee4fb119("smb:
Initializecfid->tconbefore performing network ops").Quoted 8 lines
Hi Paul,
Are you referring to the tcon reference?
If so, I think the way to fix that is to get a reference (tc_count++) on tcon right when we add it to thecfid->tcon, and put the reference insmb2_close_cached_fid.
I think that should take care of any concerns of race. Let me know if you disagree.I don't think that is sufficient.
cfid->tconneeds to be initialized prior to the network operations inopen_cached_dir. If we receive a lease break from the server and the delayed work involved in that runs *before*cfid->tconis initialized, it leaks a ref. (cached_dir_lease_breakhas a direct copy of the tcon passed as an arg and takes a ref from that, butcfid->tconis still NULL)T1 T2
open_cached_dir
<open+lease acquisition>
(gets far enough along that the file is open and lease key initialized)// receives a lease break from the server cached_dir_lease_break // takes a ref on tcon (via arg to function, not cfid->tcon) cached_dir_put_work cached_dir_offload_close // cifs_put_tcon on cfid->tcon, which is still NULL
Quoted 6 lines
open_cached_dircontinues on at this pointAnother reason you need to initialize
cfid->tconearlier is that in the error paths,cfid->tconneeds to be valid so thatsmb2_close_cached_fidcan actually close it. Beforec353ee4fb119, the tcon was initialized at the same timecfid->is_openwas set.
Fair points. I'll put it back where it was. Just protect it with the cfid_lock. Although I'm now wondering if we'll be hit by some weird race by not using cfid_mutex in lease break handler and laundromat thread too. :)
Quoted 25 lines
Quoted 15 lines
Quoted 14 lines
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 305
Quoted 9 lines
fs/smb/client/cached_dir.c +0
-1/* * We do not hold the lock for the open because in case @@ -302,9 +296,6 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, } goto oshr_free; } - cfid->is_open = true;Moving this down below is going to regress part of
66d45ca1350a("cifs: Check the lease context if we actually got a lease"), I think. If parsing the lease fails, the cfid is disposed, but sinceis_openis false, the code won't callSMB2_close.Quoted 6 lines
That's a good point. I'll submit a patch to fix this.
--
Regards,
Shyam--
~Paul
Signature · 2 lines
Regards, Shyam
PATCH 3/7 cifs: do not return an invalidated cfid
In reply to [email protected]From: Shyam Prasad N <[email protected]>
open_cached_dir should either return an existing valid cfid or create a new one. Validity of cfid depends on both cfid->has_lease and cfid->time to be true. However, if has_lease was invalidated by a worker thread in parallel, we could end up leaking both a dentry and a server handle.
This change checks if the entry was invalidated and returns a -ENOENT in such a case.
--- fs/smb/client/cached_dir.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)
fs/smb/client/cached_dir.c +9 -1
@@ -199,9 +199,11 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, }/* +* check again that the cfid is valid (with mutex held this time). * Return cached fid if it is valid (has a lease and has a time). * Otherwise, it is either a new entry or laundromat worker removed it -* from @cfids->entries. Caller will put last reference if the latter. +* from @cfids->entries. If the latter, we drop the refcount and return +* an error to the caller. */ spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) {
@@ -209,6 +211,12 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, *ret_cfid = cfid;kfree (utf16_path);return 0 ; + }else if (!cfid->has_lease) { +spin_unlock (&cfid->fid_lock); +/* drop the ref that we have */ +kref_put (&cfid->refcount, smb2_close_cached_fid); +kfree (utf16_path); +return -ENOENT ; }spin_unlock (&cfid->fid_lock);
Signature · 1 line
2.43.0
PATCH 4/7 cifs: serialize initialization and cleanup of cfid
In reply to [email protected]From: Shyam Prasad N <[email protected]>
Today we can have multiple processes calling open_cached_dir and other workers freeing the cached dir all in parallel. Although small sections of this code is locked to protect individual fields, there can be races between these threads which can be hard to debug.
This patch serializes all initialization and cleanup of the cfid struct and the associated resources: dentry and the server handle.
--- fs/smb/client/cached_dir.c | 16 ++++++++++++++++ fs/smb/client/cached_dir.h | 1 + 2 files changed, 17 insertions(+)
fs/smb/client/cached_dir.c +16 -0
@@ -198,6 +198,12 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,return -ENOENT ; } +/* +* the following is a critical section. We need to make sure that the +* callers are serialized per-cfid +*/ +mutex_lock (&cfid->cfid_mutex); +/* * check again that the cfid is valid (with mutex held this time). * Return cached fid if it is valid (has a lease and has a time).
@@ -208,11 +214,13 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) {spin_unlock (&cfid->fid_lock); +mutex_unlock (&cfid->cfid_mutex); *ret_cfid = cfid;kfree (utf16_path);return 0 ; }else if (!cfid->has_lease) {spin_unlock (&cfid->fid_lock); +mutex_unlock (&cfid->cfid_mutex);/* drop the ref that we have */ kref_put (&cfid->refcount, smb2_close_cached_fid);kfree (utf16_path);
@@ -229,6 +237,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,*/ npath =path_no_prefix (cifs_sb, path);if (IS_ERR (npath)) { +mutex_unlock (&cfid->cfid_mutex); rc =PTR_ERR (npath); goto out; }
@@ -389,6 +398,8 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, *ret_cfid = cfid;atomic_inc (&tcon->num_remote_opens); } +mutex_unlock (&cfid->cfid_mutex); +kfree (utf16_path); if (is_replayable_error(rc) &&
@@ -432,6 +443,9 @@ smb2_close_cached_fid(struct kref *ref) refcount);int rc; +/* make sure not to race with server open */ +mutex_lock (&cfid->cfid_mutex); +spin_lock (&cfid->cfids->cfid_list_lock);if (cfid->on_list) {list_del (&cfid->entry);
@@ -452,6 +466,7 @@ smb2_close_cached_fid(struct kref *ref)if (rc)/* should we retry on -EBUSY or -EAGAIN? */ cifs_dbg (VFS ,"close cached dir rc %d\n" , rc); } +mutex_unlock (&cfid->cfid_mutex);free_cached_dir (cfid); }
@@ -666,6 +681,7 @@ static struct cached_fid *init_cached_dir(const char *path)INIT_LIST_HEAD (&cfid->entry);INIT_LIST_HEAD (&cfid->dirents.entries);mutex_init (&cfid->dirents.de_mutex); +mutex_init (&cfid->cfid_mutex);spin_lock_init (&cfid->fid_lock);kref_init (&cfid->refcount);return cfid;
fs/smb/client/cached_dir.h +1 -0
@@ -42,6 +42,7 @@ struct cached_fid {struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock; +struct mutex cfid_mutex;struct cifs_tcon *tcon;struct dentry *dentry;struct work_struct put_work;
Signature · 1 line
2.43.0
Re: [PATCH 4/7] cifs: serialize initialization and cleanup of cfid
Quoted 50 lines · On 2025-06-04 15:48:13 +0530, [email protected] wrote:
From: Shyam Prasad N <[email protected]>
Today we can have multiple processes calling
open_cached_dirand other workers freeing the cached dir all in parallel. Although small sections of this code is locked to protect individual fields, there can be races between these threads which can be hard to debug.This patch serializes all initialization and cleanup of the cfid struct and the associated resources: dentry and the server handle.
Signed-off-by Shyam Prasad N <[email protected]>---
fs/smb/client/cached_dir.c| 16 ++++++++++++++++fs/smb/client/cached_dir.h| 1 +2 files changed, 17 insertions(+)fs/smb/client/cached_dir.c +8
-0@@ -198,6 +198,12 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,return -ENOENT ; } +/* +* the following is a critical section. We need to make sure that the +* callers are serialized per-cfid +*/ +mutex_lock (&cfid->cfid_mutex); +/* * check again that the cfid is valid (with mutex held this time). * Return cached fid if it is valid (has a lease and has a time). @@ -208,11 +214,13 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) {spin_unlock (&cfid->fid_lock); +mutex_unlock (&cfid->cfid_mutex); *ret_cfid = cfid;kfree (utf16_path);return 0 ; }else if (!cfid->has_lease) {spin_unlock (&cfid->fid_lock); +mutex_unlock (&cfid->cfid_mutex);/* drop the ref that we have */ kref_put (&cfid->refcount, smb2_close_cached_fid);kfree (utf16_path);
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 240
Quoted 4 lines
fs/smb/client/cached_dir.c +1
-0@@ -229,6 +237,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,*/ npath =path_no_prefix (cifs_sb, path);if (IS_ERR (npath)) { +mutex_unlock (&cfid->cfid_mutex);
Double mutex_unlock? (It's also unlocked unconditionally in the 'out' path)
Quoted 53 lines
fs/smb/client/cached_dir.c +7
-0rc =PTR_ERR (npath); goto out; }@@ -389,6 +398,8 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, *ret_cfid = cfid;atomic_inc (&tcon->num_remote_opens); } +mutex_unlock (&cfid->cfid_mutex); +kfree (utf16_path); if (is_replayable_error(rc) &&@@ -432,6 +443,9 @@ smb2_close_cached_fid(struct kref *ref) refcount);int rc; +/* make sure not to race with server open */ +mutex_lock (&cfid->cfid_mutex); +spin_lock (&cfid->cfids->cfid_list_lock);if (cfid->on_list) {list_del (&cfid->entry);@@ -452,6 +466,7 @@ smb2_close_cached_fid(struct kref *ref)if (rc)/* should we retry on -EBUSY or -EAGAIN? */ cifs_dbg (VFS ,"close cached dir rc %d\n" , rc); } +mutex_unlock (&cfid->cfid_mutex);free_cached_dir (cfid); }@@ -666,6 +681,7 @@ static struct cached_fid *init_cached_dir(const char *path)INIT_LIST_HEAD (&cfid->entry);INIT_LIST_HEAD (&cfid->dirents.entries);mutex_init (&cfid->dirents.de_mutex); +mutex_init (&cfid->cfid_mutex);spin_lock_init (&cfid->fid_lock);kref_init (&cfid->refcount);return cfid;fs/smb/client/cached_dir.h +1
-0@@ -42,6 +42,7 @@ struct cached_fid {struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock; +struct mutex cfid_mutex;struct cifs_tcon *tcon;struct dentry *dentry;struct work_struct put_work;Signature · 1 line
2.43.0
~Paul
Re: [PATCH 4/7] cifs: serialize initialization and cleanup of cfid
Quoted 50 lines · Paul Aurich
Quoted 48 lines · On 2025-06-04 15:48:13 +0530, [email protected] wrote:
From: Shyam Prasad N <[email protected]>
Today we can have multiple processes calling
open_cached_dirand other workers freeing the cached dir all in parallel. Although small sections of this code is locked to protect individual fields, there can be races between these threads which can be hard to debug.This patch serializes all initialization and cleanup of the cfid struct and the associated resources: dentry and the server handle.
Signed-off-by Shyam Prasad N <[email protected]>---
fs/smb/client/cached_dir.c| 16 ++++++++++++++++fs/smb/client/cached_dir.h| 1 +2 files changed, 17 insertions(+)fs/smb/client/cached_dir.c +8
-0@@ -198,6 +198,12 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,return -ENOENT ; } +/* +* the following is a critical section. We need to make sure that the +* callers are serialized per-cfid +*/ +mutex_lock (&cfid->cfid_mutex); +/* * check again that the cfid is valid (with mutex held this time). * Return cached fid if it is valid (has a lease and has a time). @@ -208,11 +214,13 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) {spin_unlock (&cfid->fid_lock); +mutex_unlock (&cfid->cfid_mutex); *ret_cfid = cfid;kfree (utf16_path);return 0 ; }else if (!cfid->has_lease) {spin_unlock (&cfid->fid_lock); +mutex_unlock (&cfid->cfid_mutex);/* drop the ref that we have */ kref_put (&cfid->refcount, smb2_close_cached_fid);
Quoted 8 lines
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 240
Quoted 6 lines
fs/smb/client/cached_dir.c +1
-0kfree (utf16_path);@@ -229,6 +237,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,*/ npath =path_no_prefix (cifs_sb, path);if (IS_ERR (npath)) { +mutex_unlock (&cfid->cfid_mutex);Double
mutex_unlock? (It's also unlocked unconditionally in the 'out' path)
Good catch!
Steve: Do you want me to submit a v2? Or submit another patch to fix this?
Quoted 57 lines
Quoted 52 lines
fs/smb/client/cached_dir.c +7
-0rc =PTR_ERR (npath); goto out; }@@ -389,6 +398,8 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, *ret_cfid = cfid;atomic_inc (&tcon->num_remote_opens); } +mutex_unlock (&cfid->cfid_mutex); +kfree (utf16_path); if (is_replayable_error(rc) &&@@ -432,6 +443,9 @@ smb2_close_cached_fid(struct kref *ref) refcount);int rc; +/* make sure not to race with server open */ +mutex_lock (&cfid->cfid_mutex); +spin_lock (&cfid->cfids->cfid_list_lock);if (cfid->on_list) {list_del (&cfid->entry);@@ -452,6 +466,7 @@ smb2_close_cached_fid(struct kref *ref)if (rc)/* should we retry on -EBUSY or -EAGAIN? */ cifs_dbg (VFS ,"close cached dir rc %d\n" , rc); } +mutex_unlock (&cfid->cfid_mutex);free_cached_dir (cfid); }@@ -666,6 +681,7 @@ static struct cached_fid *init_cached_dir(const char *path)INIT_LIST_HEAD (&cfid->entry);INIT_LIST_HEAD (&cfid->dirents.entries);mutex_init (&cfid->dirents.de_mutex); +mutex_init (&cfid->cfid_mutex);spin_lock_init (&cfid->fid_lock);kref_init (&cfid->refcount);return cfid;fs/smb/client/cached_dir.h +1
-1@@ -42,6 +42,7 @@ struct cached_fid {struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock; +struct mutex cfid_mutex;struct cifs_tcon *tcon;struct dentry *dentry;struct work_struct put_work; --Quoted 2 lines
2.43.0
~Paul
Signature · 2 lines
Regards, Shyam
Re: [PATCH 4/7] cifs: serialize initialization and cleanup of cfid
Quoted 1 line
Steve: Do you want me to submit a v2? Or submit another patch to fix this?
For patches that are not in mainline yet, then better to send a v2, yes
Quoted 126 lines · Shyam Prasad N
Quoted 50 lines · Paul Aurich
Quoted 48 lines · On 2025-06-04 15:48:13 +0530, [email protected] wrote:
From: Shyam Prasad N <[email protected]>
Today we can have multiple processes calling
open_cached_dirand other workers freeing the cached dir all in parallel. Although small sections of this code is locked to protect individual fields, there can be races between these threads which can be hard to debug.This patch serializes all initialization and cleanup of the cfid struct and the associated resources: dentry and the server handle.
Signed-off-by Shyam Prasad N <[email protected]>---
fs/smb/client/cached_dir.c| 16 ++++++++++++++++fs/smb/client/cached_dir.h| 1 +2 files changed, 17 insertions(+)fs/smb/client/cached_dir.c +8
-0@@ -198,6 +198,12 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,return -ENOENT ; } +/* +* the following is a critical section. We need to make sure that the +* callers are serialized per-cfid +*/ +mutex_lock (&cfid->cfid_mutex); +/* * check again that the cfid is valid (with mutex held this time). * Return cached fid if it is valid (has a lease and has a time). @@ -208,11 +214,13 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) {spin_unlock (&cfid->fid_lock); +mutex_unlock (&cfid->cfid_mutex); *ret_cfid = cfid;kfree (utf16_path);return 0 ; }else if (!cfid->has_lease) {spin_unlock (&cfid->fid_lock); +mutex_unlock (&cfid->cfid_mutex);/* drop the ref that we have */ kref_put (&cfid->refcount, smb2_close_cached_fid);Quoted 8 lines
fs/smb/client/cached_dir.c int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, · line 240
Quoted 6 lines
fs/smb/client/cached_dir.c +1
-0kfree (utf16_path);@@ -229,6 +237,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,*/ npath =path_no_prefix (cifs_sb, path);if (IS_ERR (npath)) { +mutex_unlock (&cfid->cfid_mutex);Double
mutex_unlock? (It's also unlocked unconditionally in the 'out' path)Good catch!
Steve: Do you want me to submit a v2? Or submit another patch to fix this?Quoted 48 lines
Quoted 48 lines
fs/smb/client/cached_dir.c +7
-0rc =PTR_ERR (npath); goto out; }@@ -389,6 +398,8 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, *ret_cfid = cfid;atomic_inc (&tcon->num_remote_opens); } +mutex_unlock (&cfid->cfid_mutex); +kfree (utf16_path); if (is_replayable_error(rc) &&@@ -432,6 +443,9 @@ smb2_close_cached_fid(struct kref *ref) refcount);int rc; +/* make sure not to race with server open */ +mutex_lock (&cfid->cfid_mutex); +spin_lock (&cfid->cfids->cfid_list_lock);if (cfid->on_list) {list_del (&cfid->entry);@@ -452,6 +466,7 @@ smb2_close_cached_fid(struct kref *ref)if (rc)/* should we retry on -EBUSY or -EAGAIN? */ cifs_dbg (VFS ,"close cached dir rc %d\n" , rc); } +mutex_unlock (&cfid->cfid_mutex);free_cached_dir (cfid); }@@ -666,6 +681,7 @@ static struct cached_fid *init_cached_dir(const char *path)INIT_LIST_HEAD (&cfid->entry);INIT_LIST_HEAD (&cfid->dirents.entries);mutex_init (&cfid->dirents.de_mutex); +mutex_init (&cfid->cfid_mutex);spin_lock_init (&cfid->fid_lock);kref_init (&cfid->refcount);return cfid;fs/smb/client/cached_dir.h +1
-0@@ -42,6 +42,7 @@ struct cached_fid {struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock; +struct mutex cfid_mutex;Quoted 9 lines
Quoted 6 lines
fs/smb/client/cached_dir.h +0
-1struct cifs_tcon *tcon;struct dentry *dentry;struct work_struct put_work; --2.43.0
~Paul
--
Regards,
Shyam
Signature · 3 lines
Thanks, Steve
PATCH 5/7 cifs: update the lock ordering comments with new mutex
In reply to [email protected]From: Shyam Prasad N <[email protected]>
The lock ordering rules listed as comments in cifsglob.h were missing some lock details and the newly introduced cfid_mutex.
Updated those notes in this commit.
--- fs/smb/client/cifsglob.h | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-)
fs/smb/client/cifsglob.h +9 -5
@@ -1989,13 +1989,14 @@ require use of the stronger protocol */* TCP_Server_Info-> TCP_Server_Info cifs_get_tcp_session * reconnect_mutex * TCP_Server_Info->srv_mutex TCP_Server_Info cifs_get_tcp_session -* cifs_ses->session_mutex cifs_ses sesInfoAlloc -* cifs_tcon +* cifs_ses->session_mutex cifs_ses sesInfoAlloc +* cached_fid->cfid_mutex cached_fid init_cached_dir * cifs_tcon->open_file_lock cifs_tcon->openFileList tconInfoAlloc * cifs_tcon->pending_opens * cifs_tcon->stat_lock cifs_tcon->bytes_read tconInfoAlloc * cifs_tcon->bytes_written * cifs_tcp_ses_lock cifs_tcp_ses_list sesInfoAlloc +* cached_fids->cfid_list_lock cifs_tcon->cfids->entries init_cached_dirs * GlobalMid_Lock GlobalMaxActiveXid init_cifs * GlobalCurrentXid * GlobalTotalActiveXid
@@ -2009,21 +2010,24 @@ require use of the stronger protocol */* ->oplock_credits * ->reconnect_instance * cifs_ses->ses_lock (anything that is not protected by another lock and can change) +* sesInfoAlloc * cifs_ses->iface_lock cifs_ses->iface_list sesInfoAlloc * ->iface_count * ->iface_last_update -* cifs_ses->chan_lock cifs_ses->chans +* cifs_ses->chan_lock cifs_ses->chans sesInfoAlloc * ->chans_need_reconnect * ->chans_in_reconnect * cifs_tcon->tc_lock (anything that is not protected by another lock and can change) +* tcon_info_alloc * inode->i_rwsem, taken by fs/netfs/locking.c e.g. should be taken before cifsInodeInfo locks * cifsInodeInfo->open_file_lock cifsInodeInfo->openFileList cifs_alloc_inode * cifsInodeInfo->writers_lock cifsInodeInfo->writers cifsInodeInfo_alloc * cifsInodeInfo->lock_sem cifsInodeInfo->llist cifs_init_once * ->can_cache_brlcks * cifsInodeInfo->deferred_lock cifsInodeInfo->deferred_closes cifsInodeInfo_alloc -* cached_fids->cfid_list_lock cifs_tcon->cfids->entries init_cached_dirs -* cifsFileInfo->fh_mutex cifsFileInfo cifs_new_fileinfo +* cached_fid->fid_lock (anything that is not protected by another lock and can change) +* init_cached_dir +* cifsFileInfo->fh_mutex cifsFileInfo cifs_new_fileinfo * cifsFileInfo->file_info_lock cifsFileInfo->count cifs_new_fileinfo * ->invalidHandle initiate_cifs_search * ->oplock_break_cancelled
Signature · 1 line
2.43.0
PATCH 6/7 cifs: tc_count updates should be done with tc_lock
In reply to [email protected]From: Shyam Prasad N <[email protected]>
We had problems with deadlocks using the cifs_tcp_ses_lock for protecting a lot of structures. So we broke it down into smaller spinlocks. cifs_tcon struct fields are protected by tc_lock now. Hence we should stick to using that.
--- fs/smb/client/cached_dir.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
fs/smb/client/cached_dir.c +4 -2
@@ -650,10 +650,12 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])spin_unlock (&cfid->fid_lock); cfids->num_entries--; +spin_lock (&cfid->tcon->tc_lock); ++tcon->tc_count;trace_smb3_tcon_ref (tcon->debug_id, tcon->tc_count, netfs_trace_tcon_ref_get_cached_lease_break);queue_work (cfid_put_wq, &cfid->put_work); +spin_unlock (&cfid->tcon->tc_lock);spin_unlock (&cfids->cfid_list_lock);return true; }
@@ -767,11 +769,11 @@ static void cfids_laundromat_worker(struct work_struct *work)dput (dentry);if (cfid->is_open) { -spin_lock (&cifs_tcp_ses_lock); +spin_lock (&cfid->tcon->tc_lock); ++cfid->tcon->tc_count;trace_smb3_tcon_ref (cfid->tcon->debug_id, cfid->tcon->tc_count, netfs_trace_tcon_ref_get_cached_laundromat); -spin_unlock (&cifs_tcp_ses_lock); +spin_unlock (&cfid->tcon->tc_lock);queue_work (serverclose_wq, &cfid->close_work); }else /*
Signature · 1 line
2.43.0
Re: [PATCH 6/7] cifs: tc_count updates should be done with tc_lock
Quoted 6 lines · On 2025-06-04 15:48:15 +0530, [email protected] wrote:
From: Shyam Prasad N <[email protected]>
We had problems with deadlocks using the
cifs_tcp_ses_lockfor protecting a lot of structures. So we broke it down into smaller spinlocks.cifs_tconstruct fields are protected bytc_locknow. Hence we should stick to using that.
Is it really safe to adjust this to tc_lock?
There are a number of other points in the smb client that (AFAICT) only hold cifs_tcp_ses_lock while adjusting the tc_count (which is why I used that here). See smb2_handle_cancelled_close, smb2_get_dfs_refer, smb2_reconnect_server, smb2_find_smb_tcon.
~Paul
Quoted 40 lines
Fixes 3fa640d035e5 ("smb: During unmount, ensure all cached dir instances drop their dentry")Signed-off-by Shyam Prasad N <[email protected]>---
fs/smb/client/cached_dir.c| 6 ++++--1 file changed, 4 insertions(+), 2 deletions(-)fs/smb/client/cached_dir.c +4
-2@@ -650,10 +650,12 @@ int cached_dir_lease_break(struct cifs_tcon *tcon, __u8 lease_key[16])spin_unlock (&cfid->fid_lock); cfids->num_entries--; +spin_lock (&cfid->tcon->tc_lock); ++tcon->tc_count;trace_smb3_tcon_ref (tcon->debug_id, tcon->tc_count, netfs_trace_tcon_ref_get_cached_lease_break);queue_work (cfid_put_wq, &cfid->put_work); +spin_unlock (&cfid->tcon->tc_lock);spin_unlock (&cfids->cfid_list_lock);return true; }@@ -767,11 +769,11 @@ static void cfids_laundromat_worker(struct work_struct *work)dput (dentry);if (cfid->is_open) { -spin_lock (&cifs_tcp_ses_lock); +spin_lock (&cfid->tcon->tc_lock); ++cfid->tcon->tc_count;trace_smb3_tcon_ref (cfid->tcon->debug_id, cfid->tcon->tc_count, netfs_trace_tcon_ref_get_cached_laundromat); -spin_unlock (&cifs_tcp_ses_lock); +spin_unlock (&cfid->tcon->tc_lock);queue_work (serverclose_wq, &cfid->close_work); }else /* Signature · 1 line
2.43.0
PATCH 7/7 cifs: add new field to track the last access time of cfid
In reply to [email protected]From: Shyam Prasad N <[email protected]>
The handlecache code today tracks the time at which dir lease was acquired and the laundromat thread uses that to check for old entries to cleanup.
However, if a directory is actively accessed, it should not be chosen to expire first.
This change adds a new last_access_time field to cfid and uses that to decide expiry of the cfid.
--- fs/smb/client/cached_dir.c | 6 ++++-- fs/smb/client/cached_dir.h | 1 + 2 files changed, 5 insertions(+), 2 deletions(-)
fs/smb/client/cached_dir.c +4 -2
@@ -213,6 +213,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,*/ spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) { + cfid->last_access_time = jiffies;spin_unlock (&cfid->fid_lock);mutex_unlock (&cfid->cfid_mutex); *ret_cfid = cfid;
@@ -365,6 +366,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->tcon = tcon; cfid->is_open = true; cfid->time = jiffies; + cfid->last_access_time = jiffies;spin_unlock (&cfid->fid_lock); oshr_free:
@@ -741,8 +743,8 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) {spin_lock (&cfid->fid_lock); -if (cfid->time && -time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { +if (cfid->last_access_time && +time_after (jiffies, cfid->last_access_time +HZ * dir_cache_timeout)) { cfid->on_list = false;list_move (&cfid->entry, &entry); cfids->num_entries--;
fs/smb/client/cached_dir.h +1 -0
@@ -39,6 +39,7 @@ struct cached_fid { bool on_list:1; bool file_all_info_is_valid:1;unsigned long time;/* jiffies of when lease was taken */ +unsigned long last_access_time;/* jiffies of when last accessed */ struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock;
Signature · 1 line
2.43.0
Re: [PATCH 7/7] cifs: add new field to track the last access time of cfid
Rebased version of patch attached. See attached. Let me know if any objections. It solves an obvious important issue (we shouldn't be timing out directory leases if they are the most recently used ones)
Quoted 65 lines · On Wed, Jun 4, 2025 at 5:18 AM <[email protected]> wrote:
From: Shyam Prasad N <[email protected]>
The handlecache code today tracks the time at which dir lease was acquired and the laundromat thread uses that to check for old entries to cleanup.
However, if a directory is actively accessed, it should not be chosen to expire first.
This change adds a new
last_access_timefield to cfid and uses that to decide expiry of the cfid.Signed-off-by Shyam Prasad N <[email protected]>--- fs/smb/client/cached_dir.c | 6 ++++-- fs/smb/client/cached_dir.h | 1 + 2 files changed, 5 insertions(+), 2 deletions(-)fs/smb/client/cached_dir.c +4
-2@@ -213,6 +213,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,*/ spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) { + cfid->last_access_time = jiffies;spin_unlock (&cfid->fid_lock);mutex_unlock (&cfid->cfid_mutex); *ret_cfid = cfid;@@ -365,6 +366,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->tcon = tcon; cfid->is_open = true; cfid->time = jiffies; + cfid->last_access_time = jiffies;spin_unlock (&cfid->fid_lock); oshr_free:@@ -741,8 +743,8 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) {spin_lock (&cfid->fid_lock); -if (cfid->time && -time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { +if (cfid->last_access_time && +time_after (jiffies, cfid->last_access_time +HZ * dir_cache_timeout)) { cfid->on_list = false;list_move (&cfid->entry, &entry); cfids->num_entries--;fs/smb/client/cached_dir.h +1
-1@@ -39,6 +39,7 @@ struct cached_fid { bool on_list:1; bool file_all_info_is_valid:1;unsigned long time;/* jiffies of when lease was taken */ +unsigned long last_access_time;/* jiffies of when last accessed */ struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock; --2.43.0
Signature · 3 lines
Thanks, Steve
fs/smb/client/cached_dir.c +4 -2
@@ -195,6 +195,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,* from @cfids->entries. Caller will put last reference if the latter. */ if (cfid->has_lease && cfid->time) { + cfid->last_access_time = jiffies;spin_unlock (&cfids->cfid_list_lock); *ret_cfid = cfid;kfree (utf16_path);
@@ -363,6 +364,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->file_all_info_is_valid = true; cfid->time = jiffies; + cfid->last_access_time = jiffies;spin_unlock (&cfids->cfid_list_lock);/* At this point the directory handle is fully cached */ rc =0 ;
@@ -730,8 +732,8 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) { -if (cfid->time && -time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { +if (cfid->last_access_time && +time_after (jiffies, cfid->last_access_time +HZ * dir_cache_timeout)) { cfid->on_list = false;list_move (&cfid->entry, &entry); cfids->num_entries--;
fs/smb/client/cached_dir.h +1 -0
@@ -39,6 +39,7 @@ struct cached_fid { bool on_list:1; bool file_all_info_is_valid:1;unsigned long time;/* jiffies of when lease was taken */ +unsigned long last_access_time;/* jiffies of when last accessed */ struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock;
Re: [PATCH 7/7] cifs: add new field to track the last access time of cfid
Updated patch with description
Quoted 78 lines · Steve French
Rebased version of patch attached. See attached. Let me know if any objections. It solves an obvious important issue (we shouldn't be timing out directory leases if they are the most recently used ones)
Quoted 63 lines · On Wed, Jun 4, 2025 at 5:18 AM <[email protected]> wrote:
From: Shyam Prasad N <[email protected]>
The handlecache code today tracks the time at which dir lease was acquired and the laundromat thread uses that to check for old entries to cleanup.
However, if a directory is actively accessed, it should not be chosen to expire first.
This change adds a new
last_access_timefield to cfid and uses that to decide expiry of the cfid.Signed-off-by Shyam Prasad N <[email protected]>--- fs/smb/client/cached_dir.c | 6 ++++-- fs/smb/client/cached_dir.h | 1 + 2 files changed, 5 insertions(+), 2 deletions(-)fs/smb/client/cached_dir.c +4
-2@@ -213,6 +213,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,*/ spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) { + cfid->last_access_time = jiffies;spin_unlock (&cfid->fid_lock);mutex_unlock (&cfid->cfid_mutex); *ret_cfid = cfid;@@ -365,6 +366,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->tcon = tcon; cfid->is_open = true; cfid->time = jiffies; + cfid->last_access_time = jiffies;spin_unlock (&cfid->fid_lock); oshr_free:@@ -741,8 +743,8 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) {spin_lock (&cfid->fid_lock); -if (cfid->time && -time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { +if (cfid->last_access_time && +time_after (jiffies, cfid->last_access_time +HZ * dir_cache_timeout)) { cfid->on_list = false;list_move (&cfid->entry, &entry); cfids->num_entries--;fs/smb/client/cached_dir.h +1
-1@@ -39,6 +39,7 @@ struct cached_fid { bool on_list:1; bool file_all_info_is_valid:1;unsigned long time;/* jiffies of when lease was taken */ +unsigned long last_access_time;/* jiffies of when last accessed */ struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock; --Quoted 2 lines
2.43.0
--
Thanks,Steve
Signature · 3 lines
Thanks, Steve
From e4ae27d15ddeeb8c2763f6d038e2553a48afe34a Mon Sep 17 00:00:00 2001
From: Shyam Prasad N <[email protected]>
Date: Thu, 24 Jul 2025 22:23:53 -0500
Subject: [PATCH] cifs: add new field to track the last access time of cfid
The handlecache code today tracks the time at which dir lease was acquired and the laundromat thread uses that to check for old entries to cleanup.
However, if a directory is actively accessed, it should not be chosen to expire first.
This change adds a new last_access_time field to cfid and uses that to decide expiry of the cfid.
--- fs/smb/client/cached_dir.c | 6 ++++-- fs/smb/client/cached_dir.h | 1 + 2 files changed, 5 insertions(+), 2 deletions(-)
fs/smb/client/cached_dir.c +4 -2
@@ -195,6 +195,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,* from @cfids->entries. Caller will put last reference if the latter. */ if (cfid->has_lease && cfid->time) { + cfid->last_access_time = jiffies;spin_unlock (&cfids->cfid_list_lock); *ret_cfid = cfid;kfree (utf16_path);
@@ -363,6 +364,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->file_all_info_is_valid = true; cfid->time = jiffies; + cfid->last_access_time = jiffies;spin_unlock (&cfids->cfid_list_lock);/* At this point the directory handle is fully cached */ rc =0 ;
@@ -730,8 +732,8 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) { -if (cfid->time && -time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { +if (cfid->last_access_time && +time_after (jiffies, cfid->last_access_time +HZ * dir_cache_timeout)) { cfid->on_list = false;list_move (&cfid->entry, &entry); cfids->num_entries--;
fs/smb/client/cached_dir.h +1 -0
@@ -39,6 +39,7 @@ struct cached_fid { bool on_list:1; bool file_all_info_is_valid:1;unsigned long time;/* jiffies of when lease was taken */ +unsigned long last_access_time;/* jiffies of when last accessed */ struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock;
Signature · 1 line
2.43.0
Re: [PATCH 7/7] cifs: add new field to track the last access time of cfid
One minor improvement we could do - I noticed that this helps workloads like:
"ls /mnt/somedir ; do stuff ; ls /mnt/somedir (uses cached data from first) ; ... "ls /mnt/somedir"
but it doesn't help
"ls /mnt/somedir ; do stuff ; stat /mnt/somedir ; do stuff ; stat /mnt/somedir ; etc."
When "ls" is repeated then we bump up the timeout for that directory lease, but if ls is instead only followed by "stat" or "revalidate" (not ls/query_dir) of the directory (stat/revalidate is much more common than ls) then it doesn't bump up the timeout for the directory lease.
Maybe followon patch can determine when to bump up the timeout when revalidate/stat is frequently accessing the cached directory metadata
Quoted 90 lines · Steve French
Updated patch with description
Quoted 77 lines · Steve French
Rebased version of patch attached. See attached. Let me know if any objections. It solves an obvious important issue (we shouldn't be timing out directory leases if they are the most recently used ones)
Quoted 63 lines · On Wed, Jun 4, 2025 at 5:18 AM <[email protected]> wrote:
From: Shyam Prasad N <[email protected]>
The handlecache code today tracks the time at which dir lease was acquired and the laundromat thread uses that to check for old entries to cleanup.
However, if a directory is actively accessed, it should not be chosen to expire first.
This change adds a new
last_access_timefield to cfid and uses that to decide expiry of the cfid.Signed-off-by Shyam Prasad N <[email protected]>--- fs/smb/client/cached_dir.c | 6 ++++-- fs/smb/client/cached_dir.h | 1 + 2 files changed, 5 insertions(+), 2 deletions(-)fs/smb/client/cached_dir.c +4
-2@@ -213,6 +213,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon,*/ spin_lock (&cfid->fid_lock);if (cfid->has_lease && cfid->time) { + cfid->last_access_time = jiffies;spin_unlock (&cfid->fid_lock);mutex_unlock (&cfid->cfid_mutex); *ret_cfid = cfid;@@ -365,6 +366,7 @@ int open_cached_dir(unsigned int xid, struct cifs_tcon *tcon, cfid->tcon = tcon; cfid->is_open = true; cfid->time = jiffies; + cfid->last_access_time = jiffies;spin_unlock (&cfid->fid_lock); oshr_free:@@ -741,8 +743,8 @@ static void cfids_laundromat_worker(struct work_struct *work)spin_lock (&cfids->cfid_list_lock);list_for_each_entry_safe (cfid, q, &cfids->entries, entry) {spin_lock (&cfid->fid_lock); -if (cfid->time && -time_after (jiffies, cfid->time +HZ * dir_cache_timeout)) { +if (cfid->last_access_time && +time_after (jiffies, cfid->last_access_time +HZ * dir_cache_timeout)) { cfid->on_list = false;list_move (&cfid->entry, &entry); cfids->num_entries--;fs/smb/client/cached_dir.h +1
-1@@ -39,6 +39,7 @@ struct cached_fid { bool on_list:1; bool file_all_info_is_valid:1;unsigned long time;/* jiffies of when lease was taken */ +unsigned long last_access_time;/* jiffies of when last accessed */ struct kref refcount;struct cifs_fid fid;spinlock_t fid_lock; --Quoted 2 lines
2.43.0
--
Thanks,Quoted 1 line
Steve
--
Thanks,Steve
Signature · 3 lines
Thanks, Steve