NFSD: Revert 6c41d9a9bd0298002805758216a9c44e38a8500d
authorChuck Lever <chuck.lever@oracle.com>
Sat, 16 Dec 2023 16:57:43 +0000 (11:57 -0500)
committerChuck Lever <chuck.lever@oracle.com>
Mon, 18 Dec 2023 16:22:16 +0000 (11:22 -0500)
For some reason, the wait_on_bit() in nfsd4_deleg_getattr_conflict()
is waiting forever, preventing a clean server shutdown. The
requesting client might also hang waiting for a reply to the
conflicting GETATTR.

Invoking wait_on_bit() in an nfsd thread context is a hazard. The
correct fix is to replace this wait_on_bit() call site with a
mechanism that defers the conflicting GETATTR until the CB_GETATTR
completes or is known to have failed.

That will require some surgery and extended testing and it's late
in the v6.7-rc cycle, so I'm reverting now in favor of trying again
in a subsequent kernel release.

This is my fault: I should have recognized the ramifications of
calling wait_on_bit() in here before accepting this patch.

Thanks to Dai Ngo <dai.ngo@oracle.com> for diagnosing the issue.

Reported-by: Wolfgang Walter <linux-nfs@stwm.de>
Closes: https://lore.kernel.org/linux-nfs/e3d43ecdad554fbdcaa7181833834f78@stwm.de/
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
fs/nfsd/nfs4state.c
fs/nfsd/nfs4xdr.c
fs/nfsd/state.h

index 3709e58..31909f5 100644 (file)
@@ -127,7 +127,6 @@ static void free_session(struct nfsd4_session *);
 
 static const struct nfsd4_callback_ops nfsd4_cb_recall_ops;
 static const struct nfsd4_callback_ops nfsd4_cb_notify_lock_ops;
-static const struct nfsd4_callback_ops nfsd4_cb_getattr_ops;
 
 static struct workqueue_struct *laundry_wq;
 
@@ -1190,10 +1189,6 @@ alloc_init_deleg(struct nfs4_client *clp, struct nfs4_file *fp,
        dp->dl_recalled = false;
        nfsd4_init_cb(&dp->dl_recall, dp->dl_stid.sc_client,
                      &nfsd4_cb_recall_ops, NFSPROC4_CLNT_CB_RECALL);
-       nfsd4_init_cb(&dp->dl_cb_fattr.ncf_getattr, dp->dl_stid.sc_client,
-                       &nfsd4_cb_getattr_ops, NFSPROC4_CLNT_CB_GETATTR);
-       dp->dl_cb_fattr.ncf_file_modified = false;
-       dp->dl_cb_fattr.ncf_cb_bmap[0] = FATTR4_WORD0_CHANGE | FATTR4_WORD0_SIZE;
        get_nfs4_file(fp);
        dp->dl_stid.sc_file = fp;
        return dp;
@@ -2901,56 +2896,11 @@ nfsd4_cb_recall_any_release(struct nfsd4_callback *cb)
        spin_unlock(&nn->client_lock);
 }
 
-static int
-nfsd4_cb_getattr_done(struct nfsd4_callback *cb, struct rpc_task *task)
-{
-       struct nfs4_cb_fattr *ncf =
-                       container_of(cb, struct nfs4_cb_fattr, ncf_getattr);
-
-       ncf->ncf_cb_status = task->tk_status;
-       switch (task->tk_status) {
-       case -NFS4ERR_DELAY:
-               rpc_delay(task, 2 * HZ);
-               return 0;
-       default:
-               return 1;
-       }
-}
-
-static void
-nfsd4_cb_getattr_release(struct nfsd4_callback *cb)
-{
-       struct nfs4_cb_fattr *ncf =
-                       container_of(cb, struct nfs4_cb_fattr, ncf_getattr);
-       struct nfs4_delegation *dp =
-                       container_of(ncf, struct nfs4_delegation, dl_cb_fattr);
-
-       nfs4_put_stid(&dp->dl_stid);
-       clear_bit(CB_GETATTR_BUSY, &ncf->ncf_cb_flags);
-       wake_up_bit(&ncf->ncf_cb_flags, CB_GETATTR_BUSY);
-}
-
 static const struct nfsd4_callback_ops nfsd4_cb_recall_any_ops = {
        .done           = nfsd4_cb_recall_any_done,
        .release        = nfsd4_cb_recall_any_release,
 };
 
-static const struct nfsd4_callback_ops nfsd4_cb_getattr_ops = {
-       .done           = nfsd4_cb_getattr_done,
-       .release        = nfsd4_cb_getattr_release,
-};
-
-void nfs4_cb_getattr(struct nfs4_cb_fattr *ncf)
-{
-       struct nfs4_delegation *dp =
-                       container_of(ncf, struct nfs4_delegation, dl_cb_fattr);
-
-       if (test_and_set_bit(CB_GETATTR_BUSY, &ncf->ncf_cb_flags))
-               return;
-       refcount_inc(&dp->dl_stid.sc_count);
-       nfsd4_run_cb(&ncf->ncf_getattr);
-}
-
 static struct nfs4_client *create_client(struct xdr_netobj name,
                struct svc_rqst *rqstp, nfs4_verifier *verf)
 {
@@ -5686,8 +5636,6 @@ nfs4_open_delegation(struct nfsd4_open *open, struct nfs4_ol_stateid *stp,
        struct svc_fh *parent = NULL;
        int cb_up;
        int status = 0;
-       struct kstat stat;
-       struct path path;
 
        cb_up = nfsd4_cb_channel_good(oo->oo_owner.so_client);
        open->op_recall = false;
@@ -5725,18 +5673,6 @@ nfs4_open_delegation(struct nfsd4_open *open, struct nfs4_ol_stateid *stp,
        if (open->op_share_access & NFS4_SHARE_ACCESS_WRITE) {
                open->op_delegate_type = NFS4_OPEN_DELEGATE_WRITE;
                trace_nfsd_deleg_write(&dp->dl_stid.sc_stateid);
-               path.mnt = currentfh->fh_export->ex_path.mnt;
-               path.dentry = currentfh->fh_dentry;
-               if (vfs_getattr(&path, &stat,
-                               (STATX_SIZE | STATX_CTIME | STATX_CHANGE_COOKIE),
-                               AT_STATX_SYNC_AS_STAT)) {
-                       nfs4_put_stid(&dp->dl_stid);
-                       destroy_delegation(dp);
-                       goto out_no_deleg;
-               }
-               dp->dl_cb_fattr.ncf_cur_fsize = stat.size;
-               dp->dl_cb_fattr.ncf_initial_cinfo =
-                       nfsd4_change_attribute(&stat, d_inode(currentfh->fh_dentry));
        } else {
                open->op_delegate_type = NFS4_OPEN_DELEGATE_READ;
                trace_nfsd_deleg_read(&dp->dl_stid.sc_stateid);
@@ -8489,8 +8425,6 @@ nfsd4_get_writestateid(struct nfsd4_compound_state *cstate,
  * nfsd4_deleg_getattr_conflict - Recall if GETATTR causes conflict
  * @rqstp: RPC transaction context
  * @inode: file to be checked for a conflict
- * @modified: return true if file was modified
- * @size: new size of file if modified is true
  *
  * This function is called when there is a conflict between a write
  * delegation and a change/size GETATTR from another client. The server
@@ -8499,23 +8433,21 @@ nfsd4_get_writestateid(struct nfsd4_compound_state *cstate,
  * delegation before replying to the GETATTR. See RFC 8881 section
  * 18.7.4.
  *
+ * The current implementation does not support CB_GETATTR yet. However
+ * this can avoid recalling the delegation could be added in follow up
+ * work.
+ *
  * Returns 0 if there is no conflict; otherwise an nfs_stat
  * code is returned.
  */
 __be32
-nfsd4_deleg_getattr_conflict(struct svc_rqst *rqstp, struct inode *inode,
-                            bool *modified, u64 *size)
+nfsd4_deleg_getattr_conflict(struct svc_rqst *rqstp, struct inode *inode)
 {
+       __be32 status;
        struct file_lock_context *ctx;
-       struct nfs4_delegation *dp;
-       struct nfs4_cb_fattr *ncf;
        struct file_lock *fl;
-       struct iattr attrs;
-       __be32 status;
-
-       might_sleep();
+       struct nfs4_delegation *dp;
 
-       *modified = false;
        ctx = locks_inode_context(inode);
        if (!ctx)
                return 0;
@@ -8542,34 +8474,10 @@ nfsd4_deleg_getattr_conflict(struct svc_rqst *rqstp, struct inode *inode,
 break_lease:
                        spin_unlock(&ctx->flc_lock);
                        nfsd_stats_wdeleg_getattr_inc();
-
-                       dp = fl->fl_owner;
-                       ncf = &dp->dl_cb_fattr;
-                       nfs4_cb_getattr(&dp->dl_cb_fattr);
-                       wait_on_bit(&ncf->ncf_cb_flags, CB_GETATTR_BUSY, TASK_INTERRUPTIBLE);
-                       if (ncf->ncf_cb_status) {
-                               status = nfserrno(nfsd_open_break_lease(inode, NFSD_MAY_READ));
-                               if (status != nfserr_jukebox ||
-                                               !nfsd_wait_for_delegreturn(rqstp, inode))
-                                       return status;
-                       }
-                       if (!ncf->ncf_file_modified &&
-                                       (ncf->ncf_initial_cinfo != ncf->ncf_cb_change ||
-                                       ncf->ncf_cur_fsize != ncf->ncf_cb_fsize))
-                               ncf->ncf_file_modified = true;
-                       if (ncf->ncf_file_modified) {
-                               /*
-                                * The server would not update the file's metadata
-                                * with the client's modified size.
-                                */
-                               attrs.ia_mtime = attrs.ia_ctime = current_time(inode);
-                               attrs.ia_valid = ATTR_MTIME | ATTR_CTIME;
-                               setattr_copy(&nop_mnt_idmap, inode, &attrs);
-                               mark_inode_dirty(inode);
-                               ncf->ncf_cur_fsize = ncf->ncf_cb_fsize;
-                               *size = ncf->ncf_cur_fsize;
-                               *modified = true;
-                       }
+                       status = nfserrno(nfsd_open_break_lease(inode, NFSD_MAY_READ));
+                       if (status != nfserr_jukebox ||
+                                       !nfsd_wait_for_delegreturn(rqstp, inode))
+                               return status;
                        return 0;
                }
                break;
index ec4ed62..b499fe9 100644 (file)
@@ -3505,9 +3505,7 @@ nfsd4_encode_fattr4(struct svc_rqst *rqstp, struct xdr_stream *xdr,
                u32             attrmask[3];
                unsigned long   mask[2];
        } u;
-       bool file_modified;
        unsigned long bit;
-       u64 size = 0;
 
        WARN_ON_ONCE(bmval[1] & NFSD_WRITEONLY_ATTRS_WORD1);
        WARN_ON_ONCE(!nfsd_attrs_supported(minorversion, bmval));
@@ -3534,8 +3532,7 @@ nfsd4_encode_fattr4(struct svc_rqst *rqstp, struct xdr_stream *xdr,
        }
        args.size = 0;
        if (u.attrmask[0] & (FATTR4_WORD0_CHANGE | FATTR4_WORD0_SIZE)) {
-               status = nfsd4_deleg_getattr_conflict(rqstp, d_inode(dentry),
-                                                     &file_modified, &size);
+               status = nfsd4_deleg_getattr_conflict(rqstp, d_inode(dentry));
                if (status)
                        goto out;
        }
@@ -3545,7 +3542,7 @@ nfsd4_encode_fattr4(struct svc_rqst *rqstp, struct xdr_stream *xdr,
                          AT_STATX_SYNC_AS_STAT);
        if (err)
                goto out_nfserr;
-       args.size = file_modified ? size : args.stat.size;
+       args.size = args.stat.size;
 
        if (!(args.stat.result_mask & STATX_BTIME))
                /* underlying FS does not offer btime so we can't share it */
index f96eaa8..0bbbe57 100644 (file)
@@ -125,16 +125,8 @@ struct nfs4_cb_fattr {
        /* from CB_GETATTR reply */
        u64 ncf_cb_change;
        u64 ncf_cb_fsize;
-
-       unsigned long ncf_cb_flags;
-       bool ncf_file_modified;
-       u64 ncf_initial_cinfo;
-       u64 ncf_cur_fsize;
 };
 
-/* bits for ncf_cb_flags */
-#define        CB_GETATTR_BUSY         0
-
 /*
  * Represents a delegation stateid. The nfs4_client holds references to these
  * and they are put when it is being destroyed or when the delegation is
@@ -754,6 +746,5 @@ static inline bool try_to_expire_client(struct nfs4_client *clp)
 }
 
 extern __be32 nfsd4_deleg_getattr_conflict(struct svc_rqst *rqstp,
-               struct inode *inode, bool *file_modified, u64 *size);
-extern void nfs4_cb_getattr(struct nfs4_cb_fattr *ncf);
+                               struct inode *inode);
 #endif   /* NFSD4_STATE_H */