ceph: rework dcache readdir

Previously our dcache readdir code relies on that child dentries in
directory dentry's d_subdir list are sorted by dentry's offset in
descending order. When adding dentries to the dcache, if a dentry
already exists, our readdir code moves it to head of directory
dentry's d_subdir list. This design relies on dcache internals.
Al Viro suggests using ncpfs's approach: keeping array of pointers
to dentries in page cache of directory inode. the validity of those
pointers are presented by directory inode's complete and ordered
flags. When a dentry gets pruned, we clear directory inode's complete
flag in the d_prune() callback. Before moving a dentry to other
directory, we clear the ordered flag for both old and new directory.
Signed-off-by: Yan, Zheng <zyan@redhat.com>

fs/ceph/caps.c

@@ -833,7 +833,9 @@ int __ceph_caps_used(struct ceph_inode_info *ci)
 		used |= CEPH_CAP_PIN;
 	if (ci->i_rd_ref)
 		used |= CEPH_CAP_FILE_RD;
-	if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages)
+	if (ci->i_rdcache_ref ||
+	    (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
+	     ci->vfs_inode.i_data.nrpages))
 		used |= CEPH_CAP_FILE_CACHE;
 	if (ci->i_wr_ref)
 		used |= CEPH_CAP_FILE_WR;
@@ -1651,9 +1653,10 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 	 * If we fail, it's because pages are locked.... try again later.
 	 */
 	if ((!is_delayed || mdsc->stopping) &&
-	    ci->i_wrbuffer_ref == 0 &&               /* no dirty pages... */
-	    inode->i_data.nrpages &&                 /* have cached pages */
-	    (file_wanted == 0 ||                     /* no open files */
+	    !S_ISDIR(inode->i_mode) &&		/* ignore readdir cache */
+	    ci->i_wrbuffer_ref == 0 &&		/* no dirty pages... */
+	    inode->i_data.nrpages &&		/* have cached pages */
+	    (file_wanted == 0 ||		/* no open files */
 	     (revoking & (CEPH_CAP_FILE_CACHE|
 			  CEPH_CAP_FILE_LAZYIO))) && /*  or revoking cache */
 	    !tried_invalidate) {
@@ -2805,7 +2808,8 @@ static void handle_cap_grant(struct ceph_mds_client *mdsc,
 	 * try to invalidate (once).  (If there are dirty buffers, we
 	 * will invalidate _after_ writeback.)
 	 */
-	if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
+	if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
+	    ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
 	    (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
 	    !ci->i_wrbuffer_ref) {
 		if (try_nonblocking_invalidate(inode)) {

fs/ceph/file.c

@@ -96,6 +96,7 @@ static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
 		}
 		cf->fmode = fmode;
 		cf->next_offset = 2;
+		cf->readdir_cache_idx = -1;
 		file->private_data = cf;
 		BUG_ON(inode->i_fop->release != ceph_release);
 		break;
@@ -324,7 +325,6 @@ int ceph_release(struct inode *inode, struct file *file)
 		ceph_mdsc_put_request(cf->last_readdir);
 	kfree(cf->last_name);
 	kfree(cf->dir_info);
-	dput(cf->dentry);
 	kmem_cache_free(ceph_file_cachep, cf);
 
 	/* wake up anyone waiting for caps on this inode */

fs/ceph/inode.c

@@ -390,9 +390,10 @@ struct inode *ceph_alloc_inode(struct super_block *sb)
 	ci->i_inline_version = 0;
 	ci->i_time_warp_seq = 0;
 	ci->i_ceph_flags = 0;
-	ci->i_ordered_count = 0;
-	atomic_set(&ci->i_release_count, 1);
-	atomic_set(&ci->i_complete_count, 0);
+	atomic64_set(&ci->i_ordered_count, 1);
+	atomic64_set(&ci->i_release_count, 1);
+	atomic64_set(&ci->i_complete_seq[0], 0);
+	atomic64_set(&ci->i_complete_seq[1], 0);
 	ci->i_symlink = NULL;
 
 	memset(&ci->i_dir_layout, 0, sizeof(ci->i_dir_layout));
@@ -860,9 +861,10 @@ static int fill_inode(struct inode *inode, struct page *locked_page,
 			    (issued & CEPH_CAP_FILE_EXCL) == 0 &&
 			    !__ceph_dir_is_complete(ci)) {
 				dout(" marking %p complete (empty)\n", inode);
+				i_size_write(inode, 0);
 				__ceph_dir_set_complete(ci,
-					atomic_read(&ci->i_release_count),
-					ci->i_ordered_count);
+					atomic64_read(&ci->i_release_count),
+					atomic64_read(&ci->i_ordered_count));
 			}
 
 			wake = true;
@@ -1214,6 +1216,10 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req,
 			dout("fill_trace doing d_move %p -> %p\n",
 			     req->r_old_dentry, dn);
 
+			/* d_move screws up sibling dentries' offsets */
+			ceph_dir_clear_ordered(dir);
+			ceph_dir_clear_ordered(olddir);
+
 			d_move(req->r_old_dentry, dn);
 			dout(" src %p '%pd' dst %p '%pd'\n",
 			     req->r_old_dentry,
@@ -1224,10 +1230,6 @@ int ceph_fill_trace(struct super_block *sb, struct ceph_mds_request *req,
 			   rehashing bug in vfs_rename_dir */
 			ceph_invalidate_dentry_lease(dn);
 
-			/* d_move screws up sibling dentries' offsets */
-			ceph_dir_clear_ordered(dir);
-			ceph_dir_clear_ordered(olddir);
-
 			dout("dn %p gets new offset %lld\n", req->r_old_dentry,
 			     ceph_dentry(req->r_old_dentry)->offset);
 
@@ -1335,6 +1337,49 @@ static int readdir_prepopulate_inodes_only(struct ceph_mds_request *req,
 	return err;
 }
 
+void ceph_readdir_cache_release(struct ceph_readdir_cache_control *ctl)
+{
+	if (ctl->page) {
+		kunmap(ctl->page);
+		page_cache_release(ctl->page);
+		ctl->page = NULL;
+	}
+}
+
+static int fill_readdir_cache(struct inode *dir, struct dentry *dn,
+			      struct ceph_readdir_cache_control *ctl,
+			      struct ceph_mds_request *req)
+{
+	struct ceph_inode_info *ci = ceph_inode(dir);
+	unsigned nsize = PAGE_CACHE_SIZE / sizeof(struct dentry*);
+	unsigned idx = ctl->index % nsize;
+	pgoff_t pgoff = ctl->index / nsize;
+
+	if (!ctl->page || pgoff != page_index(ctl->page)) {
+		ceph_readdir_cache_release(ctl);
+		ctl->page  = grab_cache_page(&dir->i_data, pgoff);
+		if (!ctl->page) {
+			ctl->index = -1;
+			return -ENOMEM;
+		}
+		/* reading/filling the cache are serialized by
+		 * i_mutex, no need to use page lock */
+		unlock_page(ctl->page);
+		ctl->dentries = kmap(ctl->page);
+	}
+
+	if (req->r_dir_release_cnt == atomic64_read(&ci->i_release_count) &&
+	    req->r_dir_ordered_cnt == atomic64_read(&ci->i_ordered_count)) {
+		dout("readdir cache dn %p idx %d\n", dn, ctl->index);
+		ctl->dentries[idx] = dn;
+		ctl->index++;
+	} else {
+		dout("disable readdir cache\n");
+		ctl->index = -1;
+	}
+	return 0;
+}
+
 int ceph_readdir_prepopulate(struct ceph_mds_request *req,
 			     struct ceph_mds_session *session)
 {
@@ -1347,8 +1392,11 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
 	struct inode *snapdir = NULL;
 	struct ceph_mds_request_head *rhead = req->r_request->front.iov_base;
 	struct ceph_dentry_info *di;
-	u64 r_readdir_offset = req->r_readdir_offset;
 	u32 frag = le32_to_cpu(rhead->args.readdir.frag);
+	struct ceph_readdir_cache_control cache_ctl = {};
+
+	if (req->r_aborted)
+		return readdir_prepopulate_inodes_only(req, session);
 
 	if (rinfo->dir_dir &&
 	    le32_to_cpu(rinfo->dir_dir->frag) != frag) {
@@ -1356,14 +1404,11 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
 		     frag, le32_to_cpu(rinfo->dir_dir->frag));
 		frag = le32_to_cpu(rinfo->dir_dir->frag);
 		if (ceph_frag_is_leftmost(frag))
-			r_readdir_offset = 2;
+			req->r_readdir_offset = 2;
 		else
-			r_readdir_offset = 0;
+			req->r_readdir_offset = 0;
 	}
 
-	if (req->r_aborted)
-		return readdir_prepopulate_inodes_only(req, session);
-
 	if (le32_to_cpu(rinfo->head->op) == CEPH_MDS_OP_LSSNAP) {
 		snapdir = ceph_get_snapdir(d_inode(parent));
 		parent = d_find_alias(snapdir);
@@ -1376,6 +1421,17 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
 			ceph_fill_dirfrag(d_inode(parent), rinfo->dir_dir);
 	}
 
+	if (ceph_frag_is_leftmost(frag) && req->r_readdir_offset == 2) {
+		/* note dir version at start of readdir so we can tell
+		 * if any dentries get dropped */
+		struct ceph_inode_info *ci = ceph_inode(d_inode(parent));
+		req->r_dir_release_cnt = atomic64_read(&ci->i_release_count);
+		req->r_dir_ordered_cnt = atomic64_read(&ci->i_ordered_count);
+		req->r_readdir_cache_idx = 0;
+	}
+
+	cache_ctl.index = req->r_readdir_cache_idx;
+
 	/* FIXME: release caps/leases if error occurs */
 	for (i = 0; i < rinfo->dir_nr; i++) {
 		struct ceph_vino vino;
@@ -1415,13 +1471,6 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
 			d_delete(dn);
 			dput(dn);
 			goto retry_lookup;
-		} else {
-			/* reorder parent's d_subdirs */
-			spin_lock(&parent->d_lock);
-			spin_lock_nested(&dn->d_lock, DENTRY_D_LOCK_NESTED);
-			list_move(&dn->d_child, &parent->d_subdirs);
-			spin_unlock(&dn->d_lock);
-			spin_unlock(&parent->d_lock);
 		}
 
 		/* inode */
@@ -1438,13 +1487,15 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
 			}
 		}
 
-		if (fill_inode(in, NULL, &rinfo->dir_in[i], NULL, session,
-			       req->r_request_started, -1,
-			       &req->r_caps_reservation) < 0) {
+		ret = fill_inode(in, NULL, &rinfo->dir_in[i], NULL, session,
+				 req->r_request_started, -1,
+				 &req->r_caps_reservation);
+		if (ret < 0) {
 			pr_err("fill_inode badness on %p\n", in);
 			if (d_really_is_negative(dn))
 				iput(in);
 			d_drop(dn);
+			err = ret;
 			goto next_item;
 		}
 
@@ -1460,19 +1511,28 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
 		}
 
 		di = dn->d_fsdata;
-		di->offset = ceph_make_fpos(frag, i + r_readdir_offset);
+		di->offset = ceph_make_fpos(frag, i + req->r_readdir_offset);
 
 		update_dentry_lease(dn, rinfo->dir_dlease[i],
 				    req->r_session,
 				    req->r_request_started);
+
+		if (err == 0 && cache_ctl.index >= 0) {
+			ret = fill_readdir_cache(d_inode(parent), dn,
+						 &cache_ctl, req);
+			if (ret < 0)
+				err = ret;
+		}
 next_item:
 		if (dn)
 			dput(dn);
 	}
-	if (err == 0)
-		req->r_did_prepopulate = true;
-
 out:
+	if (err == 0) {
+		req->r_did_prepopulate = true;
+		req->r_readdir_cache_idx = cache_ctl.index;
+	}
+	ceph_readdir_cache_release(&cache_ctl);
 	if (snapdir) {
 		iput(snapdir);
 		dput(parent);

fs/ceph/mds_client.h

@@ -253,6 +253,9 @@ struct ceph_mds_request {
 	bool		  r_got_unsafe, r_got_safe, r_got_result;
 
 	bool              r_did_prepopulate;
+	long long	  r_dir_release_cnt;
+	long long	  r_dir_ordered_cnt;
+	int		  r_readdir_cache_idx;
 	u32               r_readdir_offset;
 
 	struct ceph_cap_reservation r_caps_reservation;

fs/ceph/super.h

@@ -282,9 +282,9 @@ struct ceph_inode_info {
 	u32 i_time_warp_seq;
 
 	unsigned i_ceph_flags;
-	int i_ordered_count;
-	atomic_t i_release_count;
-	atomic_t i_complete_count;
+	atomic64_t i_release_count;
+	atomic64_t i_ordered_count;
+	atomic64_t i_complete_seq[2];
 
 	struct ceph_dir_layout i_dir_layout;
 	struct ceph_file_layout i_layout;
@@ -471,30 +471,36 @@ static inline struct inode *ceph_find_inode(struct super_block *sb,
 
 
 static inline void __ceph_dir_set_complete(struct ceph_inode_info *ci,
-					   int release_count, int ordered_count)
+					   long long release_count,
+					   long long ordered_count)
 {
-	atomic_set(&ci->i_complete_count, release_count);
-	if (ci->i_ordered_count == ordered_count)
-		ci->i_ceph_flags |= CEPH_I_DIR_ORDERED;
-	else
-		ci->i_ceph_flags &= ~CEPH_I_DIR_ORDERED;
+	smp_mb__before_atomic();
+	atomic64_set(&ci->i_complete_seq[0], release_count);
+	atomic64_set(&ci->i_complete_seq[1], ordered_count);
 }
 
 static inline void __ceph_dir_clear_complete(struct ceph_inode_info *ci)
 {
-	atomic_inc(&ci->i_release_count);
+	atomic64_inc(&ci->i_release_count);
+}
+
+static inline void __ceph_dir_clear_ordered(struct ceph_inode_info *ci)
+{
+	atomic64_inc(&ci->i_ordered_count);
 }
 
 static inline bool __ceph_dir_is_complete(struct ceph_inode_info *ci)
 {
-	return atomic_read(&ci->i_complete_count) ==
-		atomic_read(&ci->i_release_count);
+	return atomic64_read(&ci->i_complete_seq[0]) ==
+		atomic64_read(&ci->i_release_count);
 }
 
 static inline bool __ceph_dir_is_complete_ordered(struct ceph_inode_info *ci)
 {
-	return __ceph_dir_is_complete(ci) &&
-		(ci->i_ceph_flags & CEPH_I_DIR_ORDERED);
+	return  atomic64_read(&ci->i_complete_seq[0]) ==
+		atomic64_read(&ci->i_release_count) &&
+		atomic64_read(&ci->i_complete_seq[1]) ==
+		atomic64_read(&ci->i_ordered_count);
 }
 
 static inline void ceph_dir_clear_complete(struct inode *inode)
@@ -504,20 +510,13 @@ static inline void ceph_dir_clear_complete(struct inode *inode)
 
 static inline void ceph_dir_clear_ordered(struct inode *inode)
 {
-	struct ceph_inode_info *ci = ceph_inode(inode);
-	spin_lock(&ci->i_ceph_lock);
-	ci->i_ordered_count++;
-	ci->i_ceph_flags &= ~CEPH_I_DIR_ORDERED;
-	spin_unlock(&ci->i_ceph_lock);
+	__ceph_dir_clear_ordered(ceph_inode(inode));
 }
 
 static inline bool ceph_dir_is_complete_ordered(struct inode *inode)
 {
-	struct ceph_inode_info *ci = ceph_inode(inode);
-	bool ret;
-	spin_lock(&ci->i_ceph_lock);
-	ret = __ceph_dir_is_complete_ordered(ci);
-	spin_unlock(&ci->i_ceph_lock);
+	bool ret = __ceph_dir_is_complete_ordered(ceph_inode(inode));
+	smp_rmb();
 	return ret;
 }
 
@@ -636,16 +635,20 @@ struct ceph_file_info {
 	unsigned offset;       /* offset of last chunk, adjusted for . and .. */
 	unsigned next_offset;  /* offset of next chunk (last_name's + 1) */
 	char *last_name;       /* last entry in previous chunk */
-	struct dentry *dentry; /* next dentry (for dcache readdir) */
-	int dir_release_count;
-	int dir_ordered_count;
+	long long dir_release_count;
+	long long dir_ordered_count;
+	int readdir_cache_idx;
 
 	/* used for -o dirstat read() on directory thing */
 	char *dir_info;
 	int dir_info_len;
 };
 
-
+struct ceph_readdir_cache_control {
+	struct page  *page;
+	struct dentry **dentries;
+	int index;
+};
 
 /*
  * A "snap realm" describes a subset of the file hierarchy sharing
@@ -944,6 +947,7 @@ extern void ceph_dentry_lru_del(struct dentry *dn);
 extern void ceph_invalidate_dentry_lease(struct dentry *dentry);
 extern unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn);
 extern struct inode *ceph_get_dentry_parent_inode(struct dentry *dentry);
+extern void ceph_readdir_cache_release(struct ceph_readdir_cache_control *ctl);
 
 /*
  * our d_ops vary depending on whether the inode is live,