afs: Use the contained hashtable to search a directory

Each directory image contains a hashtable with 128 buckets to speed up
searching.  Currently, kafs does not use this, but rather iterates over all
the occupied slots in the image as it can share this with readdir.

Switch kafs to use the hashtable for lookups to reduce the latency.  Care
must be taken that the hash chains are acyclic.

Signed-off-by: David Howells <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
cc: Marc Dionne <[email protected]>
cc: [email protected]
Signed-off-by: Christian Brauner <[email protected]>

Author: dhowells

fs/afs/Makefile

@@ -11,6 +11,7 @@ kafs-y := \
 	cmservice.o \
 	dir.o \
 	dir_edit.o \
+	dir_search.o \
 	dir_silly.o \
 	dynroot.o \
 	file.o \

fs/afs/dir.c

@@ -88,8 +88,6 @@ struct afs_lookup_one_cookie {
 struct afs_lookup_cookie {
 	struct dir_context	ctx;
 	struct qstr		name;
-	bool			found;
-	bool			one_only;
 	unsigned short		nr_fids;
 	struct afs_fid		fids[50];
 };
@@ -309,7 +307,7 @@ ssize_t afs_read_single(struct afs_vnode *dvnode, struct file *file)
  * Read the directory into a folio_queue buffer in one go, scrubbing the
  * previous contents.  We return -ESTALE if the caller needs to call us again.
  */
-static ssize_t afs_read_dir(struct afs_vnode *dvnode, struct file *file)
+ssize_t afs_read_dir(struct afs_vnode *dvnode, struct file *file)
 	__acquires(&dvnode->validate_lock)
 {
 	ssize_t ret;
@@ -649,19 +647,10 @@ static bool afs_lookup_filldir(struct dir_context *ctx, const char *name,
 	BUILD_BUG_ON(sizeof(union afs_xdr_dir_block) != 2048);
 	BUILD_BUG_ON(sizeof(union afs_xdr_dirent) != 32);
 
-	if (cookie->found) {
-		if (cookie->nr_fids < 50) {
-			cookie->fids[cookie->nr_fids].vnode	= ino;
-			cookie->fids[cookie->nr_fids].unique	= dtype;
-			cookie->nr_fids++;
-		}
-	} else if (cookie->name.len == nlen &&
-		   memcmp(cookie->name.name, name, nlen) == 0) {
-		cookie->fids[1].vnode	= ino;
-		cookie->fids[1].unique	= dtype;
-		cookie->found = 1;
-		if (cookie->one_only)
-			return false;
+	if (cookie->nr_fids < 50) {
+		cookie->fids[cookie->nr_fids].vnode	= ino;
+		cookie->fids[cookie->nr_fids].unique	= dtype;
+		cookie->nr_fids++;
 	}
 
 	return cookie->nr_fids < 50;
@@ -789,6 +778,7 @@ static struct inode *afs_do_lookup(struct inode *dir, struct dentry *dentry)
 	struct afs_vnode *dvnode = AFS_FS_I(dir), *vnode;
 	struct inode *inode = NULL, *ti;
 	afs_dataversion_t data_version = READ_ONCE(dvnode->status.data_version);
+	bool supports_ibulk;
 	long ret;
 	int i;
 
@@ -805,19 +795,19 @@ static struct inode *afs_do_lookup(struct inode *dir, struct dentry *dentry)
 	cookie->nr_fids = 2; /* slot 1 is saved for the fid we actually want
 			      * and slot 0 for the directory */
 
-	if (!afs_server_supports_ibulk(dvnode))
-		cookie->one_only = true;
-
-	/* search the directory */
-	ret = afs_dir_iterate(dir, &cookie->ctx, NULL, &data_version);
+	/* Search the directory for the named entry using the hash table... */
+	ret = afs_dir_search(dvnode, &dentry->d_name, &cookie->fids[1], &data_version);
 	if (ret < 0)
 		goto out;
 
-	dentry->d_fsdata = (void *)(unsigned long)data_version;
+	supports_ibulk = afs_server_supports_ibulk(dvnode);
+	if (supports_ibulk) {
+		/* ...then scan linearly from that point for entries to lookup-ahead. */
+		cookie->ctx.pos = (ret + 1) * AFS_DIR_DIRENT_SIZE;
+		afs_dir_iterate(dir, &cookie->ctx, NULL, &data_version);
+	}
 
-	ret = -ENOENT;
-	if (!cookie->found)
-		goto out;
+	dentry->d_fsdata = (void *)(unsigned long)data_version;
 
 	/* Check to see if we already have an inode for the primary fid. */
 	inode = ilookup5(dir->i_sb, cookie->fids[1].vnode,
@@ -876,7 +866,7 @@ static struct inode *afs_do_lookup(struct inode *dir, struct dentry *dentry)
 	 * the whole operation.
 	 */
 	afs_op_set_error(op, -ENOTSUPP);
-	if (!cookie->one_only) {
+	if (supports_ibulk) {
 		op->ops = &afs_inline_bulk_status_operation;
 		afs_begin_vnode_operation(op);
 		afs_wait_for_operation(op);

fs/afs/dir_edit.c

@@ -245,7 +245,7 @@ void afs_edit_dir_add(struct afs_vnode *vnode,
 	union afs_xdr_dir_block *meta, *block;
 	union afs_xdr_dirent *de;
 	struct afs_dir_iter iter = { .dvnode = vnode };
-	unsigned int need_slots, nr_blocks, b;
+	unsigned int nr_blocks, b, entry;
 	loff_t i_size;
 	int slot;
 
@@ -263,7 +263,7 @@ void afs_edit_dir_add(struct afs_vnode *vnode,
 		return;
 
 	/* Work out how many slots we're going to need. */
-	need_slots = afs_dir_calc_slots(name->len);
+	iter.nr_slots = afs_dir_calc_slots(name->len);
 
 	if (i_size == 0)
 		goto new_directory;
@@ -281,7 +281,7 @@ void afs_edit_dir_add(struct afs_vnode *vnode,
 
 		/* Lower dir blocks have a counter in the header we can check. */
 		if (b < AFS_DIR_BLOCKS_WITH_CTR &&
-		    meta->meta.alloc_ctrs[b] < need_slots)
+		    meta->meta.alloc_ctrs[b] < iter.nr_slots)
 			continue;
 
 		block = afs_dir_get_block(&iter, b);
@@ -308,7 +308,7 @@ void afs_edit_dir_add(struct afs_vnode *vnode,
 		/* We need to try and find one or more consecutive slots to
 		 * hold the entry.
 		 */
-		slot = afs_find_contig_bits(block, need_slots);
+		slot = afs_find_contig_bits(block, iter.nr_slots);
 		if (slot >= 0) {
 			_debug("slot %u", slot);
 			goto found_space;
@@ -347,12 +347,18 @@ void afs_edit_dir_add(struct afs_vnode *vnode,
 	de->u.name[name->len] = 0;
 
 	/* Adjust the bitmap. */
-	afs_set_contig_bits(block, slot, need_slots);
-	kunmap_local(block);
+	afs_set_contig_bits(block, slot, iter.nr_slots);
 
 	/* Adjust the allocation counter. */
 	if (b < AFS_DIR_BLOCKS_WITH_CTR)
-		meta->meta.alloc_ctrs[b] -= need_slots;
+		meta->meta.alloc_ctrs[b] -= iter.nr_slots;
+
+	/* Adjust the hash chain. */
+	entry = b * AFS_DIR_SLOTS_PER_BLOCK + slot;
+	iter.bucket = afs_dir_hash_name(name);
+	de->u.hash_next = meta->meta.hashtable[iter.bucket];
+	meta->meta.hashtable[iter.bucket] = htons(entry);
+	kunmap_local(block);
 
 	inode_inc_iversion_raw(&vnode->netfs.inode);
 	afs_stat_v(vnode, n_dir_cr);
@@ -387,12 +393,14 @@ void afs_edit_dir_add(struct afs_vnode *vnode,
 void afs_edit_dir_remove(struct afs_vnode *vnode,
 			 struct qstr *name, enum afs_edit_dir_reason why)
 {
-	union afs_xdr_dir_block *meta, *block;
-	union afs_xdr_dirent *de;
+	union afs_xdr_dir_block *meta, *block, *pblock;
+	union afs_xdr_dirent *de, *pde;
 	struct afs_dir_iter iter = { .dvnode = vnode };
-	unsigned int need_slots, nr_blocks, b;
+	struct afs_fid fid;
+	unsigned int b, slot, entry;
 	loff_t i_size;
-	int slot;
+	__be16 next;
+	int found;
 
 	_enter(",,{%d,%s},", name->len, name->name);
 
@@ -403,59 +411,90 @@ void afs_edit_dir_remove(struct afs_vnode *vnode,
 		afs_invalidate_dir(vnode, afs_dir_invalid_edit_rem_bad_size);
 		return;
 	}
-	nr_blocks = i_size / AFS_DIR_BLOCK_SIZE;
 
-	meta = afs_dir_get_block(&iter, 0);
-	if (!meta)
+	if (!afs_dir_init_iter(&iter, name))
 		return;
 
-	/* Work out how many slots we're going to discard. */
-	need_slots = afs_dir_calc_slots(name->len);
-
-	/* Find a block that has sufficient slots available.  Each folio
-	 * contains two or more directory blocks.
-	 */
-	for (b = 0; b < nr_blocks; b++) {
-		block = afs_dir_get_block(&iter, b);
-		if (!block)
-			goto error;
-
-		/* Abandon the edit if we got a callback break. */
-		if (!test_bit(AFS_VNODE_DIR_VALID, &vnode->flags))
-			goto already_invalidated;
-
-		if (b > AFS_DIR_BLOCKS_WITH_CTR ||
-		    meta->meta.alloc_ctrs[b] <= AFS_DIR_SLOTS_PER_BLOCK - 1 - need_slots) {
-			slot = afs_dir_scan_block(block, name, b);
-			if (slot >= 0)
-				goto found_dirent;
-		}
+	meta = afs_dir_find_block(&iter, 0);
+	if (!meta)
+		return;
 
-		kunmap_local(block);
+	/* Find the entry in the blob. */
+	found = afs_dir_search_bucket(&iter, name, &fid);
+	if (found < 0) {
+		/* Didn't find the dirent to clobber.  Re-download. */
+		trace_afs_edit_dir(vnode, why, afs_edit_dir_delete_noent,
+				   0, 0, 0, 0, name->name);
+		afs_invalidate_dir(vnode, afs_dir_invalid_edit_rem_wrong_name);
+		goto out_unmap;
 	}
 
-	/* Didn't find the dirent to clobber.  Download the directory again. */
-	trace_afs_edit_dir(vnode, why, afs_edit_dir_delete_noent,
-			   0, 0, 0, 0, name->name);
-	afs_invalidate_dir(vnode, afs_dir_invalid_edit_rem_wrong_name);
-	goto out_unmap;
+	entry = found;
+	b    = entry / AFS_DIR_SLOTS_PER_BLOCK;
+	slot = entry % AFS_DIR_SLOTS_PER_BLOCK;
 
-found_dirent:
+	block = afs_dir_find_block(&iter, b);
+	if (!block)
+		goto error;
+	if (!test_bit(AFS_VNODE_DIR_VALID, &vnode->flags))
+		goto already_invalidated;
+
+	/* Check and clear the entry. */
 	de = &block->dirents[slot];
+	if (de->u.valid != 1)
+		goto error_unmap;
 
 	trace_afs_edit_dir(vnode, why, afs_edit_dir_delete, b, slot,
 			   ntohl(de->u.vnode), ntohl(de->u.unique),
 			   name->name);
 
-	memset(de, 0, sizeof(*de) * need_slots);
-
 	/* Adjust the bitmap. */
-	afs_clear_contig_bits(block, slot, need_slots);
-	kunmap_local(block);
+	afs_clear_contig_bits(block, slot, iter.nr_slots);
 
 	/* Adjust the allocation counter. */
 	if (b < AFS_DIR_BLOCKS_WITH_CTR)
-		meta->meta.alloc_ctrs[b] += need_slots;
+		meta->meta.alloc_ctrs[b] += iter.nr_slots;
+
+	/* Clear the constituent entries. */
+	next = de->u.hash_next;
+	memset(de, 0, sizeof(*de) * iter.nr_slots);
+	kunmap_local(block);
+
+	/* Adjust the hash chain: if iter->prev_entry is 0, the hashtable head
+	 * index is previous; otherwise it's slot number of the previous entry.
+	 */
+	if (!iter.prev_entry) {
+		__be16 prev_next = meta->meta.hashtable[iter.bucket];
+
+		if (unlikely(prev_next != htons(entry))) {
+			pr_warn("%llx:%llx:%x: not head of chain b=%x p=%x,%x e=%x %*s",
+				vnode->fid.vid, vnode->fid.vnode, vnode->fid.unique,
+				iter.bucket, iter.prev_entry, prev_next, entry,
+				name->len, name->name);
+			goto error;
+		}
+		meta->meta.hashtable[iter.bucket] = next;
+	} else {
+		unsigned int pb = iter.prev_entry / AFS_DIR_SLOTS_PER_BLOCK;
+		unsigned int ps = iter.prev_entry % AFS_DIR_SLOTS_PER_BLOCK;
+		__be16 prev_next;
+
+		pblock = afs_dir_find_block(&iter, pb);
+		if (!pblock)
+			goto error;
+		pde = &pblock->dirents[ps];
+		prev_next = pde->u.hash_next;
+		if (prev_next != htons(entry)) {
+			kunmap_local(pblock);
+			pr_warn("%llx:%llx:%x: not prev in chain b=%x p=%x,%x e=%x %*s",
+				vnode->fid.vid, vnode->fid.vnode, vnode->fid.unique,
+				iter.bucket, iter.prev_entry, prev_next, entry,
+				name->len, name->name);
+			goto error;
+		}
+		pde->u.hash_next = next;
+		kunmap_local(pblock);
+	}
 
 	netfs_single_mark_inode_dirty(&vnode->netfs.inode);
 
@@ -474,6 +513,8 @@ void afs_edit_dir_remove(struct afs_vnode *vnode,
 			   0, 0, 0, 0, name->name);
 	goto out_unmap;
 
+error_unmap:
+	kunmap_local(block);
 error:
 	trace_afs_edit_dir(vnode, why, afs_edit_dir_delete_error,
 			   0, 0, 0, 0, name->name);