path: root/fs/nfs/nfs42xattr.c

// SPDX-License-Identifier: GPL-2.0

/*
 * Copyright 2019, 2020 Amazon.com, Inc. or its affiliates. All rights reserved.
 *
 * User extended attribute client side cache functions.
 *
 * Author: Frank van der Linden <fllinden@amazon.com>
 */
#include <linux/errno.h>
#include <linux/nfs_fs.h>
#include <linux/hashtable.h>
#include <linux/refcount.h>
#include <uapi/linux/xattr.h>

#include "nfs4_fs.h"
#include "internal.h"

/*
 * User extended attributes client side caching is implemented by having
 * a cache structure attached to NFS inodes. This structure is allocated
 * when needed, and freed when the cache is zapped.
 *
 * The cache structure contains as hash table of entries, and a pointer
 * to a special-cased entry for the listxattr cache.
 *
 * Accessing and allocating / freeing the caches is done via reference
 * counting. The cache entries use a similar refcounting scheme.
 *
 * This makes freeing a cache, both from the shrinker and from the
 * zap cache path, easy. It also means that, in current use cases,
 * the large majority of inodes will not waste any memory, as they
 * will never have any user extended attributes assigned to them.
 *
 * Attribute entries are hashed in to a simple hash table. They are
 * also part of an LRU.
 *
 * There are three shrinkers.
 *
 * Two shrinkers deal with the cache entries themselves: one for
 * large entries (> PAGE_SIZE), and one for smaller entries. The
 * shrinker for the larger entries works more aggressively than
 * those for the smaller entries.
 *
 * The other shrinker frees the cache structures themselves.
 */

/*
 * 64 buckets is a good default. There is likely no reasonable
 * workload that uses more than even 64 user extended attributes.
 * You can certainly add a lot more - but you get what you ask for
 * in those circumstances.
 */
#define NFS4_XATTR_HASH_SIZE	64

#define NFSDBG_FACILITY	NFSDBG_XATTRCACHE

struct nfs4_xattr_cache;
struct nfs4_xattr_entry;

struct nfs4_xattr_bucket {
	spinlock_t lock;
	struct hlist_head hlist;
	struct nfs4_xattr_cache *cache;
	bool draining;
};

struct nfs4_xattr_cache {
	struct kref ref;
	struct nfs4_xattr_bucket buckets[NFS4_XATTR_HASH_SIZE];
	struct list_head lru;
	struct list_head dispose;
	atomic_long_t nent;
	spinlock_t listxattr_lock;
	struct inode *inode;
	struct nfs4_xattr_entry *listxattr;
};

struct nfs4_xattr_entry {
	struct kref ref;
	struct hlist_node hnode;
	struct list_head lru;
	struct list_head dispose;
	char *xattr_name;
	void *xattr_value;
	size_t xattr_size;
	struct nfs4_xattr_bucket *bucket;
	uint32_t flags;
};

#define	NFS4_XATTR_ENTRY_EXTVAL	0x0001

/*
 * LRU list of NFS inodes that have xattr caches.
 */
static struct list_lru nfs4_xattr_cache_lru;
static struct list_lru nfs4_xattr_entry_lru;
static struct list_lru nfs4_xattr_large_entry_lru;

static struct kmem_cache *nfs4_xattr_cache_cachep;

/*
 * Hashing helper functions.
 */
static void
nfs4_xattr_hash_init(struct nfs4_xattr_cache *cache)
{
	unsigned int i;

	for (i = 0; i < NFS4_XATTR_HASH_SIZE; i++) {
		INIT_HLIST_HEAD(&cache->buckets[i].hlist);
		spin_lock_init(&cache->buckets[i].lock);
		cache->buckets[i].cache = cache;
		cache->buckets[i].draining = false;
	}
}

/*
 * Locking order:
 * 1. inode i_lock or bucket lock
 * 2. list_lru lock (taken by list_lru_* functions)
 */

/*
 * Wrapper functions to add a cache entry to the right LRU.
 */
static bool
nfs4_xattr_entry_lru_add(struct nfs4_xattr_entry *entry)
{
	struct list_lru *lru;

	lru = (entry->flags & NFS4_XATTR_ENTRY_EXTVAL) ?
	    &nfs4_xattr_large_entry_lru : &nfs4_xattr_entry_lru;

	return list_lru_add_obj(lru, &entry->lru);
}

static bool
nfs4_xattr_entry_lru_del(struct nfs4_xattr_entry *entry)
{
	struct list_lru *lru;

	lru = (entry->flags & NFS4_XATTR_ENTRY_EXTVAL) ?
	    &nfs4_xattr_large_entry_lru : &nfs4_xattr_entry_lru;

	return list_lru_del_obj(lru, &entry->lru);
}

/*
 * This function allocates cache entries. They are the normal
 * extended attribute name/value pairs, but may also be a listxattr
 * cache. Those allocations use the same entry so that they can be
 * treated as one by the memory shrinker.
 *
 * xattr cache entries are allocated together with names. If the
 * value fits in to one page with the entry structure and the name,
 * it will also be part of the same allocation (kmalloc). This is
 * expected to be the vast majority of cases. Larger allocations
 * have a value pointer that is allocated separately by kvmalloc.
 *
 * Parameters:
 *
 * @name:  Name of the extended attribute. NULL for listxattr cache
 *         entry.
 * @value: Value of attribute, or listxattr cache. NULL if the
 *         value is to be copied from pages instead.
 * @pages: Pages to copy the value from, if not NULL. Passed in to
 *	   make it easier to copy the value after an RPC, even if
 *	   the value will not be passed up to application (e.g.
 *	   for a 'query' getxattr with NULL buffer).
 * @len:   Length of the value. Can be 0 for zero-length attributes.
 *         @value and @pages will be NULL if @len is 0.
 */
static struct nfs4_xattr_entry *
nfs4_xattr_alloc_entry(const char *name, const void *value,
		       struct page **pages, size_t len)
{
	struct nfs4_xattr_entry *entry;
	void *valp;
	char *namep;
	size_t alloclen, slen;
	char *buf;
	uint32_t flags;

	BUILD_BUG_ON(sizeof(struct nfs4_xattr_entry) +
	    XATTR_NA