linux-stable/fs/btrfs/delayed-ref.h
Josef Bacik e19eb11f4f btrfs: only let one thread pre-flush delayed refs in commit
I've been running a stress test that runs 20 workers in their own
subvolume, which are running an fsstress instance with 4 threads per
worker, which is 80 total fsstress threads.  In addition to this I'm
running balance in the background as well as creating and deleting
snapshots.  This test takes around 12 hours to run normally, going
slower and slower as the test goes on.

The reason for this is because fsstress is running fsync sometimes, and
because we're messing with block groups we often fall through to
btrfs_commit_transaction, so will often have 20-30 threads all calling
btrfs_commit_transaction at the same time.

These all get stuck contending on the extent tree while they try to run
delayed refs during the initial part of the commit.

This is suboptimal, really because the extent tree is a single point of
failure we only want one thread acting on that tree at once to reduce
lock contention.

Fix this by making the flushing mechanism a bit operation, to make it
easy to use test_and_set_bit() in order to make sure only one task does
this initial flush.

Once we're into the transaction commit we only have one thread doing
delayed ref running, it's just this initial pre-flush that is
problematic.  With this patch my stress test takes around 90 minutes to
run, instead of 12 hours.

The memory barrier is not necessary for the flushing bit as it's
ordered, unlike plain int. The transaction state accessed in
btrfs_should_end_transaction could be affected by that too as it's not
always used under transaction lock. Upon Nikolay's analysis in [1]
it's not necessary:

  In should_end_transaction it's read without holding any locks. (U)

  It's modified in btrfs_cleanup_transaction without holding the
  fs_info->trans_lock (U), but the STATE_ERROR flag is going to be set.

  set in cleanup_transaction under fs_info->trans_lock (L)
  set in btrfs_commit_trans to COMMIT_START under fs_info->trans_lock.(L)
  set in btrfs_commit_trans to COMMIT_DOING under fs_info->trans_lock.(L)
  set in btrfs_commit_trans to COMMIT_UNBLOCK under
  fs_info->trans_lock.(L)

  set in btrfs_commit_trans to COMMIT_COMPLETED without locks but at this
  point the transaction is finished and fs_info->running_trans is NULL (U
  but irrelevant).

  So by the looks of it we can have a concurrent READ race with a WRITE,
  due to reads not taking a lock. In this case what we want to ensure is
  we either see new or old state. I consulted with Will Deacon and he said
  that in such a case we'd want to annotate the accesses to ->state with
  (READ|WRITE)_ONCE so as to avoid a theoretical tear, in this case I
  don't think this could happen but I imagine at some point KCSAN would
  flag such an access as racy (which it is).

[1] https://lore.kernel.org/linux-btrfs/e1fd5cc1-0f28-f670-69f4-e9958b4964e6@suse.com

Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
[ add comments regarding memory barrier ]
Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-08 22:58:56 +01:00

400 lines
11 KiB
C

/* SPDX-License-Identifier: GPL-2.0 */
/*
* Copyright (C) 2008 Oracle. All rights reserved.
*/
#ifndef BTRFS_DELAYED_REF_H
#define BTRFS_DELAYED_REF_H
#include <linux/refcount.h>
/* these are the possible values of struct btrfs_delayed_ref_node->action */
#define BTRFS_ADD_DELAYED_REF 1 /* add one backref to the tree */
#define BTRFS_DROP_DELAYED_REF 2 /* delete one backref from the tree */
#define BTRFS_ADD_DELAYED_EXTENT 3 /* record a full extent allocation */
#define BTRFS_UPDATE_DELAYED_HEAD 4 /* not changing ref count on head ref */
struct btrfs_delayed_ref_node {
struct rb_node ref_node;
/*
* If action is BTRFS_ADD_DELAYED_REF, also link this node to
* ref_head->ref_add_list, then we do not need to iterate the
* whole ref_head->ref_list to find BTRFS_ADD_DELAYED_REF nodes.
*/
struct list_head add_list;
/* the starting bytenr of the extent */
u64 bytenr;
/* the size of the extent */
u64 num_bytes;
/* seq number to keep track of insertion order */
u64 seq;
/* ref count on this data structure */
refcount_t refs;
/*
* how many refs is this entry adding or deleting. For
* head refs, this may be a negative number because it is keeping
* track of the total mods done to the reference count.
* For individual refs, this will always be a positive number
*
* It may be more than one, since it is possible for a single
* parent to have more than one ref on an extent
*/
int ref_mod;
unsigned int action:8;
unsigned int type:8;
/* is this node still in the rbtree? */
unsigned int is_head:1;
unsigned int in_tree:1;
};
struct btrfs_delayed_extent_op {
struct btrfs_disk_key key;
u8 level;
bool update_key;
bool update_flags;
bool is_data;
u64 flags_to_set;
};
/*
* the head refs are used to hold a lock on a given extent, which allows us
* to make sure that only one process is running the delayed refs
* at a time for a single extent. They also store the sum of all the
* reference count modifications we've queued up.
*/
struct btrfs_delayed_ref_head {
u64 bytenr;
u64 num_bytes;
refcount_t refs;
/*
* the mutex is held while running the refs, and it is also
* held when checking the sum of reference modifications.
*/
struct mutex mutex;
spinlock_t lock;
struct rb_root_cached ref_tree;
/* accumulate add BTRFS_ADD_DELAYED_REF nodes to this ref_add_list. */
struct list_head ref_add_list;
struct rb_node href_node;
struct btrfs_delayed_extent_op *extent_op;
/*
* This is used to track the final ref_mod from all the refs associated
* with this head ref, this is not adjusted as delayed refs are run,
* this is meant to track if we need to do the csum accounting or not.
*/
int total_ref_mod;
/*
* This is the current outstanding mod references for this bytenr. This
* is used with lookup_extent_info to get an accurate reference count
* for a bytenr, so it is adjusted as delayed refs are run so that any
* on disk reference count + ref_mod is accurate.
*/
int ref_mod;
/*
* when a new extent is allocated, it is just reserved in memory
* The actual extent isn't inserted into the extent allocation tree
* until the delayed ref is processed. must_insert_reserved is
* used to flag a delayed ref so the accounting can be updated
* when a full insert is done.
*
* It is possible the extent will be freed before it is ever
* inserted into the extent allocation tree. In this case
* we need to update the in ram accounting to properly reflect
* the free has happened.
*/
unsigned int must_insert_reserved:1;
unsigned int is_data:1;
unsigned int is_system:1;
unsigned int processing:1;
};
struct btrfs_delayed_tree_ref {
struct btrfs_delayed_ref_node node;
u64 root;
u64 parent;
int level;
};
struct btrfs_delayed_data_ref {
struct btrfs_delayed_ref_node node;
u64 root;
u64 parent;
u64 objectid;
u64 offset;
};
enum btrfs_delayed_ref_flags {
/* Indicate that we are flushing delayed refs for the commit */
BTRFS_DELAYED_REFS_FLUSHING,
};
struct btrfs_delayed_ref_root {
/* head ref rbtree */
struct rb_root_cached href_root;
/* dirty extent records */
struct rb_root dirty_extent_root;
/* this spin lock protects the rbtree and the entries inside */
spinlock_t lock;
/* how many delayed ref updates we've queued, used by the
* throttling code
*/
atomic_t num_entries;
/* total number of head nodes in tree */
unsigned long num_heads;
/* total number of head nodes ready for processing */
unsigned long num_heads_ready;
u64 pending_csums;
unsigned long flags;
u64 run_delayed_start;
/*
* To make qgroup to skip given root.
* This is for snapshot, as btrfs_qgroup_inherit() will manually
* modify counters for snapshot and its source, so we should skip
* the snapshot in new_root/old_roots or it will get calculated twice
*/
u64 qgroup_to_skip;
};
enum btrfs_ref_type {
BTRFS_REF_NOT_SET,
BTRFS_REF_DATA,
BTRFS_REF_METADATA,
BTRFS_REF_LAST,
};
struct btrfs_data_ref {
/* For EXTENT_DATA_REF */
/* Root which refers to this data extent */
u64 ref_root;
/* Inode which refers to this data extent */
u64 ino;
/*
* file_offset - extent_offset
*
* file_offset is the key.offset of the EXTENT_DATA key.
* extent_offset is btrfs_file_extent_offset() of the EXTENT_DATA data.
*/
u64 offset;
};
struct btrfs_tree_ref {
/*
* Level of this tree block
*
* Shared for skinny (TREE_BLOCK_REF) and normal tree ref.
*/
int level;
/*
* Root which refers to this tree block.
*
* For TREE_BLOCK_REF (skinny metadata, either inline or keyed)
*/
u64 root;
/* For non-skinny metadata, no special member needed */
};
struct btrfs_ref {
enum btrfs_ref_type type;
int action;
/*
* Whether this extent should go through qgroup record.
*
* Normally false, but for certain cases like delayed subtree scan,
* setting this flag can hugely reduce qgroup overhead.
*/
bool skip_qgroup;
/*
* Optional. For which root is this modification.
* Mostly used for qgroup optimization.
*
* When unset, data/tree ref init code will populate it.
* In certain cases, we're modifying reference for a different root.
* E.g. COW fs tree blocks for balance.
* In that case, tree_ref::root will be fs tree, but we're doing this
* for reloc tree, then we should set @real_root to reloc tree.
*/
u64 real_root;
u64 bytenr;
u64 len;
/* Bytenr of the parent tree block */
u64 parent;
union {
struct btrfs_data_ref data_ref;
struct btrfs_tree_ref tree_ref;
};
};
extern struct kmem_cache *btrfs_delayed_ref_head_cachep;
extern struct kmem_cache *btrfs_delayed_tree_ref_cachep;
extern struct kmem_cache *btrfs_delayed_data_ref_cachep;
extern struct kmem_cache *btrfs_delayed_extent_op_cachep;
int __init btrfs_delayed_ref_init(void);
void __cold btrfs_delayed_ref_exit(void);
static inline void btrfs_init_generic_ref(struct btrfs_ref *generic_ref,
int action, u64 bytenr, u64 len, u64 parent)
{
generic_ref->action = action;
generic_ref->bytenr = bytenr;
generic_ref->len = len;
generic_ref->parent = parent;
}
static inline void btrfs_init_tree_ref(struct btrfs_ref *generic_ref,
int level, u64 root)
{
/* If @real_root not set, use @root as fallback */
if (!generic_ref->real_root)
generic_ref->real_root = root;
generic_ref->tree_ref.level = level;
generic_ref->tree_ref.root = root;
generic_ref->type = BTRFS_REF_METADATA;
}
static inline void btrfs_init_data_ref(struct btrfs_ref *generic_ref,
u64 ref_root, u64 ino, u64 offset)
{
/* If @real_root not set, use @root as fallback */
if (!generic_ref->real_root)
generic_ref->real_root = ref_root;
generic_ref->data_ref.ref_root = ref_root;
generic_ref->data_ref.ino = ino;
generic_ref->data_ref.offset = offset;
generic_ref->type = BTRFS_REF_DATA;
}
static inline struct btrfs_delayed_extent_op *
btrfs_alloc_delayed_extent_op(void)
{
return kmem_cache_alloc(btrfs_delayed_extent_op_cachep, GFP_NOFS);
}
static inline void
btrfs_free_delayed_extent_op(struct btrfs_delayed_extent_op *op)
{
if (op)
kmem_cache_free(btrfs_delayed_extent_op_cachep, op);
}
static inline void btrfs_put_delayed_ref(struct btrfs_delayed_ref_node *ref)
{
WARN_ON(refcount_read(&ref->refs) == 0);
if (refcount_dec_and_test(&ref->refs)) {
WARN_ON(ref->in_tree);
switch (ref->type) {
case BTRFS_TREE_BLOCK_REF_KEY:
case BTRFS_SHARED_BLOCK_REF_KEY:
kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
break;
case BTRFS_EXTENT_DATA_REF_KEY:
case BTRFS_SHARED_DATA_REF_KEY:
kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
break;
default:
BUG();
}
}
}
static inline u64 btrfs_ref_head_to_space_flags(
struct btrfs_delayed_ref_head *head_ref)
{
if (head_ref->is_data)
return BTRFS_BLOCK_GROUP_DATA;
else if (head_ref->is_system)
return BTRFS_BLOCK_GROUP_SYSTEM;
return BTRFS_BLOCK_GROUP_METADATA;
}
static inline void btrfs_put_delayed_ref_head(struct btrfs_delayed_ref_head *head)
{
if (refcount_dec_and_test(&head->refs))
kmem_cache_free(btrfs_delayed_ref_head_cachep, head);
}
int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
struct btrfs_ref *generic_ref,
struct btrfs_delayed_extent_op *extent_op);
int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
struct btrfs_ref *generic_ref,
u64 reserved);
int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
u64 bytenr, u64 num_bytes,
struct btrfs_delayed_extent_op *extent_op);
void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
struct btrfs_delayed_ref_root *delayed_refs,
struct btrfs_delayed_ref_head *head);
struct btrfs_delayed_ref_head *
btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
u64 bytenr);
int btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
struct btrfs_delayed_ref_head *head);
static inline void btrfs_delayed_ref_unlock(struct btrfs_delayed_ref_head *head)
{
mutex_unlock(&head->mutex);
}
void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
struct btrfs_delayed_ref_head *head);
struct btrfs_delayed_ref_head *btrfs_select_ref_head(
struct btrfs_delayed_ref_root *delayed_refs);
int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq);
void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr);
void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans);
int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
enum btrfs_reserve_flush_enum flush);
void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
struct btrfs_block_rsv *src,
u64 num_bytes);
int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans);
bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info);
/*
* helper functions to cast a node into its container
*/
static inline struct btrfs_delayed_tree_ref *
btrfs_delayed_node_to_tree_ref(struct btrfs_delayed_ref_node *node)
{
return container_of(node, struct btrfs_delayed_tree_ref, node);
}
static inline struct btrfs_delayed_data_ref *
btrfs_delayed_node_to_data_ref(struct btrfs_delayed_ref_node *node)
{
return container_of(node, struct btrfs_delayed_data_ref, node);
}
#endif