btrfs: replace pending/pinned chunks lists with io tree
authorJeff Mahoney <jeffm@suse.com>
Wed, 27 Mar 2019 12:24:12 +0000 (14:24 +0200)
committerDavid Sterba <dsterba@suse.com>
Mon, 29 Apr 2019 17:02:37 +0000 (19:02 +0200)
The pending chunks list contains chunks that are allocated in the
current transaction but haven't been created yet. The pinned chunks
list contains chunks that are being released in the current transaction.
Both describe chunks that are not reflected on disk as in use but are
unavailable just the same.

The pending chunks list is anchored by the transaction handle, which
means that we need to hold a reference to a transaction when working
with the list.

The way we use them is by iterating over both lists to perform
comparisons on the stripes they describe for each device. This is
backwards and requires that we keep a transaction handle open while
we're trimming.

This patchset adds an extent_io_tree to btrfs_device that maintains
the allocation state of the device.  Extents are set dirty when
chunks are first allocated -- when the extent maps are added to the
mapping tree. They're cleared when last removed -- when the extent
maps are removed from the mapping tree. This matches the lifespan
of the pending and pinned chunks list and allows us to do trims
on unallocated space safely without pinning the transaction for what
may be a lengthy operation. We can also use this io tree to mark
which chunks have already been trimmed so we don't repeat the operation.

Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/ctree.h
fs/btrfs/disk-io.c
fs/btrfs/extent-tree.c
fs/btrfs/extent_map.c
fs/btrfs/free-space-cache.c
fs/btrfs/transaction.c
fs/btrfs/transaction.h
fs/btrfs/volumes.c
fs/btrfs/volumes.h

index 19833b4af630abb3de27034f3d0084e2820413f1..93270e20a8e75a1ab594c222685a38d7f75f4ac5 100644 (file)
@@ -1149,12 +1149,6 @@ struct btrfs_fs_info {
        struct mutex unused_bg_unpin_mutex;
        struct mutex delete_unused_bgs_mutex;
 
-       /*
-        * Chunks that can't be freed yet (under a trim/discard operation)
-        * and will be latter freed. Protected by fs_info->chunk_mutex.
-        */
-       struct list_head pinned_chunks;
-
        /* Cached block sizes */
        u32 nodesize;
        u32 sectorsize;
index 8c10702a3f838598b07b046eff6b0da4ac7445cd..0b2b75a7efbdbb8bbb97a6f300b531e91e181b81 100644 (file)
@@ -2774,8 +2774,6 @@ int open_ctree(struct super_block *sb,
        init_waitqueue_head(&fs_info->async_submit_wait);
        init_waitqueue_head(&fs_info->delayed_iputs_wait);
 
-       INIT_LIST_HEAD(&fs_info->pinned_chunks);
-
        /* Usable values until the real ones are cached from the superblock */
        fs_info->nodesize = 4096;
        fs_info->sectorsize = 4096;
@@ -4050,15 +4048,6 @@ void close_ctree(struct btrfs_fs_info *fs_info)
 
        btrfs_free_stripe_hash_table(fs_info);
        btrfs_free_ref_cache(fs_info);
-
-       while (!list_empty(&fs_info->pinned_chunks)) {
-               struct extent_map *em;
-
-               em = list_first_entry(&fs_info->pinned_chunks,
-                                     struct extent_map, list);
-               list_del_init(&em->list);
-               free_extent_map(em);
-       }
 }
 
 int btrfs_buffer_uptodate(struct extent_buffer *buf, u64 parent_transid,
index c5f9e8359c6f7b58412c001daa4d8e507242c0e0..a9f504e7be33bfc84c7a4faf8d11d9a12922ba1c 100644 (file)
@@ -10946,10 +10946,6 @@ int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
        memcpy(&key, &block_group->key, sizeof(key));
 
        mutex_lock(&fs_info->chunk_mutex);
-       if (!list_empty(&em->list)) {
-               /* We're in the transaction->pending_chunks list. */
-               free_extent_map(em);
-       }
        spin_lock(&block_group->lock);
        block_group->removed = 1;
        /*
@@ -10976,25 +10972,6 @@ int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
         * the transaction commit has completed.
         */
        remove_em = (atomic_read(&block_group->trimming) == 0);
-       /*
-        * Make sure a trimmer task always sees the em in the pinned_chunks list
-        * if it sees block_group->removed == 1 (needs to lock block_group->lock
-        * before checking block_group->removed).
-        */
-       if (!remove_em) {
-               /*
-                * Our em might be in trans->transaction->pending_chunks which
-                * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
-                * and so is the fs_info->pinned_chunks list.
-                *
-                * So at this point we must be holding the chunk_mutex to avoid
-                * any races with chunk allocation (more specifically at
-                * volumes.c:contains_pending_extent()), to ensure it always
-                * sees the em, either in the pending_chunks list or in the
-                * pinned_chunks list.
-                */
-               list_move_tail(&em->list, &fs_info->pinned_chunks);
-       }
        spin_unlock(&block_group->lock);
 
        if (remove_em) {
@@ -11002,11 +10979,6 @@ int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
 
                em_tree = &fs_info->mapping_tree.map_tree;
                write_lock(&em_tree->lock);
-               /*
-                * The em might be in the pending_chunks list, so make sure the
-                * chunk mutex is locked, since remove_extent_mapping() will
-                * delete us from that list.
-                */
                remove_extent_mapping(em_tree, em);
                write_unlock(&em_tree->lock);
                /* once for the tree */
index 928f729c55baf61b02e7a4fb612eaebffbdc2d50..5a79a656dfa6481ac4579ae3832dc898b168f15d 100644 (file)
@@ -4,6 +4,7 @@
 #include <linux/slab.h>
 #include <linux/spinlock.h>
 #include "ctree.h"
+#include "volumes.h"
 #include "extent_map.h"
 #include "compression.h"
 
@@ -337,6 +338,37 @@ static inline void setup_extent_mapping(struct extent_map_tree *tree,
                try_merge_map(tree, em);
 }
 
+static void extent_map_device_set_bits(struct extent_map *em, unsigned bits)
+{
+       struct map_lookup *map = em->map_lookup;
+       u64 stripe_size = em->orig_block_len;
+       int i;
+
+       for (i = 0; i < map->num_stripes; i++) {
+               struct btrfs_bio_stripe *stripe = &map->stripes[i];
+               struct btrfs_device *device = stripe->dev;
+
+               set_extent_bits_nowait(&device->alloc_state, stripe->physical,
+                                stripe->physical + stripe_size - 1, bits);
+       }
+}
+
+static void extent_map_device_clear_bits(struct extent_map *em, unsigned bits)
+{
+       struct map_lookup *map = em->map_lookup;
+       u64 stripe_size = em->orig_block_len;
+       int i;
+
+       for (i = 0; i < map->num_stripes; i++) {
+               struct btrfs_bio_stripe *stripe = &map->stripes[i];
+               struct btrfs_device *device = stripe->dev;
+
+               __clear_extent_bit(&device->alloc_state, stripe->physical,
+                                  stripe->physical + stripe_size - 1, bits,
+                                  0, 0, NULL, GFP_NOWAIT, NULL);
+       }
+}
+
 /**
  * add_extent_mapping - add new extent map to the extent tree
  * @tree:      tree to insert new map in
@@ -357,6 +389,8 @@ int add_extent_mapping(struct extent_map_tree *tree,
                goto out;
 
        setup_extent_mapping(tree, em, modified);
+       if (test_bit(EXTENT_FLAG_FS_MAPPING, &em->flags))
+               extent_map_device_set_bits(em, CHUNK_ALLOCATED);
 out:
        return ret;
 }
@@ -438,6 +472,8 @@ void remove_extent_mapping(struct extent_map_tree *tree, struct extent_map *em)
        rb_erase_cached(&em->rb_node, &tree->map);
        if (!test_bit(EXTENT_FLAG_LOGGING, &em->flags))
                list_del_init(&em->list);
+       if (test_bit(EXTENT_FLAG_FS_MAPPING, &em->flags))
+               extent_map_device_clear_bits(em, CHUNK_ALLOCATED);
        RB_CLEAR_NODE(&em->rb_node);
 }
 
index 74aa552f47930699aa2a81134c3200bf2f5d5c5e..207fb50dcc7aa9b051ff161156b3b9e0359a65b7 100644 (file)
@@ -3366,10 +3366,6 @@ void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
                em = lookup_extent_mapping(em_tree, block_group->key.objectid,
                                           1);
                BUG_ON(!em); /* logic error, can't happen */
-               /*
-                * remove_extent_mapping() will delete us from the pinned_chunks
-                * list, which is protected by the chunk mutex.
-                */
                remove_extent_mapping(em_tree, em);
                write_unlock(&em_tree->lock);
                mutex_unlock(&fs_info->chunk_mutex);
index b32769998bbba5386831545b7e68b0358e16fbc3..e5404326fc55bdb7d67f9dcfa6172cf65f9a889d 100644 (file)
@@ -50,14 +50,6 @@ void btrfs_put_transaction(struct btrfs_transaction *transaction)
                        btrfs_err(transaction->fs_info,
                                  "pending csums is %llu",
                                  transaction->delayed_refs.pending_csums);
-               while (!list_empty(&transaction->pending_chunks)) {
-                       struct extent_map *em;
-
-                       em = list_first_entry(&transaction->pending_chunks,
-                                             struct extent_map, list);
-                       list_del_init(&em->list);
-                       free_extent_map(em);
-               }
                /*
                 * If any block groups are found in ->deleted_bgs then it's
                 * because the transaction was aborted and a commit did not
@@ -235,7 +227,6 @@ loop:
        spin_lock_init(&cur_trans->delayed_refs.lock);
 
        INIT_LIST_HEAD(&cur_trans->pending_snapshots);
-       INIT_LIST_HEAD(&cur_trans->pending_chunks);
        INIT_LIST_HEAD(&cur_trans->dev_update_list);
        INIT_LIST_HEAD(&cur_trans->switch_commits);
        INIT_LIST_HEAD(&cur_trans->dirty_bgs);
index 2bd76f681520f6b316efb0cfe4320fcaed62a353..4419a4a0294bc3ae08ccec63f373b007f1e3ef12 100644 (file)
@@ -51,7 +51,6 @@ struct btrfs_transaction {
        wait_queue_head_t writer_wait;
        wait_queue_head_t commit_wait;
        struct list_head pending_snapshots;
-       struct list_head pending_chunks;
        struct list_head dev_update_list;
        struct list_head switch_commits;
        struct list_head dirty_bgs;
index 2bde9e9c188eec0a079c4c158771f423dd1a2aae..2e5e48d8dd2f98a3238c72aa71558dd225facccb 100644 (file)
@@ -336,6 +336,7 @@ void btrfs_free_device(struct btrfs_device *device)
 {
        WARN_ON(!list_empty(&device->post_commit_list));
        rcu_string_free(device->name);
+       extent_io_tree_release(&device->alloc_state);
        bio_put(device->flush_bio);
        kfree(device);
 }
@@ -412,6 +413,7 @@ static struct btrfs_device *__alloc_device(void)
        btrfs_device_data_ordered_init(dev);
        INIT_RADIX_TREE(&dev->reada_zones, GFP_NOFS & ~__GFP_DIRECT_RECLAIM);
        INIT_RADIX_TREE(&dev->reada_extents, GFP_NOFS & ~__GFP_DIRECT_RECLAIM);
+       extent_io_tree_init(NULL, &dev->alloc_state, 0, NULL);
 
        return dev;
 }
@@ -1499,58 +1501,30 @@ error_bdev_put:
        return device;
 }
 
-static int contains_pending_extent(struct btrfs_transaction *transaction,
-                                  struct btrfs_device *device,
-                                  u64 *start, u64 len)
+/*
+ * Try to find a chunk that intersects [start, start + len] range and when one
+ * such is found, record the end of it in *start
+ */
+#define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
+static bool contains_pending_extent(struct btrfs_device *device, u64 *start,
+                                   u64 len)
 {
-       struct btrfs_fs_info *fs_info = device->fs_info;
-       struct extent_map *em;
-       struct list_head *search_list = &fs_info->pinned_chunks;
-       int ret = 0;
-       u64 physical_start = *start;
+       u64 physical_start, physical_end;
 
-       if (transaction)
-               search_list = &transaction->pending_chunks;
-again:
-       list_for_each_entry(em, search_list, list) {
-               struct map_lookup *map;
-               int i;
+       lockdep_assert_held(&device->fs_info->chunk_mutex);
 
-               map = em->map_lookup;
-               for (i = 0; i < map->num_stripes; i++) {
-                       u64 end;
+       if (!find_first_extent_bit(&device->alloc_state, *start,
+                                  &physical_start, &physical_end,
+                                  CHUNK_ALLOCATED, NULL)) {
 
-                       if (map->stripes[i].dev != device)
-                               continue;
-                       if (map->stripes[i].physical >= physical_start + len ||
-                           map->stripes[i].physical + em->orig_block_len <=
-                           physical_start)
-                               continue;
-                       /*
-                        * Make sure that while processing the pinned list we do
-                        * not override our *start with a lower value, because
-                        * we can have pinned chunks that fall within this
-                        * device hole and that have lower physical addresses
-                        * than the pending chunks we processed before. If we
-                        * do not take this special care we can end up getting
-                        * 2 pending chunks that start at the same physical
-                        * device offsets because the end offset of a pinned
-                        * chunk can be equal to the start offset of some
-                        * pending chunk.
-                        */
-                       end = map->stripes[i].physical + em->orig_block_len;
-                       if (end > *start) {
-                               *start = end;
-                               ret = 1;
-                       }
+               if (in_range(physical_start, *start, len) ||
+                   in_range(*start, physical_start,
+                            physical_end - physical_start)) {
+                       *start = physical_end + 1;
+                       return true;
                }
        }
-       if (search_list != &fs_info->pinned_chunks) {
-               search_list = &fs_info->pinned_chunks;
-               goto again;
-       }
-
-       return ret;
+       return false;
 }
 
 
@@ -1661,15 +1635,12 @@ again:
                         * Have to check before we set max_hole_start, otherwise
                         * we could end up sending back this offset anyway.
                         */
-                       if (contains_pending_extent(transaction, device,
-                                                   &search_start,
+                       if (contains_pending_extent(device, &search_start,
                                                    hole_size)) {
-                               if (key.offset >= search_start) {
+                               if (key.offset >= search_start)
                                        hole_size = key.offset - search_start;
-                               } else {
-                                       WARN_ON_ONCE(1);
+                               else
                                        hole_size = 0;
-                               }
                        }
 
                        if (hole_size > max_hole_size) {
@@ -1710,8 +1681,7 @@ next:
        if (search_end > search_start) {
                hole_size = search_end - search_start;
 
-               if (contains_pending_extent(transaction, device, &search_start,
-                                           hole_size)) {
+               if (contains_pending_extent(device, &search_start, hole_size)) {
                        btrfs_release_path(path);
                        goto again;
                }
@@ -4756,7 +4726,7 @@ int btrfs_shrink_device(struct btrfs_device *device, u64 new_size)
         * in-memory chunks are synced to disk so that the loop below sees them
         * and relocates them accordingly.
         */
-       if (contains_pending_extent(trans->transaction, device, &start, diff)) {
+       if (contains_pending_extent(device, &start, diff)) {
                mutex_unlock(&fs_info->chunk_mutex);
                ret = btrfs_commit_transaction(trans);
                if (ret)
@@ -5189,9 +5159,6 @@ static int __btrfs_alloc_chunk(struct btrfs_trans_handle *trans,
                free_extent_map(em);
                goto error;
        }
-
-       list_add_tail(&em->list, &trans->transaction->pending_chunks);
-       refcount_inc(&em->refs);
        write_unlock(&em_tree->lock);
 
        ret = btrfs_make_block_group(trans, 0, type, start, chunk_size);
@@ -5224,8 +5191,6 @@ error_del_extent:
        free_extent_map(em);
        /* One for the tree reference */
        free_extent_map(em);
-       /* One for the pending_chunks list reference */
-       free_extent_map(em);
 error:
        kfree(devices_info);
        return ret;
index d3658a4e65dbb4765e49a5036edc5db69bb7bf2a..6c466ac27c2eddbd448c8a7378a69191e1653728 100644 (file)
@@ -133,6 +133,8 @@ struct btrfs_device {
        /* Counter to record the change of device stats */
        atomic_t dev_stats_ccnt;
        atomic_t dev_stat_values[BTRFS_DEV_STAT_VALUES_MAX];
+
+       struct extent_io_tree alloc_state;
 };
 
 /*