btrfs: periodic block_group reclaim
Hello Boris Burkov,
Commit e4ca3932ae90 ("btrfs: periodic block_group reclaim") from Feb 2, 2024 (linux-next), leads to the following Smatch static checker warning:
fs/btrfs/space-info.c:2139 do_reclaim_sweep() warn: iterator 'bg' changed during iteration
fs/btrfs/space-info.c
2125 static bool do_reclaim_sweep(struct btrfs_space_info *space_info, int raid)
2126 {
2127 struct btrfs_block_group *bg;
2128 int thresh_pct;
2129 bool will_reclaim = false;
2130 bool urgent;
2131
2132 spin_lock(&space_info->lock);
2133 urgent = is_reclaim_urgent(space_info);
2134 thresh_pct = btrfs_calc_reclaim_threshold(space_info);
2135 spin_unlock(&space_info->lock);
2136
2137 down_read(&space_info->groups_sem);
2138 again:
--> 2139 list_for_each_entry(bg, &space_info->block_groups[raid], list) {
2140 u64 thresh;
2141 bool reclaim = false;
2142
2143 btrfs_get_block_group(bg);We bump the refcount here.
2144 spin_lock(&bg->lock);
2145 thresh = mult_perc(bg->length, thresh_pct);
2146 if (bg->used < thresh && bg->reclaim_mark) {
2147 will_reclaim = true;
2148 reclaim = true;
2149 }
2150 bg->reclaim_mark++;
2151 spin_unlock(&bg->lock);
2152 if (reclaim)
2153 btrfs_mark_bg_to_reclaim(bg);
2154 btrfs_put_block_group(bg);
^^This decrements the "bg" so now presumably a different thread could free it. The race window between this btrfs_put_block_group() and the btrfs_get_block_group() on the next iteration is pretty small. Most likely this get/put pair could be removed. I can't see a reason for it.
2155 }
2156
2157 /*
2158 * In situations where we are very motivated to reclaim (low unalloc)
2159 * use two passes to make the reclaim mark check best effort.
2160 *
2161 * If we have any staler groups, we don't touch the fresher ones, but if we
2162 * really need a block group, do take a fresh one.
2163 */
2164 if (!will_reclaim && urgent) {
2165 urgent = false;
2166 goto again;
2167 }
2168
2169 up_read(&space_info->groups_sem);
2170 return will_reclaim;
2171 }This email is a free service from the Smatch-CI project [smatch.sf.net].
regards,
dan carpenter
Quoted 45 lines · Dan Carpenter
Hello Boris Burkov,
Commit
e4ca3932ae90("btrfs: periodicblock_groupreclaim") from Feb 2, 2024 (linux-next), leads to the following Smatch static checker warning:fs/btrfs/space-info.c:2139 do_reclaim_sweep() warn: iterator 'bg' changed during iteration
fs/btrfs/space-info.c2125 static bool do_reclaim_sweep(struct btrfs_space_info *space_info, int raid) 2126 { 2127 struct btrfs_block_group *bg; 2128 int thresh_pct; 2129 bool will_reclaim = false; 2130 bool urgent; 2131 2132 spin_lock(&space_info->lock); 2133 urgent = is_reclaim_urgent(space_info); 2134 thresh_pct = btrfs_calc_reclaim_threshold(space_info); 2135 spin_unlock(&space_info->lock); 2136 2137 down_read(&space_info->groups_sem); 2138 again: --> 2139 list_for_each_entry(bg, &space_info->block_groups[raid], list) { 2140 u64 thresh; 2141 bool reclaim = false; 2142 2143 btrfs_get_block_group(bg);We bump the refcount here.
2144 spin_lock(&bg->lock); 2145 thresh = mult_perc(bg->length, thresh_pct); 2146 if (bg->used < thresh && bg->reclaim_mark) { 2147 will_reclaim = true; 2148 reclaim = true; 2149 } 2150 bg->reclaim_mark++; 2151 spin_unlock(&bg->lock); 2152 if (reclaim) 2153 btrfs_mark_bg_to_reclaim(bg); 2154 btrfs_put_block_group(bg); ^^
Quoted 6 lines
This decrements the "bg" so now presumably a different thread could free it. The race window between this
btrfs_put_block_group()and thebtrfs_get_block_group()on the next iteration is pretty small. Most likely this get/put pair could be removed. I can't see a reason for it.
Hi,
This should be protected by read locking the groups_sem for the duration of the loop. The code which unlinks block_group->list write locks groups_sem and has a reference to the bg during that time.
There is at least one other loop that uses this same pattern (find_free_extent() in fs/btrfs/extent-tree.c). I agree that it doesn't make the most sense to get/put the refcount if it doesn't really protect against anything and the rwsem is the actual protection.
There was a patch on the list a few months back that switched to a per-cpu refcount for block_group->refs, which suggests that getting rid of some of these logically redundant refcounts could matter for performance too. I'll take a look soon.
Thanks,
Boris
Quoted 22 lines
2155 } 2156 2157 /* 2158 * In situations where we are very motivated to reclaim (low unalloc) 2159 * use two passes to make the reclaim mark check best effort. 2160 * 2161 * If we have any staler groups, we don't touch the fresher ones, but if we 2162 * really need a block group, do take a fresh one. 2163 */ 2164 if (!will_reclaim && urgent) { 2165 urgent = false; 2166 goto again; 2167 } 2168 2169 up_read(&space_info->groups_sem); 2170 return will_reclaim; 2171 }This email is a free service from the Smatch-CI project [smatch.sf.net].
regards,
dan carpenter