I found that bchfs_fallocate() only evicts pagecache when FALLOC_FL_ZERO_RANGE
is set:
if (mode & FALLOC_FL_ZERO_RANGE) {
truncate_pagecache_range(&inode->v, offset, end - 1);
}
The hole detection code already skips ZERO_RANGE:
if (!(mode & FALLOC_FL_ZERO_RANGE)) {
bch2_clamp_data_hole(...); // Only runs for basic fallocate
}
The syzbot reproducer uses mode=0 (basic fallocate), which doesn't evict pages
but does run hole detection and hits the deadlock.
The fix should be to remove pagecache_block_get() from
bch2_fallocate_dispatch() and add it only where pages are actually evicted:
In bchfs_fallocate():
if (mode & FALLOC_FL_ZERO_RANGE) {
bch2_pagecache_block_get(inode);
truncate_pagecache_range(...);
// Do allocation work
bch2_pagecache_block_put(inode);
}
Similarly add it to bchfs_fpunch() and bchfs_fcollapse_finsert() around their
page eviction calls.
This way basic fallocate never holds the lock, avoiding the deadlock. Does this
approach look right?
