On Tue, Mar 12, 2013 at 03:47:28PM +0800, majianpeng wrote: > Hi all, > I think bio_detain(in dm-cache-target.c) may cause some bug. Hi Jianpeng, I don't really see what you're getting at, but will try and explain bio_detain a bit. The bio prison contains a bunch of cells. When you call bio_detain the following steps are performed: i) look up the cell ii) if the cell exists add the bio to that cell iii) if there is no cell create it and insert the bio as the holder It does do more than I'd like a single function to do, but we need this group of operations to be atomic to avoid race conditions. Now if you look at cache_map() where bio_detain is used: r = bio_detain(cache, block, bio, cell, (cell_free_fn) free_prison_cell, cache, &cell); if (r) { if (r < 0) defer_bio(cache, bio); return DM_MAPIO_SUBMITTED; } We call bio_detain checking the return value, there are 3 cases: 0 - this bio has been inserted as the holder, we try and process it immediately and do not pass it to the worker thread. 1 - the bio has been inserted into the cell, but it's not the holder. When the holder of that particular cell completes it will release all bios in the cell, putting them on the deferred list for processing by the worker thread. < 0 - we couldn't complete, hand over to the worker thread who has a bit more leeway regarding memory allocation etc. In typical use, a cell is only held while we decide what to do with the bio (see calls to cell_defer in cache_map). But sometimes we decide to move a block from/to the cache, in which case the bio_prison prevents any io going to that block while the migration is pending. I hope that helps, - Joe -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel