| Message ID | 20180502144659.118628-3-colyli@suse.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 05/02/2018 04:46 PM, Coly Li wrote: > Commit c7b7bd07404c5 ("bcache: add io_disable to struct cached_dev") tries > to stop bcache device by calling bcache_device_stop() when too many I/O > errors happened on backing device. But if there is internal I/O happening > on cache device (writeback scan, garbage collection, etc), a regular I/O > request triggers the internal I/Os may still holds a refcount of dc->count, > and the refcount may only be dropped after the internal I/O stopped. > > By this patch, bch_cached_dev_error() will check if the backing device is > attached to a cache set, if yes that CACHE_SET_IO_DISABLE will be set to > flags of this cache set. Then internal I/Os on cache device will be > rejected and stopped immediately, and the bcache device can be stopped. > > For people who are not familiar with the interesting refcount dependance, > let me explain a bit more how the fix works. Example the writeback thread > will scan cache device for dirty data writeback purpose. Before it stopps, > it holds a refcount of dc->count. When CACHE_SET_IO_DISABLE bit is set, > the internal I/O will stopped and the while-loop in bch_writeback_thread() > quits and calls cached_dev_put() to drop dc->count. If this is the last > refcount to drop, then cached_dev_detach_finish() will be called. In this > call back function, in turn closure_put(dc->disk.cl) is called to drop a > refcount of closure dc->disk.cl. If this is the last refcount of this > closure to drop, then cached_dev_flush() will be called. Then the cached > device is freed. So if CACHE_SET_IO_DISABLE is not set, the bache device > can not be stopped until all inernal cache device I/O stopped. For large > size cache device, and writeback thread competes locks with gc thread, > there might be a quite long time to wait. > > Fixes: c7b7bd07404c5 ("bcache: add io_disable to struct cached_dev") > Signed-off-by: Coly Li <colyli@suse.de> > --- > drivers/md/bcache/super.c | 17 +++++++++++++++++ > 1 file changed, 17 insertions(+) > > diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c > index 8196b19fada2..a0d5a3ccc7d0 100644 > --- a/drivers/md/bcache/super.c > +++ b/drivers/md/bcache/super.c > @@ -1369,6 +1369,8 @@ int bch_flash_dev_create(struct cache_set *c, uint64_t size) > > bool bch_cached_dev_error(struct cached_dev *dc) > { > + struct cache_set *c; > + > if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags)) > return false; > > @@ -1379,6 +1381,21 @@ bool bch_cached_dev_error(struct cached_dev *dc) > pr_err("stop %s: too many IO errors on backing device %s\n", > dc->disk.disk->disk_name, dc->backing_dev_name); > > + /* > + * If the cached device is still attached to a cache set, > + * even dc->io_disable is true and no more I/O requests > + * accepted, cache device internal I/O (writeback scan or > + * garbage collection) may still prevent bcache device from > + * being stopped. So here CACHE_SET_IO_DISABLE should be > + * set to c->flags too, to make the internal I/O to cache > + * device rejected and stopped immediately. > + * If c is NULL, that means the bcache device is not attached > + * to any cache set, then no CACHE_SET_IO_DISABLE bit to set. > + */ > + c = dc->disk.c; > + if (c && test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags)) > + pr_warn("CACHE_SET_IO_DISABLE already set"); > + > bcache_device_stop(&dc->disk); > return true; > } > Neat. Reviewed-by: Hannes Reinecke <hare@suse.com> Cheers, Hannes
diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c index 8196b19fada2..a0d5a3ccc7d0 100644 --- a/drivers/md/bcache/super.c +++ b/drivers/md/bcache/super.c @@ -1369,6 +1369,8 @@ int bch_flash_dev_create(struct cache_set *c, uint64_t size) bool bch_cached_dev_error(struct cached_dev *dc) { + struct cache_set *c; + if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags)) return false; @@ -1379,6 +1381,21 @@ bool bch_cached_dev_error(struct cached_dev *dc) pr_err("stop %s: too many IO errors on backing device %s\n", dc->disk.disk->disk_name, dc->backing_dev_name); + /* + * If the cached device is still attached to a cache set, + * even dc->io_disable is true and no more I/O requests + * accepted, cache device internal I/O (writeback scan or + * garbage collection) may still prevent bcache device from + * being stopped. So here CACHE_SET_IO_DISABLE should be + * set to c->flags too, to make the internal I/O to cache + * device rejected and stopped immediately. + * If c is NULL, that means the bcache device is not attached + * to any cache set, then no CACHE_SET_IO_DISABLE bit to set. + */ + c = dc->disk.c; + if (c && test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags)) + pr_warn("CACHE_SET_IO_DISABLE already set"); + bcache_device_stop(&dc->disk); return true; }
Commit c7b7bd07404c5 ("bcache: add io_disable to struct cached_dev") tries to stop bcache device by calling bcache_device_stop() when too many I/O errors happened on backing device. But if there is internal I/O happening on cache device (writeback scan, garbage collection, etc), a regular I/O request triggers the internal I/Os may still holds a refcount of dc->count, and the refcount may only be dropped after the internal I/O stopped. By this patch, bch_cached_dev_error() will check if the backing device is attached to a cache set, if yes that CACHE_SET_IO_DISABLE will be set to flags of this cache set. Then internal I/Os on cache device will be rejected and stopped immediately, and the bcache device can be stopped. For people who are not familiar with the interesting refcount dependance, let me explain a bit more how the fix works. Example the writeback thread will scan cache device for dirty data writeback purpose. Before it stopps, it holds a refcount of dc->count. When CACHE_SET_IO_DISABLE bit is set, the internal I/O will stopped and the while-loop in bch_writeback_thread() quits and calls cached_dev_put() to drop dc->count. If this is the last refcount to drop, then cached_dev_detach_finish() will be called. In this call back function, in turn closure_put(dc->disk.cl) is called to drop a refcount of closure dc->disk.cl. If this is the last refcount of this closure to drop, then cached_dev_flush() will be called. Then the cached device is freed. So if CACHE_SET_IO_DISABLE is not set, the bache device can not be stopped until all inernal cache device I/O stopped. For large size cache device, and writeback thread competes locks with gc thread, there might be a quite long time to wait. Fixes: c7b7bd07404c5 ("bcache: add io_disable to struct cached_dev") Signed-off-by: Coly Li <colyli@suse.de> --- drivers/md/bcache/super.c | 17 +++++++++++++++++ 1 file changed, 17 insertions(+)