On Thu, Aug 29, 2024 at 12:48:05PM +0200, Jan Kara wrote: > On Wed 14-08-24 17:25:29, Josef Bacik wrote: > > With page faults we can trigger readahead on the file, and then > > subsequent faults can find these pages and insert them into the file > > without emitting an fanotify event. To avoid this case, disable > > readahead if we have pre-content watches on the file. This way we are > > guaranteed to get an event for every range we attempt to access on a > > pre-content watched file. > > > > Reviewed-by: Christian Brauner <brauner@xxxxxxxxxx> > > Signed-off-by: Josef Bacik <josef@xxxxxxxxxxxxxx> > > ... > > > @@ -674,6 +675,14 @@ void page_cache_sync_ra(struct readahead_control *ractl, > > { > > bool do_forced_ra = ractl->file && (ractl->file->f_mode & FMODE_RANDOM); > > > > + /* > > + * If we have pre-content watches we need to disable readahead to make > > + * sure that we don't find 0 filled pages in cache that we never emitted > > + * events for. > > + */ > > + if (ractl->file && fsnotify_file_has_pre_content_watches(ractl->file)) > > + return; > > + > > There are callers which don't pass struct file to readahead (either to > page_cache_sync_ra() or page_cache_async_ra()). Luckily these are very few > - cramfs for a block device (we don't care) and btrfs from code paths like > send-receive or defrag. Now if you tell me you're fine breaking these > corner cases for btrfs, I'll take your word for it but it looks like a > nasty trap to me. Now doing things like defrag or send-receive on offline > files on HSM managed filesystem doesn't look like a terribly good idea > anyway so perhaps we just want btrfs to check and refuse such things? > We can't have HSM on a send subvolume because they have to be read only. I hadn't thought of defrag, I'll respin and add a patch to disallow defrag on a file that has content watches. Thanks, Josef