On Tue, Jan 07, 2025 at 07:10:12AM +0100, Christoph Hellwig wrote: > On Mon, Jan 06, 2025 at 06:37:06PM +0000, John Garry wrote: > >> + /* > >> + * On COW inodes we are forced to always rewrite an entire file system > >> + * block or RT extent. > >> + * > >> + * Because applications assume they can do sector sized direct writes > >> + * on XFS we fall back to buffered I/O for sub-block direct I/O in that > >> + * case. Because that needs to copy the entire block into the buffer > >> + * cache it is highly inefficient and can easily lead to page cache > >> + * invalidation races. > >> + * > >> + * Tell applications to avoid this case by reporting the natively > >> + * supported direct I/O read alignment. > > > > Maybe I mis-read the complete comment, but did you really mean "natively > > supported direct I/O write alignment"? You have been talking about writes > > only, but then finally mention read alignment. > > No, this is indeed intended to talk about the different (smaller) read > alignment we are now reporting. But I guess the wording is confusing > enough that I should improve it? How about: /* * For COW inodes, we can only perform out of place writes of entire * file allocation units (clusters). For a sub-cluster directio write, * we must fall back to buffered I/O to perform the RMW. At best this * is highly inefficient; at worst it leads to page cache invalidation * races. Tell applications to avoid this by reporting separately the * read and (larger) write alignments. */ --D
