On Thursday 21 August 2014 at 19:05:32, Ivan Shapovalov wrote: > On Thursday 21 August 2014 at 00:30:45, Edward Shishkin wrote: > > > > On 08/20/2014 10:34 PM, Ivan Shapovalov wrote: > > > On Wednesday 20 August 2014 at 01:39:42, Edward Shishkin wrote: > > >> On 08/20/2014 12:32 AM, Ivan Shapovalov wrote: > > >>> From `git log` I've seen that VFS people intend to replace ->aio_read() and > > >>> ->aio_write() of struct file_operations with new methods ->read_iter() and > > >>> ->write_iter(). > > >>> > > >>> (Along with a couple of related new helpers, differing from previous just in > > >>> calling _iter methods instead of aio_ ones.) > > >>> > > >>> From other filesystems it seems that these are simple drop-in replacements > > >>> (however, well, I have zero familiarity with VFS). So here is a question: > > >>> is there any intentional reason that generic_file_aio_write() is not used > > >>> in reiser4? > > >> > > >> Currently reiser4 is a set of two filesystems which differ in methods > > >> of handling regular files. For VFS we provide "dispatchers", which pass > > >> management to appropriate plugin (UNIX_FILE or CRYPTCOMPRESS). > > >> > > >> UNIX_FILE plugin doesn't use generic write for performance reasons > > >> (I'll try to find the respective mailing thread). CRYPTCOMPRESS doesn't > > >> use it for compatibility reasons: I don't know how how to rewrite it > > >> gracefully using the generic write method. > > >> > > >> Edward. > > > > > > Thanks for explanation! So, does this patch make any sense? > > > > > > I haven't looked at this carefully yet, but likely it is correct. > > > > Thanks, > > Edward. > > Turned out it isn't.. The iter_file_splice_write() requires ->write_iter > to be set, or a NULL dereference happens. > > At first I've thought that we're out of luck and will need to use the fallback > splice implementation (default_file_splice_write), but just setting > ->write_iter to generic_file_write_iter strangely worked. > > (By "it works" I mean "splice finishes successfully and does not cause data > corruptions"). ...aand... it doesn't. At least on reg40 (sic, NOT ccreg40): 1) `kde-cp` and FF 32.0.3's downloader produce empty files; 2) some applications (at least two games: luftrausers and x-plane 10, both proprietary) deadlock somewhere inside of generic_file_write_iter(). Removing ".splice_write = iter_file_splice_write" and ".write_iter = generic_file_write_iter" fixes the described cases. Unfortunately, no time to dig deeper (MIPT is evil). Hopefully you will look into it someday... :) Thanks, -- Ivan Shapovalov / intelfx /
Attachment:
signature.asc
Description: This is a digitally signed message part.
