On Sun, Apr 04 2021 at 13:25:00 -0400, Tavian Barnes quoth thus: > On Sun, 4 Apr 2021 at 00:42, Clay Harris <bugs@xxxxxxxxxxx> wrote: > > On Tue, Mar 30 2021 at 14:17:21 +0300, Lennert Buytenhek quoth thus: > > > > > ... > > > > > > - Make IORING_OP_GETDENTS read from the directory's current position > > > if the specified offset value is -1 (IORING_FEAT_RW_CUR_POS). > > > (Requested / pointed out by Tavian Barnes.) > > > > This seems like a good feature. As I understand it, this would > > allow submitting pairs of IORING_OP_GETDENTS with IOSQE_IO_HARDLINK > > wherein the first specifies the current offset and the second specifies > > offset -1, thereby halfing the number of kernel round trips for N getdents64. > > Yep, that was my main motivation for this suggestion. > > > If the entire directory fits into the first buffer, the second would > > indicate EOF. This would certainly seem like a win, but note there > > are diminishing returns as the directory size increases, versus just > > doubling the buffer size. > > True, but most directories are small, so I expect it would be a > benefit most of the time. Even for big directories you still get two > buffers filled with one syscall, same as if you did a conventional > getdents64() with twice as big a buffer. > > > An alternate / additional idea you may wish to consider is changing > > getdents64 itself. > > > > Ordinary read functions must return 0 length to indicate EOF, because > > they can return arbitrary data. This is not the case for getdents64. > > > > 1) Define a struct linux_dirent of minimum size containing an abnormal > > value as a sentinel. d_off = 0 or -1 should work. > > > > 2) Implement a flag for getdents64. > > Sadly getdents64() doesn't take a flags argument. We'd probably need > a new syscall. > > > IF > > the flag is set AND > > we are returning a non-zero length buffer AND > > there is room in the buffer for the sentinel structure AND > > a getdents64 call using the d_off of the last struct in the > > buffer would return EOF > > THEN > > append the sentinel struct to the buffer. > > > > > > Using the arrangement, we would still handle a 0 length return as an > > EOF, but if we see the sentinel struct, we can skip the additional call > > altogether. The saves all of the pairing of buffers and extra logic, > > and unless we're unlucky and the sentinel structure did not fit in > > the buffer at EOF, would always reduce the number of getdents64 > > calls by one. > > > > Moreover, if the flag was available outside of io_uring, for smaller > > directories, this feature would cut the number of directory reads > > of readdir(3) by up to half. > > If we need a new syscall anyway, the calling convention could be > adjusted to indicate EOF more easily than that, e.g. > > int getdents2(int fd, void *buf, size_t *size, unsigned long flags); > > With 0 being EOF, 1 being not-EOF, and -1 for error, or something. Good point! I was hedging the idea above with "if the flag was available outside of io_uring" to allow an internal-only flag. A new syscall would certainly be more useful as it would improve every readdir(3) call, and of course could be called from inside io_uring more efficiently than the current getdents64. Alas, adding a new syscall like this is a little beyond my level of expertise. Would you (or anyone else reading this) have any interest in implementing a getdents2? > -- > Tavian Barnes
