On 3/3/22 16:31, Jens Axboe wrote:
On 3/3/22 7:40 AM, Jens Axboe wrote:
On 3/3/22 7:36 AM, Jens Axboe wrote:
The only potential oddity here is that the fd passed back is not a
legitimate fd. io_uring does support poll(2) on its file descriptor, so
that could cause some confusion even if I don't think anyone actually
does poll(2) on io_uring.
Side note - the only implication here is that we then likely can't make
the optimized behavior the default, it has to be an IORING_SETUP_REG
flag which tells us that the application is aware of this limitation.
Though I guess close(2) might mess with that too... Hmm.
Not sure I can find a good approach for that. Tried out your patch and
made some fixes:
- Missing free on final tctx free
- Rename registered_files to registered_rings
- Fix off-by-ones in checking max registration count
- Use kcalloc
- Rename ENTER_FIXED_FILE -> ENTER_REGISTERED_RING
- Don't pass in tctx to io_uring_unreg_ringfd()
- Get rid of forward declaration for adding tctx node
- Get rid of extra file pointer in io_uring_enter()
- Fix deadlock in io_ringfd_register()
- Use io_uring_rsrc_update rather than add a new struct type
Patch I ran below.
Ran some testing here, and on my laptop, running:
axboe@m1pro-kvm ~/g/fio (master)> t/io_uring -N1 -s1 -f0
polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128
Engine=io_uring, sq_ring=128, cq_ring=128
submitter=0, tid=673
IOPS=6627K, IOS/call=1/1, inflight=()
IOPS=6995K, IOS/call=1/1, inflight=()
IOPS=6992K, IOS/call=1/1, inflight=()
IOPS=7005K, IOS/call=1/1, inflight=()
IOPS=6999K, IOS/call=1/1, inflight=()
and with registered ring
axboe@m1pro-kvm ~/g/fio (master)> t/io_uring -N1 -s1 -f1
polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128
Engine=io_uring, sq_ring=128, cq_ring=128
submitter=0, tid=687
ring register 0
IOPS=7714K, IOS/call=1/1, inflight=()
IOPS=8030K, IOS/call=1/1, inflight=()
IOPS=8025K, IOS/call=1/1, inflight=()
IOPS=8015K, IOS/call=1/1, inflight=()
IOPS=8037K, IOS/call=1/1, inflight=()
which is about a 15% improvement, pretty massive...
Is the bench single threaded (including io-wq)? Because if it
is, get/put shouldn't do any atomics and I don't see where the
result comes from.
static unsigned long __fget_light(unsigned int fd, fmode_t mask)
{
struct files_struct *files = current->files;
struct file *file;
if (atomic_read(&files->count) == 1) {
file = files_lookup_fd_raw(files, fd);
if (!file || unlikely(file->f_mode & mask))
return 0;
return (unsigned long)file;
} else {
file = __fget(fd, mask, 1);
if (!file)
return 0;
return FDPUT_FPUT | (unsigned long)file;
}
}
--
Pavel Begunkov
