[I shortened the CC list as that long came up only due to mmap and optimizations] On 6/12/24 16:56, Bernd Schubert wrote: > On 6/12/24 16:07, Miklos Szeredi wrote: >> On Wed, 12 Jun 2024 at 15:33, Bernd Schubert <bschubert@xxxxxxx> wrote: >> >>> I didn't do that yet, as we are going to use the ring buffer for requests, >>> i.e. the ring buffer immediately gets all the data from network, there is >>> no copy. Even if the ring buffer would get data from local disk - there >>> is no need to use a separate application buffer anymore. And with that >>> there is just no extra copy >> >> Let's just tackle this shared request buffer, as it seems to be a >> central part of your design. >> >> You say the shared buffer is used to immediately get the data from the >> network (or various other sources), which is completely viable. >> >> And then the kernel will do the copy from the shared buffer. Single copy, fine. >> >> But if the buffer wasn't shared? What would be the difference? >> Single copy also. >> >> Why is the shared buffer better? I mean it may even be worse due to >> cache aliasing issues on certain architectures. copy_to_user() / >> copy_from_user() are pretty darn efficient. > > Right now we have: > > - Application thread writes into the buffer, then calls io_uring_cmd_done > > I can try to do without mmap and set a pointer to the user buffer in the > 80B section of the SQE. I'm not sure if the application is allowed to > write into that buffer, possibly/probably we will be forced to use > io_uring_cmd_complete_in_task() in all cases (without 19/19 we have that > anyway). My greatest fear here is that the extra task has performance > implications for sync requests. > > >> >> Why is it better to have that buffer managed by kernel? Being locked >> in memory (being unswappable) is probably a disadvantage as well. And >> if locking is required, it can be done on the user buffer. > > Well, let me try to give the buffer in the 80B section. > >> >> And there are all the setup and teardown complexities... > > If the buffer in the 80B section works setup becomes easier, mmap and > ioctls go away. Teardown, well, we still need the workaround as we need > to handle io_uring_cmd_done, but if you could live with that for the > instance, I would ask Jens or Pavel or Ming for help if we could solve > that in io-uring itself. > Is the ring workaround in fuse_dev_release() acceptable for you? Or do > you have any another idea about it? > >> >> Note: the ring buffer used by io_uring is different. It literally >> allows communication without invoking any system calls in certain >> cases. That shared buffer doesn't add anything like that. At least I >> don't see what it actually adds. >> >> Hmm? > > The application can write into the buffer. We won't shared queue buffers > if we could solve the same with a user pointer. Wanted to send out a new series today, https://github.com/bsbernd/linux/tree/fuse-uring-for-6.10-rfc3-without-mmap but then just noticed a tear down issue. 1525.905504] KASAN: null-ptr-deref in range [0x00000000000001a0-0x00000000000001a7] [ 1525.910431] CPU: 15 PID: 183 Comm: kworker/15:1 Tainted: G O 6.10.0+ #48 [ 1525.916449] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.3-debian-1.16.3-2 04/01/2014 [ 1525.922470] Workqueue: events io_fallback_req_func [ 1525.925840] RIP: 0010:__lock_acquire+0x74/0x7b80 [ 1525.929010] Code: 89 bc 24 80 00 00 00 0f 85 1c 5f 00 00 83 3d 6e 80 b0 02 00 0f 84 1d 12 00 00 83 3d 65 c7 67 02 00 74 27 48 89 f8 48 c1 e8 03 <42> 80 3c 30 00 74 0d e8 50 44 42 00 48 8b bc 24 80 00 00 00 48 c7 [ 1525.942211] RSP: 0018:ffff88810b2af490 EFLAGS: 00010002 [ 1525.945672] RAX: 0000000000000034 RBX: 0000000000000000 RCX: 0000000000000001 [ 1525.950421] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 00000000000001a0 [ 1525.955200] RBP: 0000000000000000 R08: 0000000000000001 R09: 0000000000000000 [ 1525.959979] R10: dffffc0000000000 R11: fffffbfff07b1cbe R12: 0000000000000000 [ 1525.964252] R13: 0000000000000001 R14: dffffc0000000000 R15: 0000000000000001 [ 1525.968225] FS: 0000000000000000(0000) GS:ffff88875b200000(0000) knlGS:0000000000000000 [ 1525.973932] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 1525.976694] CR2: 00005555b6a381f0 CR3: 000000012f5f1000 CR4: 00000000000006f0 [ 1525.980030] Call Trace: [ 1525.981371] <TASK> [ 1525.982567] ? __die_body+0x66/0xb0 [ 1525.984376] ? die_addr+0xc1/0x100 [ 1525.986111] ? exc_general_protection+0x1c6/0x330 [ 1525.988401] ? asm_exc_general_protection+0x22/0x30 [ 1525.990864] ? __lock_acquire+0x74/0x7b80 [ 1525.992901] ? mark_lock+0x9f/0x360 [ 1525.994635] ? __lock_acquire+0x1420/0x7b80 [ 1525.996629] ? attach_entity_load_avg+0x47d/0x550 [ 1525.998765] ? hlock_conflict+0x5a/0x1f0 [ 1526.000515] ? __bfs+0x2dc/0x5a0 [ 1526.001993] lock_acquire+0x1fb/0x3d0 [ 1526.004727] ? gup_fast_fallback+0x13f/0x1d80 [ 1526.006586] ? gup_fast_fallback+0x13f/0x1d80 [ 1526.008412] gup_fast_fallback+0x158/0x1d80 [ 1526.010170] ? gup_fast_fallback+0x13f/0x1d80 [ 1526.011999] ? __lock_acquire+0x2b07/0x7b80 [ 1526.013793] __iov_iter_get_pages_alloc+0x36e/0x980 [ 1526.015876] ? do_raw_spin_unlock+0x5a/0x8a0 [ 1526.017734] iov_iter_get_pages2+0x56/0x70 [ 1526.019491] fuse_copy_fill+0x48e/0x980 [fuse] [ 1526.021400] fuse_copy_args+0x174/0x6a0 [fuse] [ 1526.023199] fuse_uring_prepare_send+0x319/0x6c0 [fuse] [ 1526.025178] fuse_uring_send_req_in_task+0x42/0x100 [fuse] [ 1526.027163] io_fallback_req_func+0xb4/0x170 [ 1526.028737] ? process_scheduled_works+0x75b/0x1160 [ 1526.030445] process_scheduled_works+0x85c/0x1160 [ 1526.032073] worker_thread+0x8ba/0xce0 [ 1526.033388] kthread+0x23e/0x2b0 [ 1526.035404] ? pr_cont_work_flush+0x290/0x290 [ 1526.036958] ? kthread_blkcg+0xa0/0xa0 [ 1526.038321] ret_from_fork+0x30/0x60 [ 1526.039600] ? kthread_blkcg+0xa0/0xa0 [ 1526.040942] ret_from_fork_asm+0x11/0x20 [ 1526.042353] </TASK> We probably need to call iov_iter_get_pages2() immediately on submitting the buffer from fuse server and not only when needed. I had planned to do that as optimization later on, I think it is also needed to avoid io_uring_cmd_complete_in_task(). The part I don't like here is that with mmap we had a complex initialization - but then either it worked or did not. No exceptions at IO time. And run time was just a copy into the buffer. Without mmap initialization is much simpler, but now complexity shifts to IO time. Thanks, Bernd