基础
qemu 参数 ,discard=on,aio=native,cache.direct=on
原来，cache.direct=on 的含义是由于 aio 无法处理 buffer io ？
qemu-system-x86_64: -drive file=/home/martins3/hack/vm/2403-nix/img/virtio_blk_1,format=qcow2,if=none,id=virtio-blk1,aio=native: aio=native was specified, but it requires cache.direct=on, which was not specified.
/* Currently Linux does AIO only for files opened with O_DIRECT */
if (s->use_linux_aio && !(s->open_flags & O_DIRECT)) {
error_setg(errp, "aio=native was specified, but it requires "
"cache.direct=on, which was not specified.");
ret = -EINVAL;
goto fail;
}
if (s->use_linux_aio) {
s->has_laio_fdsync = laio_has_fdsync(s->fd);
}
也就是 iouring 可以读写 buffer ，但是 aio 必须使用 O_DIRECT 的模式
进行 io 的时候，页面是什么时候 pin 住的
如果是 direct io ，那么需要在提交的时候，将用户态的页面 pin 住
@[
bio_iov_iter_get_pages+5
iomap_dio_bio_iter+587
__iomap_dio_rw+937
iomap_dio_rw+18
xfs_file_dio_read+185
xfs_file_read_iter+265
__io_read+175
io_read+63
__io_issue_sqe+59
io_issue_sqe+55
io_submit_sqes+274
__do_sys_io_uring_enter+516
do_syscall_64+126
entry_SYSCALL_64_after_hwframe+118
]: 1200961
@[
bio_iov_iter_get_pages+5
iomap_dio_bio_iter+587
__iomap_dio_rw+937
iomap_dio_rw+18
xfs_file_dio_read+185
xfs_file_read_iter+265
aio_read+306
io_submit_one+239
__x64_sys_io_submit+148
do_syscall_64+126
entry_SYSCALL_64_after_hwframe+118
]: 659511
如果是 buffer io read ，需要在提交的时候 pin 页面用于 dma
@[
bio_iov_iter_get_pages+5
iomap_dio_bio_iter+485
__iomap_dio_rw+937
iomap_dio_rw+18
xfs_file_dio_read+185
xfs_file_read_iter+265
__io_read+175
io_read+63
__io_issue_sqe+59
io_issue_sqe+55
io_submit_sqes+274
__do_sys_io_uring_enter+456
do_syscall_64+126
entry_SYSCALL_64_after_hwframe+118
]: 55964
在完成的时候，来拷贝页面，而且需要注意到，如果提交的时候，内存不足，那么是在提交的过程中 来分配，不会由于没有内存，就返回 EAGAIN 的
6.69% __do_sys_io_uring_enter
- 69.55% io_submit_sqes
- 68.71% io_issue_sqe
- 68.16% __io_issue_sqe
- io_read
- __io_read
- 67.42% 0xffffffffc10b411f
- 66.17% 0xffffffffc10b3444
- 66.13% filemap_read
- 65.07% filemap_get_pages
- 59.00% force_page_cache_ra
- 58.89% page_cache_ra_unbounded
- 33.80% filemap_add_folio
- 24.68% __mem_cgroup_charge
- 23.76% charge_memcg
+ 23.22% try_charge_memcg
0.78% get_mem_cgroup_from_mm
+ 7.80% __filemap_add_folio
+ 1.15% __folio_batch_add_and_move
- 20.45% read_pages
+ 18.42% iomap_readahead
- 1.85% blk_finish_plug
- __blk_flush_plug
- blk_mq_flush_plug_list
- blk_mq_dispatch_queue_requests
0.81% 0xffffffffc04659c0
2.31% up_read
+ 1.73% folio_alloc_noprof
- 3.28% filemap_get_read_batch
2.58% xas_load
1.88% filemap_update_page
- 0.60% touch_atime
atime_needs_update
0.57% 0xffffffffc10b3454
- 14.42% __io_run_local_work
- 13.70% __io_run_local_work_loop
- 13.65% io_req_task_submit
- io_issue_sqe
- 13.46% __io_issue_sqe
- 13.41% io_read
- 12.91% __io_read
- 12.36% 0xffffffffc10b411f
- 11.90% 0xffffffffc10b3444
- 11.86% filemap_read
- 6.82% copy_page_to_iter
_copy_to_iter
- 3.36% filemap_get_pages
- filemap_get_read_batch
2.66% xas_load
buffer write 的 perf 结果，可以看到大多数工作都是在 io_wq_worker 中完成的， 而且工作主要是拷贝内存，提交任务在另外系统中:
- 45.84% 0.00% iou-wrk-128624 [unknown] [k] 0x0000000004365b10
0x4365b10
0
ret_from_fork_asm
ret_from_fork
io_wq_worker
- io_worker_handle_work
- 45.02% io_wq_submit_work
- io_issue_sqe
- 39.81% __io_issue_sqe
- 39.56% io_write
- 38.71% ext4_buffered_write_iter
- 37.56% generic_perform_write
+ 27.76% ext4_da_write_begin
4.98% copy_folio_from_iter_atomic
+ 4.14% ext4_da_do_write_end.isra.0
- 4.94% __io_req_task_work_add
+ 4.43% try_to_wake_up
io_cqring_wait
当 iouring 等待动作完成的时候，就是等待在这里，所以经常可以看到 process 卡到这里:
[<0>] io_cqring_wait+0x674/0x6a0 # 具体在 __io_cqring_wait_schedule
[<0>] __do_sys_io_uring_enter+0x141/0x410
[<0>] do_syscall_64+0x84/0x250
[<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e
例如使用 direct=0 ext4 10G 的文件 1000s 然后观察到的结果:
- 87.20% entry_SYSCALL_64_after_hwframe
- do_syscall_64
- 84.85% __do_sys_io_uring_enter
- 60.97% io_submit_sqes
- 60.14% io_issue_sqe
- 59.54% __io_issue_sqe
- io_read
- __io_read
- 58.76% filemap_read
- 58.07% filemap_get_pages
- 52.92% force_page_cache_ra
- 52.81% page_cache_ra_unbounded
- 42.89% read_pages
- 27.04% ext4_mpage_readpages
- 21.12% submit_bio_noacct_nocheck
+ 20.61% __submit_bio
- 2.30% bio_alloc_bioset
- 1.09% mempool_alloc_noprof
0.95% kmem_cache_alloc_noprof
- 1.02% bio_associate_blkg
0.92% bio_associate_blkg_from_css
- 2.14% ext4_map_blocks
- 1.19% __check_block_validity.constprop.0
1.01% ext4_sb_block_valid
0.82% ext4_es_lookup_extent
+ 15.52% blk_finish_plug
+ 6.32% filemap_add_folio
+ 2.66% folio_alloc_noprof
- 2.95% filemap_get_read_batch
2.50% xas_load
1.40% filemap_update_page
- 21.16% io_cqring_wait
- 9.73% schedule
+ 9.57% __schedule
- 9.18% __io_run_local_work
- 8.13% __io_run_local_work_loop
- io_req_task_submit
- 7.91% io_issue_sqe
- __io_issue_sqe
- 7.46% io_read
- __io_read
- 6.19% filemap_read
- 2.34% filemap_get_pages
- filemap_get_read_batch
1.33% xas_load
- 2.32% copy_page_to_iter
2.27% _copy_to_iter
- 0.63% touch_atime
0.51% atime_needs_update
0.59% llist_reverse_order
- 1.46% __io_run_local_work
- 1.31% __io_run_local_work_loop
- io_req_task_submit
- io_issue_sqe
- __io_issue_sqe
- 1.23% io_read
- __io_read
- 1.14% filemap_read
- 0.56% copy_page_to_iter
_copy_to_iter
+ 1.14% exit_to_user_mode_loop
0.63% arch_exit_to_user_mode_prepare.isra.0
struct kiocb 和 struct io_kiocb
组装的位置在 __kernel_read，然后传递给这种注册的 read_iter 和 write_iter
ret = file->f_op->read_iter(&kiocb, &iter);
简而言之 kiocb 就是为了为了让为了执行 callback
struct kiocb {
struct file *ki_filp; // 文件指针
loff_t ki_pos; // 文件偏移
void (*ki_complete)(struct kiocb *iocb, long ret); // 完成回调
void *private; // 私有数据
int ki_flags; // 标志位
u16 ki_ioprio; // IO 优先级
// ...
};
struct io_kiocb {
union {
struct file *file; // 文件指针
struct io_cmd_data cmd;
};
u8 opcode; // 操作码 (如 IORING_OP_READ, IORING_OP_WRITE)
io_req_flags_t flags; // REQ_F_* 标志
struct io_cqe cqe; // 完成队列条目
struct io_ring_ctx *ctx; // io_uring 上下文
struct io_uring_task *tctx; // per-task 上下文
atomic_t refs; // 引用计数
struct io_kiocb *link; // 链接的请求（用于链式操作）
void *async_data; // 异步数据
// ...
};
常见的就是 aio 和 io uring 了，但是也可以看到一些其他的
- erofs
经典的数据流的过程，这个 rq_submit 已经到了驱动了，但是下层又是文件系统，所以是 rq submit 到 iocb
static void erofs_fileio_rq_submit(struct erofs_fileio_rq *rq) { struct iov_iter iter; int ret; if (!rq) return; rq->iocb.ki_pos = rq->bio.bi_iter.bi_sector << SECTOR_SHIFT; rq->iocb.ki_ioprio = get_current_ioprio(); rq->iocb.ki_complete = erofs_fileio_ki_complete; if (test_opt(&EROFS_SB(rq->sb)->opt, DIRECT_IO) && rq->iocb.ki_filp->f_mode & FMODE_CAN_ODIRECT) rq->iocb.ki_flags = IOCB_DIRECT; iov_iter_bvec(&iter, ITER_DEST, rq->bvecs, rq->bio.bi_vcnt, rq->bio.bi_iter.bi_size); ret = vfs_iocb_iter_read(rq->iocb.ki_filp, &rq->iocb, &iter); if (ret != -EIOCBQUEUED) erofs_fileio_ki_complete(&rq->iocb, ret); } - mm/page_io.c:swap_writepage_fs
buffered io 和 direct io 的 callback 是什么?
- blk_update_request 中来处理 bi_end_io 的 callback
- blkdev_bio_end_io_async (direct io : 如果直接在 blkdev 上)
- io_complete_rw
- iomap_dio_bio_end_io (direct io : 如果在 xfs 上)
- io_complete_rw
- mpage_end_io (buffer io)
- __read_end_io
- folio_wake_bit
- io_async_buf_func : read buffer io 的 callback
- io_req_task_queue (将完成的任务挂上去)
- io_async_buf_func : read buffer io 的 callback
- folio_wake_bit
- __read_end_io
- blkdev_bio_end_io_async (direct io : 如果直接在 blkdev 上)
direct io 经典 callback :
@[
io_complete_rw+5
blkdev_bio_end_io_async+78
blk_update_request+415
scsi_end_request+39
scsi_io_completion+83
blk_done_softirq+74
handle_softirqs+241
__irq_exit_rcu+194
common_interrupt+133
asm_common_interrupt+38
cpuidle_enter_state+211
cpuidle_enter+45
cpuidle_idle_call+241
do_idle+120
cpu_startup_entry+41
start_secondary+296
common_startup_64+318
]: 301
buffer io 经典的 callback : io_async_buf_func
@[
folio_wake_bit+1
__read_end_io+282
blk_update_request+415
scsi_end_request+39
scsi_io_completion+83
blk_done_softirq+74
handle_softirqs+241
__irq_exit_rcu+194
common_interrupt+133
asm_common_interrupt+38
cpuidle_enter_state+211
cpuidle_enter+45
cpuidle_idle_call+241
do_idle+120
cpu_startup_entry+41
start_secondary+296
common_startup_64+318
]: 867
31) | folio_wake_bit() {
31) 0.107 us | _raw_spin_lock_irqsave();
31) | __wake_up_locked_key() {
31) | __wake_up_common() {
31) | io_async_buf_func) {
31) | io_req_task_queue() {
31) | __io_req_task_work_add() {
31) 0.087 us | __rcu_read_lock();
31) 0.077 us | __rcu_read_unlock();
31) 0.490 us | }
31) 0.688 us | }
31) 0.911 us | }
31) 1.165 us | }
31) 1.336 us | }
31) 0.090 us | _raw_spin_unlock_irqrestore();
31) 1.948 us | }
31) 2.124 us | }
存储系统中断返回的 callback 分析
- blk_mq_end_request
- blk_update_request
- bio_endio
- end_bio_bh_io_sync
- bio_put
- end_bio_bh_io_sync
- blk_stat_add
- blk_mq_free_request
- bio_endio
- blk_update_request
可以看到，同时会释放 request / bio ，然后统计数据。
总体来说，这是符合预期的，做出来的功能: 释放，统计，唤醒等待的进程
# tracer: function_graph
#
# CPU DURATION FUNCTION CALLS
# | | | | | | |
2) | dummy_complete_rq [virtio_dummy]() {
2) | blk_mq_end_request() {
2) | blk_update_request() {
2) | blk_account_io_completion.part.0() {
2) 0.120 us | preempt_count_add();
2) 0.087 us | preempt_count_sub();
2) 0.755 us | }
2) | bio_endio() {
2) 0.476 us | blk_throtl_bio_endio();
2) | bio_uninit() {
2) 0.082 us | __rcu_read_lock();
2) 0.090 us | __rcu_read_unlock();
2) 0.587 us | }
2) | end_bio_bh_io_sync() {
2) | end_buffer_read_sync() {
2) | wake_up_bit() { // 省掉一大片
2) 6.722 us | }
2) 6.914 us | }
2) | bio_put() {
2) | bio_free() {
2) 0.083 us | bio_uninit();
2) 0.084 us | bvec_free();
2) | mempool_free() {
2) | mempool_free_slab() {
2) | kmem_cache_free() {
2) 0.118 us | __slab_free();
2) 0.437 us | }
2) 0.611 us | }
2) 0.774 us | }
2) 1.359 us | }
2) 1.590 us | }
2) 8.833 us | }
2) + 10.441 us | }
2) + 11.684 us | }
2) 0.173 us | ktime_get();
2) | blk_stat_add() {
2) | blk_throtl_stat_add() {
2) 0.083 us | throtl_track_latency();
2) 0.317 us | }
2) 0.081 us | __rcu_read_lock();
2) 0.082 us | preempt_count_add();
2) 0.082 us | preempt_count_sub();
2) 0.082 us | __rcu_read_unlock();
2) 1.642 us | }
2) 0.081 us | preempt_count_add();
2) 0.093 us | update_io_ticks(); // TODO 为什么这里需要 preempt 的操作
2) 0.082 us | preempt_count_sub();
2) | blk_mq_free_request() {
2) | __blk_mq_free_request() {
2) 0.262 us | blk_mq_put_tag();
2) | blk_queue_exit() {
2) 0.083 us | __rcu_read_lock();
2) 0.075 us | __rcu_read_unlock();
2) 0.472 us | }
2) 0.075 us | __rcu_read_unlock();
2) 0.472 us | }
2) 1.374 us | }
2) 1.600 us | }
2) + 16.309 us | }
2) + 17.881 us | }
aio 提交 buffer io 等待位置
[<0>] folio_wait_bit_common+0x12b/0x320
[<0>] filemap_update_page+0x2fe/0x310
[<0>] filemap_get_pages+0x2ba/0x470
[<0>] filemap_read+0xff/0x440
[<0>] aio_read+0x133/0x210
[<0>] io_submit_one+0xee/0x370
[<0>] __x64_sys_io_submit+0x94/0x1e0
[<0>] do_syscall_64+0x84/0x250
[<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e
- 86.83% entry_SYSCALL_64_after_hwframe
- do_syscall_64
- 49.49% __x64_sys_io_submit
- 48.07% io_submit_one
- 43.62% aio_write
- 41.73% ext4_buffered_write_iter
- 40.49% generic_perform_write
- 32.87% ext4_da_write_begin
- 28.05% __filemap_get_folio
- 18.73% filemap_add_folio
+ 8.39% __filemap_add_folio
+ 8.20% __mem_cgroup_charge
- 1.95% __folio_batch_add_and_move
- folio_batch_move_lru
- 1.35% lru_add
0.95% lru_gen_add_folio
- 6.08% filemap_get_entry
5.92% xas_load
+ 2.92% folio_alloc_noprof
+ 4.47% ext4_block_write_begin
+ 4.16% ext4_da_do_write_end.isra.0
2.44% copy_folio_from_iter_atomic
0.63% balance_dirty_pages_ratelimited_flags
1.03% aio_complete
0.88% __io_submit_one.isra.0
0.62% _copy_from_user
0.56% kmem_cache_alloc_noprof
0.80% lookup_ioctx
+ 33.08% __x64_sys_fadvise64
+ 3.22% __x64_sys_io_getevents
0.58% arch_exit_to_user_mode_prepare.isra.0
然后分析对比一下 ./aio.md 和 ./aio-buffer-io.md
本站所有文章转发 CSDN 将按侵权追究法律责任，其它情况随意。