在上一篇《一次降低进程IO延迟的性能优化实践——基于block层bfq调度器》基础上,本文主要总结实现该IO性能优化过程遇到的 IO卡死、IO重复派发、内核crash等问题。
1:IO重复派发触发了crash
在初版代码编写完成后,启动fio测试+cat读取文件,有很大概率触发了内核crash,现场如下
- PID: 11602 TASK: ffff95f3092ddf00 CPU: 3 COMMAND: "cat"
- #0 [ffffa67081ceb390] machine_kexec at ffffffff8525bf3e
- #1 [ffffa67081ceb3e8] __crash_kexec at ffffffff8536072d
- #2 [ffffa67081ceb4b0] panic at ffffffff852b5dc7
- #3 [ffffa67081ceb530] __warn.cold.12 at ffffffff852b5fee
- #4 [ffffa67081ceb538] blk_mq_start_request at ffffffff856075d0
- #5 [ffffa67081ceb560] blk_mq_start_request at ffffffff856075d0
- #6 [ffffa67081ceb590] do_error_trap at ffffffff8521f9de
- #7 [ffffa67081ceb5d0] do_invalid_op at ffffffff8521fe36
- #8 [ffffa67081ceb5f0] invalid_op at ffffffff85c00d84
- [exception RIP: blk_mq_start_request+496]
- RIP: ffffffff856075d0 RSP: ffffa67081ceb6a0 RFLAGS: 00010202
- RAX: 0000000000000001 RBX: ffff95f28fc57810 RCX: 0000000000000018
- RDX: 00000000004b1dc2 RSI: ffff95f28fc57810 RDI: ffff95f297722758
- RBP: ffff95f38f868000 R8: ffffa67081ceb7e8 R9: 0000000000000000
- R10: 0000000000000000 R11: 0000000000000011 R12: ffff95f296143000
- R13: ffff95f2987fe000 R14: ffff95f2987fe050 R15: ffffa67081ceb788
- ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018
- #9 [ffffa67081ceb6b8] scsi_queue_rq at ffffffff857d1a51
- #10 [ffffa67081ceb708] blk_mq_dispatch_rq_list at ffffffff85609f4c
- #11 [ffffa67081ceb7d8] blk_mq_do_dispatch_sched at ffffffff8560f4ba
- #12 [ffffa67081ceb830] __blk_mq_sched_dispatch_requests at ffffffff8560ff99
- #13 [ffffa67081ceb890] blk_mq_sched_dispatch_requests at ffffffff85610020
- #14 [ffffa67081ceb8a0] __blk_mq_run_hw_queue at ffffffff856076a1
- #15 [ffffa67081ceb8b8] __blk_mq_delay_run_hw_queue at ffffffff85607f61
- #16 [ffffa67081ceb8e0] blk_mq_sched_insert_requests at ffffffff85610351
- #17 [ffffa67081ceb918] blk_mq_flush_plug_list at ffffffff8560b4d6
- #18 [ffffa67081ceb998] blk_flush_plug_list at ffffffff855ffbe7
- #19 [ffffa67081ceb9e8] blk_mq_make_request at ffffffff8560ad38
- #20 [ffffa67081ceba78] generic_make_request at ffffffff855fe85f
- #21 [ffffa67081cebad0] submit_bio at ffffffff855feadc
- #22 [ffffa67081cebb10] ext4_mpage_readpages at ffffffffc08eead1 [ext4]
- #23 [ffffa67081cebbf8] read_pages at ffffffff8543743b
- #24 [ffffa67081cebc70] __do_page_cache_readahead at ffffffff85437721
- ………………….
触发crash的源码位置如下
- void blk_mq_start_request(struct request *rq)
- {
- struct request_queue *q = rq->q;
- blk_mq_sched_started_request(rq);
- trace_block_rq_issue(q, rq);
- if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
- rq->io_start_time_ns = ktime_get_ns();
- rq_aux(rq)->stats_sectors = blk_rq_sectors(rq);
- rq->rq_flags |= RQF_STATS;
- rq_qos_issue(q, rq);
- }
- WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);//这里crash
- blk_add_timer(rq);
- //标记rq->state 为MQ_RQ_IN_FLIGHT,表示IO请求派发给磁盘驱动了
- WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
- }
- static inline enum mq_rq_state blk_mq_rq_state(struct request *rq)
- {
- return READ_ONCE(rq->state);
- }
crash过程是:在把rq派发给磁盘驱动过程执行blk_mq_start_request()函数中,rq->state不是MQ_RQ_IDLE,然后就主动触发WARN_ON_ONCE而crash。按照经验,crash现场的RDI寄存器就是blk_mq_start_request()函数传输rq指针,看下这个rq的参数:
- crash> request ffff95f297722758
- __data_len = 0, //date_len 有问题
- tag = -275282040, //tag 有问题
- __sector = 18446638524612970376, //扇区地址明显有问题
- bio = 0x0, //这个bio有问题
- biotail = 0x0,
- rq_disk = 0x0, /rq_disk 不可能是NULL
- state = MQ_RQ_IDLE,
到这里怀疑rdi:0xffff95f297722758应该不是blk_mq_start_request()函数传参rq指针,因为打印的rq结构体变量根本不符合常理,对于不符合常理的就要另找他法。
因为这个case比较容易复现,大概率跟我在_bfq_dispatch_request()添加的代码有关。于是在blk_mq_start_request()和__bfq_dispatch_request()中添加一下调试信息,如下红色代码:
- void blk_mq_start_request(struct request *rq)
- {
- struct request_queue *q = rq->q;
- blk_mq_sched_started_request(rq);
- trace_block_rq_issue(q, rq);
- if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
- rq->io_start_time_ns = ktime_get_ns();
- rq_aux(rq)->stats_sectors = blk_rq_sectors(rq);
- rq->rq_flags |= RQF_STATS;
- rq_qos_issue(q, rq);
- }
- printk("%s %s %d rq:0x%llx rq->rq_disk:0x%llx \n",__func__,current->comm,current->pid,(u64)rq,(u64)rq->rq_disk);
- WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
- blk_add_timer(rq);
- //标记rq->state 为MQ_RQ_IN_FLIGHT,表示IO请求派发给磁盘驱动了
- WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
- }
- static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
- {
- ..................
- if(bfqd->bfq_high_io_prio_mode)
- {
- //在 bfq_high_io_prio_mode 非0时间的5s内,如果遇到非high prio io,并且驱动队列IO个数大于限制,则把不派发该IO,而是临时添加到bfq_high_prio_tmp_list链表
- if((bfqd->rq_in_driver >= 16) && (bfqd->bfq_high_prio_tmp_list_rq_count < 100)){
- //把rq从原有链表删掉并把rq移动到bfq_high_prio_tmp_list链表尾,派发时是从bfq_high_prio_tmp_list链表头取出rq,保证先到先派发
- list_add_tail(&rq->queuelist,&bfqd->bfq_high_prio_tmp_list);
- bfqd->bfq_high_prio_tmp_list_rq_count ++;
- p_process_io_info_tmp->block_io_count ++;
- printk("%s %s %d rq:0x%llx bfqq:0x%llx pid:%d bfqq->dispatched:%d bfq_high_prio_tmp_list_rq_count:%d rq_in_driver:%d !!!!!!!!!!!!\n",__func__,current->comm,current->pid,(u64)rq,(u64)bfqq,bfqq->pid,bfqq->dispatched,bfqd->bfq_high_prio_tmp_list_rq_count,bfqd->rq_in_driver);
- goto exit1;
- }
- }
- ..................
- }
等下次触发crash,内核打印 blk_mq_start_request cat 15092 rq:0xffff8eff2401d990 rq->rq_disk:0xffff8efe1b1b4000,看下它的成员信息:
- crash> request 0xffff8eff2401d990
- struct request {
- __data_len = 1048576,
- tag = 86,
- __sector = 3468288,
- bio = 0xffff8efd875e8300,
- biotail = 0xffff8efd875e8300,
- rq_disk = 0xffff8efe1b1b4000,
- state = MQ_RQ_IN_FLIGHT,
看来,这次的rq指针是正确的,刚才通过rdi获取blk_mq_start_request()函数传参是有问题的。这个rq->state是MQ_RQ_IN_FLIGHT,就是说该rq已经派发给磁盘驱动了,在传输完成前又派发给磁盘驱动,显然重复了。再看下crash前的内核打印,印证了我的想法
- //rq:0xffff8eff2401d990 这里被插入 bfq_high_prio_tmp_list_rq_count 链表
- [ 132.559190] __bfq_dispatch_request cat 15092 rq:0xffff8eff2401d990 bfqq:0xffff8efe1ba0b200 pid:15092 bfqq->dispatched:17 bfq_high_prio_tmp_list_rq_count:1 rq_in_driver:16 !!!!!!!!!!!!1
- //rq:0xffff8eff2401d990 被派发
- [ 132.559244] blk_mq_start_request cat 15092 rq:0xffff8eff2401d990 rq->rq_disk:0xffff8efe1b1b4000
- //rq:0xffff8eff2401d990 又被派发
- [ 132.561350] blk_mq_start_request cat 15092 rq:0xffff8eff2401d990 rq->rq_disk:0xffff8efe1b1b4000
- [ 132.561398] WARNING: CPU: 1 PID: 15092 at block/blk-mq.c:696 blk_mq_start_request+0x128/0x263
- [ 132.561401] Kernel panic - not syncing: panic_on_warn set ...
- [ 132.561409] CPU: 1 PID: 15092 Comm: cat Kdump: loaded Tainted: G E ---------r- - 4.18.0 #2
- [ 132.561412] Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 02/27/2020
- [ 132.561414] Call Trace:
- [ 132.561431] dump_stack+0x5c/0x80
- [ 132.561437] panic+0xe7/0x2a9
- [ 132.561443] ? blk_mq_start_request+0x128/0x263
- [ 132.561447] __warn.cold.12+0x31/0x33
- [ 132.561450] ? blk_mq_start_request+0x128/0x263
- [ 132.561454] ? blk_mq_start_request+0x128/0x263
- [ 132.561457] report_bug+0xb1/0xd0-
显然,rq:0xffff8eff2401d990就是被连续派发了两次,就得看看我添加的代码哪里有问题了?
- static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
- {
- ...............
- rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
- if (rq) {
- if(bfqd->queue->high_io_prio_enable)
- {
- if(rq->rq_flags & RQF_HIGH_PRIO){//高优先级IO
- //第一次遇到high prio io,置1 bfq_high_io_prio_mode,启动5s定时器,定时到了对bfq_high_io_prio_mode清0
- if(bfqd->bfq_high_io_prio_mode == 0){
- bfqd->bfq_high_io_prio_mode = 1;
- hrtimer_start(&bfqd->bfq_high_prio_timer, ms_to_ktime(5000),HRTIMER_MODE_REL);
- }
- p_process_io_info_tmp->high_prio_io_count ++;
- p_process_io_info_tmp->dispatch_io_count++;
- }
- else非高优先级IO
- {
- p_process_io_info_tmp->high_not_prio_io_count ++;
- if(bfqd->bfq_high_io_prio_mode)
- {
- //在 bfq_high_io_prio_mode 非0时间的5s内,如果遇到非high prio io,并且驱动队列IO个数大于限制,则把不派发该IO,而是临时添加到bfq_high_prio_tmp_list链表
- if((bfqd->rq_in_driver >= 16) && (bfqd->bfq_high_prio_tmp_list_rq_count < 100)){
- //把rq从原有链表删掉并把rq移动到bfq_high_prio_tmp_list链表尾,派发时是从bfq_high_prio_tmp_list链表头取出rq,保证先到先派发
- list_add_tail(&rq->queuelist,&bfqd->bfq_high_prio_tmp_list);
- bfqd->bfq_high_prio_tmp_list_rq_count ++;
- p_process_io_info_tmp->block_io_count ++;
- ///bug就出在这里,这里的rq添加到 bfq_high_prio_tmp_list 链表后,本次就不应该再派发了!!!!!!!!!但是却 goto exit1 ,该函数return rq返回该rq并派发了!!!!!!!!!!!!正确的做法是rq = NULL,赋值rq为NULL
- goto exit1;
- }
- }
- }
- }
- /*如果 bfq_high_prio_tmp_list 链表上有rq要派发,不执行这里的rq_in_driver++,在下边的exit那里会执行,当echo 0 >/sys/block/sdb/process_high_io_prio 置1再置0后,这个if判断就起作用了。没这个判断,这里会bfqd->rq_in_driver++,下边的if里再bfqd->rq_in_driver++,导致rq_in_driver泄漏*/
- if((rq->rq_flags & RQF_HIGH_PRIO) || list_empty(&bfqd->bfq_high_prio_tmp_list)){
- inc_in_driver_start_rq:
- bfqd->rq_in_driver++;
- start_rq:
- rq->rq_flags |= RQF_STARTED;
- }
- }
- exit:
- //1:如果是高优先级IO该if不成立,直接跳过。 2:如果非高优先级IO,则把rq添加到bfq_high_prio_tmp_list尾,从链表头选一个rq派发 3:如果rq是NULL,则也从bfq_high_prio_tmp_list选一个rq派发
- if(!direct_dispatch && ((rq && !(rq->rq_flags & RQF_HIGH_PRIO)) || !rq)){
- if(!list_empty(&bfqd->bfq_high_prio_tmp_list)){
- if(rq){
- list_add_tail(&rq->queuelist,&bfqd->bfq_high_prio_tmp_list);
- bfqd->bfq_high_prio_tmp_list_rq_count ++;
- if(p_process_io_info_tmp)
- p_process_io_info_tmp->block_io_count2++;
- }
- rq = list_first_entry(&bfqd->bfq_high_prio_tmp_list, struct request, queuelist);
- list_del_init(&rq->queuelist);
- bfqd->bfq_high_prio_tmp_list_rq_count --;
- bfqd->rq_in_driver++;
- rq->rq_flags |= RQF_STARTED;
- }
- }
- exit1:
- return rq;
- }
问题就出在红色代码goto exit1哪里,那里的rq添加到 bfq_high_prio_tmp_list 链表后,本次就不应该再派发了,但是却 goto exit1 ,该函数return rq返回该rq并派发了。后续再从bfq_high_prio_tmp_list 链表链表取出该rq,就会导致rq重复派发了。解决方法很简单,先rq = NULL再goto exit1,这样就避免第一次派发该rq了。
2:派发IO时遇到卡死
2.1 因bfq_has_work()返回false导致一直卡死
上一个问题解决了,新的问题又来了。启动fio压测竟然卡死了,kill -9 fio进程也不行。系统有很多D进程,启动crash工具看下D进程信息
- crash> ps -m | grep UN
- [0 00:06:40.712] [UN] PID: 2767 TASK: ffff8cb3ff450000 CPU: 0 COMMAND: "fio"
- [0 00:06:40.718] [UN] PID: 2780 TASK: ffff8cb3c9d5c740 CPU: 3 COMMAND: "fio"
- [0 00:06:40.719] [UN] PID: 2773 TASK: ffff8cb3c9d317c0 CPU: 2 COMMAND: "fio"
- [0 00:06:40.727] [UN] PID: 2769 TASK: ffff8cb3c9d0df00 CPU: 3 COMMAND: "fio"
- [0 00:06:40.731] [UN] PID: 2778 TASK: ffff8cb3c9d5df00 CPU: 3 COMMAND: "fio"
- [0 00:06:40.735] [UN] PID: 2772 TASK: ffff8cb3c9d08000 CPU: 3 COMMAND: "fio"
- [0 00:06:40.738] [UN] PID: 2775 TASK: ffff8cb3c9d32f80 CPU: 3 COMMAND: "fio"
- [0 00:06:40.742] [UN] PID: 2770 TASK: ffff8cb3c9d0af80 CPU: 3 COMMAND: "fio"
- [0 00:06:40.744] [UN] PID: 2768 TASK: ffff8cb3c9d097c0 CPU: 2 COMMAND: "fio"
- [0 00:06:40.757] [UN] PID: 2777 TASK: ffff8cb3c9d30000 CPU: 2 COMMAND: "fio"
- [0 00:06:40.768] [UN] PID: 2782 TASK: ffff8cb3c9d597c0 CPU: 3 COMMAND: "fio"
- [0 00:06:40.769] [UN] PID: 2764 TASK: ffff8cb3ff454740 CPU: 0 COMMAND: "fio"
看下栈回溯
- crash> bt 2764
- PID: 2764 TASK: ffff8cb3ff454740 CPU: 0 COMMAND: "fio"
- #0 [ffffb2348279bb70] __schedule at ffffffffa84c8826
- #1 [ffffb2348279bc08] schedule at ffffffffa84c8cb8
- #2 [ffffb2348279bc18] rwsem_down_write_slowpath at ffffffffa7d105ed
- #3 [ffffb2348279bc90] bfq_has_work at ffffffffc08054d2 [bfq]
- #4 [ffffb2348279bca0] _cond_resched at ffffffffa84c8d95
- #5 [ffffb2348279bcd8] ext4_file_write_iter at ffffffffc08c29bb [ext4]
- #6 [ffffb2348279bd38] aio_write at ffffffffa7f31206
- #7 [ffffb2348279be40] io_submit_one at ffffffffa7f31581
- #8 [ffffb2348279beb8] __x64_sys_io_submit at ffffffffa7f31b82
- #9 [ffffb2348279bf38] do_syscall_64 at ffffffffa7c0419b
- #10 [ffffb2348279bf50] entry_SYSCALL_64_after_hwframe at ffffffffa86000ad
- crash> bt 2780
- PID: 2780 TASK: ffff8cb3c9d5c740 CPU: 3 COMMAND: "fio"
- #0 [ffffb23482983b30] __schedule at ffffffffa84c8826
- #1 [ffffb23482983bc8] schedule at ffffffffa84c8cb8
- #2 [ffffb23482983bd8] rwsem_down_read_slowpath at ffffffffa84cbd05
- #3 [ffffb23482983c88] ext4_direct_IO at ffffffffc08d6e5d [ext4]
- #4 [ffffb23482983cf0] generic_file_read_iter at ffffffffa7e2da7f
- #5 [ffffb23482983d38] aio_read at ffffffffa7f313a5
- #6 [ffffb23482983e40] io_submit_one at ffffffffa7f3165b
- #7 [ffffb23482983eb8] __x64_sys_io_submit at ffffffffa7f31b82
- #8 [ffffb23482983f38] do_syscall_64 at ffffffffa7c0419b
- #9 [ffffb23482983f50] entry_SYSCALL_64_after_hwframe at ffffffffa86000ad
- crash> bt 2776
- PID: 2776 TASK: ffff8cb3c9d34740 CPU: 2 COMMAND: "fio"
- #0 [ffffb23482953958] __schedule at ffffffffa84c8826
- #1 [ffffb234829539f0] schedule at ffffffffa84c8cb8
- #2 [ffffb23482953a00] io_schedule at ffffffffa84c90d2
- #3 [ffffb23482953a10] bit_wait_io at ffffffffa84c94dd
- #4 [ffffb23482953a20] __wait_on_bit_lock at ffffffffa84c934d
- #5 [ffffb23482953a58] out_of_line_wait_on_bit_lock at ffffffffa84c9421
- #6 [ffffb23482953aa8] do_get_write_access at ffffffffc083ae68 [jbd2]
- #7 [ffffb23482953b08] jbd2_journal_get_write_access at ffffffffc083b10c [jbd2]
- #8 [ffffb23482953b28] __ext4_journal_get_write_access at ffffffffc08b63f6 [ext4]
- #9 [ffffb23482953b58] ext4_reserve_inode_write at ffffffffc08d35a6 [ext4]
- #10 [ffffb23482953b80] ext4_mark_inode_dirty at ffffffffc08d37d1 [ext4]
- #11 [ffffb23482953bf0] ext4_dirty_inode at ffffffffc08d8a15 [ext4]
- #12 [ffffb23482953c08] __mark_inode_dirty at ffffffffa7f0aa6a
- #13 [ffffb23482953c40] generic_update_time at ffffffffa7ef76e6
- #14 [ffffb23482953c50] file_update_time at ffffffffa7ef7b01
- #15 [ffffb23482953c98] __generic_file_write_iter at ffffffffa7e2dd38
- #16 [ffffb23482953cd8] ext4_file_write_iter at ffffffffc08c2761 [ext4]
- #17 [ffffb23482953d38] aio_write at ffffffffa7f31206
- #18 [ffffb23482953e40] io_submit_one at ffffffffa7f31581
- #19 [ffffb23482953eb8] __x64_sys_io_submit at ffffffffa7f31b82
- #20 [ffffb23482953f38] do_syscall_64 at ffffffffa7c0419b
- #21 [ffffb23482953f50] entry_SYSCALL_64_after_hwframe at ffffffffa86000ad
有几个fio进程的栈回溯竟然是bfq_has_work,这里边没有调用什么锁呀?很奇怪,难道卡死根源跟bfq_has_work有关。看下它的源码:
- //返回0则blk_mq_do_dispatch_sched()中就无法派发继续派发IO了
- static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
- {
- struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
-
- //list_empty_careful(&bfqd->dispatch)返回NULL,说明该链表上有rq派发,返回1
- return !list_empty_careful(&bfqd->dispatch) ||
- //bfq_tot_busy_queues(bfqd)大于0说明还有active bfqq,则派发该bfqq上的rq,此时返回1
- bfq_tot_busy_queues(bfqd) > 0;
- }
一般是派发blk-mq派发blk_mq_do_dispatch_sched()函数中会调用bfq_has_work()函数,源码如下:
- static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
- {
- struct request_queue *q = hctx->queue;
- struct elevator_queue *e = q->elevator;
- LIST_HEAD(rq_list);
- int ret = 0;
- do {
- struct request *rq;
- //调用bfq_has_work
- if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
- break;
- if (!list_empty_careful(&hctx->dispatch)) {
- ret = -EAGAIN;
- break;
- }
- if (!blk_mq_get_dispatch_budget(hctx))
- break;
- //调用bfq调度器函数 bfq_dispatch_request
- rq = e->type->ops.dispatch_request(hctx);
- if (!rq) {
- blk_mq_put_dispatch_budget(hctx);
- blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
- break;
- }
- list_add(&rq->queuelist, &rq_list);
- /*取出rq_list链表上的req派发给磁盘驱动,如果因驱动队列繁忙或者nvme硬件繁忙导致派发失败,则把req添加hctx->dispatch等稍后派发遇到req派发失败返回false,退出while循环*/
- } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
- return ret;
- }
当 bfq_has_work 返回0原本说明bfq没有IO可派发了,blk_mq_do_dispatch_sched()就不再派发IO了。但是我对bfq派发IO的bfq_dispatch_request函数做了优化,增加了一个 bfq_high_prio_tmp_list链表保存普通优先级的rq。当bfq空闲时,bfq_tot_busy_queues(bfqd)返回0,但是bfq_high_prio_tmp_list链表上还有rq要派发,此时还需要继续派发rq。fio暂存在 bfq_high_prio_tmp_list链表上的rq得不到派发,fio进程就卡主,不能再派发新rq,除非老的rq派发完成。简单说,这种情况下,要想判断bfq是否还有rq没派发,必须判断bfq_high_prio_tmp_list链表上是否有IO。于是在bfq_has_work()函数中添加如下红色代码:
- static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
- {
- struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
- return !list_empty_careful(&bfqd->dispatch) ||
- !list_empty(&bfqd->bfq_high_prio_tmp_list) ||
- bfq_tot_busy_queues(bfqd) > 0;
- }
ok,这个问题解决了,但是新的问题又来了。
2.2 bfqq->dispatched泄漏导致的卡死
这个问题的表现也是派发IO的fio或者cat进程卡死,同样也是有很多D进程,ps -eLlf | grep fio |awk '{print $6}' | while read line;do echo "*********";cat /proc/$line/stack;done 看下栈回溯,主要是以下两类:
- *********
- [<0>] rwsem_down_write_slowpath+0x32d/0x4e0
- [<0>] ext4_file_write_iter+0x3cb/0x3e0 [ext4]
- [<0>] aio_write+0xf6/0x1c0
- [<0>] io_submit_one+0x131/0x3c0
- [<0>] __x64_sys_io_submit+0xa2/0x180
- [<0>] do_syscall_64+0x5b/0x1a0
- [<0>] entry_SYSCALL_64_after_hwframe+0x65/0xca
- *********
- [<0>] blk_mq_get_tag+0x119/0x270
- [<0>] __blk_mq_alloc_request+0xb1/0x100
- [<0>] blk_mq_make_request+0x14e/0x5d0
- [<0>] generic_make_request+0xcf/0x310
- [<0>] submit_bio+0x3c/0x160
- [<0>] do_blockdev_direct_IO+0x21e6/0x2e60
- [<0>] ext4_direct_IO+0x247/0x730 [ext4]
- [<0>] generic_file_direct_write+0x93/0x160
- [<0>] __generic_file_write_iter+0xb7/0x1c0
- [<0>] ext4_file_write_iter+0x171/0x3e0 [ext4]
- [<0>] aio_write+0xf6/0x1c0
- [<0>] io_submit_one+0x131/0x3c0
- [<0>] __x64_sys_io_submit+0xa2/0x180
- [<0>] do_syscall_64+0x5b/0x1a0
- [<0>] entry_SYSCALL_64_after_hwframe+0x65/0xca
分析根源应该是有进程 __blk_mq_alloc_request->blk_mq_get_tag 分配tag失败导致的。
在派发IO的__bfq_dispatch_request()函数最后添加如下红色代码调试信息。
- static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
- {
- exit:
- ..............
- printk("5:%s %s %d dispatch rq:0x%llx bfq_high_io_prio_count:%d rq_in_driver:%d\n",__func__,current->comm,current->pid,(u64)rq,bfqd->bfq_high_io_prio_count,bfqd->rq_in_driver);
- return rq;
- }
卡死时刷屏打印如下信息:
5:__bfq_dispatch_request kworker/3:1H 497 dispatch rq:0x0 bfq_high_io_prio_count:0 rq_in_driver:0
这是blk-mq驱动了内核线程在疯狂的派发rq,但是派发的rq一直是NULL。正常情况应该会退出派发的!
看下 497 派发IO的函数流程,为什么会一直派发rq呢?执行这个命令stap --all-modules -ve 'probe module("bfq").function("bfq_dispatch_request") {printf("%s %d\n",execname(),tid()) print_backtrace()}',刷屏打印
- kworker/3:1H 497
- 0xffffffffc06b3950 : bfq_dispatch_request+0x0/0x9f0 [bfq]
- 0xffffffffa480f385 : blk_mq_do_dispatch_sched+0xc5/0x160 [kernel]
- 0xffffffffa480feb9 : __blk_mq_sched_dispatch_requests+0x189/0x1e0 [kernel]
- 0xffffffffa480ff40 : blk_mq_sched_dispatch_requests+0x30/0x60 [kernel]
- 0xffffffffa48076a1 : __blk_mq_run_hw_queue+0x51/0xd0 [kernel]
- 0xffffffffa44d3477 : process_one_work+0x1a7/0x360 [kernel]
- 0xffffffffa44d3b40 : worker_thread+0x30/0x390 [kernel]
- 0xffffffffa44d9502 : kthread+0x112/0x130 [kernel]
- 0xffffffffa4e00255 : ret_from_fork+0x35/0x40 [kernel]
- 0xffffffffa4e00255 : ret_from_fork+0x35/0x40 [kernel] (inexact)
为什么 kworker/3:1H 进程会刷屏执行 __blk_mq_run_hw_queue 而最终疯狂派发 rq 呢?继续执行stap --all-modules -ve 'probe kernel.function("blk_mq_do_dispatch_sched").return {if(tid()== 497) {printf("%s %d\n",execname(),tid()) print_backtrace()}}'调试,刷屏打印:
- kworker/3:1H 497
- Returning from: 0xffffffffa480f2c0 : blk_mq_do_dispatch_sched+0x0/0x160 [kernel]
- Returning to : 0xffffffffa480feb9 : __blk_mq_sched_dispatch_requests+0x189/0x1e0 [kernel]
- 0xffffffffa480ff40 : blk_mq_sched_dispatch_requests+0x30/0x60 [kernel]
- 0xffffffffa48076a1 : __blk_mq_run_hw_queue+0x51/0xd0 [kernel]
- 0xffffffffa44d3477 : process_one_work+0x1a7/0x360 [kernel]
- 0xffffffffa44d3b40 : worker_thread+0x30/0x390 [kernel]
- 0xffffffffa44d9502 : kthread+0x112/0x130 [kernel]
- 0xffffffffa4e00255 : ret_from_fork+0x35/0x40 [kernel]
- 0xffffffffa4e00255 : ret_from_fork+0x35/0x40 [kernel] (inexact)
源码分析这是blk-mq驱动启动的内核线程,而启动的根源在blk_mq_run_work_fn()函数,继续用如下命令调试stap --all-modules -ve 'probe kernel.function("blk_mq_run_work_fn") {if(tid()== 497) {printf("%s %d\n",execname(),tid()) print_backtrace()}}',刷屏打印:
- kworker/3:1H 497
- 0xffffffffa4807720 : blk_mq_run_work_fn+0x0/0x20 [kernel]
- 0xffffffffa44d3477 : process_one_work+0x1a7/0x360 [kernel]
- 0xffffffffa44d3b40 : worker_thread+0x30/0x390 [kernel]
- 0xffffffffa44d9502 : kthread+0x112/0x130 [kernel]
- 0xffffffffa4e00255 : ret_from_fork+0x35/0x40 [kernel]
- 0xffffffffa4e00255 : ret_from_fork+0x35/0x40 [kernel] (inexact)
这个打印验证了想法。并且,分析可能性最大是__blk_mq_delay_run_hw_queue函数里执行的__blk_mq_run_hw_queue函数。用如下命令验证stap --all-modules -ve 'probe kernel.function("__blk_mq_delay_run_hw_queue") {{printf("%s %d\n",execname(),tid()) print_backtrace()}}',刷屏打印:
- kworker/3:1H 497
- 0xffffffffa4807e20 : __blk_mq_delay_run_hw_queue+0x0/0x160 [kernel]
- 0xffffffffa4807fd8 : blk_mq_delay_run_hw_queues+0x38/0x50 [kernel]
- 0xffffffffa480f412 : blk_mq_do_dispatch_sched+0x152/0x160 [kernel]
- 0xffffffffa480feb9 : __blk_mq_sched_dispatch_requests+0x189/0x1e0 [kernel]
- 0xffffffffa480ff40 : blk_mq_sched_dispatch_requests+0x30/0x60 [kernel]
- 0xffffffffa48076a1 : __blk_mq_run_hw_queue+0x51/0xd0 [kernel]
- 0xffffffffa44d3477 : process_one_work+0x1a7/0x360 [kernel]
- 0xffffffffa44d3b40 : worker_thread+0x30/0x390 [kernel]
- 0xffffffffa44d9502 : kthread+0x112/0x130 [kernel]
- 0xffffffffa4e00255 : ret_from_fork+0x35/0x40 [kernel]
- 0xffffffffa4e00255 : ret_from_fork+0x35/0x40 [kernel] (inexact)
综合这些调试信息,基本可以确定:blk_mq_do_dispatch_sched()函数因为派发的rq 是NULL,而频繁执行 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY)->blk_mq_delay_run_hw_queue->__blk_mq_delay_run_hw_queue->kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,msecs_to_jiffies(msecs)) 而再次触发 mq 异步派发进程,就是 kworker/3:1H497 进程。这个逻辑好像没问题,但是为什么会频繁触发 blk-mq 异步派发进程 kworker/3:1H 497 呢?看下blk_mq_do_dispatch_sched()函数派发IO的代码:
- static int blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
- {
- struct request_queue *q = hctx->queue;
- struct elevator_queue *e = q->elevator;
- LIST_HEAD(rq_list);
- int ret = 0;
- do {
- struct request *rq;
- if (e->type->ops.has_work && !e->type->ops.has_work(hctx))//bfq_has_work
- break;
- if (!list_empty_careful(&hctx->dispatch)) {
- ret = -EAGAIN;
- break;
- }
- if (!blk_mq_get_dispatch_budget(hctx))
- break;
- rq = e->type->ops.dispatch_request(hctx);//调用bfq调度器函数 bfq_dispatch_request
- if (!rq) {
- //如果bfq_dispatch_request返回rq是NULL,则执行blk_mq_delay_run_hw_queues()启动blk-mq异步派发IO内核线程
- blk_mq_put_dispatch_budget(hctx);
- blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY);
- break;
- }
- list_add(&rq->queuelist, &rq_list);
- /*取出rq_list链表上的req派发给磁盘驱动,如果因驱动队列繁忙或者nvme硬件繁忙导致派发失败,则把rq添加hctx->dispatch等稍后派发遇到rq派发失败返回false,退出while循环*/
- } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
- return ret;
- }
跟踪下bfq_has_work()函数,stap --all-modules -ve 'probe module("bfq").function("bfq_has_work").return {{printf("%s %d %d bfqd:0x%x\n",execname(),tid(),$return,$hctx->queue->elevator->elevator_data)}}',刷屏打印如下:
- kworker/3:1H 497 1 bfqd:0xffffa0657f07e800
- kworker/3:1H 497 1 bfqd:0xffffa0657f07e800
是在没什么思路,那就把bfq算法核心数据bfqq或bfqd结构体成员信息打印出来,看能否发现什么异常!启动crash,
- crash> bfq_data 0xffffa0657f07e800
- struct bfq_data {
- queue = 0xffffa0659740eda8,
- dispatch = {
- next = 0xffffa0657f07e808,
- prev = 0xffffa0657f07e808
- },
- ...........
- bfq_high_prio_tmp_list = {
- next = 0xffffa0657f07ec28,
- prev = 0xffffa0657f07ec28
- },
这两个暂存IO的链表都是空的,那bfq_has_work函数返回1只能可能是 bfq_tot_busy_queues 返回true 了,测试一下果然是。stap --all-modules -ve 'probe module("bfq").function("bfq_tot_busy_queues").return {{printf("%s %d %d\n",execname(),tid(),$return)}}'刷屏打印:
- kworker/3:1H 497 21
- kworker/3:1H 497 21
- kworker/3:1H 497 21
- kworker/3:1H 497 21
- kworker/3:1H 497 21
- kworker/3:1H 497 21
此时,怀疑有很多IO的派发都有问题。我在内核检测哪些rq添加到bfq算法队列后30s还没传输完成,结果打印:
- [10168.410008] rq:0xffffa0659b96e110 long time do not dispatch
- [10168.410008] rq:0xffffa0659b95f790 long time do not dispatch
- [10168.410008] rq:0xffffa0659b950010 long time do not dispatch
- [10168.410009] rq:0xffffa0659b968350 long time do not dispatch
- [10168.410009] rq:0xffffa065974b8350 long time do not dispatch
- [10168.410009] rq:0xffffa0659b958e90 long time do not dispatch
- [10168.411852] 5:__bfq_dispatch_request kworker/3:1H 497 dispatch rq:0x0 bfq_high_io_prio_count:0 rq_in_driver:0
- [10168.415764] 5:__bfq_dispatch_request kworker/3:1H 497 dispatch rq:0x0 bfq_high_io_prio_count:0 rq_in_driver:0
- [10168.419817] 5:__bfq_dispatch_request kworker/3:1H 497 dispatch rq:0x0 bfq_high_io_prio_count:0 rq_in_driver:0
- [10168.423622] 5:__bfq_dispatch_request kworker/3:1H 497 dispatch rq:0x0 bfq_high_io_prio_count:0 rq_in_driver:0
- [10168.427652] 5:__bfq_dispatch_request kworker/3:1H 497 dispatch rq:0x0 bfq_high_io_prio_count:0 rq_in_driver:0
有时一个很大的疑问,还是重点看下 __bfq_dispatch_request 函数为什么派发的rq总是0把!怀疑 里边返回的 bfq_select_queue 有问题。因为__bfq_dispatch_request函数中是先执行bfq_select_queue选择一个bfqq,再从bfqq中跳一个rq派发,是否bfq_select_queue选择的bfqq就有问题呢?当有很多怀疑点时,就抓住核心的疑问穷追不舍!
用stap --all-modules -ve 'probe module("bfq").function("bfq_select_queue").return {{printf("%s %d %d\n",execname(),tid(),$return)}}'这个命令调试,打印
- kworker/3:1H 497 0
- kworker/3:1H 497 0
- kworker/3:1H 497 0
- kworker/3:1H 497 0
- kworker/3:1H 497 0
- kworker/3:1H 497 0
- kworker/3:1H 497 0
- kworker/3:1H 497 0
果然 bfq_select_queue 返回的bfqq 有问题。那就通过bfqd->in_service_queue看下当前正在派发IO的bfqq是哪个!前文调试已经知道bfqd指针是0xffffa0657f07e800。
- crash> bfq_data 0xffffa0657f07e800 | grep in_service_queue
- in_service_queue = 0xffffa06597e1c000,
- crash> bfq_queue 0xffffa06597e1c000 | grep pid
- pid = 1272,
- crash> bt 1272
- PID: 1272 TASK: ffffa065a692df00 CPU: 0 COMMAND: "jbd2/sdb-8"
- #0 [ffffbcc1c21efa48] __schedule at ffffffffa4cc8826
- #1 [ffffbcc1c21efae0] schedule at ffffffffa4cc8cb8
- #2 [ffffbcc1c21efaf0] io_schedule at ffffffffa4cc90d2
- #3 [ffffbcc1c21efb00] blk_mq_get_tag at ffffffffa480dca9
- #4 [ffffbcc1c21efb78] __blk_mq_alloc_request at ffffffffa4807ba1
- #5 [ffffbcc1c21efb98] blk_mq_make_request at ffffffffa480ab5e
- #6 [ffffbcc1c21efc28] generic_make_request at ffffffffa47fe85f
- #7 [ffffbcc1c21efc80] submit_bio at ffffffffa47feadc
- #8 [ffffbcc1c21efcc0] submit_bh_wbc at ffffffffa471673a
- #9 [ffffbcc1c21efcf8] jbd2_journal_commit_transaction at ffffffffc06e28a4 [jbd2]
- #10 [ffffbcc1c21efea0] kjournald2 at ffffffffc06e792d [jbd2]
- #11 [ffffbcc1c21eff10] kthread at ffffffffa44d9502
- #12 [ffffbcc1c21eff50] ret_from_fork at ffffffffa4e00255
当前正在派发rq的bfqq的进程竟然卡死了!继续看下bfq_select_queue函数里有哪些疑问?看下他的函数源码:
- static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
- {
- ................
- if (bfq_bfqq_wait_request(bfqq) ||
- (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
- ..........
- //如果进程有异步bfqq,则取出这个异步bfqq
- if (async_bfqq &&
- icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
- bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
- bfq_bfqq_budget_left(async_bfqq))
- bfqq = bfqq->bic->bfqq[0];
- else if (bfq_bfqq_has_waker(bfqq) &&
- bfq_bfqq_busy(bfqq->waker_bfqq) &&
- bfqq->next_rq &&
- bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
- bfqq->waker_bfqq) <=
- bfq_bfqq_budget_left(bfqq->waker_bfqq)
- )
- //取出bfqq->waker_bfqq
- bfqq = bfqq->waker_bfqq;
- //bfqd->in_service_queue这个bfqq绑定的进程空闲时没有大量连续快速向bfqq->sort_list插入IO请求特性
- else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
- //bfqd->in_service_queue这个bfqq没有权重提升
- (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
- //bfqd->in_service_queue这个bfqq绑定的进程在派发IO请求时,没有快速插入IO请求的特性
- !bfq_bfqq_has_short_ttime(bfqq)))
- /*该if成立说明bfqd->in_service_queue这个bfqq初步符合被inject bfqq抢占的条件,在bfq_choose_bfqq_for_injection()里,如果遍历st->active tree上的bfqq,符合bfqd->rq_in_driver < limit条件,就返回这个bfqq,抢占bfqd->in_service_queue*/
- bfqq = bfq_choose_bfqq_for_injection(bfqd);
- else
- bfqq = NULL;
- goto keep_queue;
- }
- expire:
- //bfqq过期失效
- bfq_bfqq_expire(bfqd, bfqq, false, reason);
- new_queue:
- bfqq = bfq_set_in_service_queue(bfqd);
- if (bfqq) {
- //找到bfqq则goto check_queue分支
- goto check_queue;
- }
- keep_queue:
- return bfqq;
- }
用stap --all-modules -ve 'probe module("bfq").function("bfq_bfqq_expire") {{printf("%s %d 0x%x\n",execname(),tid(),$bfqq)}}'看下是否执行了bfq_bfqq_expire()函数,什么打印都没有。再用stap --all-modules -ve 'probe module("bfq").function("idling_boosts_thr_without_issues").return {{printf("%s %d 0x%x\n",execname(),tid(),$return)}}'看下是否调用了idling_boosts_thr_without_issues函数,刷屏打印:
- kworker/3:1H 497 0x0
- kworker/3:1H 497 0x0
- kworker/3:1H 497 0x0
- kworker/3:1H 497 0x0
- kworker/3:1H 497 0x0
看来执行到了if (!idling_boosts_thr_without_issues(bfqd, bfqq)…)那个if判断,我认为这个if不成立,而是执行了else分支bfqq = NULL,然后goto keep_queue返回bfqq = NULL,这样就导致bfq_select_queue()函数一直返回NULL呀。怎么验证,启动crash工具,前文知道当前的bfqq指针是0xffffa06597e1c000:
- crash> bfq_queue 0xffffa06597e1c000 | grep wr_coeff
- wr_coeff = 30,
- crash> bfq_data 0xffffa0657f07e800 | grep wr_busy_queues
- wr_busy_queues = 1,
- crash> bfq_queue 0xffffa06597e1c000 -x | grep flags
- flags = 0xf2,
- crash> bfq_queue 0xffffa06597e1c000 | grep dispatched
- dispatched = 2, //bfqq
BFQQF_has_short_ttime 是bit5 ,而现在 bfqq:0xffffa06597e1c000 的 flags bit5是1,因此 if(!idling_boosts_thr_without_issues(bfqd, bfqq) &&(bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||!bfq_bfqq_has_short_ttime(bfqq))) 不成立,因此 否else 分支, bfqq=NULL,这就是 bfq_select_queue 返回的bfqq是NULL。神奇了,为什么会这样呢?
难道我在bfqq添加的代码影响到了 bfqq 算法?那段代码要成立,得先有更外边的 if (bfq_bfqq_wait_request(bfqq) || (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) 成立!而BFQQF_wait_request 是bit2,但bfqq的flags的bit2是0。bfqq->dispatched 是2,那应该是这个导致if ((bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) 成立。
验证一下 bfq_better_to_idle()返回true,stap --all-modules -ve 'probe module("bfq").function("bfq_better_to_idle").return {{printf("%s %d bfqq:0x%x 0x%x\n",execname(),tid(),$bfqq,$return)}}',刷屏打印:
- kworker/3:1H 497 bfqq:0xffffa06597e1c000 0x1
- kworker/3:1H 497 bfqq:0xffffa06597e1c000 0x1
- kworker/3:1H 497 bfqq:0xffffa06597e1c000 0x1
- kworker/3:1H 497 bfqq:0xffffa06597e1c000 0x1
- kworker/3:1H 497 bfqq:0xffffa06597e1c000 0x1
- kworker/3:1H 497 bfqq:0xffffa06597e1c000 0x1
- kworker/3:1H 497 bfqq:0xffffa06597e1c000 0x1
看来,如果 bfqq:0xffffa06597e1c000 的 dispatched 是0,那if就不会成立了吗。但事实是bfqq->dispatched 始终是2!
看来问题的根源是 bfqq:0xffffa06597e1c000 的 dispatched 始终是2,大于0 呀?神奇了,难道 我的代码导致 bfqq:0xffffa06597e1c000 的 dispatched 泄漏了,导致始终大于0?仔细分析我在__bfq_dispatch_request()中添加的代码,果然发现了问题,如下红色代码:
- static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
- {
- ....................
- rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
-
- if (rq) {
- if(bfqd->queue->high_io_prio_enable)
- {
- if(rq->rq_flags & RQF_HIGH_PRIO){//高优先级IO
- if(bfqd->bfq_high_io_prio_mode == 0){
- bfqd->bfq_high_io_prio_mode = 1;
- hrtimer_start(&bfqd->bfq_high_prio_timer, ms_to_ktime(5000),HRTIMER_MODE_REL);
- }
- }
- else//非高优先级IO
- {
- if(bfqd->bfq_high_io_prio_mode)
- {
- //在 bfq_high_io_prio_mode 非0时间的5s内,如果遇到非high prio io,并且驱动队列IO个数大于限制,则把不派发该IO,而是临时添加到bfq_high_prio_tmp_list链表
- if(bfqd->rq_in_driver >= HIGH_PRIO_IO_LIMIT){
- list_add_tail(&rq->queuelist,&bfqd->bfq_high_prio_tmp_list);
- bfqq->dispatched --;
- bfqd->bfq_high_io_prio_count ++;
- return NULL;
- }
- }
- }
- }
- if(list_empty(&bfqd->bfq_high_prio_tmp_list)){
- inc_in_driver_start_rq:
- bfqd->rq_in_driver++;
- start_rq:
- rq->rq_flags |= RQF_STARTED;
- }
- }
- exit:
- //1:如果是高优先级IO该if不成立,直接跳过。 2:如果非高优先级IO,则把rq添加到bfq_high_prio_tmp_list尾,从链表头选一个rq派发 3:如果rq是NULL,则也从bfq_high_prio_tmp_list选一个rq派发
- if(((rq && !(rq->rq_flags & RQF_HIGH_PRIO)) || !rq)){
- if(!list_empty(&bfqd->bfq_high_prio_tmp_list)){
- if(rq){
- list_add_tail(&rq->queuelist,&bfqd->bfq_high_prio_tmp_list);
- bfqq->dispatched --;
- bfqd->bfq_high_io_prio_count ++;
- }
- rq = list_first_entry(&bfqd->bfq_high_prio_tmp_list, struct request, queuelist);
- list_del_init(&rq->queuelist);
- bfqd->bfq_high_io_prio_count --;
- bfqq = RQ_BFQQ(rq);
- if(bfqq)
- bfqq->dispatched++;
- bfqd->rq_in_driver++;
- rq->rq_flags |= RQF_STARTED;
- }
- }
- return rq;
- }
如果rq有 RQF_HIGH_PRIO属性,rq在派发时先有__bfq_dispatch_request->bfq_dispatch_rq_from_bfqq()默认的bfqq->dispatched++。回到__bfq_dispatch_request函数,如果 bfq_high_prio_tmp_list 链表空,那if(!list_empty(&bfqd->bfq_high_prio_tmp_list))不成立,就不会执行 rq->rq_flags |= RQF_STARTED 。再下边的 if(((rq && !(rq->rq_flags & RQF_HIGH_PRIO)) || !rq)) 也不成立。于是再次错过了rq->rq_flags |= RQF_STARTED。
等rq传输完成,执行到bfq_finish_requeue_request函数
- static void bfq_finish_requeue_request(struct request *rq)
- {
- //由传输完成的IO请求rq得到bfqq
- struct bfq_queue *bfqq = RQ_BFQQ(rq);
- struct bfq_data *bfqd;
-
- if (likely(rq->rq_flags & RQF_STARTED)) {
- unsigned long flags;
- spin_lock_irqsave(&bfqd->lock, flags);
- if (rq == bfqd->waited_rq)
- bfq_update_inject_limit(bfqd, bfqq);
- //IO传输完成重点执行的函数在这里
- bfq_completed_request(bfqq, bfqd);
- bfq_finish_requeue_request_body(bfqq);
- spin_unlock_irqrestore(&bfqd->lock, flags);
- }
- }
- static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
- {
- u64 now_ns;
- u32 delta_us;
- bfq_update_hw_tag(bfqd);
- //已经派发但是还没传输完成的reqIO请求个数
- bfqd->rq_in_driver--;
- //还没有传输完成的IO请求个数,为0表示所有的IO请求都传输完成了,跟bfqd->rq_in_driver类似
- bfqq->dispatched--;
- }
因为 rq 没有 RQF_STARTED 标记,导致没有执行bfqq->dispatched--,这就导致bfqq->dispatched泄漏了。解决方法很简单,rq 有 RQF_HIGH_PRIO属性标记并且 bfq_high_prio_tmp_list 链表空时,也要执行 rq->rq_flags |= RQF_STARTED。把if(list_empty(&bfqd->bfq_high_prio_tmp_list))改成if((rq->rq_flags & RQF_HIGH_PRIO) || list_empty(&bfqd->bfq_high_prio_tmp_list))即可!
就是一个细节逻辑分析疏忽,导致了这么复杂的排查过程,服了!
最后,关于blk-mq内核派发rq的kworker/0:1H内核线程多了一层理解。blk_mq_do_dispatch_sched函数中,因为以后很多个rq暂存在 bfq_high_prio_tmp_list链表, if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) 不成立。于是执行 rq = e->type->ops.dispatch_request(hctx) 即 __bfq_dispatch_request()。
如果 进程在 执行__bfq_dispatch_request时,因为rq没有RQF_HIGH_PRIO属性,导致__bfq_dispatch_request返回NULL,即 rq = e->type->ops.dispatch_request(hctx) 返回NULL,那就执行 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY) ,在kworker/0:1H 内核线程延迟派发rq。然后2ms后再次执行 blk_mq_do_dispatch_sched,重复上述流程,直到bfq_high_prio_tmp_list链表上的rq全派发完。然后bfq_high_prio_tmp_list链表空,kworker/0:1H 线程最后一次执行 blk_mq_do_dispatch_sched(),bfq_has_work返回NULL,if (e->type->ops.has_work && !e->type->ops.has_work(hctx)) 成立,最终退出rq派发。
相当于我利用了 blk-mq的 blk_mq_delay_run_hw_queues(q, BLK_MQ_BUDGET_DELAY) 延迟派发的特性,从而保证没有进程执行 __blk_mq_sched_dispatch_requests->blk_mq_do_dispatch_sched->blk_mq_dispatch_rq_list 派发rq时,也可以由内核线程 kworker/0:1H 延迟派发完所有的rq。这样我就不用担心rq暂存到bfq_high_prio_tmp_list链表后,会导致这些rq无法被进程主动派发了!
3:bfqq->dispatched和rq暂存bfq_high_prio_tmp_list链表的深入分析
我在bfq添加的代码,有多处 执行 bfqq->dispatched -- 和 bfqq->dispatched ++。本身rq在rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq) 里已经执行 bfqq->dispatched ++。我在bfq添加的 bfqq->dispatched -- 和 bfqq->dispatched ++ 是否会影响bfq算法呢?我原本的意思是,rq如果 添加到 bfq_high_prio_tmp_list链表,那就bfqq->dispatched --,等rq真正派发时再 bfqq->dispatched ++。但是这样有问题,如果rq在bfq_high_prio_tmp_list链表停留时间过长,因为提前 bfqq->dispatched --,如果这是bfqq的最后一个rq,就相当于bfqq的所有rq全派发完成了。
但实际并没有,只是rq暂存在 bfq_high_prio_tmp_list链表而已。如果 bfqq->dispatched 是0了,那估计会影响bfqq过期失效,从st->active tree剔除。这样,等该bfqq暂存在 bfq_high_prio_tmp_list链表上的rq终于派发了,再 bfqq->dispatched ++。这样就有问题了,因为该bfqq可能已经被新进程拥有了!这样分析,我的代码里不应该 bfqq->dispatched ++ 和 bfqq->dispatched --。不对,分析错了。因为先有 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq) 里的 bfqq->dispatched ++,然后再有我的代码里的 bfqq->dispatched --,这就相当于该bfqq上的rq并没有派发呀,rq还保存在bfqq上,这样bfqq也不会过期失效的!!!!!!
但是,我的bfq代码是否可以放到 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq); 前边呢?因为 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq) 执行过后,相当于rq就从bfqq上的链表剔除了,而我把该rq长时间保存在 bfq_high_prio_tmp_list链表,可能会影响bfq算法呀。因为正常 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq) 执行过的rq很快就会传输成功呀。而我是把rq暂存在bfq_high_prio_tmp_list链表,可能要过一段时间才会传输完成。
并且 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq) 选中要派发的rq一定来自bfqq->next_rq ,并且还会执行 bfq_dispatch_rq_from_bfqq->bfq_bfqq_served 把rq传输消耗的配额累加到rq所属bfqq的entity->service,然后我把rq添加到bfq_high_prio_tmp_list链表。如果这个bfqq的配额正好消耗光了,那bfqq就会过期失效。等从bfq_high_prio_tmp_list链表再取出这个rq,rq所属的bfqq已经过期失效了,然后的代码里却 bfqq->dispatched++ 。然后派发给驱动,等rq传输完成,执行bfq_completed_request(),还要 bfqq->dispatched--。这样就会有问题了,因为bfqq已经过期失效了!
问题来了,rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq)从bfqq取出rq,然后把rq添加到bfq_high_prio_tmp_list链表后,rq和原属的bfqq要不要彻底脱离关系???不脱离关系,那rq在bfq_high_prio_tmp_list链表暂存时,bfqq可能因配额消耗光而失效。这样从bfq_high_prio_tmp_list链表取出该rq后,使用rq的bfqq已经过期失效了?不能再按照原流程处理了!那怎么办?rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq)从bfqq取出rq,然后把rq添加到bfq_high_prio_tmp_list链表后:先执行bfqq->dispatched--,这制造一个假象,这个rq传输完成了!因为正常bfqq->dispatched--就说明rq传输完成了。然后执行 rq->elv.priv[0] = NULL 和 rq->elv.priv[1] = NULL ,令rq所属的bfqq是NULL,这样rq和bfqq就脱离关系了!接着,从bfq_high_prio_tmp_list链表取出该rq后,不再执行bfqq->dispatched++,因为rq不再属于哪个 bfqq了,接着派发该rq。然后在该rq传输完成后,执行bfq_finish_requeue_request()函数,因rq所属bfqq是NULL,则直接返回,不会再执行bfq_completed_request()令bfqq->dispatched--了。
但是这个方案也有一个问题,因为正常情况,rq传输完成后,会执行 bfq_finish_requeue_request->bfq_completed_request(),更新很多bfqq参数,这些与bfq算法紧密相关。而我的bfq优化算法,一旦rq加入 bfq_high_prio_tmp_list链表,就要令rq所属bfqq是NULL,然后rq传输完成后就执行不了 bfq_finish_requeue_request->bfq_completed_request() 了,影响了bfqq参数更新,肯定会对bfq算法造成影响。左右为难,没有一个完美的解法。
不对,想来想去,还是有解法的!执行 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq) 后,然后执行我添加的bfq代码时,把rq添加到bfq_high_prio_tmp_list链表。但是把bfqq->dispatched++ 和 bfqq->dispatched-- 都去掉,其他代码不修改。之后 rq所属bfqq可能过期失效,从st->active tree 移动到 st->idle tree。但是该bfqq可能会被完全释放吗?不会,第一,bfqq所属的进程派发的rq,还有保存在bfq_high_prio_tmp_list链表,进程必须等这些rq派发完才会退出。我之前说添加到bfq_high_prio_tmp_list链表的rq的bfqq可能被释放,bfqq会被新的进程有用,这个说法是错误的。什么情况下bfqq才会被释放呢? 在 bfq_put_queue()函数释放bfqq,但是前提是 bfqq->ref 是0。每向bfqq插入一个rq则bfqq->ref ++,看来只有bfqq上的rq全派发完才有可能 bfqq->ref是0。然后才有概率 bfq_forget_entity()-> bfq_put_queue()中因 bfqq->ref为0 而释放掉 bfqq。因此,即便 bfqq 的rq有插入 bfq_high_prio_tmp_list 链表的,然后bfqq上的rq全派发完了,bfqq过期失效,也不会释放bfqq。应该是这样!
因此,我的分析:把 rq = bfq_dispatch_rq_from_bfqq(里边有bfqq->dispatched++ )上的rq插入bfq_high_prio_tmp_list链表后,不再 bfqq->dispatched--,就相当于该rq还是属于bfqq,只不过换了一个保存位置而已。只不过延迟派发给驱动而已。想想,即便没有我的代码,rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq) 选中的rq直接派发给驱动,在磁盘阵列驱动繁忙时,rq也是暂存在磁盘驱动队列,这个rq也无法直接派发给磁盘硬件。rq暂存在磁盘驱动队列,我的bfq代码是把rq暂存在 bfq_high_prio_tmp_list 链表,都是延迟派发,有什么区别呢?