Alexey Ivanov | cc01a9c | 2019-01-16 09:50:46 -0800 | [diff] [blame] | 1 | #!/usr/bin/python |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 2 | # @lint-avoid-python-3-compatibility-imports |
| 3 | # |
| 4 | # biotop block device (disk) I/O by process. |
| 5 | # For Linux, uses BCC, eBPF. |
| 6 | # |
xingfeng2510 | 14dacd8 | 2022-03-17 22:53:00 +0800 | [diff] [blame] | 7 | # USAGE: biotop.py [-h] [-C] [-r MAXROWS] [-p PID] [interval] [count] |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 8 | # |
| 9 | # This uses in-kernel eBPF maps to cache process details (PID and comm) by I/O |
| 10 | # request, as well as a starting timestamp for calculating I/O latency. |
| 11 | # |
| 12 | # Copyright 2016 Netflix, Inc. |
| 13 | # Licensed under the Apache License, Version 2.0 (the "License") |
| 14 | # |
| 15 | # 06-Feb-2016 Brendan Gregg Created this. |
xingfeng2510 | 14dacd8 | 2022-03-17 22:53:00 +0800 | [diff] [blame] | 16 | # 17-Mar-2022 Rocky Xing Added PID filter support. |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 17 | |
| 18 | from __future__ import print_function |
| 19 | from bcc import BPF |
| 20 | from time import sleep, strftime |
| 21 | import argparse |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 22 | from subprocess import call |
| 23 | |
| 24 | # arguments |
| 25 | examples = """examples: |
| 26 | ./biotop # block device I/O top, 1 second refresh |
| 27 | ./biotop -C # don't clear the screen |
xingfeng2510 | 14dacd8 | 2022-03-17 22:53:00 +0800 | [diff] [blame] | 28 | ./biotop -p 181 # only trace PID 181 |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 29 | ./biotop 5 # 5 second summaries |
| 30 | ./biotop 5 10 # 5 second summaries, 10 times only |
| 31 | """ |
| 32 | parser = argparse.ArgumentParser( |
| 33 | description="Block device (disk) I/O by process", |
| 34 | formatter_class=argparse.RawDescriptionHelpFormatter, |
Nathan Scott | 1a197db | 2018-01-21 09:14:27 +1100 | [diff] [blame] | 35 | epilog=examples) |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 36 | parser.add_argument("-C", "--noclear", action="store_true", |
| 37 | help="don't clear the screen") |
| 38 | parser.add_argument("-r", "--maxrows", default=20, |
| 39 | help="maximum rows to print, default 20") |
xingfeng2510 | 14dacd8 | 2022-03-17 22:53:00 +0800 | [diff] [blame] | 40 | parser.add_argument("-p", "--pid", type=int, metavar="PID", |
| 41 | help="trace this PID only") |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 42 | parser.add_argument("interval", nargs="?", default=1, |
| 43 | help="output interval, in seconds") |
| 44 | parser.add_argument("count", nargs="?", default=99999999, |
| 45 | help="number of outputs") |
Nathan Scott | f5fb9af | 2018-01-17 09:39:59 +1100 | [diff] [blame] | 46 | parser.add_argument("--ebpf", action="store_true", |
| 47 | help=argparse.SUPPRESS) |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 48 | args = parser.parse_args() |
| 49 | interval = int(args.interval) |
| 50 | countdown = int(args.count) |
| 51 | maxrows = int(args.maxrows) |
| 52 | clear = not int(args.noclear) |
| 53 | |
| 54 | # linux stats |
| 55 | loadavg = "/proc/loadavg" |
| 56 | diskstats = "/proc/diskstats" |
| 57 | |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 58 | # load BPF program |
Nathan Scott | ca4ba55 | 2018-01-16 11:02:58 +1100 | [diff] [blame] | 59 | bpf_text = """ |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 60 | #include <uapi/linux/ptrace.h> |
Jerome Marchand | ee81072 | 2022-02-23 16:04:30 +0100 | [diff] [blame] | 61 | #include <linux/blk-mq.h> |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 62 | |
Andreas Gerstmayr | 7766826 | 2021-02-25 19:33:08 +0100 | [diff] [blame] | 63 | // for saving the timestamp and __data_len of each request |
| 64 | struct start_req_t { |
| 65 | u64 ts; |
| 66 | u64 data_len; |
| 67 | }; |
| 68 | |
Brendan Gregg | 6321d00 | 2016-02-07 00:54:44 -0800 | [diff] [blame] | 69 | // for saving process info by request |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 70 | struct who_t { |
| 71 | u32 pid; |
| 72 | char name[TASK_COMM_LEN]; |
| 73 | }; |
| 74 | |
Brendan Gregg | 6321d00 | 2016-02-07 00:54:44 -0800 | [diff] [blame] | 75 | // the key for the output summary |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 76 | struct info_t { |
| 77 | u32 pid; |
Brendan Gregg | 51add78 | 2016-11-29 11:11:31 -0800 | [diff] [blame] | 78 | int rwflag; |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 79 | int major; |
| 80 | int minor; |
| 81 | char name[TASK_COMM_LEN]; |
| 82 | }; |
| 83 | |
Brendan Gregg | 6321d00 | 2016-02-07 00:54:44 -0800 | [diff] [blame] | 84 | // the value of the output summary |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 85 | struct val_t { |
| 86 | u64 bytes; |
| 87 | u64 us; |
| 88 | u32 io; |
| 89 | }; |
| 90 | |
Andreas Gerstmayr | 7766826 | 2021-02-25 19:33:08 +0100 | [diff] [blame] | 91 | BPF_HASH(start, struct request *, struct start_req_t); |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 92 | BPF_HASH(whobyreq, struct request *, struct who_t); |
| 93 | BPF_HASH(counts, struct info_t, struct val_t); |
| 94 | |
| 95 | // cache PID and comm by-req |
| 96 | int trace_pid_start(struct pt_regs *ctx, struct request *req) |
| 97 | { |
| 98 | struct who_t who = {}; |
xingfeng2510 | 14dacd8 | 2022-03-17 22:53:00 +0800 | [diff] [blame] | 99 | u32 pid; |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 100 | |
| 101 | if (bpf_get_current_comm(&who.name, sizeof(who.name)) == 0) { |
xingfeng2510 | 14dacd8 | 2022-03-17 22:53:00 +0800 | [diff] [blame] | 102 | pid = bpf_get_current_pid_tgid() >> 32; |
| 103 | if (FILTER_PID) |
| 104 | return 0; |
| 105 | |
| 106 | who.pid = pid; |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 107 | whobyreq.update(&req, &who); |
| 108 | } |
| 109 | |
| 110 | return 0; |
| 111 | } |
| 112 | |
| 113 | // time block I/O |
| 114 | int trace_req_start(struct pt_regs *ctx, struct request *req) |
| 115 | { |
Andreas Gerstmayr | 7766826 | 2021-02-25 19:33:08 +0100 | [diff] [blame] | 116 | struct start_req_t start_req = { |
| 117 | .ts = bpf_ktime_get_ns(), |
| 118 | .data_len = req->__data_len |
| 119 | }; |
| 120 | start.update(&req, &start_req); |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 121 | return 0; |
| 122 | } |
| 123 | |
| 124 | // output |
| 125 | int trace_req_completion(struct pt_regs *ctx, struct request *req) |
| 126 | { |
Andreas Gerstmayr | 7766826 | 2021-02-25 19:33:08 +0100 | [diff] [blame] | 127 | struct start_req_t *startp; |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 128 | |
| 129 | // fetch timestamp and calculate delta |
Andreas Gerstmayr | 7766826 | 2021-02-25 19:33:08 +0100 | [diff] [blame] | 130 | startp = start.lookup(&req); |
| 131 | if (startp == 0) { |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 132 | return 0; // missed tracing issue |
| 133 | } |
| 134 | |
| 135 | struct who_t *whop; |
xingfeng2510 | 14dacd8 | 2022-03-17 22:53:00 +0800 | [diff] [blame] | 136 | u32 pid; |
| 137 | |
| 138 | whop = whobyreq.lookup(&req); |
| 139 | pid = whop != 0 ? whop->pid : 0; |
| 140 | if (FILTER_PID) { |
| 141 | start.delete(&req); |
| 142 | if (whop != 0) { |
| 143 | whobyreq.delete(&req); |
| 144 | } |
| 145 | return 0; |
| 146 | } |
| 147 | |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 148 | struct val_t *valp, zero = {}; |
Andreas Gerstmayr | 7766826 | 2021-02-25 19:33:08 +0100 | [diff] [blame] | 149 | u64 delta_us = (bpf_ktime_get_ns() - startp->ts) / 1000; |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 150 | |
| 151 | // setup info_t key |
| 152 | struct info_t info = {}; |
Tejun Heo | ddfcc29 | 2022-03-10 08:37:21 -1000 | [diff] [blame] | 153 | info.major = req->__RQ_DISK__->major; |
| 154 | info.minor = req->__RQ_DISK__->first_minor; |
Brendan Gregg | 51add78 | 2016-11-29 11:11:31 -0800 | [diff] [blame] | 155 | /* |
| 156 | * The following deals with a kernel version change (in mainline 4.7, although |
| 157 | * it may be backported to earlier kernels) with how block request write flags |
| 158 | * are tested. We handle both pre- and post-change versions here. Please avoid |
| 159 | * kernel version tests like this as much as possible: they inflate the code, |
| 160 | * test, and maintenance burden. |
| 161 | */ |
| 162 | #ifdef REQ_WRITE |
| 163 | info.rwflag = !!(req->cmd_flags & REQ_WRITE); |
Ryan Leary | bc43a29 | 2017-01-07 15:34:31 -0500 | [diff] [blame] | 164 | #elif defined(REQ_OP_SHIFT) |
Brendan Gregg | 51add78 | 2016-11-29 11:11:31 -0800 | [diff] [blame] | 165 | info.rwflag = !!((req->cmd_flags >> REQ_OP_SHIFT) == REQ_OP_WRITE); |
Ryan Leary | bc43a29 | 2017-01-07 15:34:31 -0500 | [diff] [blame] | 166 | #else |
| 167 | info.rwflag = !!((req->cmd_flags & REQ_OP_MASK) == REQ_OP_WRITE); |
Brendan Gregg | 51add78 | 2016-11-29 11:11:31 -0800 | [diff] [blame] | 168 | #endif |
| 169 | |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 170 | if (whop == 0) { |
| 171 | // missed pid who, save stats as pid 0 |
yonghong-song | 82f4302 | 2019-10-31 08:16:12 -0700 | [diff] [blame] | 172 | valp = counts.lookup_or_try_init(&info, &zero); |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 173 | } else { |
| 174 | info.pid = whop->pid; |
| 175 | __builtin_memcpy(&info.name, whop->name, sizeof(info.name)); |
yonghong-song | 82f4302 | 2019-10-31 08:16:12 -0700 | [diff] [blame] | 176 | valp = counts.lookup_or_try_init(&info, &zero); |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 177 | } |
| 178 | |
Philip Gladstone | ba64f03 | 2019-09-20 01:12:01 -0400 | [diff] [blame] | 179 | if (valp) { |
| 180 | // save stats |
| 181 | valp->us += delta_us; |
Andreas Gerstmayr | 7766826 | 2021-02-25 19:33:08 +0100 | [diff] [blame] | 182 | valp->bytes += startp->data_len; |
Philip Gladstone | ba64f03 | 2019-09-20 01:12:01 -0400 | [diff] [blame] | 183 | valp->io++; |
| 184 | } |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 185 | |
| 186 | start.delete(&req); |
| 187 | whobyreq.delete(&req); |
| 188 | |
| 189 | return 0; |
| 190 | } |
Nathan Scott | ca4ba55 | 2018-01-16 11:02:58 +1100 | [diff] [blame] | 191 | """ |
| 192 | |
| 193 | if args.ebpf: |
| 194 | print(bpf_text) |
| 195 | exit() |
| 196 | |
Tejun Heo | ddfcc29 | 2022-03-10 08:37:21 -1000 | [diff] [blame] | 197 | if BPF.kernel_struct_has_field(b'request', b'rq_disk'): |
| 198 | bpf_text = bpf_text.replace('__RQ_DISK__', 'rq_disk') |
| 199 | else: |
| 200 | bpf_text = bpf_text.replace('__RQ_DISK__', 'q->disk') |
| 201 | |
xingfeng2510 | 14dacd8 | 2022-03-17 22:53:00 +0800 | [diff] [blame] | 202 | if args.pid is not None: |
| 203 | bpf_text = bpf_text.replace('FILTER_PID', 'pid != %d' % args.pid) |
| 204 | else: |
| 205 | bpf_text = bpf_text.replace('FILTER_PID', '0') |
| 206 | |
Nathan Scott | ca4ba55 | 2018-01-16 11:02:58 +1100 | [diff] [blame] | 207 | b = BPF(text=bpf_text) |
Hengqi Chen | 8c80b29 | 2021-12-11 17:36:17 +0800 | [diff] [blame] | 208 | if BPF.get_kprobe_functions(b'__blk_account_io_start'): |
| 209 | b.attach_kprobe(event="__blk_account_io_start", fn_name="trace_pid_start") |
| 210 | else: |
| 211 | b.attach_kprobe(event="blk_account_io_start", fn_name="trace_pid_start") |
Jerome Marchand | 74e25ed | 2018-12-10 08:54:50 +0100 | [diff] [blame] | 212 | if BPF.get_kprobe_functions(b'blk_start_request'): |
| 213 | b.attach_kprobe(event="blk_start_request", fn_name="trace_req_start") |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 214 | b.attach_kprobe(event="blk_mq_start_request", fn_name="trace_req_start") |
Hengqi Chen | 8c80b29 | 2021-12-11 17:36:17 +0800 | [diff] [blame] | 215 | if BPF.get_kprobe_functions(b'__blk_account_io_done'): |
| 216 | b.attach_kprobe(event="__blk_account_io_done", fn_name="trace_req_completion") |
| 217 | else: |
| 218 | b.attach_kprobe(event="blk_account_io_done", fn_name="trace_req_completion") |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 219 | |
| 220 | print('Tracing... Output every %d secs. Hit Ctrl-C to end' % interval) |
| 221 | |
| 222 | # cache disk major,minor -> diskname |
| 223 | disklookup = {} |
| 224 | with open(diskstats) as stats: |
| 225 | for line in stats: |
| 226 | a = line.split() |
| 227 | disklookup[a[0] + "," + a[1]] = a[2] |
| 228 | |
| 229 | # output |
| 230 | exiting = 0 |
| 231 | while 1: |
| 232 | try: |
| 233 | sleep(interval) |
| 234 | except KeyboardInterrupt: |
| 235 | exiting = 1 |
| 236 | |
| 237 | # header |
| 238 | if clear: |
| 239 | call("clear") |
| 240 | else: |
| 241 | print() |
| 242 | with open(loadavg) as stats: |
| 243 | print("%-8s loadavg: %s" % (strftime("%H:%M:%S"), stats.read())) |
xingfeng2510 | 03e4948 | 2022-03-17 13:07:16 +0800 | [diff] [blame] | 244 | print("%-7s %-16s %1s %-3s %-3s %-8s %5s %7s %6s" % ("PID", "COMM", |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 245 | "D", "MAJ", "MIN", "DISK", "I/O", "Kbytes", "AVGms")) |
| 246 | |
| 247 | # by-PID output |
| 248 | counts = b.get_table("counts") |
| 249 | line = 0 |
| 250 | for k, v in reversed(sorted(counts.items(), |
| 251 | key=lambda counts: counts[1].bytes)): |
| 252 | |
| 253 | # lookup disk |
| 254 | disk = str(k.major) + "," + str(k.minor) |
| 255 | if disk in disklookup: |
| 256 | diskname = disklookup[disk] |
| 257 | else: |
| 258 | diskname = "?" |
| 259 | |
| 260 | # print line |
| 261 | avg_ms = (float(v.us) / 1000) / v.io |
xingfeng2510 | 03e4948 | 2022-03-17 13:07:16 +0800 | [diff] [blame] | 262 | print("%-7d %-16s %1s %-3d %-3d %-8s %5s %7s %6.2f" % (k.pid, |
jeromemarchand | b96ebcd | 2018-10-10 01:58:15 +0200 | [diff] [blame] | 263 | k.name.decode('utf-8', 'replace'), "W" if k.rwflag else "R", |
| 264 | k.major, k.minor, diskname, v.io, v.bytes / 1024, avg_ms)) |
Brendan Gregg | 6f075b9 | 2016-02-07 00:46:34 -0800 | [diff] [blame] | 265 | |
| 266 | line += 1 |
| 267 | if line >= maxrows: |
| 268 | break |
| 269 | counts.clear() |
| 270 | |
| 271 | countdown -= 1 |
| 272 | if exiting or countdown == 0: |
| 273 | print("Detaching...") |
| 274 | exit() |