Erwan Velu | 8c12794 | 2020-05-20 20:25:32 +0200 | [diff] [blame] | 1 | #!/usr/bin/python |
| 2 | # @lint-avoid-python-3-compatibility-imports |
| 3 | # |
| 4 | # dirtop file reads and writes by directory. |
| 5 | # For Linux, uses BCC, eBPF. |
| 6 | # |
| 7 | # USAGE: dirtop.py -d 'directory1,directory2' [-h] [-C] [-r MAXROWS] [interval] [count] |
| 8 | # |
| 9 | # This uses in-kernel eBPF maps to store per process summaries for efficiency. |
| 10 | # |
| 11 | # Copyright 2016 Netflix, Inc. |
| 12 | # Licensed under the Apache License, Version 2.0 (the "License") |
| 13 | # |
| 14 | # 13-Mar-2020 Erwan Velu Created dirtop from filetop |
| 15 | # 06-Feb-2016 Brendan Gregg Created filetop. |
| 16 | |
| 17 | from __future__ import print_function |
| 18 | from bcc import BPF |
| 19 | from time import sleep, strftime |
| 20 | import argparse |
| 21 | import os |
| 22 | import stat |
| 23 | from subprocess import call |
| 24 | |
| 25 | # arguments |
| 26 | examples = """examples: |
| 27 | ./dirtop -d '/hdfs/uuid/*/yarn' # directory I/O top, 1 second refresh |
| 28 | ./dirtop -d '/hdfs/uuid/*/yarn' -C # don't clear the screen |
| 29 | ./dirtop -d '/hdfs/uuid/*/yarn' 5 # 5 second summaries |
| 30 | ./dirtop -d '/hdfs/uuid/*/yarn' 5 10 # 5 second summaries, 10 times only |
| 31 | ./dirtop -d '/hdfs/uuid/*/yarn,/hdfs/uuid/*/data' # Running dirtop on two set of directories |
| 32 | """ |
| 33 | parser = argparse.ArgumentParser( |
| 34 | description="File reads and writes by process", |
| 35 | formatter_class=argparse.RawDescriptionHelpFormatter, |
| 36 | epilog=examples) |
| 37 | parser.add_argument("-C", "--noclear", action="store_true", |
| 38 | help="don't clear the screen") |
| 39 | parser.add_argument("-r", "--maxrows", default=20, |
| 40 | help="maximum rows to print, default 20") |
| 41 | parser.add_argument("-s", "--sort", default="all", |
| 42 | choices=["all", "reads", "writes", "rbytes", "wbytes"], |
| 43 | help="sort column, default all") |
| 44 | parser.add_argument("-p", "--pid", type=int, metavar="PID", dest="tgid", |
| 45 | help="trace this PID only") |
| 46 | parser.add_argument("interval", nargs="?", default=1, |
| 47 | help="output interval, in seconds") |
| 48 | parser.add_argument("count", nargs="?", default=99999999, |
| 49 | help="number of outputs") |
| 50 | parser.add_argument("--ebpf", action="store_true", |
| 51 | help=argparse.SUPPRESS) |
| 52 | parser.add_argument("-d", "--root-directories", type=str, required=True, dest="rootdirs", |
| 53 | help="select the directories to observe, separated by commas") |
| 54 | args = parser.parse_args() |
| 55 | interval = int(args.interval) |
| 56 | countdown = int(args.count) |
| 57 | maxrows = int(args.maxrows) |
| 58 | clear = not int(args.noclear) |
| 59 | debug = 0 |
| 60 | |
| 61 | # linux stats |
| 62 | loadavg = "/proc/loadavg" |
| 63 | |
| 64 | # define BPF program |
| 65 | bpf_text = """ |
| 66 | # include <uapi/linux/ptrace.h> |
| 67 | # include <linux/blkdev.h> |
| 68 | |
| 69 | // the key for the output summary |
| 70 | struct info_t { |
| 71 | unsigned long inode_id; |
| 72 | }; |
| 73 | |
| 74 | // the value of the output summary |
| 75 | struct val_t { |
| 76 | u64 reads; |
| 77 | u64 writes; |
| 78 | u64 rbytes; |
| 79 | u64 wbytes; |
| 80 | }; |
| 81 | |
| 82 | BPF_HASH(counts, struct info_t, struct val_t); |
| 83 | |
| 84 | static int do_entry(struct pt_regs *ctx, struct file *file, |
| 85 | char __user *buf, size_t count, int is_read) |
| 86 | { |
| 87 | u32 tgid = bpf_get_current_pid_tgid() >> 32; |
| 88 | if (TGID_FILTER) |
| 89 | return 0; |
| 90 | |
| 91 | // The directory inodes we look at |
| 92 | u32 dir_ids[INODES_NUMBER] = DIRECTORY_INODES; |
| 93 | struct info_t info = {.inode_id = 0}; |
| 94 | struct dentry *pde = file->f_path.dentry; |
| 95 | for (int i=0; i<50; i++) { |
| 96 | // If we don't have any parent, we reached the root |
| 97 | if (!pde->d_parent) { |
| 98 | break; |
| 99 | } |
| 100 | pde = pde->d_parent; |
| 101 | // Does the files is part of the directory we look for |
| 102 | for(int dir_id=0; dir_id<INODES_NUMBER; dir_id++) { |
| 103 | if (pde->d_inode->i_ino == dir_ids[dir_id]) { |
| 104 | // Yes, let's export the top directory inode |
| 105 | info.inode_id = pde->d_inode->i_ino; |
| 106 | break; |
| 107 | } |
| 108 | } |
| 109 | } |
| 110 | // If we didn't found any, let's abort |
| 111 | if (info.inode_id == 0) { |
| 112 | return 0; |
| 113 | } |
| 114 | |
| 115 | struct val_t *valp, zero = {}; |
| 116 | valp = counts.lookup_or_try_init(&info, &zero); |
| 117 | if (valp) { |
| 118 | if (is_read) { |
| 119 | valp->reads++; |
| 120 | valp->rbytes += count; |
| 121 | } else { |
| 122 | valp->writes++; |
| 123 | valp->wbytes += count; |
| 124 | } |
| 125 | } |
| 126 | return 0; |
| 127 | } |
| 128 | |
| 129 | int trace_read_entry(struct pt_regs *ctx, struct file *file, |
| 130 | char __user *buf, size_t count) |
| 131 | { |
| 132 | return do_entry(ctx, file, buf, count, 1); |
| 133 | } |
| 134 | |
| 135 | int trace_write_entry(struct pt_regs *ctx, struct file *file, |
| 136 | char __user *buf, size_t count) |
| 137 | { |
| 138 | return do_entry(ctx, file, buf, count, 0); |
| 139 | } |
| 140 | |
| 141 | """ |
| 142 | |
| 143 | |
| 144 | def get_searched_ids(root_directories): |
| 145 | """Export the inode numbers of the selected directories.""" |
| 146 | from glob import glob |
| 147 | inode_to_path = {} |
| 148 | inodes = "{" |
| 149 | total_dirs = 0 |
| 150 | for root_directory in root_directories.split(','): |
William Findlay | 1efba05 | 2020-10-16 13:06:36 -0400 | [diff] [blame] | 151 | try: |
| 152 | searched_dirs = glob(root_directory, recursive=True) |
| 153 | except TypeError: |
| 154 | searched_dirs = glob(root_directory) |
Erwan Velu | 8c12794 | 2020-05-20 20:25:32 +0200 | [diff] [blame] | 155 | if not searched_dirs: |
| 156 | continue |
| 157 | |
| 158 | for mydir in searched_dirs: |
| 159 | total_dirs = total_dirs + 1 |
| 160 | # If we pass more than 15 dirs, ebpf program fails |
| 161 | if total_dirs > 15: |
| 162 | print('15 directories limit reached') |
| 163 | break |
| 164 | inode_id = os.lstat(mydir)[stat.ST_INO] |
| 165 | if inode_id in inode_to_path: |
| 166 | if inode_to_path[inode_id] == mydir: |
| 167 | print('Skipping {} as already considered'.format(mydir)) |
| 168 | else: |
| 169 | inodes = "{},{}".format(inodes, inode_id) |
| 170 | inode_to_path[inode_id] = mydir |
| 171 | print('Considering {} with inode_id {}'.format(mydir, inode_id)) |
| 172 | |
| 173 | inodes = inodes + '}' |
| 174 | if len(inode_to_path) == 0: |
| 175 | print('Cannot find any valid directory') |
| 176 | exit() |
| 177 | return inodes.replace('{,', '{'), inode_to_path |
| 178 | |
| 179 | |
| 180 | if args.tgid: |
| 181 | bpf_text = bpf_text.replace('TGID_FILTER', 'tgid != %d' % args.tgid) |
| 182 | else: |
| 183 | bpf_text = bpf_text.replace('TGID_FILTER', '0') |
| 184 | |
| 185 | inodes, inodes_to_path = get_searched_ids(args.rootdirs) |
| 186 | bpf_text = bpf_text.replace("DIRECTORY_INODES", inodes) |
| 187 | bpf_text = bpf_text.replace( |
| 188 | "INODES_NUMBER", '{}'.format(len(inodes.split(',')))) |
| 189 | |
| 190 | if debug or args.ebpf: |
| 191 | print(bpf_text) |
| 192 | if args.ebpf: |
| 193 | exit() |
| 194 | |
| 195 | # initialize BPF |
| 196 | b = BPF(text=bpf_text) |
| 197 | b.attach_kprobe(event="vfs_read", fn_name="trace_read_entry") |
| 198 | b.attach_kprobe(event="vfs_write", fn_name="trace_write_entry") |
| 199 | |
| 200 | DNAME_INLINE_LEN = 32 # linux/dcache.h |
| 201 | |
| 202 | print('Tracing... Output every %d secs. Hit Ctrl-C to end' % interval) |
| 203 | |
| 204 | |
| 205 | def sort_fn(counts): |
| 206 | """Define how to sort the columns""" |
| 207 | if args.sort == "all": |
| 208 | return (counts[1].rbytes + counts[1].wbytes + counts[1].reads + counts[1].writes) |
| 209 | else: |
| 210 | return getattr(counts[1], args.sort) |
| 211 | |
| 212 | |
| 213 | # output |
| 214 | exiting = 0 |
| 215 | while 1: |
| 216 | try: |
| 217 | sleep(interval) |
| 218 | except KeyboardInterrupt: |
| 219 | exiting = 1 |
| 220 | |
| 221 | # header |
| 222 | if clear: |
| 223 | call("clear") |
| 224 | else: |
| 225 | print() |
| 226 | with open(loadavg) as stats: |
| 227 | print("%-8s loadavg: %s" % (strftime("%H:%M:%S"), stats.read())) |
| 228 | |
| 229 | print("%-6s %-6s %-8s %-8s %s" % |
| 230 | ("READS", "WRITES", "R_Kb", "W_Kb", "PATH")) |
| 231 | # by-TID output |
| 232 | counts = b.get_table("counts") |
| 233 | line = 0 |
| 234 | reads = {} |
| 235 | writes = {} |
| 236 | reads_Kb = {} |
| 237 | writes_Kb = {} |
| 238 | for k, v in reversed(sorted(counts.items(), |
| 239 | key=sort_fn)): |
| 240 | # If it's the first time we see this inode |
| 241 | if k.inode_id not in reads: |
| 242 | # let's create a new entry |
| 243 | reads[k.inode_id] = v.reads |
| 244 | writes[k.inode_id] = v.writes |
| 245 | reads_Kb[k.inode_id] = v.rbytes / 1024 |
| 246 | writes_Kb[k.inode_id] = v.wbytes / 1024 |
| 247 | else: |
| 248 | # unless add the current performance metrics |
| 249 | # to the previous ones |
| 250 | reads[k.inode_id] += v.reads |
| 251 | writes[k.inode_id] += v.writes |
| 252 | reads_Kb[k.inode_id] += v.rbytes / 1024 |
| 253 | writes_Kb[k.inode_id] += v.wbytes / 1024 |
| 254 | |
| 255 | for node_id in reads: |
| 256 | print("%-6d %-6d %-8d %-8d %s" % |
| 257 | (reads[node_id], writes[node_id], reads_Kb[node_id], writes_Kb[node_id], inodes_to_path[node_id])) |
| 258 | line += 1 |
| 259 | if line >= maxrows: |
| 260 | break |
| 261 | |
| 262 | counts.clear() |
| 263 | |
| 264 | countdown -= 1 |
| 265 | if exiting or countdown == 0: |
| 266 | print("Detaching...") |
| 267 | exit() |