Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 1 | #!/usr/bin/env python |
| 2 | # |
| 3 | # kvmexit.py |
| 4 | # |
| 5 | # Display the exit_reason and its statistics of each vm exit |
| 6 | # for all vcpus of all virtual machines. For example: |
| 7 | # $./kvmexit.py |
| 8 | # PID TID KVM_EXIT_REASON COUNT |
| 9 | # 1273551 1273568 EXIT_REASON_MSR_WRITE 6 |
| 10 | # 1274253 1274261 EXIT_REASON_EXTERNAL_INTERRUPT 1 |
| 11 | # 1274253 1274261 EXIT_REASON_HLT 12 |
| 12 | # ... |
| 13 | # |
| 14 | # Besides, we also allow users to specify one pid, tid(s), or one |
| 15 | # pid and its vcpu. See kvmexit_example.txt for more examples. |
| 16 | # |
| 17 | # @PID: each vitual machine's pid in the user space. |
| 18 | # @TID: the user space's thread of each vcpu of that virtual machine. |
| 19 | # @KVM_EXIT_REASON: the reason why the vm exits. |
| 20 | # @COUNT: the counts of the @KVM_EXIT_REASONS. |
| 21 | # |
| 22 | # REQUIRES: Linux 4.7+ (BPF_PROG_TYPE_TRACEPOINT support) |
| 23 | # |
| 24 | # Copyright (c) 2021 ByteDance Inc. All rights reserved. |
| 25 | # |
| 26 | # Author(s): |
| 27 | # Fei Li <lifei.shirley@bytedance.com> |
| 28 | |
| 29 | |
| 30 | from __future__ import print_function |
Hengqi Chen | e12ec04 | 2021-09-23 16:58:44 +0800 | [diff] [blame] | 31 | from time import sleep |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 32 | from bcc import BPF |
| 33 | import argparse |
| 34 | import multiprocessing |
| 35 | import os |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 36 | import subprocess |
| 37 | |
| 38 | # |
| 39 | # Process Arguments |
| 40 | # |
| 41 | def valid_args_list(args): |
| 42 | args_list = args.split(",") |
| 43 | for arg in args_list: |
| 44 | try: |
| 45 | int(arg) |
| 46 | except: |
| 47 | raise argparse.ArgumentTypeError("must be valid integer") |
| 48 | return args_list |
| 49 | |
| 50 | # arguments |
| 51 | examples = """examples: |
| 52 | ./kvmexit # Display kvm_exit_reason and its statistics in real-time until Ctrl-C |
| 53 | ./kvmexit 5 # Display in real-time after sleeping 5s |
| 54 | ./kvmexit -p 3195281 # Collpase all tids for pid 3195281 with exit reasons sorted in descending order |
| 55 | ./kvmexit -p 3195281 20 # Collpase all tids for pid 3195281 with exit reasons sorted in descending order, and display after sleeping 20s |
| 56 | ./kvmexit -p 3195281 -v 0 # Display only vcpu0 for pid 3195281, descending sort by default |
| 57 | ./kvmexit -p 3195281 -a # Display all tids for pid 3195281 |
| 58 | ./kvmexit -t 395490 # Display only for tid 395490 with exit reasons sorted in descending order |
| 59 | ./kvmexit -t 395490 20 # Display only for tid 395490 with exit reasons sorted in descending order after sleeping 20s |
| 60 | ./kvmexit -T '395490,395491' # Display for a union like {395490, 395491} |
| 61 | """ |
| 62 | parser = argparse.ArgumentParser( |
| 63 | description="Display kvm_exit_reason and its statistics at a timed interval", |
| 64 | formatter_class=argparse.RawDescriptionHelpFormatter, |
| 65 | epilog=examples) |
| 66 | parser.add_argument("duration", nargs="?", default=99999999, type=int, help="show delta for next several seconds") |
| 67 | parser.add_argument("-p", "--pid", type=int, help="trace this PID only") |
| 68 | exgroup = parser.add_mutually_exclusive_group() |
| 69 | exgroup.add_argument("-t", "--tid", type=int, help="trace this TID only") |
| 70 | exgroup.add_argument("-T", "--tids", type=valid_args_list, help="trace a comma separated series of tids with no space in between") |
| 71 | exgroup.add_argument("-v", "--vcpu", type=int, help="trace this vcpu only") |
| 72 | exgroup.add_argument("-a", "--alltids", action="store_true", help="trace all tids for this pid") |
| 73 | args = parser.parse_args() |
| 74 | duration = int(args.duration) |
| 75 | |
| 76 | # |
| 77 | # Setup BPF |
| 78 | # |
| 79 | |
| 80 | # load BPF program |
| 81 | bpf_text = """ |
| 82 | #include <linux/delay.h> |
| 83 | |
| 84 | #define REASON_NUM 69 |
| 85 | #define TGID_NUM 1024 |
| 86 | |
| 87 | struct exit_count { |
| 88 | u64 exit_ct[REASON_NUM]; |
| 89 | }; |
| 90 | BPF_PERCPU_ARRAY(init_value, struct exit_count, 1); |
| 91 | BPF_TABLE("percpu_hash", u64, struct exit_count, pcpu_kvm_stat, TGID_NUM); |
| 92 | |
| 93 | struct cache_info { |
| 94 | u64 cache_pid_tgid; |
| 95 | struct exit_count cache_exit_ct; |
| 96 | }; |
| 97 | BPF_PERCPU_ARRAY(pcpu_cache, struct cache_info, 1); |
| 98 | |
Hengqi Chen | 8f3b2db | 2023-06-27 12:07:49 +0000 | [diff] [blame] | 99 | TRACEPOINT_PROBE(kvm, kvm_exit) { |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 100 | int cache_miss = 0; |
| 101 | int zero = 0; |
Hengqi Chen | 8f3b2db | 2023-06-27 12:07:49 +0000 | [diff] [blame] | 102 | u32 er = args->exit_reason; |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 103 | if (er >= REASON_NUM) { |
| 104 | return 0; |
| 105 | } |
| 106 | |
| 107 | u64 cur_pid_tgid = bpf_get_current_pid_tgid(); |
| 108 | u32 tgid = cur_pid_tgid >> 32; |
| 109 | u32 pid = cur_pid_tgid; |
| 110 | |
| 111 | if (THREAD_FILTER) |
| 112 | return 0; |
| 113 | |
| 114 | struct exit_count *tmp_info = NULL, *initial = NULL; |
| 115 | struct cache_info *cache_p; |
| 116 | cache_p = pcpu_cache.lookup(&zero); |
| 117 | if (cache_p == NULL) { |
| 118 | return 0; |
| 119 | } |
| 120 | |
| 121 | if (cache_p->cache_pid_tgid == cur_pid_tgid) { |
| 122 | //a. If the cur_pid_tgid hit this physical cpu consecutively, save it to pcpu_cache |
| 123 | tmp_info = &cache_p->cache_exit_ct; |
| 124 | } else { |
| 125 | //b. If another pid_tgid matches this pcpu for the last hit, OR it is the first time to hit this physical cpu. |
| 126 | cache_miss = 1; |
| 127 | |
| 128 | // b.a Try to load the last cache struct if exists. |
| 129 | tmp_info = pcpu_kvm_stat.lookup(&cur_pid_tgid); |
| 130 | |
| 131 | // b.b If it is the first time for the cur_pid_tgid to hit this pcpu, employ a |
| 132 | // per_cpu array to initialize pcpu_kvm_stat's exit_count with each exit reason's count is zero |
| 133 | if (tmp_info == NULL) { |
| 134 | initial = init_value.lookup(&zero); |
| 135 | if (initial == NULL) { |
| 136 | return 0; |
| 137 | } |
| 138 | |
| 139 | pcpu_kvm_stat.update(&cur_pid_tgid, initial); |
| 140 | tmp_info = pcpu_kvm_stat.lookup(&cur_pid_tgid); |
| 141 | // To pass the verifier |
| 142 | if (tmp_info == NULL) { |
| 143 | return 0; |
| 144 | } |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | if (er < REASON_NUM) { |
| 149 | tmp_info->exit_ct[er]++; |
| 150 | if (cache_miss == 1) { |
| 151 | if (cache_p->cache_pid_tgid != 0) { |
| 152 | // b.*.a Let's save the last hit cache_info into kvm_stat. |
| 153 | pcpu_kvm_stat.update(&cache_p->cache_pid_tgid, &cache_p->cache_exit_ct); |
| 154 | } |
| 155 | // b.* As the cur_pid_tgid meets current pcpu_cache_array for the first time, save it. |
| 156 | cache_p->cache_pid_tgid = cur_pid_tgid; |
| 157 | bpf_probe_read(&cache_p->cache_exit_ct, sizeof(*tmp_info), tmp_info); |
| 158 | } |
| 159 | return 0; |
| 160 | } |
| 161 | |
| 162 | return 0; |
| 163 | } |
| 164 | """ |
| 165 | |
| 166 | # format output |
| 167 | exit_reasons = ( |
| 168 | "EXCEPTION_NMI", |
| 169 | "EXTERNAL_INTERRUPT", |
| 170 | "TRIPLE_FAULT", |
| 171 | "INIT_SIGNAL", |
| 172 | "N/A", |
| 173 | "N/A", |
| 174 | "N/A", |
| 175 | "INTERRUPT_WINDOW", |
| 176 | "NMI_WINDOW", |
| 177 | "TASK_SWITCH", |
| 178 | "CPUID", |
| 179 | "N/A", |
| 180 | "HLT", |
| 181 | "INVD", |
| 182 | "INVLPG", |
| 183 | "RDPMC", |
| 184 | "RDTSC", |
| 185 | "N/A", |
| 186 | "VMCALL", |
| 187 | "VMCLEAR", |
| 188 | "VMLAUNCH", |
| 189 | "VMPTRLD", |
| 190 | "VMPTRST", |
| 191 | "VMREAD", |
| 192 | "VMRESUME", |
| 193 | "VMWRITE", |
| 194 | "VMOFF", |
| 195 | "VMON", |
| 196 | "CR_ACCESS", |
| 197 | "DR_ACCESS", |
| 198 | "IO_INSTRUCTION", |
| 199 | "MSR_READ", |
| 200 | "MSR_WRITE", |
| 201 | "INVALID_STATE", |
| 202 | "MSR_LOAD_FAIL", |
| 203 | "N/A", |
| 204 | "MWAIT_INSTRUCTION", |
| 205 | "MONITOR_TRAP_FLAG", |
| 206 | "N/A", |
| 207 | "MONITOR_INSTRUCTION", |
| 208 | "PAUSE_INSTRUCTION", |
| 209 | "MCE_DURING_VMENTRY", |
| 210 | "N/A", |
| 211 | "TPR_BELOW_THRESHOLD", |
| 212 | "APIC_ACCESS", |
| 213 | "EOI_INDUCED", |
| 214 | "GDTR_IDTR", |
| 215 | "LDTR_TR", |
| 216 | "EPT_VIOLATION", |
| 217 | "EPT_MISCONFIG", |
| 218 | "INVEPT", |
| 219 | "RDTSCP", |
| 220 | "PREEMPTION_TIMER", |
| 221 | "INVVPID", |
| 222 | "WBINVD", |
| 223 | "XSETBV", |
| 224 | "APIC_WRITE", |
| 225 | "RDRAND", |
| 226 | "INVPCID", |
| 227 | "VMFUNC", |
| 228 | "ENCLS", |
| 229 | "RDSEED", |
| 230 | "PML_FULL", |
| 231 | "XSAVES", |
| 232 | "XRSTORS", |
| 233 | "N/A", |
| 234 | "N/A", |
| 235 | "UMWAIT", |
| 236 | "TPAUSE" |
| 237 | ) |
| 238 | |
| 239 | # |
| 240 | # Do some checks |
| 241 | # |
| 242 | try: |
| 243 | # Currently, only adapte on intel architecture |
| 244 | cmd = "cat /proc/cpuinfo | grep vendor_id | head -n 1" |
| 245 | arch_info = subprocess.check_output(cmd, shell=True).strip() |
| 246 | if b"Intel" in arch_info: |
| 247 | pass |
| 248 | else: |
| 249 | raise Exception("Currently we only support Intel architecture, please do expansion if needs more.") |
| 250 | |
| 251 | # Check if kvm module is loaded |
| 252 | if os.access("/dev/kvm", os.R_OK | os.W_OK): |
| 253 | pass |
| 254 | else: |
| 255 | raise Exception("Please insmod kvm module to use kvmexit tool.") |
| 256 | except Exception as e: |
| 257 | raise Exception("Failed to do precondition check, due to: %s." % e) |
| 258 | |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 259 | def find_tid(tgt_dir, tgt_vcpu): |
| 260 | for tid in os.listdir(tgt_dir): |
| 261 | path = tgt_dir + "/" + tid + "/comm" |
| 262 | fp = open(path, "r") |
| 263 | comm = fp.read() |
| 264 | if (comm.find(tgt_vcpu) != -1): |
| 265 | return tid |
| 266 | return -1 |
| 267 | |
| 268 | # set process/thread filter |
| 269 | thread_context = "" |
| 270 | header_format = "" |
| 271 | need_collapse = not args.alltids |
| 272 | if args.tid is not None: |
| 273 | thread_context = "TID %s" % args.tid |
| 274 | thread_filter = 'pid != %s' % args.tid |
| 275 | elif args.tids is not None: |
| 276 | thread_context = "TIDS %s" % args.tids |
| 277 | thread_filter = "pid != " + " && pid != ".join(args.tids) |
| 278 | header_format = "TIDS " |
| 279 | elif args.pid is not None: |
| 280 | thread_context = "PID %s" % args.pid |
| 281 | thread_filter = 'tgid != %s' % args.pid |
| 282 | if args.vcpu is not None: |
| 283 | thread_context = "PID %s VCPU %s" % (args.pid, args.vcpu) |
| 284 | # transfer vcpu to tid |
| 285 | tgt_dir = '/proc/' + str(args.pid) + '/task' |
| 286 | tgt_vcpu = "CPU " + str(args.vcpu) |
| 287 | args.tid = find_tid(tgt_dir, tgt_vcpu) |
| 288 | if args.tid == -1: |
| 289 | raise Exception("There's no v%s for PID %d." % (tgt_vcpu, args.pid)) |
| 290 | thread_filter = 'pid != %s' % args.tid |
| 291 | elif args.alltids: |
| 292 | thread_context = "PID %s and its all threads" % args.pid |
| 293 | header_format = "TID " |
| 294 | else: |
| 295 | thread_context = "all threads" |
| 296 | thread_filter = '0' |
| 297 | header_format = "PID TID " |
| 298 | bpf_text = bpf_text.replace('THREAD_FILTER', thread_filter) |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 299 | b = BPF(text=bpf_text) |
| 300 | |
| 301 | |
| 302 | # header |
| 303 | print("Display kvm exit reasons and statistics for %s" % thread_context, end="") |
| 304 | if duration < 99999999: |
| 305 | print(" after sleeping %d secs." % duration) |
| 306 | else: |
| 307 | print("... Hit Ctrl-C to end.") |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 308 | |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 309 | try: |
| 310 | sleep(duration) |
| 311 | except KeyboardInterrupt: |
Hengqi Chen | e12ec04 | 2021-09-23 16:58:44 +0800 | [diff] [blame] | 312 | print() |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 313 | |
| 314 | |
| 315 | # Currently, sort multiple tids in descending order is not supported. |
| 316 | if (args.pid or args.tid): |
| 317 | ct_reason = [] |
| 318 | if args.pid: |
| 319 | tgid_exit = [0 for i in range(len(exit_reasons))] |
| 320 | |
| 321 | # output |
Hengqi Chen | e12ec04 | 2021-09-23 16:58:44 +0800 | [diff] [blame] | 322 | print("%s%-35s %s" % (header_format, "KVM_EXIT_REASON", "COUNT")) |
| 323 | |
Fei Li | 4236563 | 2020-08-31 21:35:33 +0800 | [diff] [blame] | 324 | pcpu_kvm_stat = b["pcpu_kvm_stat"] |
| 325 | pcpu_cache = b["pcpu_cache"] |
| 326 | for k, v in pcpu_kvm_stat.items(): |
| 327 | tgid = k.value >> 32 |
| 328 | pid = k.value & 0xffffffff |
| 329 | for i in range(0, len(exit_reasons)): |
| 330 | sum1 = 0 |
| 331 | for inner_cpu in range(0, multiprocessing.cpu_count()): |
| 332 | cachePIDTGID = pcpu_cache[0][inner_cpu].cache_pid_tgid |
| 333 | # Take priority to check if it is in cache |
| 334 | if cachePIDTGID == k.value: |
| 335 | sum1 += pcpu_cache[0][inner_cpu].cache_exit_ct.exit_ct[i] |
| 336 | # If not in cache, find from kvm_stat |
| 337 | else: |
| 338 | sum1 += v[inner_cpu].exit_ct[i] |
| 339 | if sum1 == 0: |
| 340 | continue |
| 341 | |
| 342 | if (args.pid and args.pid == tgid and need_collapse): |
| 343 | tgid_exit[i] += sum1 |
| 344 | elif (args.tid and args.tid == pid): |
| 345 | ct_reason.append((sum1, i)) |
| 346 | elif not need_collapse or args.tids: |
| 347 | print("%-8u %-35s %-8u" % (pid, exit_reasons[i], sum1)) |
| 348 | else: |
| 349 | print("%-8u %-8u %-35s %-8u" % (tgid, pid, exit_reasons[i], sum1)) |
| 350 | |
| 351 | # Display only for the target tid in descending sort |
| 352 | if (args.tid and args.tid == pid): |
| 353 | ct_reason.sort(reverse=True) |
| 354 | for i in range(0, len(ct_reason)): |
| 355 | if ct_reason[i][0] == 0: |
| 356 | continue |
| 357 | print("%-35s %-8u" % (exit_reasons[ct_reason[i][1]], ct_reason[i][0])) |
| 358 | break |
| 359 | |
| 360 | |
| 361 | # Aggregate all tids' counts for this args.pid in descending sort |
| 362 | if args.pid and need_collapse: |
| 363 | for i in range(0, len(exit_reasons)): |
| 364 | ct_reason.append((tgid_exit[i], i)) |
| 365 | ct_reason.sort(reverse=True) |
| 366 | for i in range(0, len(ct_reason)): |
| 367 | if ct_reason[i][0] == 0: |
| 368 | continue |
| 369 | print("%-35s %-8u" % (exit_reasons[ct_reason[i][1]], ct_reason[i][0])) |