1 #include <stdlib.h> 2 #include <stdio.h> 3 #include <unistd.h> 4 #include <string.h> 5 #include <sys/ioctl.h> 6 #include <linux/perf_event.h> 7 #include <asm/unistd.h> 8 #include <sys/mman.h> 9 #include <sys/stat.h> 10 #include <poll.h> 11 #include <signal.h> 12 #include <fcntl.h> 13 #include <elf.h> 14 #include <string.h> 15 16 #include <vector> 17 #include <string> 18 #include <map> 19 #include <unordered_map> 20 #include <unordered_set> 21 #include <algorithm> 22 using namespace std; 23 24 25 #define MAXN 512 26 #define MAXCPU 128 27 #define error(msg) do { perror(msg); exit(1); } while(0) 28 //--------------------------------Tree for call chain and report------------------------------- 29 // 30 struct TNode { 31 int c=0; 32 unordered_map<string, TNode*> s; 33 struct TNode *add(string n) { 34 c++; 35 if (s[n]==nullptr) s[n] = new TNode(); 36 return s[n]; 37 } 38 int printit(FILE *fp, int k) { 39 if (s.size()) { 40 using tt = tuple<int, string, TNode*>; 41 vector<tt> xx; 42 for (auto x: s) xx.push_back(make_tuple(x.second->c, x.first, x.second)); 43 sort(begin(xx), end(xx), greater<tt>()); 44 for (auto x: xx) { 45 auto count = get<0>(x); 46 if (100.0*count/c<1) continue; 47 auto name = get<1>(x); 48 auto nx = get<2>(x); 49 fprintf(fp, "<li>\n"); 50 fprintf(fp, "<input type=\"checkbox\" id=\"c%d\" />\n", k); 51 fprintf(fp, "<label class=\"tree_label\" for=\"c%d\">%s(%.3f%% %d/%d)</label>\n", k, name.c_str(), 100.0*count/c, count, c); 52 fprintf(fp, "<ul>\n"); 53 // printf("%s(%.3f%% %d/%d)\n", name.c_str(), 100.0*count/c, count, c); 54 k = nx->printit(fp, k+1); 55 fprintf(fp, "</ul>\n"); 56 fprintf(fp, "</li>\n"); 57 } 58 } 59 return k; 60 } 61 }; 62 63 //--------------------------------symbols------------------------------------------- 64 using STORE_T = map<unsigned long long, pair<string, unsigned long long>>; 65 using K_STORE_T = map<unsigned long long, string>; 66 67 /* 68 * load FUNC symbols refering to the section indicated by the offset, relocate the virtual address 69 */ 70 void parse_elf64(FILE *fp, unsigned long long v_addr, unsigned long long v_size, unsigned long long v_offset, STORE_T& store) { 71 Elf64_Ehdr ehdr; 72 int rc = fread(&ehdr, sizeof(ehdr), 1, fp); 73 if (rc != 1) return; 74 int n, s, i; 75 unsigned long long offset; 76 77 // load program headers 78 unsigned long long p_vaddr, p_size; 79 n = ehdr.e_phnum; 80 s = ehdr.e_phentsize; 81 offset = ehdr.e_phoff; 82 Elf64_Phdr phdr; 83 for (i=0; i<n; i++) { 84 rc = fseek(fp, offset, SEEK_SET); 85 if (rc<0) { perror("fail to seek"); return; } 86 rc = fread(&phdr, sizeof(phdr), 1, fp); 87 if (rc != 1) { perror("fail to read program header"); return; } 88 if (phdr.p_flags&PF_X) { 89 if (phdr.p_offset == v_offset) { 90 p_vaddr = phdr.p_vaddr; 91 p_size = phdr.p_memsz; if (p_size==0) p_size = 0xffffffff; 92 break; 93 } 94 } 95 offset+=s; 96 } 97 if (i>=n) { printf("No program header match offset found, fail to load\n"); return; } 98 99 // load section headers 100 n = ehdr.e_shnum; 101 s = ehdr.e_shentsize; 102 offset = ehdr.e_shoff; 103 Elf64_Shdr shdr; 104 vector<Elf64_Shdr> headers; 105 for (int i=0; i<n; i++) { 106 rc = fseek(fp, offset, SEEK_SET); 107 if (rc<0) { perror("fail to seek"); return; } 108 rc = fread(&shdr, sizeof(shdr), 1, fp); 109 if (rc != 1) { perror("fail to read sec header"); return; } 110 headers.push_back(shdr); 111 offset+=s; 112 } 113 Elf64_Sym symb; 114 unsigned long long faddr, fsize; 115 unsigned long long size, item_size; 116 int link, ix, flink, k; 117 char fname[128]; 118 for (int i=0; i<n; i++) { 119 switch(headers[i].sh_type) { 120 case SHT_SYMTAB: 121 case SHT_DYNSYM: 122 offset = headers[i].sh_offset; 123 size = headers[i].sh_size; 124 item_size = headers[i].sh_entsize; 125 link = headers[i].sh_link; 126 if (link<=0) break; 127 for (k=0; k+item_size<=size; k+=item_size) { 128 rc = fseek(fp, offset+k, SEEK_SET); if (rc<0) continue; 129 rc = fread(&symb, sizeof(symb), 1, fp); if (rc != 1) continue; 130 if (ELF64_ST_TYPE(symb.st_info) != STT_FUNC ) continue; 131 flink = symb.st_shndx; if (flink==0) continue; 132 fsize = symb.st_size; // if (fsize==0) continue; 133 faddr = symb.st_value; if (faddr>p_vaddr+p_size) continue; 134 ix = symb.st_name; if (ix==0) continue; 135 rc = fseek(fp, headers[link].sh_offset+ix, SEEK_SET); if (rc<0) continue; 136 if (fgets(fname, sizeof(fname), fp)==NULL) continue; 137 faddr = faddr-p_vaddr+v_addr; 138 if (store.count(faddr)) { 139 if (store[faddr].second<fsize) store[faddr] = make_pair(string(fname), fsize); 140 } else store[faddr] = make_pair(string(fname), fsize); 141 } 142 break; 143 default: 144 break; 145 } 146 } 147 } 148 149 int load_symbol_from_file(const char *path, unsigned long long addr, unsigned long long size, unsigned long long offset, STORE_T& store) { 150 printf("loading symble from %s\n", path); 151 FILE *fp = fopen(path, "rb"); 152 if (fp==NULL) { perror("fail to open file"); return -1; } 153 char ident[EI_NIDENT], c; 154 int err=0; 155 int rc = fread(ident, sizeof(ident), 1, fp); 156 if (rc != 1) { perror("fail to read ident"); err=-1; goto end; } 157 if (ident[0]!=0x7f) { printf("not a elf file\n"); err=-1; goto end; } 158 c=ident[4]; 159 rc = fseek(fp, 0, SEEK_SET); if (rc<0) { perror("fail to rewind"); goto end; } 160 if (c == ELFCLASS32) { 161 printf("32bit elf not supported yet\n"); err=-2; goto end; 162 } else if (c == ELFCLASS64) { 163 parse_elf64(fp, addr, size, offset, store); 164 } 165 166 end: 167 fclose(fp); 168 return err; 169 } 170 171 static unsigned long long parse_hex(char *p, int *n) { 172 unsigned long long r=0; 173 int i=0; 174 *n = 0; 175 while(p[i]==' '||p[i]=='\t') i++; 176 if (p[i]==0) return 0; 177 if (p[i+1]=='x') i+=2; 178 int v; 179 while(p[i]) { 180 if (p[i]>='0'&&p[i]<='9') v=p[i]-'0'; 181 else if (p[i]>='a'&&p[i]<='f') v=10+p[i]-'a'; 182 else if (p[i]>='A'&&p[i]<='F') v=10+p[i]-'A'; 183 else break; 184 r=(r<<4)+v; 185 i++; 186 } 187 *n = i; 188 return r; 189 } 190 191 STORE_T* load_symbol_pid(int pid) { 192 printf("loading symbols for %d\n", pid); 193 char bb[256]; 194 sprintf(bb, "/proc/%d/maps", pid); 195 FILE* fp = fopen(bb, "r"); 196 if (fp==NULL) return NULL; 197 STORE_T *store = new STORE_T(); 198 unsigned long long start, end, offset, inode; 199 char *p; 200 int i, c, j; 201 char fname[128], xx[64], xxx[32], mod[16], idx[16]; 202 while(1) { 203 p=fgets(bb, sizeof(bb), fp); if (p==NULL) break; 204 if (sscanf(p, "%s %s %s %s %lld %s", xx, mod, xxx, idx, &inode, fname)!=6) continue; 205 i=0; c=0; 206 start = parse_hex(xx, &c); if (c==0) continue; i+=c; if (p[i]!='-') continue; i++; 207 end = parse_hex(xx+i, &c); if (c==0) continue; 208 // parse type 209 for (j=0; j<8; j++) if (mod[j]=='x') break; if (j>=8) continue; 210 if (fname[0]!='/') continue; 211 offset = parse_hex(xxx, &c); if (c==0) continue; 212 // remaining should contains '/' indicating this mmap is refering to a file 213 sprintf(bb, "/proc/%d/root%s", pid, fname); 214 load_symbol_from_file(bb, start, end-start, offset, *store); 215 } 216 fclose(fp); 217 if (store->size()==0) { 218 delete store; 219 store = NULL; 220 } 221 return store; 222 } 223 224 /* parse kernel func symbols from /proc/kallsyms */ 225 K_STORE_T* load_kernel() { 226 FILE* fp = fopen("/proc/kallsyms", "r"); 227 if (fp == NULL) return NULL; 228 char *p; 229 unsigned long long addr; 230 int c; 231 K_STORE_T* store = new K_STORE_T(); 232 char bb[128], adr[128], type[8], name[128]; 233 while(1) { 234 p = fgets(bb, sizeof(bb), fp); if (p==NULL) break; 235 if (sscanf(p, "%s %s %s", adr, type, name)!=3) continue;; 236 if (type[0]!='t'&&type[0]!='T') continue; 237 addr=parse_hex(adr, &c); if (c==0) continue; 238 (*store)[addr] = string(name); 239 } 240 return store; 241 fclose(fp); 242 } 243 244 //------------------------------perf profiler------------------------- 245 static long perf_event_open(struct perf_event_attr *perf_event, pid_t pid, int cpu, int group_fd, unsigned long flags) { 246 return syscall(__NR_perf_event_open, perf_event, pid, cpu, group_fd, flags); 247 } 248 unordered_map<int, STORE_T*> pid_symbols; 249 K_STORE_T* kernel_symbols = NULL; 250 251 struct pollfd polls[MAXCPU]; 252 // res for cleanup 253 static long long psize; 254 map<int, pair<void*, long long>> res; 255 TNode* gnode = NULL; 256 257 unordered_map<unsigned long long, string> unknowns; 258 void int_exit(int _) { 259 for (auto x: res) { 260 auto y = x.second; 261 void* addr = y.first; 262 munmap(addr, (1+MAXN)*psize); 263 close(x.first); 264 } 265 res.clear(); 266 if (gnode!=NULL) { 267 FILE* fp = fopen("./report.html", "w"); 268 if (fp) { 269 fprintf(fp, "<head> <link rel=\"stylesheet\" href=\"report.css\"> <script src=\"report.js\"> </script> </head>\n"); 270 fprintf(fp, "<ul class=\"tree\">\n"); 271 gnode->printit(fp, 0); 272 fprintf(fp, "</ul>\n"); 273 fclose(fp); 274 printf("report done\n"); 275 } 276 gnode = NULL; 277 } 278 printf("---------------------unknowns-----------------\n"); 279 for (auto x=unknowns.begin(); x!=unknowns.end(); x++) { 280 printf("0x%llx --?> %s\n", (*x).first, (*x).second.c_str()); 281 } 282 exit(0); 283 } 284 /* 285 perf call chain process 286 For now, if a address would not be located to some function, the address would be skipped. 287 */ 288 int process_event(char *base, unsigned long long size, unsigned long long offset) { 289 struct perf_event_header* p = NULL; 290 int pid, xpid; 291 unsigned long long time; 292 offset%=size; 293 // assuming the header would fit within size 294 p = (struct perf_event_header*) (base+offset); 295 offset+=sizeof(*p); if (offset>=size) offset-=size; 296 if (p->type != PERF_RECORD_SAMPLE) return p->size; 297 // pid, tip; 298 pid = *((int *)(base+offset)); offset+=8; if (offset>=size) offset-=size; 299 unsigned long long nr = *((unsigned long long*)(base+offset)); offset+=8; if (offset>=size) offset-=size; 300 if (nr>128) return -1; 301 unsigned long long addr, o, addr0; 302 if (nr) { 303 if (gnode==NULL) gnode=new TNode(); 304 char bb[64]; 305 TNode* r = gnode; 306 if (pid_symbols.count(pid)==0) pid_symbols[pid] = load_symbol_pid(pid); 307 STORE_T* px = pid_symbols[pid]; 308 addr0 = *((unsigned long long *)(base+offset)); 309 char user_mark = 0; 310 for (int i=nr-1; i>=0; i--) { 311 o = i*8+offset; if (o>=size) o-=size; 312 addr = *((unsigned long long*)(base+o)); 313 if (addr==0) continue; // something wrong? 314 if ((addr>>56)==(addr0>>56) && (p->misc&PERF_RECORD_MISC_KERNEL)) { 315 // skip the cross line command, no idear how to correctly resolve it now. 316 if (user_mark) { user_mark=0; continue; } 317 // check in kernel 318 if (kernel_symbols&&!kernel_symbols->empty()) { 319 auto x = kernel_symbols->upper_bound(addr); 320 if (x==kernel_symbols->begin()) { 321 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb)); 322 r = r->add(string("unknown")); 323 } else { 324 x--; 325 r = r->add((*x).second); 326 } 327 } else { 328 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb)); 329 r = r->add(string("unknown")); 330 } 331 } else { 332 if (px) { 333 auto x = px->upper_bound(addr); 334 if (x==px->begin()) { 335 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb)); 336 auto y = (*x).second; 337 r = r->add(y.first+"?"); 338 } else { 339 x--; 340 auto y = (*x).second; 341 if (y.second && addr>(*x).first+y.second) { 342 // r = r->add(y.first); 343 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb)); 344 x++; 345 if (x==px->end()) r = r->add(y.first+"??"); 346 else { 347 auto z = (*x).second; 348 r = r->add(y.first+"?"+z.first); 349 } 350 } else { 351 r = r->add(y.first); 352 } 353 } 354 } else { 355 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb)); 356 // r = r->add(string("unknown")); 357 } 358 user_mark=1; 359 } 360 } 361 } 362 return p->size; 363 } 364 365 int main(int argc, char *argv[]) { 366 kernel_symbols = load_kernel(); 367 if (argc<2) { printf("Need pid\n"); return 1; } 368 int pid = atoi(argv[1]); if (pid<=0) { printf("invalid pid %s\n", argv[1]); return 1; } 369 // find cgroup 370 char xb[256], xb2[256]; 371 int i, j, k, fd; 372 void* addr; 373 sprintf(xb, "/proc/%d/cgroup", pid); 374 FILE* fp = fopen(xb, "r"); 375 if (fp==NULL) error("fail to open cgroup file"); 376 char *p; 377 xb2[0]=0; 378 int cgroup_name_len=0; 379 while(1) { 380 p = fgets(xb, sizeof(xb), fp); if (p==NULL) break; 381 i=0; while(p[i]&&p[i]!=':') i++; if (p[i]==0) continue; 382 if (strstr(p, "perf_event")) { 383 i++; while(p[i]!=':'&&p[i]) i++; if (p[i]!=':') continue; i++; 384 j=i; while(p[j]!='\r'&&p[j]!='\n'&&p[j]!=0) j++; p[j]=0; 385 sprintf(xb2, "/sys/fs/cgroup/perf_event%s", p+i); 386 cgroup_name_len=j-i; 387 break; 388 } else if (p[i+1]==':') { 389 i+=2; j=i; while(p[j]!='\r'&&p[j]!='\n'&&p[j]!=0) j++; p[j]=0; 390 sprintf(xb2, "/sys/fs/cgroup/%s", p+i); 391 cgroup_name_len=j-i; 392 } 393 } 394 fclose(fp); 395 if (xb2[0]==0) error("no proper cgroup found\n"); 396 if (cgroup_name_len<2) { 397 printf("cgroup %s seems to be root, not allowed\n", xb2); 398 return -1; 399 } 400 printf("try to use cgroup %s\n", xb2); 401 int cgroup_id = open(xb2, O_CLOEXEC); 402 if (cgroup_id<=0) { perror("error open cgroup dir"); return 1; } 403 // start perf event 404 psize = sysconf(_SC_PAGE_SIZE); // getpagesize(); 405 int cpu_num = sysconf(_SC_NPROCESSORS_ONLN); 406 struct perf_event_attr attr; 407 memset(&attr, 0, sizeof(attr)); 408 attr.type = PERF_TYPE_SOFTWARE; 409 attr.size = sizeof(attr); 410 attr.config = PERF_COUNT_SW_CPU_CLOCK; 411 attr.sample_freq = 777; // adjust it 412 attr.freq = 1; 413 attr.wakeup_events = 16; 414 attr.sample_type = PERF_SAMPLE_TID|PERF_SAMPLE_CALLCHAIN; 415 for (i=0, k=0; i<cpu_num&&i<MAXCPU; i++) { 416 printf("attaching cpu %d\n", i); 417 fd = perf_event_open(&attr, cgroup_id, i, -1, PERF_FLAG_FD_CLOEXEC|PERF_FLAG_PID_CGROUP); 418 if (fd<0) { perror("fail to open perf event"); continue; } 419 addr = mmap(NULL, (1+MAXN)*psize, PROT_READ, MAP_SHARED, fd, 0); 420 if (addr == MAP_FAILED) { perror("mmap failed"); close(fd); continue; } 421 res[fd] = make_pair(addr, 0); 422 polls[k].fd = fd; 423 polls[k].events = POLLIN; 424 polls[k].revents = 0; 425 k++; 426 } 427 if (k==0) { printf("no cpu event attached at all\n"); return 1; } 428 429 signal(SIGINT, int_exit); 430 signal(SIGTERM, int_exit); 431 432 unsigned long long head; 433 int event_size; 434 struct perf_event_mmap_page *mp; 435 while (poll(polls, k, -1)>0) { 436 // printf("wake\n"); 437 for (i=0; i<k; i++) { 438 if ((polls[i].revents&POLLIN)==0) continue; 439 fd = polls[i].fd; 440 addr = res[fd].first; 441 mp = (struct perf_event_mmap_page *)addr; 442 head = res[fd].second; 443 ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, 1); 444 if (head>mp->data_head) head=mp->data_head; 445 head = mp->data_head-((mp->data_head-head)%mp->data_size); 446 while(head<mp->data_head) { 447 event_size = process_event((char*)addr+mp->data_offset, mp->data_size, head); 448 if (event_size<0) { 449 // resync 450 head=mp->data_head; 451 break; 452 } 453 head += event_size; 454 } 455 res[fd].second = mp->data_head; 456 ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, 0); 457 } 458 } 459 460 int_exit(0); 461 return 0; 462 } 463