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