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 <algorithm>
21 using namespace std;
22
23
24 #define MAXN 128
25 #define MAXCPU 1024
26 #define error(msg) do { perror(msg); exit(1); } while(0)
27 //--------------------------------Tree for call chain and report-------------------------------
28 //
29 struct TNode {
30 int c=0;
31 unordered_map<string, TNode*> s;
addTNode32 struct TNode *add(string n) {
33 c++;
34 if (s[n]==nullptr) s[n] = new TNode();
35 return s[n];
36 }
printitTNode37 int printit(FILE *fp, int k) {
38 if (s.size()) {
39 using tt = tuple<int, string, TNode*>;
40 vector<tt> xx;
41 for (auto x: s) xx.push_back(make_tuple(x.second->c, x.first, x.second));
42 sort(begin(xx), end(xx), greater<tt>());
43 for (auto x: xx) {
44 auto count = get<0>(x);
45 if (100.0*count/c<1) continue;
46 auto name = get<1>(x);
47 auto nx = get<2>(x);
48 fprintf(fp, "<li>\n");
49 fprintf(fp, "<input type=\"checkbox\" id=\"c%d\" />\n", k);
50 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);
51 fprintf(fp, "<ul>\n");
52 // printf("%s(%.3f%% %d/%d)\n", name.c_str(), 100.0*count/c, count, c);
53 k = nx->printit(fp, k+1);
54 fprintf(fp, "</ul>\n");
55 fprintf(fp, "</li>\n");
56 }
57 }
58 return k;
59 }
60 };
61
62 //--------------------------------symbols-------------------------------------------
63 using STORE_T = map<unsigned long long, pair<string, unsigned long long>>;
64 using K_STORE_T = map<unsigned long long, string>;
65
66 /*
67 * load FUNC symbols refering to the section indicated by the offset, relocate the virtual address
68 */
parse_elf64(FILE * fp,unsigned long long v_addr,unsigned long long v_size,unsigned long long v_offset,STORE_T & store)69 void parse_elf64(FILE *fp, unsigned long long v_addr, unsigned long long v_size, unsigned long long v_offset, STORE_T& store) {
70 // printf("read elf with offset 0x%llx, addr 0x%llx\n", v_offset, v_addr);
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 store[faddr] = make_pair(string(fname), fsize);
139 }
140 break;
141 default:
142 break;
143 }
144 }
145 }
146
load_symbol_from_file(const char * path,unsigned long long addr,unsigned long long size,unsigned long long offset,STORE_T & store)147 int load_symbol_from_file(const char *path, unsigned long long addr, unsigned long long size, unsigned long long offset, STORE_T& store) {
148 printf("loading symble from %s\n", path);
149 FILE *fp = fopen(path, "rb");
150 if (fp==NULL) { perror("fail to open file"); return -1; }
151 char ident[EI_NIDENT], c;
152 int err=0;
153 int rc = fread(ident, sizeof(ident), 1, fp);
154 if (rc != 1) { perror("fail to read ident"); err=-1; goto end; }
155 if (ident[0]!=0x7f) { printf("not a elf file\n"); err=-1; goto end; }
156 c=ident[4];
157 rc = fseek(fp, 0, SEEK_SET); if (rc<0) { perror("fail to rewind"); goto end; }
158 if (c == ELFCLASS32) {
159 printf("32bit elf not supported yet\n"); err=-2; goto end;
160 } else if (c == ELFCLASS64) {
161 parse_elf64(fp, addr, size, offset, store);
162 }
163
164 end:
165 fclose(fp);
166 return err;
167 }
168
parse_hex(char * p,int * n)169 static unsigned long long parse_hex(char *p, int *n) {
170 unsigned long long r=0;
171 int i=0;
172 *n = 0;
173 while(p[i]==' '||p[i]=='\t') i++;
174 if (p[i]==0) return 0;
175 if (p[i+1]=='x') i+=2;
176 int v;
177 while(p[i]) {
178 if (p[i]>='0'&&p[i]<='9') v=p[i]-'0';
179 else if (p[i]>='a'&&p[i]<='f') v=10+p[i]-'a';
180 else if (p[i]>='A'&&p[i]<='F') v=10+p[i]-'A';
181 else break;
182 r=(r<<4)+v;
183 i++;
184 }
185 *n = i;
186 return r;
187 }
188
load_symbol_pid(int pid)189 STORE_T* load_symbol_pid(int pid) {
190 printf("loading symbols for %d\n", pid);
191 char bb[128];
192 sprintf(bb, "/proc/%d/maps", pid);
193 FILE* fp = fopen(bb, "r");
194 if (fp==NULL) return NULL;
195 STORE_T *store = new STORE_T();
196 unsigned long long start, end, offset;
197 char *p;
198 int i, c, j;
199 while(1) {
200 p=fgets(bb, sizeof(bb), fp); if (p==NULL) break;
201 i=0; c=0;
202 start = parse_hex(p, &c); if (start==0) continue; i+=c; if (p[i]!='-') continue; i++;
203 end = parse_hex(p+i, &c); if (end==0) continue; i+=c;
204 // parse type
205 for (j=0; j<8; j++) { if (p[i]=='x') break; i++; } if (j>=8) continue;
206 while(p[i]!=' '&&p[i]!='\t'&&p[i]!=0) i++; if (p[i]==0) continue;
207 offset = parse_hex(p+i, &c); if (c==0) continue;
208 // remaining should contains '/' indicating this mmap is refering to a file
209 while(p[i]&&p[i]!='/') i++; if (p[i]==0) continue;
210 sprintf(bb, "/proc/%d/map_files/%llx-%llx", pid, start, end);
211 load_symbol_from_file(bb, start, end-start, offset, *store);
212 }
213 fclose(fp);
214 return store;
215 }
216
217 /* parse kernel func symbols from /proc/kallsyms */
load_kernel()218 K_STORE_T* load_kernel() {
219 FILE* fp = fopen("/proc/kallsyms", "r");
220 if (fp == NULL) return NULL;
221 char *p;
222 unsigned long long addr;
223 int c;
224 K_STORE_T* store = new K_STORE_T();
225 char bb[128], adr[128], type[8], name[128];
226 while(1) {
227 p = fgets(bb, sizeof(bb), fp); if (p==NULL) break;
228 if (sscanf(p, "%s %s %s", adr, type, name)!=3) continue;;
229 if (type[0]!='t'&&type[0]!='T') continue;
230 addr=parse_hex(adr, &c); if (c==0) continue;
231 (*store)[addr] = string(name);
232 }
233 return store;
234 fclose(fp);
235 }
236
237 //------------------------------perf profiler-------------------------
perf_event_open(struct perf_event_attr * perf_event,pid_t pid,int cpu,int group_fd,unsigned long flags)238 static long perf_event_open(struct perf_event_attr *perf_event, pid_t pid, int cpu, int group_fd, unsigned long flags) {
239 return syscall(__NR_perf_event_open, perf_event, pid, cpu, group_fd, flags);
240 }
241 unordered_map<int, STORE_T*> pid_symbols;
242 K_STORE_T* kernel_symbols = NULL;
243
244 struct pollfd polls[MAXCPU];
245 // res for cleanup
246 static long long psize;
247 map<int, pair<void*, long long>> res;
248 TNode* gnode = NULL;
249
int_exit(int _)250 void int_exit(int _) {
251 for (auto x: res) {
252 auto y = x.second;
253 void* addr = y.first;
254 munmap(addr, (1+MAXN)*psize);
255 close(x.first);
256 }
257 res.clear();
258 if (gnode!=NULL) {
259 FILE* fp = fopen("./report.html", "w");
260 if (fp) {
261 fprintf(fp, "<head> <link rel=\"stylesheet\" href=\"report.css\"> <script src=\"report.js\"> </script> </head>\n");
262 fprintf(fp, "<ul class=\"tree\">\n");
263 gnode->printit(fp, 0);
264 fprintf(fp, "</ul>\n");
265 fclose(fp);
266 printf("report done\n");
267 }
268 gnode = NULL;
269 }
270 }
271 /*
272 perf call chain process
273 For now, if a address would not be located to some function, the address would be skipped.
274 */
process_event(char * base,unsigned long long size,unsigned long long offset)275 int process_event(char *base, unsigned long long size, unsigned long long offset) {
276 struct perf_event_header* p = NULL;
277 int pid, xpid;
278 unsigned long long time;
279 offset%=size;
280 // assuming the header would fit within size
281 p = (struct perf_event_header*) (base+offset);
282 offset+=sizeof(*p); if (offset>=size) offset-=size;
283 if (p->type != PERF_RECORD_SAMPLE) return p->size;
284 // pid, tip;
285 pid = *((int *)(base+offset)); offset+=8; if (offset>=size) offset-=size;
286 unsigned long long nr = *((unsigned long long*)(base+offset)); offset+=8; if (offset>=size) offset-=size;
287 unsigned long long addr, o, addr0;
288 if (nr) {
289 if (gnode==NULL) gnode=new TNode();
290 char bb[64];
291 TNode* r = gnode;
292 if (pid_symbols.count(pid)==0) pid_symbols[pid] = load_symbol_pid(pid);
293 STORE_T* px = pid_symbols[pid];
294 addr0 = *((unsigned long long *)(base+offset));
295 char user_mark = 0;
296 for (int i=nr-1; i>=0; i--) {
297 o = i*8+offset; if (o>=size) o-=size;
298 addr = *((unsigned long long*)(base+o));
299 if ((addr>>56)==(addr0>>56) && (p->misc&PERF_RECORD_MISC_KERNEL)) {
300 // skip the cross line command, no idear how to correctly resolve it now.
301 if (user_mark) { user_mark=0; continue; }
302 // check in kernel
303 if (kernel_symbols&&!kernel_symbols->empty()) {
304 auto x = kernel_symbols->upper_bound(addr);
305 if (x==kernel_symbols->begin()) {
306 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
307 } else {
308 x--;
309 r = r->add((*x).second);
310 }
311 } else {
312 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
313 }
314 } else {
315 if (px) {
316 auto x = px->upper_bound(addr);
317 if (x==px->begin()) {
318 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
319 } else {
320 x--;
321 auto y = (*x).second;
322 if (addr>(*x).first+y.second) {
323 // r = r->add(y.first);
324 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
325 } else {
326 r = r->add(y.first);
327 }
328 }
329 } else {
330 // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
331 }
332 user_mark=1;
333 }
334 }
335 }
336 return p->size;
337 }
338
main(int argc,char * argv[])339 int main(int argc, char *argv[]) {
340 if (argc != 2) { printf("need pid\n"); return 1; }
341 load_kernel();
342 int pid = atoi(argv[1]); if (pid<=0) { printf("invalid pid %s\n", argv[1]); return 1; }
343 // find cgroup
344 char xb[256], xb2[256];
345 int i, j, k;
346 int fd;
347 void *addr;
348 sprintf(xb, "/proc/%d/cgroup", pid);
349 FILE* fp = fopen(xb, "r");
350 if (fp==NULL) error("fail to open cgroup file");
351 char *p;
352 xb2[0]=0;
353 int cgroup_name_len=0;
354 while(1) {
355 p = fgets(xb, sizeof(xb), fp); if (p==NULL) break;
356 i=0; while(p[i]&&p[i]!=':') i++; if (p[i]==0) continue;
357 if (strstr(p, "perf_event")) {
358 i++; while(p[i]!=':'&&p[i]) i++; if (p[i]!=':') continue; i++;
359 j=i; while(p[j]!='\r'&&p[j]!='\n'&&p[j]!=0) j++; p[j]=0;
360 sprintf(xb2, "/sys/fs/cgroup/perf_event%s", p+i);
361 cgroup_name_len=j-i;
362 break;
363 } else if (p[i+1]==':') {
364 i+=2; j=i; while(p[j]!='\r'&&p[j]!='\n'&&p[j]!=0) j++; p[j]=0;
365 sprintf(xb2, "/sys/fs/cgroup/%s", p+i);
366 cgroup_name_len=j-i;
367 }
368 }
369 fclose(fp);
370 if (xb2[0]==0) error("no proper cgroup found\n");
371 if (cgroup_name_len<2) {
372 printf("cgroup %s seems to be root, not allowed\n", xb2);
373 return -1;
374 }
375 printf("try to use cgroup %s\n", xb2);
376 int cgroup_id = open(xb2, O_CLOEXEC);
377 if (cgroup_id<=0) { perror("error open cgroup dir"); return 1; }
378 // start perf event
379 psize = sysconf(_SC_PAGE_SIZE); // getpagesize();
380 int cpu_num = sysconf(_SC_NPROCESSORS_ONLN);
381 struct perf_event_attr attr;
382 memset(&attr, 0, sizeof(attr));
383 attr.type = PERF_TYPE_SOFTWARE;
384 attr.size = sizeof(attr);
385 attr.config = PERF_COUNT_SW_CONTEXT_SWITCHES;
386 // attr.config = PERF_COUNT_SW_CPU_MIGRATIONS;
387 attr.sample_period = 1;
388 attr.wakeup_events = 32;
389 attr.sample_type = PERF_SAMPLE_TID|PERF_SAMPLE_CALLCHAIN;
390 attr.context_switch = 1;
391 // attr.sample_id_all = 1;
392 for (i=0, k=0; i<cpu_num&&i<MAXCPU; i++) {
393 printf("attaching cpu %d\n", i);
394 fd = perf_event_open(&attr, cgroup_id, i, -1, PERF_FLAG_FD_CLOEXEC|PERF_FLAG_PID_CGROUP);
395 if (fd<0) { perror("fail to open perf event"); continue; }
396 addr = mmap(NULL, (1+MAXN)*psize, PROT_READ, MAP_SHARED, fd, 0);
397 if (addr == MAP_FAILED) { perror("mmap failed"); close(fd); continue; }
398 res[fd] = make_pair(addr, 0);
399 polls[k].fd = fd;
400 polls[k].events = POLLIN;
401 polls[k].revents = 0;
402 k++;
403 }
404 if (k==0) { printf("no cpu event attached at all\n"); return 1; }
405
406 signal(SIGINT, int_exit);
407 signal(SIGTERM, int_exit);
408
409 unsigned long long head;
410 struct perf_event_mmap_page *mp;
411 while (poll(polls, k, -1)>0) {
412 for (i=0; i<k; i++) {
413 if ((polls[i].revents&POLLIN)==0) continue;
414 fd = polls[i].fd;
415 addr = res[fd].first;
416 mp = (struct perf_event_mmap_page *)addr;
417 head = res[fd].second;
418 if (head==mp->data_head) continue;
419 ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, 1);
420 head = mp->data_head-((mp->data_head-head)%mp->data_size);
421 while(head<mp->data_head) head+=process_event((char*)addr+mp->data_offset, mp->data_size, head);
422 ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, 0);
423 res[fd].second = mp->data_head;
424 }
425 }
426
427 int_exit(0);
428 return 0;
429 }
430