xref: /linux-tools/perf/profiler/profiler-kprobe.cpp (revision 1009c19dc94f77e7a7c049e186f796e4881717f9)
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;
32     struct TNode *add(string n) {
33         c++;
34         if (s[n]==nullptr) s[n] = new TNode();
35         return s[n];
36     }
37     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  */
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                     // printf("0x%llx +%lld > %s\n", faddr, fsize, fname);
139                     store[faddr] = make_pair(string(fname), fsize);
140                 }
141                 break;
142             default:
143                 break;
144         }
145     }
146 }
147 
148 int load_symbol_from_file(const char *path, unsigned long long addr, unsigned long long size, unsigned long long offset, STORE_T& store) {
149     printf("loading symble from %s\n", path);
150     FILE *fp = fopen(path, "rb");
151     if (fp==NULL) { perror("fail to open file"); return -1; }
152     char ident[EI_NIDENT], c;
153     int err=0;
154     int rc = fread(ident, sizeof(ident), 1, fp);
155     if (rc != 1) { perror("fail to read ident"); err=-1; goto end; }
156     if (ident[0]!=0x7f) { printf("not a elf file\n"); err=-1; goto  end; }
157     c=ident[4];
158     rc = fseek(fp, 0, SEEK_SET); if (rc<0) { perror("fail to rewind"); goto end; }
159     if (c == ELFCLASS32) {
160         printf("32bit elf not supported yet\n"); err=-2; goto end;
161     } else if (c == ELFCLASS64) {
162         parse_elf64(fp, addr, size, offset, store);
163     } else {
164         printf("unknown elf type %d\n", c);
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, STORE_T* in) {
193     printf("loading symbols for %d\n", pid);
194     char bb[128];
195     sprintf(bb, "/proc/%d/maps", pid);
196     FILE* fp = fopen(bb, "r");
197     if (fp==NULL) return NULL;
198     STORE_T *store = in;
199     unsigned long long start, end, offset=0;
200     if (store==NULL) store=new STORE_T();
201     char *p;
202     int i, c, j;
203     while(1) {
204         p=fgets(bb, sizeof(bb), fp); if (p==NULL) break;
205         i=0; c=0;
206         start = parse_hex(p, &c); if (start==0) continue; i+=c; if (p[i]!='-') continue; i++;
207         end = parse_hex(p+i, &c); if (end==0) continue; i+=c;
208         // parse type
209         for (j=0; j<8; j++) { if (p[i]=='x') break; i++; } if (j>=8) continue;
210         while(p[i]!=' '&&p[i]!='\t'&&p[i]!=0) i++; if (p[i]==0) continue;
211         offset = parse_hex(p+i, &c); if (c==0) continue;
212         // remaining should contains '/' indicating this mmap is refering to a file
213         while(p[i]&&p[i]!='/') i++; if (p[i]==0) continue;
214         sprintf(bb, "/proc/%d/map_files/%llx-%llx", pid, start, end);
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 unordered_map<int, pair<string,string>> pid_infos;
247 K_STORE_T* kernel_symbols = NULL;
248 
249 struct pollfd polls[MAXCPU];
250 // res for cleanup
251 static long long psize;
252 map<int, pair<void*, long long>> res;
253 TNode* gnode = NULL;
254 int exiting = 0;
255 
256 void int_exit(int x) {
257     exiting = 1;
258     for (auto x: res) {
259         auto y = x.second;
260         void* addr = y.first;
261         munmap(addr, (1+MAXN)*psize);
262         close(x.first);
263     }
264     res.clear();
265     if (gnode!=NULL) {
266         FILE* fp = fopen("./report.html", "w");
267         if (fp) {
268             fprintf(fp, "<head> <link rel=\"stylesheet\" href=\"report.css\"> <script src=\"report.js\"> </script> </head>\n");
269             fprintf(fp, "<ul class=\"tree\">\n");
270             gnode->printit(fp, 0);
271             fprintf(fp, "</ul>\n");
272             fclose(fp);
273             printf("report done\n");
274         }
275         gnode = NULL;
276     }
277     exit(x);
278 }
279 /*
280 perf call chain process
281 For now, if a address would not be located to some function, the address would be skipped.
282  */
283 int process_event(char *base, unsigned long long size, unsigned long long offset) {
284     struct perf_event_header* p = NULL;
285     int pid, xpid;
286     unsigned long long time;
287     offset%=size;
288     // assuming the header would fit within size
289     p = (struct perf_event_header*) (base+offset);
290     offset+=sizeof(*p); if (offset>=size) offset-=size;
291     if (p->type != PERF_RECORD_SAMPLE) return p->size;
292     // pid, tip;
293     pid = *((int *)(base+offset));  offset+=8; if (offset>=size) offset-=size;
294     unsigned long long nr = *((unsigned long long*)(base+offset)); offset+=8; if (offset>=size) offset-=size;
295     unsigned long long addr, o, addr0;
296     if (nr) {
297         if (gnode==NULL) gnode=new TNode();
298         char bb[64];
299         TNode* r = gnode;
300         if (pid_symbols.count(pid)==0) {
301             pid_symbols[pid] = load_symbol_pid(pid, NULL);
302             // load command
303             char b[128];
304             char comm[128]="NA";
305             char host[128]="NA";
306             sprintf(b, "/proc/%d/comm", pid);
307             FILE* fp = fopen(b, "r");
308             if (fp) {
309                 fscanf(fp, "%s", comm);
310                 fclose(fp);
311             }
312             sprintf(b, "/proc/%d/root/etc/hostname", pid);
313             fp = fopen(b, "r");
314             if (fp) {
315                 fscanf(fp, "%s", host);
316                 fclose(fp);
317             }
318             pid_infos[pid]=make_pair(string(comm), string(host));
319         }
320         STORE_T* px = pid_symbols[pid];
321         addr0 = *((unsigned long long *)(base+offset));
322         char user_mark = 0;
323         char need_reload= 0;
324         auto  pinfo = pid_infos[pid];
325         for (int i=nr-1; i>=0; i--) {
326             o = i*8+offset; if (o>=size) o-=size;
327             addr = *((unsigned long long*)(base+o));
328             if ((addr>>56)==(addr0>>56) && (p->misc&PERF_RECORD_MISC_KERNEL)) {
329                 // skip the cross line command, no idear how to correctly resolve it now.
330                 if (user_mark) { user_mark=0; continue; }
331                 // check in kernel
332                 if (kernel_symbols&&!kernel_symbols->empty()) {
333                     auto x = kernel_symbols->upper_bound(addr);
334                     if (x==kernel_symbols->begin()) {
335                         // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
336                     } else {
337                         x--;
338                         r = r->add((*x).second);
339                     }
340                 } else {
341                     // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
342                 }
343             } else {
344                 if (px) {
345                     auto x = px->upper_bound(addr);
346                     if (x==px->begin()) {
347                         // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
348                         r = r->add(string("unknwon[")+pinfo.first+'@'+pinfo.second+"]");
349                     } else {
350                         x--;
351                         auto y = (*x).second;
352                         if (addr>(*x).first+y.second) {
353                             // r = r->add(y.first);
354                             // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
355                             r = r->add(string("unknwon[")+pinfo.first+'@'+pinfo.second+"]");
356                         } else {
357                             r = r->add(y.first+"["+pinfo.first+'@'+pinfo.second+"]");
358                         }
359                     }
360                 } else {
361                     // sprintf(bb, "0x%llx", addr); r = r->add(string(bb));
362                 }
363                 user_mark=1;
364             }
365         }
366     }
367     return p->size;
368 }
369 
370 int main(int argc, char *argv[]) {
371     kernel_symbols = load_kernel();
372     if (argc<2) { printf("Need kprobe function name, e.g. %s <some rarely called kernel function>\n", argv[0]); return 1; }
373     int type;
374     char *func = argv[1];
375     FILE *fp = fopen("/sys/bus/event_source/devices/kprobe/type", "r");
376     if (fp == NULL) { printf("fail to find type for kprobe\n"); return 1; }
377     type = 0;
378     fscanf(fp, "%d", &type);
379     fclose(fp);
380     if (type <= 0) { printf("unexpected type %d\n", type); return 1; }
381     // start perf event
382     psize = sysconf(_SC_PAGE_SIZE); // getpagesize();
383     int cpu_num = sysconf(_SC_NPROCESSORS_ONLN);
384 	struct perf_event_attr attr;
385     memset(&attr, 0, sizeof(attr));
386     attr.type = type;
387     attr.size = sizeof(attr);
388     attr.config = 0; // (1<<0) for kreprobe
389     attr.sample_period = 2;
390     attr.wakeup_events = 2;
391     attr.sample_type = PERF_SAMPLE_TID|PERF_SAMPLE_CALLCHAIN;
392     attr.kprobe_func = (__u64)func; // "do_sys_open"; // "bprm_execve";
393     attr.probe_offset = 0;
394     int fd, i, k;
395     void* addr;
396     for (i=0, k=0; i<cpu_num&&i<MAXCPU; i++) {
397         printf("attaching cpu %d\n", i);
398         fd = perf_event_open(&attr, -1, i, -1, PERF_FLAG_FD_CLOEXEC);
399         if (fd<0) { perror("fail to open perf event"); continue; }
400         addr = mmap(NULL, (1+MAXN)*psize, PROT_READ, MAP_SHARED, fd, 0);
401         if (addr == MAP_FAILED) { perror("mmap failed"); close(fd); continue; }
402         res[fd] = make_pair(addr, 0);
403         polls[k].fd = fd;
404         polls[k].events = POLLIN;
405         polls[k].revents = 0;
406         k++;
407     }
408     if (k==0) { printf("no cpu event attached at all\n"); return 1; }
409 
410 	signal(SIGINT, int_exit);
411 	signal(SIGTERM, int_exit);
412 
413     unsigned long long head;
414     struct perf_event_mmap_page *mp;
415     while (poll(polls, k, -1)>0) {
416         for (i=0; i<k; i++) {
417             if (exiting) break;
418             if ((polls[i].revents&POLLIN)==0) continue;
419             fd = polls[i].fd;
420             addr = res[fd].first;
421             mp = (struct perf_event_mmap_page *)addr;
422             head = res[fd].second;
423             if (head==mp->data_head) continue;
424             ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, 1);
425             head = mp->data_head-((mp->data_head-head)%mp->data_size);
426             while(head<mp->data_head) head+=process_event((char*)addr+mp->data_offset, mp->data_size, head);
427             ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, 0);
428             res[fd].second = mp->data_head;
429         }
430     }
431 
432     int_exit(0);
433     return 0;
434 }
435