1 // SPDX-License-Identifier: GPL-2.0+
3 * Copyright 2023 Google LLC
4 * Written by Simon Glass <sjg@chromium.org>
8 * Decode and dump U-Boot trace information into formats that can be used
9 * by trace-cmd, kernelshark or flamegraph.pl
11 * See doc/develop/trace.rst for more information
23 #include <sys/param.h>
24 #include <sys/types.h>
30 #include <linux/list.h>
32 /* Set to 1 to emit version 7 file (currently this doesn't work) */
35 /* enable some debug features */
38 /* from linux/kernel.h */
39 #define __ALIGN_MASK(x, mask) (((x) + (mask)) & ~(mask))
40 #define ALIGN(x, a) __ALIGN_MASK((x), (typeof(x))(a) - 1)
43 * container_of - cast a member of a structure out to the containing structure
44 * @ptr: the pointer to the member.
45 * @type: the type of the container struct this is embedded in.
46 * @member: the name of the member within the struct.
48 * (this is needed by list.h)
50 #define container_of(ptr, type, member) ({ \
51 const typeof( ((type *)0)->member ) *__mptr = (ptr); \
52 (type *)( (char *)__mptr - offsetof(type,member) );})
55 FUNCF_TRACE = 1 << 0, /* Include this function in trace */
56 TRACE_PAGE_SIZE = 4096, /* Assumed page size for trace */
57 TRACE_PID = 1, /* PID to use for U-Boot */
58 LEN_STACK_SIZE = 4, /* number of nested length fix-ups */
59 TRACE_PAGE_MASK = TRACE_PAGE_SIZE - 1,
60 MAX_STACK_DEPTH = 50, /* Max nested function calls */
61 MAX_LINE_LEN = 500, /* Max characters per line */
65 * enum out_format_t - supported output formats
67 * @OUT_FMT_DEFAULT: Use the default for the output file
68 * @OUT_FMT_FUNCTION: Write ftrace 'function' records
69 * @OUT_FMT_FUNCGRAPH: Write ftrace funcgraph_entry and funcgraph_exit records
70 * @OUT_FMT_FLAMEGRAPH_CALLS: Write a file suitable for flamegraph.pl
71 * @OUT_FMT_FLAMEGRAPH_TIMING: Write a file suitable for flamegraph.pl with the
72 * counts set to the number of microseconds used by each function
78 OUT_FMT_FLAMEGRAPH_CALLS,
79 OUT_FMT_FLAMEGRAPH_TIMING,
82 /* Section types for v7 format (trace-cmd format) */
87 /* Option types (trace-cmd format) */
106 /* types of trace records (trace-cmd format) */
108 __TRACE_FIRST_TYPE = 0,
124 * struct flame_node - a node in the call-stack tree
126 * Each stack frame detected in the trace is given a node corresponding to a
127 * function call in the call stack. Functions can appear multiple times when
128 * they are called by a different set of parent functions.
130 * @parent: Parent node (the call stack for the function that called this one)
131 * @child_head: List of children of this node (functions called from here)
132 * @sibling: Next node in the list of children
133 * @func: Function this node refers to (NULL for root node)
134 * @count: Number of times this call-stack occurred
135 * @duration: Number of microseconds taken to run this function, excluding all
136 * of the functions it calls
139 struct flame_node *parent;
140 struct list_head child_head;
141 struct list_head sibling_node;
142 struct func_info *func;
148 * struct flame_state - state information for building the flame graph
150 * @node: Current node being processed (corresponds to a function call)
151 * @stack: Stack of call-start time for this function as well as the
152 * accumulated total time of all child calls (so we can subtract them from the
153 * function's call time. This is an 'empty' stack, meaning that @stack_ptr
154 * points to the next available stack position
155 * @stack_ptr: points to first empty position in the stack
156 * @nodes: Number of nodes created (running count)
159 struct flame_node *node;
163 } stack[MAX_STACK_DEPTH];
169 * struct func_info - information recorded for each function
171 * @offset: Function offset in the image, measured from the text_base
172 * @name: Function name
173 * @code_size: Total code size of the function
174 * @flags: Either 0 or FUNCF_TRACE
177 unsigned long offset;
179 unsigned long code_size;
184 * enum trace_line_type - whether to include or exclude a function
186 * @TRACE_LINE_INCLUDE: Include the function
187 * @TRACE_LINE_EXCLUDE: Exclude the function
189 enum trace_line_type {
195 * struct trace_configline_info - information about a config-file line
199 * @name: identifier name / wildcard
200 * @regex: Regex to use if name starts with '/'
202 struct trace_configline_info {
203 struct trace_configline_info *next;
204 enum trace_line_type type;
210 * struct tw_len - holds information about a length value that need fix-ups
212 * This is used to record a placeholder for a u32 or u64 length which is written
213 * to the output file but needs to be updated once the length is actually known
215 * This allows us to write tw->ptr - @len_base to position @ptr in the file
217 * @ptr: Position of the length value in the file
218 * @base: Base position for the calculation
219 * @size: Size of the length value, in bytes (4 or 8)
228 * struct twriter - Writer for trace records
230 * Maintains state used when writing the output file in trace-cmd format
232 * @ptr: Current file position
233 * @len_stack: Stack of length values that need fixing up
234 * @len: Number of items on @len_stack
235 * @str_buf: Buffer of strings (for v7 format)
236 * @str_ptr: Current write-position in the buffer for strings
241 struct tw_len len_stack[LEN_STACK_SIZE];
248 /* The contents of the trace config file */
249 struct trace_configline_info *trace_config_head;
251 /* list of all functions in System.map file, sorted by offset in the image */
252 struct func_info *func_list;
254 int func_count; /* number of functions */
255 struct trace_call *call_list; /* list of all calls in the input trace file */
256 int call_count; /* number of calls */
257 int verbose; /* Verbosity level 0=none, 1=warn, 2=notice, 3=info, 4=debug */
258 ulong text_offset; /* text address of first function */
259 ulong text_base; /* CONFIG_TEXT_BASE from trace file */
261 /* debugging helpers */
262 static void outf(int level, const char *fmt, ...)
263 __attribute__ ((format (__printf__, 2, 3)));
264 #define error(fmt, b...) outf(0, fmt, ##b)
265 #define warn(fmt, b...) outf(1, fmt, ##b)
266 #define notice(fmt, b...) outf(2, fmt, ##b)
267 #define info(fmt, b...) outf(3, fmt, ##b)
268 #define debug(fmt, b...) outf(4, fmt, ##b)
270 static void outf(int level, const char *fmt, ...)
272 if (verbose >= level) {
276 vfprintf(stderr, fmt, args);
281 static void usage(void)
284 "Usage: proftool [-cmtv] <cmd> <profdata>\n"
287 " dump-ftrace\t\tDump out records in ftrace format for use by trace-cmd\n"
288 " dump-flamegraph\tWrite a file for use with flamegraph.pl\n"
291 " -c <cfg>\tSpecify config file\n"
292 " -f <subtype>\tSpecify output subtype\n"
293 " -m <map>\tSpecify Systen.map file\n"
294 " -o <fname>\tSpecify output file\n"
295 " -t <fname>\tSpecify trace data file (from U-Boot 'trace calls')\n"
296 " -v <0-4>\tSpecify verbosity\n"
298 "Subtypes for dump-ftrace:\n"
299 " function - write function-call records (caller/callee)\n"
300 " funcgraph - write function entry/exit records (graph)\n"
302 "Subtypes for dump-flamegraph\n"
303 " calls - create a flamegraph of stack frames\n"
304 " timing - create a flamegraph of microseconds for each stack frame\n");
309 * h_cmp_offset - bsearch() function to compare two functions bny their offset
311 * @v1: Pointer to first function (struct func_info)
312 * @v2: Pointer to second function (struct func_info)
313 * Returns: < 0 if v1 offset < v2 offset, 0 if equal, > 0 otherwise
315 static int h_cmp_offset(const void *v1, const void *v2)
317 const struct func_info *f1 = v1, *f2 = v2;
319 return (f1->offset / FUNC_SITE_SIZE) - (f2->offset / FUNC_SITE_SIZE);
323 * read_system_map() - read the System.map file to create a list of functions
325 * This also reads the text_offset value, since we assume that the first text
326 * symbol is at that address
329 * Returns: 0 if OK, non-zero on error
331 static int read_system_map(FILE *fin)
333 unsigned long offset, start = 0;
334 struct func_info *func;
335 char buff[MAX_LINE_LEN];
337 char symname[MAX_LINE_LEN + 1];
341 for (linenum = 1, alloced = func_count = 0;; linenum++) {
344 if (fgets(buff, sizeof(buff), fin))
345 fields = sscanf(buff, "%lx %c %100s\n", &offset,
349 } else if (feof(fin)) {
351 } else if (fields < 2) {
352 error("Map file line %d: invalid format\n", linenum);
356 /* Must be a text symbol */
357 symtype = tolower(symtype);
358 if (symtype != 't' && symtype != 'w')
361 if (func_count == alloced) {
363 func_list = realloc(func_list,
364 sizeof(struct func_info) * alloced);
370 func = &func_list[func_count++];
371 memset(func, '\0', sizeof(*func));
372 func->offset = offset - start;
373 func->name = strdup(symname);
374 func->flags = FUNCF_TRACE; /* trace by default */
376 /* Update previous function's code size */
378 func[-1].code_size = func->offset - func[-1].offset;
380 notice("%d functions found in map file, start addr %lx\n", func_count,
387 static int read_data(FILE *fin, void *buff, int size)
391 err = fread(buff, 1, size, fin);
395 error("Cannot read trace file at pos %lx\n", ftell(fin));
402 * find_func_by_offset() - Look up a function by its offset
404 * @offset: Offset to search for, from text_base
405 * Returns: function, if found, else NULL
407 * This does a fast search for a function given its offset from text_base
410 static struct func_info *find_func_by_offset(uint offset)
412 struct func_info key, *found;
415 found = bsearch(&key, func_list, func_count, sizeof(struct func_info),
422 * find_caller_by_offset() - finds the function which contains the given offset
424 * @offset: Offset to search for, from text_base
425 * Returns: function, if found, else NULL
427 * If the offset falls between two functions, then it is assumed to belong to
428 * the first function (with the lowest offset). This is a way of figuring out
429 * which function owns code at a particular offset
431 static struct func_info *find_caller_by_offset(uint offset)
433 int low; /* least function that could be a match */
434 int high; /* greated function that could be a match */
435 struct func_info key;
438 high = func_count - 1;
440 while (high > low + 1) {
441 int mid = (low + high) / 2;
444 result = h_cmp_offset(&key, &func_list[mid]);
450 return &func_list[mid];
453 return low >= 0 ? &func_list[low] : NULL;
457 * read_calls() - Read the list of calls from the trace data
459 * The calls are stored consecutively in the trace output produced by U-Boot
461 * @fin: File to read from
462 * @count: Number of calls to read
463 * Returns: 0 if OK, -1 on error
465 static int read_calls(FILE *fin, size_t count)
467 struct trace_call *call_data;
470 notice("call count: %zu\n", count);
471 call_list = (struct trace_call *)calloc(count, sizeof(*call_data));
473 error("Cannot allocate call_list\n");
478 call_data = call_list;
479 for (i = 0; i < count; i++, call_data++) {
480 if (read_data(fin, call_data, sizeof(*call_data)))
487 * read_trace() - Read the U-Boot trace file
489 * Read in the calls from the trace file. The function list is ignored at
493 * Returns 0 if OK, non-zero on error
495 static int read_trace(FILE *fin)
497 struct trace_output_hdr hdr;
502 err = read_data(fin, &hdr, sizeof(hdr));
507 text_base = hdr.text_base;
510 case TRACE_CHUNK_FUNCS:
511 /* Ignored at present */
514 case TRACE_CHUNK_CALLS:
515 if (read_calls(fin, hdr.rec_count))
524 * read_map_file() - Read the System.map file
526 * This reads the file into the func_list array
528 * @fname: Filename to read
529 * Returns 0 if OK, non-zero on error
531 static int read_map_file(const char *fname)
536 fmap = fopen(fname, "r");
538 error("Cannot open map file '%s'\n", fname);
542 err = read_system_map(fmap);
549 * read_trace_file() - Open and read the U-Boot trace file
551 * Read in the calls from the trace file. The function list is ignored at
555 * Returns 0 if OK, non-zero on error
557 static int read_trace_file(const char *fname)
562 fprof = fopen(fname, "rb");
564 error("Cannot open trace data file '%s'\n",
568 err = read_trace(fprof);
576 static int regex_report_error(regex_t *regex, int err, const char *op,
581 regerror(err, regex, buf, sizeof(buf));
582 error("Regex error '%s' in %s '%s'\n", buf, op, name);
586 static void check_trace_config_line(struct trace_configline_info *item)
588 struct func_info *func, *end;
591 debug("Checking trace config line '%s'\n", item->name);
592 for (func = func_list, end = func + func_count; func < end; func++) {
593 err = regexec(&item->regex, func->name, 0, NULL, 0);
594 debug(" - regex '%s', string '%s': %d\n", item->name,
596 if (err == REG_NOMATCH)
600 regex_report_error(&item->regex, err, "match",
605 /* It matches, so perform the action */
606 switch (item->type) {
607 case TRACE_LINE_INCLUDE:
608 info(" include %s at %lx\n", func->name,
609 text_offset + func->offset);
610 func->flags |= FUNCF_TRACE;
613 case TRACE_LINE_EXCLUDE:
614 info(" exclude %s at %lx\n", func->name,
615 text_offset + func->offset);
616 func->flags &= ~FUNCF_TRACE;
622 /** check_trace_config() - Check trace-config file, reporting any problems */
623 static void check_trace_config(void)
625 struct trace_configline_info *line;
627 for (line = trace_config_head; line; line = line->next)
628 check_trace_config_line(line);
632 * read_trace_config() - read the trace-config file
634 * This file consists of lines like:
636 * include-func <regex>
637 * exclude-func <regex>
639 * where <regex> is a regular expression matched against function names. It
640 * allows some functions to be dropped from the trace when producing ftrace
643 * @fin: File to process
644 * Returns: 0 if OK, -1 on error
646 static int read_trace_config(FILE *fin)
650 struct trace_configline_info **tailp = &trace_config_head;
652 while (fgets(buff, sizeof(buff), fin)) {
653 int len = strlen(buff);
654 struct trace_configline_info *line;
660 if (len && buff[len - 1] == '\n')
661 buff[len - 1] = '\0';
663 /* skip blank lines and comments */
664 for (s = buff; *s == ' ' || *s == '\t'; s++)
666 if (!*s || *s == '#')
669 line = (struct trace_configline_info *)calloc(1, sizeof(*line));
671 error("Cannot allocate config line\n");
675 tok = strtok_r(s, " \t", &saveptr);
677 error("Invalid trace config data on line %d\n",
681 if (0 == strcmp(tok, "include-func")) {
682 line->type = TRACE_LINE_INCLUDE;
683 } else if (0 == strcmp(tok, "exclude-func")) {
684 line->type = TRACE_LINE_EXCLUDE;
686 error("Unknown command in trace config data line %d\n",
691 tok = strtok_r(NULL, " \t", &saveptr);
693 error("Missing pattern in trace config data line %d\n",
698 err = regcomp(&line->regex, tok, REG_NOSUB);
700 int r = regex_report_error(&line->regex, err,
706 /* link this new one to the end of the list */
707 line->name = strdup(tok);
714 error("Cannot read from trace config file at position %ld\n",
721 static int read_trace_config_file(const char *fname)
726 fin = fopen(fname, "r");
728 error("Cannot open trace_config file '%s'\n", fname);
731 err = read_trace_config(fin);
737 * tputh() - Write a 16-bit little-endian value to a file
739 * @fout: File to write to
740 * @val: Value to write
741 * Returns: number of bytes written (2)
743 static int tputh(FILE *fout, unsigned int val)
746 fputc(val >> 8, fout);
752 * tputl() - Write a 32-bit little-endian value to a file
754 * @fout: File to write to
755 * @val: Value to write
756 * Returns: number of bytes written (4)
758 static int tputl(FILE *fout, ulong val)
761 fputc(val >> 8, fout);
762 fputc(val >> 16, fout);
763 fputc(val >> 24, fout);
769 * tputh() - Write a 64-bit little-endian value to a file
771 * @fout: File to write to
772 * @val: Value to write
773 * Returns: number of bytes written (8)
775 static int tputq(FILE *fout, unsigned long long val)
778 tputl(fout, val >> 32U);
784 * tputh() - Write a string to a file
786 * The string is written without its terminator
788 * @fout: File to write to
789 * @val: Value to write
790 * Returns: number of bytes written
792 static int tputs(FILE *fout, const char *str)
800 * add_str() - add a name string to the string table
802 * This is used by the v7 format
804 * @tw: Writer context
805 * @name: String to write
806 * Returns: Updated value of string pointer, or -1 if out of memory
808 static int add_str(struct twriter *tw, const char *name)
813 len = strlen(name) + 1;
814 str_ptr = tw->str_ptr;
817 if (tw->str_ptr > abuf_size(&tw->str_buf)) {
820 new_size = ALIGN(tw->str_ptr, 4096);
821 if (!abuf_realloc(&tw->str_buf, new_size))
829 * push_len() - Push a new length request onto the stack
831 * @tw: Writer context
832 * @base: Base position of the length calculation
833 * @msg: Indicates the type of caller, for debugging
834 * @size: Size of the length value, either 4 bytes or 8
835 * Returns number of bytes written to the file (=@size on success), -ve on error
837 * This marks a place where a length must be written, covering data that is
838 * about to be written. It writes a placeholder value.
840 * Once the data is written, calling pop_len() will update the placeholder with
841 * the correct length based on how many bytes have been written
843 static int push_len(struct twriter *tw, int base, const char *msg, int size)
847 if (tw->len_count >= LEN_STACK_SIZE) {
848 fprintf(stderr, "Length-stack overflow: %s\n", msg);
851 if (size != 4 && size != 8) {
852 fprintf(stderr, "Length-stack invalid size %d: %s\n", size,
857 lp = &tw->len_stack[tw->len_count++];
862 return size == 8 ? tputq(tw->fout, 0) : tputl(tw->fout, 0);
866 * pop_len() - Update a length value once the length is known
868 * Pops a value of the length stack and updates the file at that position with
869 * the number of bytes written between now and then. Once done, the file is
870 * seeked to the current (tw->ptr) position again, so writing can continue as
873 * @tw: Writer context
874 * @msg: Indicates the type of caller, for debugging
875 * Returns 0 if OK, -1 on error
877 static int pop_len(struct twriter *tw, const char *msg)
882 if (!tw->len_count) {
883 fprintf(stderr, "Length-stack underflow: %s\n", msg);
887 lp = &tw->len_stack[--tw->len_count];
888 if (fseek(tw->fout, lp->ptr, SEEK_SET))
890 len = tw->ptr - lp->base;
891 ret = lp->size == 8 ? tputq(tw->fout, len) : tputl(tw->fout, len);
894 if (fseek(tw->fout, tw->ptr, SEEK_SET))
901 * start_header() - Start a v7 section
903 * Writes a header in v7 format
905 * @tw: Writer context
906 * @id: ID of header to write (SECTION_...)
907 * @flags: Flags value to write
908 * @name: Name of section
909 * Returns: number of bytes written
911 static int start_header(struct twriter *tw, int id, uint flags,
921 lptr += tputh(tw->fout, id);
922 lptr += tputh(tw->fout, flags);
923 str_id = add_str(tw, name);
926 lptr += tputl(tw->fout, str_id);
928 /* placeholder for size */
929 ret = push_len(tw, base, "v7 header", 8);
938 * start_page() - Start a new page of output data
940 * The output is arranged in 4KB pages with a base timestamp at the start of
941 * each. This starts a new page, making sure it is aligned to 4KB in the output
944 * @tw: Writer context
945 * @timestamp: Base timestamp for the page
947 static int start_page(struct twriter *tw, ulong timestamp)
952 /* move to start of next page */
953 start = ALIGN(tw->ptr, TRACE_PAGE_SIZE);
954 ret = fseek(tw->fout, start, SEEK_SET);
956 fprintf(stderr, "Cannot seek to page start\n");
962 tw->ptr += tputq(tw->fout, timestamp);
963 ret = push_len(tw, start + 16, "page", 8);
972 * finish_page() - finish a page
974 * Sets the lengths correctly and moves to the start of the next page
976 * @tw: Writer context
977 * Returns: 0 on success, -1 on error
979 static int finish_page(struct twriter *tw)
983 ret = pop_len(tw, "page");
986 end = ALIGN(tw->ptr, TRACE_PAGE_SIZE);
989 * Write a byte so that the data actually makes to the file, in the case
990 * that we never write any more pages
992 if (tw->ptr != end) {
993 if (fseek(tw->fout, end - 1, SEEK_SET)) {
994 fprintf(stderr, "cannot seek to start of next page\n");
1005 * output_headers() - Output v6 headers to the file
1007 * Writes out the various formats so that trace-cmd and kernelshark can make
1010 * This updates tw->ptr as it goes
1012 * @tw: Writer context
1013 * Returns: 0 on success, -ve on error
1015 static int output_headers(struct twriter *tw)
1017 FILE *fout = tw->fout;
1021 tw->ptr += fprintf(fout, "%c%c%ctracing6%c%c%c", 0x17, 0x08, 0x44,
1022 0 /* terminator */, 0 /* little endian */,
1023 4 /* 32-bit long values */);
1025 /* host-machine page size 4KB */
1026 tw->ptr += tputl(fout, 4 << 10);
1028 tw->ptr += fprintf(fout, "header_page%c", 0);
1030 snprintf(str, sizeof(str),
1031 "\tfield: u64 timestamp;\toffset:0;\tsize:8;\tsigned:0;\n"
1032 "\tfield: local_t commit;\toffset:8;\tsize:8;\tsigned:1;\n"
1033 "\tfield: int overwrite;\toffset:8;\tsize:1;\tsigned:1;\n"
1034 "\tfield: char data;\toffset:16;\tsize:4080;\tsigned:1;\n");
1036 tw->ptr += tputq(fout, len);
1037 tw->ptr += tputs(fout, str);
1040 /* no compression */
1041 tw->ptr += fprintf(fout, "none%cversion%c\n", 0, 0);
1043 ret = start_header(tw, SECTION_OPTIONS, 0, "options");
1045 fprintf(stderr, "Cannot start option header\n");
1049 tw->ptr += tputh(fout, OPTION_DONE);
1050 tw->ptr += tputl(fout, 8);
1051 tw->ptr += tputl(fout, 0);
1052 ret = pop_len(tw, "t7 header");
1054 fprintf(stderr, "Cannot finish option header\n");
1059 tw->ptr += fprintf(fout, "header_event%c", 0);
1060 snprintf(str, sizeof(str),
1061 "# compressed entry header\n"
1062 "\ttype_len : 5 bits\n"
1063 "\ttime_delta : 27 bits\n"
1064 "\tarray : 32 bits\n"
1066 "\tpadding : type == 29\n"
1067 "\ttime_extend : type == 30\n"
1068 "\ttime_stamp : type == 31\n"
1069 "\tdata max type_len == 28\n");
1071 tw->ptr += tputq(fout, len);
1072 tw->ptr += tputs(fout, str);
1074 /* number of ftrace-event-format files */
1075 tw->ptr += tputl(fout, 3);
1077 snprintf(str, sizeof(str),
1081 "\tfield:unsigned short common_type;\toffset:0;\tsize:2;\tsigned:0;\n"
1082 "\tfield:unsigned char common_flags;\toffset:2;\tsize:1;\tsigned:0;\n"
1083 "\tfield:unsigned char common_preempt_count;\toffset:3;\tsize:1;signed:0;\n"
1084 "\tfield:int common_pid;\toffset:4;\tsize:4;\tsigned:1;\n"
1086 "\tfield:unsigned long ip;\toffset:8;\tsize:8;\tsigned:0;\n"
1087 "\tfield:unsigned long parent_ip;\toffset:16;\tsize:8;\tsigned:0;\n"
1089 "print fmt: \" %%ps <-- %%ps\", (void *)REC->ip, (void *)REC->parent_ip\n");
1091 tw->ptr += tputq(fout, len);
1092 tw->ptr += tputs(fout, str);
1094 snprintf(str, sizeof(str),
1095 "name: funcgraph_entry\n"
1098 "\tfield:unsigned short common_type;\toffset:0;\tsize:2;\tsigned:0;\n"
1099 "\tfield:unsigned char common_flags;\toffset:2;\tsize:1;\tsigned:0;\n"
1100 "\tfield:unsigned char common_preempt_count;\toffset:3;\tsize:1;signed:0;\n"
1101 "\tfield:int common_pid;\toffset:4;\tsize:4;\tsigned:1;\n"
1103 "\tfield:unsigned long func;\toffset:8;\tsize:8;\tsigned:0;\n"
1104 "\tfield:int depth;\toffset:16;\tsize:4;\tsigned:1;\n"
1106 "print fmt: \"--> %%ps (%%d)\", (void *)REC->func, REC->depth\n");
1108 tw->ptr += tputq(fout, len);
1109 tw->ptr += tputs(fout, str);
1111 snprintf(str, sizeof(str),
1112 "name: funcgraph_exit\n"
1115 "\tfield:unsigned short common_type;\toffset:0;\tsize:2;\tsigned:0;\n"
1116 "\tfield:unsigned char common_flags;\toffset:2;\tsize:1;\tsigned:0;\n"
1117 "\tfield:unsigned char common_preempt_count;\toffset:3;\tsize:1;signed:0;\n"
1118 "\tfield:int common_pid;\toffset:4;\tsize:4;\tsigned:1;\n"
1120 "\tfield:unsigned long func;\toffset:8;\tsize:8;\tsigned:0;\n"
1121 "\tfield:int depth;\toffset:16;\tsize:4;\tsigned:1;\n"
1122 "\tfield:unsigned int overrun;\toffset:20;\tsize:4;\tsigned:0;\n"
1123 "\tfield:unsigned long long calltime;\toffset:24;\tsize:8;\tsigned:0;\n"
1124 "\tfield:unsigned long long rettime;\toffset:32;\tsize:8;\tsigned:0;\n"
1126 "print fmt: \"<-- %%ps (%%d) (start: %%llx end: %%llx) over: %%d\", (void *)REC->func, REC->depth, REC->calltime, REC->rettime, REC->depth\n");
1128 tw->ptr += tputq(fout, len);
1129 tw->ptr += tputs(fout, str);
1135 * write_symbols() - Write the symbols out
1137 * Writes the symbol information in the following format to mimic the Linux
1138 * /proc/kallsyms file:
1140 * <address> T <name>
1142 * This updates tw->ptr as it goes
1144 * @tw: Writer context
1145 * Returns: 0 on success, -ve on error
1147 static int write_symbols(struct twriter *tw)
1153 ret = push_len(tw, tw->ptr + 4, "syms", 4);
1157 for (i = 0; i < func_count; i++) {
1158 struct func_info *func = &func_list[i];
1160 snprintf(str, sizeof(str), "%016lx T %s\n",
1161 text_offset + func->offset, func->name);
1162 tw->ptr += tputs(tw->fout, str);
1164 ret = pop_len(tw, "syms");
1173 * write_options() - Write the options out
1175 * Writes various options which are needed or useful. We use OPTION_TSC2NSEC
1176 * to indicates that values in the output need to be multiplied by 1000 since
1177 * U-Boot's trace values are in microseconds.
1179 * This updates tw->ptr as it goes
1181 * @tw: Writer context
1182 * Returns: 0 on success, -ve on error
1184 static int write_options(struct twriter *tw)
1186 FILE *fout = tw->fout;
1190 /* trace_printk, 0 for now */
1191 tw->ptr += tputl(fout, 0);
1194 snprintf(str, sizeof(str), "%d u-boot\n", TRACE_PID);
1196 tw->ptr += tputq(fout, len);
1197 tw->ptr += tputs(fout, str);
1199 /* number of CPUs */
1200 tw->ptr += tputl(fout, 1);
1202 tw->ptr += fprintf(fout, "options %c", 0);
1205 tw->ptr += tputh(fout, OPTION_TRACECLOCK);
1206 tw->ptr += tputl(fout, 0);
1209 tw->ptr += tputh(fout, OPTION_UNAME);
1210 snprintf(str, sizeof(str), "U-Boot");
1212 tw->ptr += tputl(fout, len);
1213 tw->ptr += tputs(fout, str);
1216 tw->ptr += tputh(fout, OPTION_VERSION);
1217 snprintf(str, sizeof(str), "unknown");
1219 tw->ptr += tputl(fout, len);
1220 tw->ptr += tputs(fout, str);
1223 tw->ptr += tputh(fout, OPTION_TRACEID);
1224 tw->ptr += tputl(fout, 8);
1225 tw->ptr += tputq(fout, 0x123456780abcdef0);
1227 /* time conversion */
1228 tw->ptr += tputh(fout, OPTION_TSC2NSEC);
1229 tw->ptr += tputl(fout, 16);
1230 tw->ptr += tputl(fout, 1000); /* multiplier */
1231 tw->ptr += tputl(fout, 0); /* shift */
1232 tw->ptr += tputq(fout, 0); /* offset */
1234 /* cpustat - bogus data for now, but at least it mentions the CPU */
1235 tw->ptr += tputh(fout, OPTION_CPUSTAT);
1236 snprintf(str, sizeof(str),
1240 "commit overrun: 0\n"
1242 "oldest event ts: 963732.447752\n"
1243 "now ts: 963832.146824\n"
1244 "dropped events: 0\n"
1245 "read events: 42379\n");
1247 tw->ptr += tputl(fout, len);
1248 tw->ptr += tputs(fout, str);
1250 tw->ptr += tputh(fout, OPTION_DONE);
1256 * calc_min_depth() - Calculate the minimum call depth from the call list
1258 * Starting with a depth of 0, this works through the call list, adding 1 for
1259 * each function call and subtracting 1 for each function return. Most likely
1260 * the value ends up being negative, since the trace does not start at the
1261 * very top of the call stack, e.g. main(), but some function called by that.
1263 * This value can be used to calculate the depth value for the first call,
1264 * such that it never goes negative for subsequent returns.
1266 * Returns: minimum call depth (e.g. -2)
1268 static int calc_min_depth(void)
1270 struct trace_call *call;
1271 int depth, min_depth, i;
1273 /* Calculate minimum depth */
1276 for (i = 0, call = call_list; i < call_count; i++, call++) {
1277 switch (TRACE_CALL_TYPE(call)) {
1283 if (depth < min_depth)
1293 * write_pages() - Write the pages of trace data
1295 * This works through all the calls, writing out as many pages of data as are
1298 * @tw: Writer context
1299 * @out_format: Output format to use
1300 * @missing_countp: Returns number of missing functions (not found in function
1302 * @skip_countp: Returns number of skipped functions (excluded from trace)
1304 * Returns: 0 on success, -ve on error
1306 static int write_pages(struct twriter *tw, enum out_format_t out_format,
1307 int *missing_countp, int *skip_countp)
1309 ulong func_stack[MAX_STACK_DEPTH];
1310 int stack_ptr; /* next free position in stack */
1311 int upto, depth, page_upto, i;
1312 int missing_count = 0, skip_count = 0;
1313 struct trace_call *call;
1314 ulong last_timestamp;
1315 FILE *fout = tw->fout;
1326 /* maintain a stack of start times for calling functions */
1330 * The first thing in the trace may not be the top-level function, so
1331 * set the initial depth so that no function goes below depth 0
1333 depth = -calc_min_depth();
1334 for (i = 0, call = call_list; i < call_count; i++, call++) {
1335 bool entry = TRACE_CALL_TYPE(call) == FUNCF_ENTRY;
1336 struct func_info *func;
1341 func = find_func_by_offset(call->func);
1343 warn("Cannot find function at %lx\n",
1344 text_offset + call->func);
1346 if (missing_count > 20) {
1347 /* perhaps trace does not match System.map */
1348 fprintf(stderr, "Too many missing functions\n");
1354 if (!(func->flags & FUNCF_TRACE)) {
1355 debug("Funcion '%s' is excluded from trace\n",
1361 if (out_format == OUT_FMT_FUNCTION)
1363 else /* 2 header words and then 3 or 8 others */
1364 rec_words = 2 + (entry ? 3 : 8);
1366 /* convert timestamp from us to ns */
1367 timestamp = call->flags & FUNCF_TIMESTAMP_MASK;
1369 if (page_upto + rec_words * 4 > TRACE_PAGE_SIZE) {
1370 if (finish_page(tw))
1376 if (start_page(tw, timestamp))
1379 last_timestamp = timestamp;
1381 page_upto = tw->ptr & TRACE_PAGE_MASK;
1384 "new page, last_timestamp=%ld, upto=%d\n",
1385 last_timestamp, upto);
1389 delta = timestamp - last_timestamp;
1391 fprintf(stderr, "Time went backwards\n");
1395 if (err_count > 20) {
1396 fprintf(stderr, "Too many errors, giving up\n");
1400 if (delta > 0x07fffff) {
1402 * hard to imagine how this could happen since it means
1403 * that no function calls were made for a long time
1405 fprintf(stderr, "cannot represent time delta %x\n",
1410 if (out_format == OUT_FMT_FUNCTION) {
1411 struct func_info *caller_func;
1414 fprintf(stderr, "%d: delta=%d, stamp=%ld\n",
1415 upto, delta, timestamp);
1417 " last_delta %x to %x: last_timestamp=%lx, "
1418 "timestamp=%lx, call->flags=%x, upto=%d\n",
1419 last_delta, delta, last_timestamp,
1420 timestamp, call->flags, upto);
1423 /* type_len is 6, meaning 4 * 6 = 24 bytes */
1424 tw->ptr += tputl(fout, rec_words | (uint)delta << 5);
1425 tw->ptr += tputh(fout, TRACE_FN);
1426 tw->ptr += tputh(fout, 0); /* flags */
1427 tw->ptr += tputl(fout, TRACE_PID); /* PID */
1429 tw->ptr += tputq(fout, text_offset + func->offset);
1430 caller_func = find_caller_by_offset(call->caller);
1432 tw->ptr += tputq(fout,
1433 text_offset + caller_func->offset);
1435 tw->ptr += tputl(fout, rec_words | delta << 5);
1436 tw->ptr += tputh(fout, entry ? TRACE_GRAPH_ENT
1438 tw->ptr += tputh(fout, 0); /* flags */
1439 tw->ptr += tputl(fout, TRACE_PID); /* PID */
1441 tw->ptr += tputq(fout, text_offset + func->offset);
1442 tw->ptr += tputl(fout, depth); /* depth */
1445 if (stack_ptr < MAX_STACK_DEPTH)
1446 func_stack[stack_ptr] = timestamp;
1449 ulong func_duration = 0;
1452 if (stack_ptr && stack_ptr <= MAX_STACK_DEPTH) {
1453 ulong start = func_stack[--stack_ptr];
1455 func_duration = timestamp - start;
1457 tw->ptr += tputl(fout, 0); /* overrun */
1458 tw->ptr += tputq(fout, 0); /* calltime */
1460 tw->ptr += tputq(fout, func_duration);
1465 last_timestamp = timestamp;
1466 page_upto += 4 + rec_words * 4;
1468 if (stack_ptr == MAX_STACK_DEPTH)
1471 if (in_page && finish_page(tw))
1473 *missing_countp = missing_count;
1474 *skip_countp = skip_count;
1480 * write_flyrecord() - Write the flyrecord information
1482 * Writes the header and pages of data for the "flyrecord" section. It also
1483 * writes out the counter-type info, selecting "[local]"
1485 * @tw: Writer context
1486 * @out_format: Output format to use
1487 * @missing_countp: Returns number of missing functions (not found in function
1489 * @skip_countp: Returns number of skipped functions (excluded from trace)
1491 * Returns: 0 on success, -ve on error
1493 static int write_flyrecord(struct twriter *tw, enum out_format_t out_format,
1494 int *missing_countp, int *skip_countp)
1496 int start, ret, len;
1497 FILE *fout = tw->fout;
1500 tw->ptr += fprintf(fout, "flyrecord%c", 0);
1503 start = ALIGN(tw->ptr + 16, TRACE_PAGE_SIZE);
1504 tw->ptr += tputq(fout, start);
1506 /* use a placeholder for the size */
1507 ret = push_len(tw, start, "flyrecord", 8);
1512 snprintf(str, sizeof(str),
1513 "[local] global counter uptime perf mono mono_raw boot x86-tsc\n");
1515 tw->ptr += tputq(fout, len);
1516 tw->ptr += tputs(fout, str);
1518 debug("trace text base %lx, map file %lx\n", text_base, text_offset);
1520 ret = write_pages(tw, out_format, missing_countp, skip_countp);
1522 fprintf(stderr, "Cannot output pages\n");
1526 ret = pop_len(tw, "flyrecord");
1528 fprintf(stderr, "Cannot finish flyrecord header\n");
1536 * make_ftrace() - Write out an ftrace file
1538 * See here for format:
1540 * https://github.com/rostedt/trace-cmd/blob/master/Documentation/trace-cmd/trace-cmd.dat.v7.5.txt
1542 * @fout: Output file
1543 * @out_format: Output format to use
1544 * Returns: 0 on success, -ve on error
1546 static int make_ftrace(FILE *fout, enum out_format_t out_format)
1548 int missing_count, skip_count;
1549 struct twriter tws, *tw = &tws;
1552 memset(tw, '\0', sizeof(*tw));
1553 abuf_init(&tw->str_buf);
1557 ret = output_headers(tw);
1559 fprintf(stderr, "Cannot output headers\n");
1562 /* number of event systems files */
1563 tw->ptr += tputl(fout, 0);
1565 ret = write_symbols(tw);
1567 fprintf(stderr, "Cannot write symbols\n");
1571 ret = write_options(tw);
1573 fprintf(stderr, "Cannot write options\n");
1577 ret = write_flyrecord(tw, out_format, &missing_count, &skip_count);
1579 fprintf(stderr, "Cannot write flyrecord\n");
1583 info("ftrace: %d functions not found, %d excluded\n", missing_count,
1590 * create_node() - Create a new node in the flamegraph tree
1592 * @msg: Message to use for debugging if something goes wrong
1593 * Returns: Pointer to newly created node, or NULL on error
1595 static struct flame_node *create_node(const char *msg)
1597 struct flame_node *node;
1599 node = calloc(1, sizeof(*node));
1601 fprintf(stderr, "Out of memory for %s\n", msg);
1604 INIT_LIST_HEAD(&node->child_head);
1610 * process_call(): Add a call to the flamegraph info
1612 * For function calls, if this call stack has been seen before, this increments
1613 * the call count, creating a new node if needed.
1615 * For function returns, it adds up the time spent in this call stack,
1616 * subtracting the time spent by child functions.
1618 * @state: Current flamegraph state
1619 * @entry: true if this is a function entry, false if a function exit
1620 * @timestamp: Timestamp from the trace file (in microseconds)
1621 * @func: Function that was called/returned from
1623 * Returns: 0 on success, -ve on error
1625 static int process_call(struct flame_state *state, bool entry, ulong timestamp,
1626 struct func_info *func)
1628 struct flame_node *node = state->node;
1629 int stack_ptr = state->stack_ptr;
1632 struct flame_node *child, *chd;
1634 /* see if we have this as a child node already */
1636 list_for_each_entry(chd, &node->child_head, sibling_node) {
1637 if (chd->func == func) {
1643 /* create a new node */
1644 child = create_node("child");
1647 list_add_tail(&child->sibling_node, &node->child_head);
1649 child->parent = node;
1652 debug("entry %s: move from %s to %s\n", func->name,
1653 node->func ? node->func->name : "(root)",
1656 if (stack_ptr < MAX_STACK_DEPTH) {
1657 state->stack[stack_ptr].timestamp = timestamp;
1658 state->stack[stack_ptr].child_total = 0;
1660 debug("%d: %20s: entry at %ld\n", stack_ptr, func->name,
1664 } else if (node->parent) {
1665 ulong total_duration = 0, child_duration = 0;
1666 struct stack_info *stk;
1668 debug("exit %s: move from %s to %s\n", func->name,
1669 node->func->name, node->parent->func ?
1670 node->parent->func->name : "(root)");
1671 if (stack_ptr && stack_ptr <= MAX_STACK_DEPTH) {
1672 stk = &state->stack[--stack_ptr];
1675 * get total duration of the function which just
1678 total_duration = timestamp - stk->timestamp;
1679 child_duration = stk->child_total;
1682 state->stack[stack_ptr - 1].child_total += total_duration;
1684 debug("%d: %20s: exit at %ld, total %ld, child %ld, child_total=%ld\n",
1685 stack_ptr, func->name, timestamp,
1686 total_duration, child_duration,
1689 node->duration += total_duration - child_duration;
1690 node = node->parent;
1693 state->stack_ptr = stack_ptr;
1700 * make_flame_tree() - Create a tree of stack traces
1702 * Set up a tree, with the root node having the top-level functions as children
1703 * and the leaf nodes being leaf functions. Each node has a count of how many
1704 * times this function appears in the trace
1706 * @out_format: Output format to use
1707 * @treep: Returns the resulting flamegraph tree
1708 * Returns: 0 on success, -ve on error
1710 static int make_flame_tree(enum out_format_t out_format,
1711 struct flame_node **treep)
1713 struct flame_state state;
1714 struct flame_node *tree;
1715 struct trace_call *call;
1718 /* maintain a stack of start times, etc. for 'calling' functions */
1719 state.stack_ptr = 0;
1721 tree = create_node("tree");
1727 for (i = 0, call = call_list; i < call_count; i++, call++) {
1728 bool entry = TRACE_CALL_TYPE(call) == FUNCF_ENTRY;
1729 ulong timestamp = call->flags & FUNCF_TIMESTAMP_MASK;
1730 struct func_info *func;
1732 func = find_func_by_offset(call->func);
1734 warn("Cannot find function at %lx\n",
1735 text_offset + call->func);
1739 if (process_call(&state, entry, timestamp, func))
1742 fprintf(stderr, "%d nodes\n", state.nodes);
1749 * output_tree() - Output a flamegraph tree
1751 * Writes the tree out to a file in a format suitable for flamegraph.pl
1753 * This works by maintaining a string shared across all recursive calls. The
1754 * function name for this node is added to the existing string, to make up the
1755 * full call-stack description. For example, on entry, @str might contain:
1757 * "initf_bootstage;bootstage_mark_name"
1760 * with @base pointing to the \0 at the end of the string. This function adds
1761 * a ';' following by the name of the current function, e.g. "timer_get_boot_us"
1762 * as well as the output value, to get the full line:
1764 * initf_bootstage;bootstage_mark_name;timer_get_boot_us 123
1766 * @fout: Output file
1767 * @out_format: Output format to use
1768 * @node: Node to output (pass the whole tree at first)
1769 * @str: String to use to build the output line (e.g. 500 charas long)
1770 * @maxlen: Maximum length of string
1771 * @base: Current base position in the string
1772 * @treep: Returns the resulting flamegraph tree
1773 * Returns 0 if OK, -1 on error
1775 static int output_tree(FILE *fout, enum out_format_t out_format,
1776 const struct flame_node *node, char *str, int maxlen,
1779 const struct flame_node *child;
1783 if (out_format == OUT_FMT_FLAMEGRAPH_CALLS) {
1784 fprintf(fout, "%s %d\n", str, node->count);
1787 * Write out the number of microseconds used by this
1788 * call stack. Since the time taken by child calls is
1789 * subtracted from this total, it can reach 0, meaning
1790 * that this function took no time beyond what its
1791 * children used. For this case, write 1 rather than 0,
1792 * so that this call stack appears in the flamegraph.
1793 * This has the effect of inflating the timing slightly,
1794 * but only by at most 1 microsecond per function,
1795 * assuming that is the timestamp resolution
1797 fprintf(fout, "%s %ld\n", str,
1798 node->duration ? node->duration : 1);
1805 list_for_each_entry(child, &node->child_head, sibling_node) {
1808 len = strlen(child->func->name);
1809 if (pos + len + 1 >= maxlen) {
1810 fprintf(stderr, "String too short (%d chars)\n",
1814 strcpy(str + pos, child->func->name);
1815 if (output_tree(fout, out_format, child, str, maxlen,
1824 * make_flamegraph() - Write out a flame graph
1826 * @fout: Output file
1827 * @out_format: Output format to use, e.g. function counts or timing
1828 * Returns 0 if OK, -1 on error
1830 static int make_flamegraph(FILE *fout, enum out_format_t out_format)
1832 struct flame_node *tree;
1835 if (make_flame_tree(out_format, &tree))
1839 if (output_tree(fout, out_format, tree, str, sizeof(str), 0))
1846 * prof_tool() - Performs requested action
1848 * @argc: Number of arguments (used to obtain the command
1849 * @argv: List of arguments
1850 * @trace_fname: Filename of input file (trace data from U-Boot)
1851 * @map_fname: Filename of map file (System.map from U-Boot)
1852 * @trace_config_fname: Trace-configuration file, or NULL if none
1853 * @out_fname: Output filename
1855 static int prof_tool(int argc, char *const argv[],
1856 const char *trace_fname, const char *map_fname,
1857 const char *trace_config_fname, const char *out_fname,
1858 enum out_format_t out_format)
1862 if (read_map_file(map_fname))
1864 if (trace_fname && read_trace_file(trace_fname))
1866 if (trace_config_fname && read_trace_config_file(trace_config_fname))
1869 check_trace_config();
1871 for (; argc; argc--, argv++) {
1872 const char *cmd = *argv;
1874 if (!strcmp(cmd, "dump-ftrace")) {
1877 if (out_format != OUT_FMT_FUNCTION &&
1878 out_format != OUT_FMT_FUNCGRAPH)
1879 out_format = OUT_FMT_FUNCTION;
1880 fout = fopen(out_fname, "w");
1882 fprintf(stderr, "Cannot write file '%s'\n",
1886 err = make_ftrace(fout, out_format);
1888 } else if (!strcmp(cmd, "dump-flamegraph")) {
1891 if (out_format != OUT_FMT_FLAMEGRAPH_CALLS &&
1892 out_format != OUT_FMT_FLAMEGRAPH_TIMING)
1893 out_format = OUT_FMT_FLAMEGRAPH_CALLS;
1894 fout = fopen(out_fname, "w");
1896 fprintf(stderr, "Cannot write file '%s'\n",
1900 err = make_flamegraph(fout, out_format);
1903 warn("Unknown command '%s'\n", cmd);
1910 int main(int argc, char *argv[])
1912 enum out_format_t out_format = OUT_FMT_DEFAULT;
1913 const char *map_fname = "System.map";
1914 const char *trace_fname = NULL;
1915 const char *config_fname = NULL;
1916 const char *out_fname = NULL;
1920 while ((opt = getopt(argc, argv, "c:f:m:o:t:v:")) != -1) {
1923 config_fname = optarg;
1926 if (!strcmp("function", optarg)) {
1927 out_format = OUT_FMT_FUNCTION;
1928 } else if (!strcmp("funcgraph", optarg)) {
1929 out_format = OUT_FMT_FUNCGRAPH;
1930 } else if (!strcmp("calls", optarg)) {
1931 out_format = OUT_FMT_FLAMEGRAPH_CALLS;
1932 } else if (!strcmp("timing", optarg)) {
1933 out_format = OUT_FMT_FLAMEGRAPH_TIMING;
1936 "Invalid format: use function, funcgraph, calls, timing\n");
1947 trace_fname = optarg;
1950 verbose = atoi(optarg);
1956 argc -= optind; argv += optind;
1960 if (!out_fname || !map_fname || !trace_fname) {
1962 "Must provide trace data, System.map file and output file\n");
1966 debug("Debug enabled\n");
1967 return prof_tool(argc, argv, trace_fname, map_fname, config_fname,
1968 out_fname, out_format);