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
77 /* Section types for v7 format (trace-cmd format) */
82 /* Option types (trace-cmd format) */
101 /* types of trace records (trace-cmd format) */
103 __TRACE_FIRST_TYPE = 0,
119 * struct flame_node - a node in the call-stack tree
121 * Each stack frame detected in the trace is given a node corresponding to a
122 * function call in the call stack. Functions can appear multiple times when
123 * they are called by a different set of parent functions.
125 * @parent: Parent node (the call stack for the function that called this one)
126 * @child_head: List of children of this node (functions called from here)
127 * @sibling: Next node in the list of children
128 * @func: Function this node refers to (NULL for root node)
129 * @count: Number of times this call-stack occurred
132 struct flame_node *parent;
133 struct list_head child_head;
134 struct list_head sibling_node;
135 struct func_info *func;
140 * struct flame_state - state information for building the flame graph
142 * @node: Current node being processed (corresponds to a function call)
143 * @nodes: Number of nodes created (running count)
146 struct flame_node *node;
151 * struct func_info - information recorded for each function
153 * @offset: Function offset in the image, measured from the text_base
154 * @name: Function name
155 * @code_size: Total code size of the function
156 * @flags: Either 0 or FUNCF_TRACE
159 unsigned long offset;
161 unsigned long code_size;
166 * enum trace_line_type - whether to include or exclude a function
168 * @TRACE_LINE_INCLUDE: Include the function
169 * @TRACE_LINE_EXCLUDE: Exclude the function
171 enum trace_line_type {
177 * struct trace_configline_info - information about a config-file line
181 * @name: identifier name / wildcard
182 * @regex: Regex to use if name starts with '/'
184 struct trace_configline_info {
185 struct trace_configline_info *next;
186 enum trace_line_type type;
192 * struct tw_len - holds information about a length value that need fix-ups
194 * This is used to record a placeholder for a u32 or u64 length which is written
195 * to the output file but needs to be updated once the length is actually known
197 * This allows us to write tw->ptr - @len_base to position @ptr in the file
199 * @ptr: Position of the length value in the file
200 * @base: Base position for the calculation
201 * @size: Size of the length value, in bytes (4 or 8)
210 * struct twriter - Writer for trace records
212 * Maintains state used when writing the output file in trace-cmd format
214 * @ptr: Current file position
215 * @len_stack: Stack of length values that need fixing up
216 * @len: Number of items on @len_stack
217 * @str_buf: Buffer of strings (for v7 format)
218 * @str_ptr: Current write-position in the buffer for strings
223 struct tw_len len_stack[LEN_STACK_SIZE];
230 /* The contents of the trace config file */
231 struct trace_configline_info *trace_config_head;
233 /* list of all functions in System.map file, sorted by offset in the image */
234 struct func_info *func_list;
236 int func_count; /* number of functions */
237 struct trace_call *call_list; /* list of all calls in the input trace file */
238 int call_count; /* number of calls */
239 int verbose; /* Verbosity level 0=none, 1=warn, 2=notice, 3=info, 4=debug */
240 ulong text_offset; /* text address of first function */
241 ulong text_base; /* CONFIG_TEXT_BASE from trace file */
243 /* debugging helpers */
244 static void outf(int level, const char *fmt, ...)
245 __attribute__ ((format (__printf__, 2, 3)));
246 #define error(fmt, b...) outf(0, fmt, ##b)
247 #define warn(fmt, b...) outf(1, fmt, ##b)
248 #define notice(fmt, b...) outf(2, fmt, ##b)
249 #define info(fmt, b...) outf(3, fmt, ##b)
250 #define debug(fmt, b...) outf(4, fmt, ##b)
252 static void outf(int level, const char *fmt, ...)
254 if (verbose >= level) {
258 vfprintf(stderr, fmt, args);
263 static void usage(void)
266 "Usage: proftool [-cmtv] <cmd> <profdata>\n"
269 " dump-ftrace\t\tDump out records in ftrace format for use by trace-cmd\n"
270 " dump-flamegraph\tWrite a file for use with flamegraph.pl\n"
273 " -c <cfg>\tSpecify config file\n"
274 " -f <subtype>\tSpecify output subtype\n"
275 " -m <map>\tSpecify Systen.map file\n"
276 " -o <fname>\tSpecify output file\n"
277 " -t <fname>\tSpecify trace data file (from U-Boot 'trace calls')\n"
278 " -v <0-4>\tSpecify verbosity\n"
280 "Subtypes for dump-ftrace:\n"
281 " function - write function-call records (caller/callee)\n"
282 " funcgraph - write function entry/exit records (graph)\n");
287 * h_cmp_offset - bsearch() function to compare two functions bny their offset
289 * @v1: Pointer to first function (struct func_info)
290 * @v2: Pointer to second function (struct func_info)
291 * Returns: < 0 if v1 offset < v2 offset, 0 if equal, > 0 otherwise
293 static int h_cmp_offset(const void *v1, const void *v2)
295 const struct func_info *f1 = v1, *f2 = v2;
297 return (f1->offset / FUNC_SITE_SIZE) - (f2->offset / FUNC_SITE_SIZE);
301 * read_system_map() - read the System.map file to create a list of functions
303 * This also reads the text_offset value, since we assume that the first text
304 * symbol is at that address
307 * Returns: 0 if OK, non-zero on error
309 static int read_system_map(FILE *fin)
311 unsigned long offset, start = 0;
312 struct func_info *func;
313 char buff[MAX_LINE_LEN];
315 char symname[MAX_LINE_LEN + 1];
319 for (linenum = 1, alloced = func_count = 0;; linenum++) {
322 if (fgets(buff, sizeof(buff), fin))
323 fields = sscanf(buff, "%lx %c %100s\n", &offset,
327 } else if (feof(fin)) {
329 } else if (fields < 2) {
330 error("Map file line %d: invalid format\n", linenum);
334 /* Must be a text symbol */
335 symtype = tolower(symtype);
336 if (symtype != 't' && symtype != 'w')
339 if (func_count == alloced) {
341 func_list = realloc(func_list,
342 sizeof(struct func_info) * alloced);
348 func = &func_list[func_count++];
349 memset(func, '\0', sizeof(*func));
350 func->offset = offset - start;
351 func->name = strdup(symname);
352 func->flags = FUNCF_TRACE; /* trace by default */
354 /* Update previous function's code size */
356 func[-1].code_size = func->offset - func[-1].offset;
358 notice("%d functions found in map file, start addr %lx\n", func_count,
365 static int read_data(FILE *fin, void *buff, int size)
369 err = fread(buff, 1, size, fin);
373 error("Cannot read trace file at pos %lx\n", ftell(fin));
380 * find_func_by_offset() - Look up a function by its offset
382 * @offset: Offset to search for, from text_base
383 * Returns: function, if found, else NULL
385 * This does a fast search for a function given its offset from text_base
388 static struct func_info *find_func_by_offset(uint offset)
390 struct func_info key, *found;
393 found = bsearch(&key, func_list, func_count, sizeof(struct func_info),
400 * find_caller_by_offset() - finds the function which contains the given offset
402 * @offset: Offset to search for, from text_base
403 * Returns: function, if found, else NULL
405 * If the offset falls between two functions, then it is assumed to belong to
406 * the first function (with the lowest offset). This is a way of figuring out
407 * which function owns code at a particular offset
409 static struct func_info *find_caller_by_offset(uint offset)
411 int low; /* least function that could be a match */
412 int high; /* greated function that could be a match */
413 struct func_info key;
416 high = func_count - 1;
418 while (high > low + 1) {
419 int mid = (low + high) / 2;
422 result = h_cmp_offset(&key, &func_list[mid]);
428 return &func_list[mid];
431 return low >= 0 ? &func_list[low] : NULL;
435 * read_calls() - Read the list of calls from the trace data
437 * The calls are stored consecutively in the trace output produced by U-Boot
439 * @fin: File to read from
440 * @count: Number of calls to read
441 * Returns: 0 if OK, -1 on error
443 static int read_calls(FILE *fin, size_t count)
445 struct trace_call *call_data;
448 notice("call count: %zu\n", count);
449 call_list = (struct trace_call *)calloc(count, sizeof(*call_data));
451 error("Cannot allocate call_list\n");
456 call_data = call_list;
457 for (i = 0; i < count; i++, call_data++) {
458 if (read_data(fin, call_data, sizeof(*call_data)))
465 * read_trace() - Read the U-Boot trace file
467 * Read in the calls from the trace file. The function list is ignored at
471 * Returns 0 if OK, non-zero on error
473 static int read_trace(FILE *fin)
475 struct trace_output_hdr hdr;
480 err = read_data(fin, &hdr, sizeof(hdr));
485 text_base = hdr.text_base;
488 case TRACE_CHUNK_FUNCS:
489 /* Ignored at present */
492 case TRACE_CHUNK_CALLS:
493 if (read_calls(fin, hdr.rec_count))
502 * read_map_file() - Read the System.map file
504 * This reads the file into the func_list array
506 * @fname: Filename to read
507 * Returns 0 if OK, non-zero on error
509 static int read_map_file(const char *fname)
514 fmap = fopen(fname, "r");
516 error("Cannot open map file '%s'\n", fname);
520 err = read_system_map(fmap);
527 * read_trace_file() - Open and read the U-Boot trace file
529 * Read in the calls from the trace file. The function list is ignored at
533 * Returns 0 if OK, non-zero on error
535 static int read_trace_file(const char *fname)
540 fprof = fopen(fname, "rb");
542 error("Cannot open trace data file '%s'\n",
546 err = read_trace(fprof);
554 static int regex_report_error(regex_t *regex, int err, const char *op,
559 regerror(err, regex, buf, sizeof(buf));
560 error("Regex error '%s' in %s '%s'\n", buf, op, name);
564 static void check_trace_config_line(struct trace_configline_info *item)
566 struct func_info *func, *end;
569 debug("Checking trace config line '%s'\n", item->name);
570 for (func = func_list, end = func + func_count; func < end; func++) {
571 err = regexec(&item->regex, func->name, 0, NULL, 0);
572 debug(" - regex '%s', string '%s': %d\n", item->name,
574 if (err == REG_NOMATCH)
578 regex_report_error(&item->regex, err, "match",
583 /* It matches, so perform the action */
584 switch (item->type) {
585 case TRACE_LINE_INCLUDE:
586 info(" include %s at %lx\n", func->name,
587 text_offset + func->offset);
588 func->flags |= FUNCF_TRACE;
591 case TRACE_LINE_EXCLUDE:
592 info(" exclude %s at %lx\n", func->name,
593 text_offset + func->offset);
594 func->flags &= ~FUNCF_TRACE;
600 /** check_trace_config() - Check trace-config file, reporting any problems */
601 static void check_trace_config(void)
603 struct trace_configline_info *line;
605 for (line = trace_config_head; line; line = line->next)
606 check_trace_config_line(line);
610 * read_trace_config() - read the trace-config file
612 * This file consists of lines like:
614 * include-func <regex>
615 * exclude-func <regex>
617 * where <regex> is a regular expression matched against function names. It
618 * allows some functions to be dropped from the trace when producing ftrace
621 * @fin: File to process
622 * Returns: 0 if OK, -1 on error
624 static int read_trace_config(FILE *fin)
628 struct trace_configline_info **tailp = &trace_config_head;
630 while (fgets(buff, sizeof(buff), fin)) {
631 int len = strlen(buff);
632 struct trace_configline_info *line;
638 if (len && buff[len - 1] == '\n')
639 buff[len - 1] = '\0';
641 /* skip blank lines and comments */
642 for (s = buff; *s == ' ' || *s == '\t'; s++)
644 if (!*s || *s == '#')
647 line = (struct trace_configline_info *)calloc(1, sizeof(*line));
649 error("Cannot allocate config line\n");
653 tok = strtok_r(s, " \t", &saveptr);
655 error("Invalid trace config data on line %d\n",
659 if (0 == strcmp(tok, "include-func")) {
660 line->type = TRACE_LINE_INCLUDE;
661 } else if (0 == strcmp(tok, "exclude-func")) {
662 line->type = TRACE_LINE_EXCLUDE;
664 error("Unknown command in trace config data line %d\n",
669 tok = strtok_r(NULL, " \t", &saveptr);
671 error("Missing pattern in trace config data line %d\n",
676 err = regcomp(&line->regex, tok, REG_NOSUB);
678 int r = regex_report_error(&line->regex, err,
684 /* link this new one to the end of the list */
685 line->name = strdup(tok);
692 error("Cannot read from trace config file at position %ld\n",
699 static int read_trace_config_file(const char *fname)
704 fin = fopen(fname, "r");
706 error("Cannot open trace_config file '%s'\n", fname);
709 err = read_trace_config(fin);
715 * tputh() - Write a 16-bit little-endian value to a file
717 * @fout: File to write to
718 * @val: Value to write
719 * Returns: number of bytes written (2)
721 static int tputh(FILE *fout, unsigned int val)
724 fputc(val >> 8, fout);
730 * tputl() - Write a 32-bit little-endian value to a file
732 * @fout: File to write to
733 * @val: Value to write
734 * Returns: number of bytes written (4)
736 static int tputl(FILE *fout, ulong val)
739 fputc(val >> 8, fout);
740 fputc(val >> 16, fout);
741 fputc(val >> 24, fout);
747 * tputh() - Write a 64-bit little-endian value to a file
749 * @fout: File to write to
750 * @val: Value to write
751 * Returns: number of bytes written (8)
753 static int tputq(FILE *fout, unsigned long long val)
756 tputl(fout, val >> 32U);
762 * tputh() - Write a string to a file
764 * The string is written without its terminator
766 * @fout: File to write to
767 * @val: Value to write
768 * Returns: number of bytes written
770 static int tputs(FILE *fout, const char *str)
778 * add_str() - add a name string to the string table
780 * This is used by the v7 format
782 * @tw: Writer context
783 * @name: String to write
784 * Returns: Updated value of string pointer, or -1 if out of memory
786 static int add_str(struct twriter *tw, const char *name)
791 len = strlen(name) + 1;
792 str_ptr = tw->str_ptr;
795 if (tw->str_ptr > abuf_size(&tw->str_buf)) {
798 new_size = ALIGN(tw->str_ptr, 4096);
799 if (!abuf_realloc(&tw->str_buf, new_size))
807 * push_len() - Push a new length request onto the stack
809 * @tw: Writer context
810 * @base: Base position of the length calculation
811 * @msg: Indicates the type of caller, for debugging
812 * @size: Size of the length value, either 4 bytes or 8
813 * Returns number of bytes written to the file (=@size on success), -ve on error
815 * This marks a place where a length must be written, covering data that is
816 * about to be written. It writes a placeholder value.
818 * Once the data is written, calling pop_len() will update the placeholder with
819 * the correct length based on how many bytes have been written
821 static int push_len(struct twriter *tw, int base, const char *msg, int size)
825 if (tw->len_count >= LEN_STACK_SIZE) {
826 fprintf(stderr, "Length-stack overflow: %s\n", msg);
829 if (size != 4 && size != 8) {
830 fprintf(stderr, "Length-stack invalid size %d: %s\n", size,
835 lp = &tw->len_stack[tw->len_count++];
840 return size == 8 ? tputq(tw->fout, 0) : tputl(tw->fout, 0);
844 * pop_len() - Update a length value once the length is known
846 * Pops a value of the length stack and updates the file at that position with
847 * the number of bytes written between now and then. Once done, the file is
848 * seeked to the current (tw->ptr) position again, so writing can continue as
851 * @tw: Writer context
852 * @msg: Indicates the type of caller, for debugging
853 * Returns 0 if OK, -1 on error
855 static int pop_len(struct twriter *tw, const char *msg)
860 if (!tw->len_count) {
861 fprintf(stderr, "Length-stack underflow: %s\n", msg);
865 lp = &tw->len_stack[--tw->len_count];
866 if (fseek(tw->fout, lp->ptr, SEEK_SET))
868 len = tw->ptr - lp->base;
869 ret = lp->size == 8 ? tputq(tw->fout, len) : tputl(tw->fout, len);
872 if (fseek(tw->fout, tw->ptr, SEEK_SET))
879 * start_header() - Start a v7 section
881 * Writes a header in v7 format
883 * @tw: Writer context
884 * @id: ID of header to write (SECTION_...)
885 * @flags: Flags value to write
886 * @name: Name of section
887 * Returns: number of bytes written
889 static int start_header(struct twriter *tw, int id, uint flags,
899 lptr += tputh(tw->fout, id);
900 lptr += tputh(tw->fout, flags);
901 str_id = add_str(tw, name);
904 lptr += tputl(tw->fout, str_id);
906 /* placeholder for size */
907 ret = push_len(tw, base, "v7 header", 8);
916 * start_page() - Start a new page of output data
918 * The output is arranged in 4KB pages with a base timestamp at the start of
919 * each. This starts a new page, making sure it is aligned to 4KB in the output
922 * @tw: Writer context
923 * @timestamp: Base timestamp for the page
925 static int start_page(struct twriter *tw, ulong timestamp)
930 /* move to start of next page */
931 start = ALIGN(tw->ptr, TRACE_PAGE_SIZE);
932 ret = fseek(tw->fout, start, SEEK_SET);
934 fprintf(stderr, "Cannot seek to page start\n");
940 tw->ptr += tputq(tw->fout, timestamp);
941 ret = push_len(tw, start + 16, "page", 8);
950 * finish_page() - finish a page
952 * Sets the lengths correctly and moves to the start of the next page
954 * @tw: Writer context
955 * Returns: 0 on success, -1 on error
957 static int finish_page(struct twriter *tw)
961 ret = pop_len(tw, "page");
964 end = ALIGN(tw->ptr, TRACE_PAGE_SIZE);
967 * Write a byte so that the data actually makes to the file, in the case
968 * that we never write any more pages
970 if (tw->ptr != end) {
971 if (fseek(tw->fout, end - 1, SEEK_SET)) {
972 fprintf(stderr, "cannot seek to start of next page\n");
983 * output_headers() - Output v6 headers to the file
985 * Writes out the various formats so that trace-cmd and kernelshark can make
988 * This updates tw->ptr as it goes
990 * @tw: Writer context
991 * Returns: 0 on success, -ve on error
993 static int output_headers(struct twriter *tw)
995 FILE *fout = tw->fout;
999 tw->ptr += fprintf(fout, "%c%c%ctracing6%c%c%c", 0x17, 0x08, 0x44,
1000 0 /* terminator */, 0 /* little endian */,
1001 4 /* 32-bit long values */);
1003 /* host-machine page size 4KB */
1004 tw->ptr += tputl(fout, 4 << 10);
1006 tw->ptr += fprintf(fout, "header_page%c", 0);
1008 snprintf(str, sizeof(str),
1009 "\tfield: u64 timestamp;\toffset:0;\tsize:8;\tsigned:0;\n"
1010 "\tfield: local_t commit;\toffset:8;\tsize:8;\tsigned:1;\n"
1011 "\tfield: int overwrite;\toffset:8;\tsize:1;\tsigned:1;\n"
1012 "\tfield: char data;\toffset:16;\tsize:4080;\tsigned:1;\n");
1014 tw->ptr += tputq(fout, len);
1015 tw->ptr += tputs(fout, str);
1018 /* no compression */
1019 tw->ptr += fprintf(fout, "none%cversion%c\n", 0, 0);
1021 ret = start_header(tw, SECTION_OPTIONS, 0, "options");
1023 fprintf(stderr, "Cannot start option header\n");
1027 tw->ptr += tputh(fout, OPTION_DONE);
1028 tw->ptr += tputl(fout, 8);
1029 tw->ptr += tputl(fout, 0);
1030 ret = pop_len(tw, "t7 header");
1032 fprintf(stderr, "Cannot finish option header\n");
1037 tw->ptr += fprintf(fout, "header_event%c", 0);
1038 snprintf(str, sizeof(str),
1039 "# compressed entry header\n"
1040 "\ttype_len : 5 bits\n"
1041 "\ttime_delta : 27 bits\n"
1042 "\tarray : 32 bits\n"
1044 "\tpadding : type == 29\n"
1045 "\ttime_extend : type == 30\n"
1046 "\ttime_stamp : type == 31\n"
1047 "\tdata max type_len == 28\n");
1049 tw->ptr += tputq(fout, len);
1050 tw->ptr += tputs(fout, str);
1052 /* number of ftrace-event-format files */
1053 tw->ptr += tputl(fout, 3);
1055 snprintf(str, sizeof(str),
1059 "\tfield:unsigned short common_type;\toffset:0;\tsize:2;\tsigned:0;\n"
1060 "\tfield:unsigned char common_flags;\toffset:2;\tsize:1;\tsigned:0;\n"
1061 "\tfield:unsigned char common_preempt_count;\toffset:3;\tsize:1;signed:0;\n"
1062 "\tfield:int common_pid;\toffset:4;\tsize:4;\tsigned:1;\n"
1064 "\tfield:unsigned long ip;\toffset:8;\tsize:8;\tsigned:0;\n"
1065 "\tfield:unsigned long parent_ip;\toffset:16;\tsize:8;\tsigned:0;\n"
1067 "print fmt: \" %%ps <-- %%ps\", (void *)REC->ip, (void *)REC->parent_ip\n");
1069 tw->ptr += tputq(fout, len);
1070 tw->ptr += tputs(fout, str);
1072 snprintf(str, sizeof(str),
1073 "name: funcgraph_entry\n"
1076 "\tfield:unsigned short common_type;\toffset:0;\tsize:2;\tsigned:0;\n"
1077 "\tfield:unsigned char common_flags;\toffset:2;\tsize:1;\tsigned:0;\n"
1078 "\tfield:unsigned char common_preempt_count;\toffset:3;\tsize:1;signed:0;\n"
1079 "\tfield:int common_pid;\toffset:4;\tsize:4;\tsigned:1;\n"
1081 "\tfield:unsigned long func;\toffset:8;\tsize:8;\tsigned:0;\n"
1082 "\tfield:int depth;\toffset:16;\tsize:4;\tsigned:1;\n"
1084 "print fmt: \"--> %%ps (%%d)\", (void *)REC->func, REC->depth\n");
1086 tw->ptr += tputq(fout, len);
1087 tw->ptr += tputs(fout, str);
1089 snprintf(str, sizeof(str),
1090 "name: funcgraph_exit\n"
1093 "\tfield:unsigned short common_type;\toffset:0;\tsize:2;\tsigned:0;\n"
1094 "\tfield:unsigned char common_flags;\toffset:2;\tsize:1;\tsigned:0;\n"
1095 "\tfield:unsigned char common_preempt_count;\toffset:3;\tsize:1;signed:0;\n"
1096 "\tfield:int common_pid;\toffset:4;\tsize:4;\tsigned:1;\n"
1098 "\tfield:unsigned long func;\toffset:8;\tsize:8;\tsigned:0;\n"
1099 "\tfield:int depth;\toffset:16;\tsize:4;\tsigned:1;\n"
1100 "\tfield:unsigned int overrun;\toffset:20;\tsize:4;\tsigned:0;\n"
1101 "\tfield:unsigned long long calltime;\toffset:24;\tsize:8;\tsigned:0;\n"
1102 "\tfield:unsigned long long rettime;\toffset:32;\tsize:8;\tsigned:0;\n"
1104 "print fmt: \"<-- %%ps (%%d) (start: %%llx end: %%llx) over: %%d\", (void *)REC->func, REC->depth, REC->calltime, REC->rettime, REC->depth\n");
1106 tw->ptr += tputq(fout, len);
1107 tw->ptr += tputs(fout, str);
1113 * write_symbols() - Write the symbols out
1115 * Writes the symbol information in the following format to mimic the Linux
1116 * /proc/kallsyms file:
1118 * <address> T <name>
1120 * This updates tw->ptr as it goes
1122 * @tw: Writer context
1123 * Returns: 0 on success, -ve on error
1125 static int write_symbols(struct twriter *tw)
1131 ret = push_len(tw, tw->ptr + 4, "syms", 4);
1135 for (i = 0; i < func_count; i++) {
1136 struct func_info *func = &func_list[i];
1138 snprintf(str, sizeof(str), "%016lx T %s\n",
1139 text_offset + func->offset, func->name);
1140 tw->ptr += tputs(tw->fout, str);
1142 ret = pop_len(tw, "syms");
1151 * write_options() - Write the options out
1153 * Writes various options which are needed or useful. We use OPTION_TSC2NSEC
1154 * to indicates that values in the output need to be multiplied by 1000 since
1155 * U-Boot's trace values are in microseconds.
1157 * This updates tw->ptr as it goes
1159 * @tw: Writer context
1160 * Returns: 0 on success, -ve on error
1162 static int write_options(struct twriter *tw)
1164 FILE *fout = tw->fout;
1168 /* trace_printk, 0 for now */
1169 tw->ptr += tputl(fout, 0);
1172 snprintf(str, sizeof(str), "%d u-boot\n", TRACE_PID);
1174 tw->ptr += tputq(fout, len);
1175 tw->ptr += tputs(fout, str);
1177 /* number of CPUs */
1178 tw->ptr += tputl(fout, 1);
1180 tw->ptr += fprintf(fout, "options %c", 0);
1183 tw->ptr += tputh(fout, OPTION_TRACECLOCK);
1184 tw->ptr += tputl(fout, 0);
1187 tw->ptr += tputh(fout, OPTION_UNAME);
1188 snprintf(str, sizeof(str), "U-Boot");
1190 tw->ptr += tputl(fout, len);
1191 tw->ptr += tputs(fout, str);
1194 tw->ptr += tputh(fout, OPTION_VERSION);
1195 snprintf(str, sizeof(str), "unknown");
1197 tw->ptr += tputl(fout, len);
1198 tw->ptr += tputs(fout, str);
1201 tw->ptr += tputh(fout, OPTION_TRACEID);
1202 tw->ptr += tputl(fout, 8);
1203 tw->ptr += tputq(fout, 0x123456780abcdef0);
1205 /* time conversion */
1206 tw->ptr += tputh(fout, OPTION_TSC2NSEC);
1207 tw->ptr += tputl(fout, 16);
1208 tw->ptr += tputl(fout, 1000); /* multiplier */
1209 tw->ptr += tputl(fout, 0); /* shift */
1210 tw->ptr += tputq(fout, 0); /* offset */
1212 /* cpustat - bogus data for now, but at least it mentions the CPU */
1213 tw->ptr += tputh(fout, OPTION_CPUSTAT);
1214 snprintf(str, sizeof(str),
1218 "commit overrun: 0\n"
1220 "oldest event ts: 963732.447752\n"
1221 "now ts: 963832.146824\n"
1222 "dropped events: 0\n"
1223 "read events: 42379\n");
1225 tw->ptr += tputl(fout, len);
1226 tw->ptr += tputs(fout, str);
1228 tw->ptr += tputh(fout, OPTION_DONE);
1234 * calc_min_depth() - Calculate the minimum call depth from the call list
1236 * Starting with a depth of 0, this works through the call list, adding 1 for
1237 * each function call and subtracting 1 for each function return. Most likely
1238 * the value ends up being negative, since the trace does not start at the
1239 * very top of the call stack, e.g. main(), but some function called by that.
1241 * This value can be used to calculate the depth value for the first call,
1242 * such that it never goes negative for subsequent returns.
1244 * Returns: minimum call depth (e.g. -2)
1246 static int calc_min_depth(void)
1248 struct trace_call *call;
1249 int depth, min_depth, i;
1251 /* Calculate minimum depth */
1254 for (i = 0, call = call_list; i < call_count; i++, call++) {
1255 switch (TRACE_CALL_TYPE(call)) {
1261 if (depth < min_depth)
1271 * write_pages() - Write the pages of trace data
1273 * This works through all the calls, writing out as many pages of data as are
1276 * @tw: Writer context
1277 * @out_format: Output format to use
1278 * @missing_countp: Returns number of missing functions (not found in function
1280 * @skip_countp: Returns number of skipped functions (excluded from trace)
1282 * Returns: 0 on success, -ve on error
1284 static int write_pages(struct twriter *tw, enum out_format_t out_format,
1285 int *missing_countp, int *skip_countp)
1287 ulong func_stack[MAX_STACK_DEPTH];
1288 int stack_ptr; /* next free position in stack */
1289 int upto, depth, page_upto, i;
1290 int missing_count = 0, skip_count = 0;
1291 struct trace_call *call;
1292 ulong last_timestamp;
1293 FILE *fout = tw->fout;
1304 /* maintain a stack of start times for calling functions */
1308 * The first thing in the trace may not be the top-level function, so
1309 * set the initial depth so that no function goes below depth 0
1311 depth = -calc_min_depth();
1312 for (i = 0, call = call_list; i < call_count; i++, call++) {
1313 bool entry = TRACE_CALL_TYPE(call) == FUNCF_ENTRY;
1314 struct func_info *func;
1319 func = find_func_by_offset(call->func);
1321 warn("Cannot find function at %lx\n",
1322 text_offset + call->func);
1324 if (missing_count > 20) {
1325 /* perhaps trace does not match System.map */
1326 fprintf(stderr, "Too many missing functions\n");
1332 if (!(func->flags & FUNCF_TRACE)) {
1333 debug("Funcion '%s' is excluded from trace\n",
1339 if (out_format == OUT_FMT_FUNCTION)
1341 else /* 2 header words and then 3 or 8 others */
1342 rec_words = 2 + (entry ? 3 : 8);
1344 /* convert timestamp from us to ns */
1345 timestamp = call->flags & FUNCF_TIMESTAMP_MASK;
1347 if (page_upto + rec_words * 4 > TRACE_PAGE_SIZE) {
1348 if (finish_page(tw))
1354 if (start_page(tw, timestamp))
1357 last_timestamp = timestamp;
1359 page_upto = tw->ptr & TRACE_PAGE_MASK;
1362 "new page, last_timestamp=%ld, upto=%d\n",
1363 last_timestamp, upto);
1367 delta = timestamp - last_timestamp;
1369 fprintf(stderr, "Time went backwards\n");
1373 if (err_count > 20) {
1374 fprintf(stderr, "Too many errors, giving up\n");
1378 if (delta > 0x07fffff) {
1380 * hard to imagine how this could happen since it means
1381 * that no function calls were made for a long time
1383 fprintf(stderr, "cannot represent time delta %x\n",
1388 if (out_format == OUT_FMT_FUNCTION) {
1389 struct func_info *caller_func;
1392 fprintf(stderr, "%d: delta=%d, stamp=%ld\n",
1393 upto, delta, timestamp);
1395 " last_delta %x to %x: last_timestamp=%lx, "
1396 "timestamp=%lx, call->flags=%x, upto=%d\n",
1397 last_delta, delta, last_timestamp,
1398 timestamp, call->flags, upto);
1401 /* type_len is 6, meaning 4 * 6 = 24 bytes */
1402 tw->ptr += tputl(fout, rec_words | (uint)delta << 5);
1403 tw->ptr += tputh(fout, TRACE_FN);
1404 tw->ptr += tputh(fout, 0); /* flags */
1405 tw->ptr += tputl(fout, TRACE_PID); /* PID */
1407 tw->ptr += tputq(fout, text_offset + func->offset);
1408 caller_func = find_caller_by_offset(call->caller);
1410 tw->ptr += tputq(fout,
1411 text_offset + caller_func->offset);
1413 tw->ptr += tputl(fout, rec_words | delta << 5);
1414 tw->ptr += tputh(fout, entry ? TRACE_GRAPH_ENT
1416 tw->ptr += tputh(fout, 0); /* flags */
1417 tw->ptr += tputl(fout, TRACE_PID); /* PID */
1419 tw->ptr += tputq(fout, text_offset + func->offset);
1420 tw->ptr += tputl(fout, depth); /* depth */
1423 if (stack_ptr < MAX_STACK_DEPTH)
1424 func_stack[stack_ptr] = timestamp;
1427 ulong func_duration = 0;
1430 if (stack_ptr && stack_ptr <= MAX_STACK_DEPTH) {
1431 ulong start = func_stack[--stack_ptr];
1433 func_duration = timestamp - start;
1435 tw->ptr += tputl(fout, 0); /* overrun */
1436 tw->ptr += tputq(fout, 0); /* calltime */
1438 tw->ptr += tputq(fout, func_duration);
1443 last_timestamp = timestamp;
1444 page_upto += 4 + rec_words * 4;
1446 if (stack_ptr == MAX_STACK_DEPTH)
1449 if (in_page && finish_page(tw))
1451 *missing_countp = missing_count;
1452 *skip_countp = skip_count;
1458 * write_flyrecord() - Write the flyrecord information
1460 * Writes the header and pages of data for the "flyrecord" section. It also
1461 * writes out the counter-type info, selecting "[local]"
1463 * @tw: Writer context
1464 * @out_format: Output format to use
1465 * @missing_countp: Returns number of missing functions (not found in function
1467 * @skip_countp: Returns number of skipped functions (excluded from trace)
1469 * Returns: 0 on success, -ve on error
1471 static int write_flyrecord(struct twriter *tw, enum out_format_t out_format,
1472 int *missing_countp, int *skip_countp)
1474 int start, ret, len;
1475 FILE *fout = tw->fout;
1478 tw->ptr += fprintf(fout, "flyrecord%c", 0);
1481 start = ALIGN(tw->ptr + 16, TRACE_PAGE_SIZE);
1482 tw->ptr += tputq(fout, start);
1484 /* use a placeholder for the size */
1485 ret = push_len(tw, start, "flyrecord", 8);
1490 snprintf(str, sizeof(str),
1491 "[local] global counter uptime perf mono mono_raw boot x86-tsc\n");
1493 tw->ptr += tputq(fout, len);
1494 tw->ptr += tputs(fout, str);
1496 debug("trace text base %lx, map file %lx\n", text_base, text_offset);
1498 ret = write_pages(tw, out_format, missing_countp, skip_countp);
1500 fprintf(stderr, "Cannot output pages\n");
1504 ret = pop_len(tw, "flyrecord");
1506 fprintf(stderr, "Cannot finish flyrecord header\n");
1514 * make_ftrace() - Write out an ftrace file
1516 * See here for format:
1518 * https://github.com/rostedt/trace-cmd/blob/master/Documentation/trace-cmd/trace-cmd.dat.v7.5.txt
1520 * @fout: Output file
1521 * @out_format: Output format to use
1522 * Returns: 0 on success, -ve on error
1524 static int make_ftrace(FILE *fout, enum out_format_t out_format)
1526 int missing_count, skip_count;
1527 struct twriter tws, *tw = &tws;
1530 memset(tw, '\0', sizeof(*tw));
1531 abuf_init(&tw->str_buf);
1535 ret = output_headers(tw);
1537 fprintf(stderr, "Cannot output headers\n");
1540 /* number of event systems files */
1541 tw->ptr += tputl(fout, 0);
1543 ret = write_symbols(tw);
1545 fprintf(stderr, "Cannot write symbols\n");
1549 ret = write_options(tw);
1551 fprintf(stderr, "Cannot write options\n");
1555 ret = write_flyrecord(tw, out_format, &missing_count, &skip_count);
1557 fprintf(stderr, "Cannot write flyrecord\n");
1561 info("ftrace: %d functions not found, %d excluded\n", missing_count,
1568 * create_node() - Create a new node in the flamegraph tree
1570 * @msg: Message to use for debugging if something goes wrong
1571 * Returns: Pointer to newly created node, or NULL on error
1573 static struct flame_node *create_node(const char *msg)
1575 struct flame_node *node;
1577 node = calloc(1, sizeof(*node));
1579 fprintf(stderr, "Out of memory for %s\n", msg);
1582 INIT_LIST_HEAD(&node->child_head);
1588 * process_call(): Add a call to the flamegraph info
1590 * For function calls, if this call stack has been seen before, this increments
1591 * the call count, creating a new node if needed.
1593 * For function returns, it adds up the time spent in this call stack,
1594 * subtracting the time spent by child functions.
1596 * @state: Current flamegraph state
1597 * @entry: true if this is a function entry, false if a function exit
1598 * @timestamp: Timestamp from the trace file (in microseconds)
1599 * @func: Function that was called/returned from
1601 * Returns: 0 on success, -ve on error
1603 static int process_call(struct flame_state *state, bool entry, ulong timestamp,
1604 struct func_info *func)
1606 struct flame_node *node = state->node;
1609 struct flame_node *child, *chd;
1611 /* see if we have this as a child node already */
1613 list_for_each_entry(chd, &node->child_head, sibling_node) {
1614 if (chd->func == func) {
1620 /* create a new node */
1621 child = create_node("child");
1624 list_add_tail(&child->sibling_node, &node->child_head);
1626 child->parent = node;
1629 debug("entry %s: move from %s to %s\n", func->name,
1630 node->func ? node->func->name : "(root)",
1634 } else if (node->parent) {
1635 debug("exit %s: move from %s to %s\n", func->name,
1636 node->func->name, node->parent->func ?
1637 node->parent->func->name : "(root)");
1638 node = node->parent;
1647 * make_flame_tree() - Create a tree of stack traces
1649 * Set up a tree, with the root node having the top-level functions as children
1650 * and the leaf nodes being leaf functions. Each node has a count of how many
1651 * times this function appears in the trace
1653 * @treep: Returns the resulting flamegraph tree
1654 * Returns: 0 on success, -ve on error
1656 static int make_flame_tree(struct flame_node **treep)
1658 struct flame_state state;
1659 struct flame_node *tree;
1660 struct trace_call *call;
1661 int missing_count = 0;
1665 * The first thing in the trace may not be the top-level function, so
1666 * set the initial depth so that no function goes below depth 0
1668 depth = -calc_min_depth();
1670 tree = create_node("tree");
1676 for (i = 0, call = call_list; i < call_count; i++, call++) {
1677 bool entry = TRACE_CALL_TYPE(call) == FUNCF_ENTRY;
1678 ulong timestamp = call->flags & FUNCF_TIMESTAMP_MASK;
1679 struct func_info *func;
1686 func = find_func_by_offset(call->func);
1688 warn("Cannot find function at %lx\n",
1689 text_offset + call->func);
1694 if (process_call(&state, entry, timestamp, func))
1697 fprintf(stderr, "%d nodes\n", state.nodes);
1704 * output_tree() - Output a flamegraph tree
1706 * Writes the tree out to a file in a format suitable for flamegraph.pl
1708 * This works by maintaining a string shared across all recursive calls. The
1709 * function name for this node is added to the existing string, to make up the
1710 * full call-stack description. For example, on entry, @str might contain:
1712 * "initf_bootstage;bootstage_mark_name"
1715 * with @base pointing to the \0 at the end of the string. This function adds
1716 * a ';' following by the name of the current function, e.g. "timer_get_boot_us"
1717 * as well as the output value, to get the full line:
1719 * initf_bootstage;bootstage_mark_name;timer_get_boot_us 123
1721 * @fout: Output file
1722 * @node: Node to output (pass the whole tree at first)
1723 * @str: String to use to build the output line (e.g. 500 charas long)
1724 * @maxlen: Maximum length of string
1725 * @base: Current base position in the string
1726 * @treep: Returns the resulting flamegraph tree
1727 * Returns 0 if OK, -1 on error
1729 static int output_tree(FILE *fout, const struct flame_node *node,
1730 char *str, int maxlen, int base)
1732 const struct flame_node *child;
1736 fprintf(fout, "%s %d\n", str, node->count);
1741 list_for_each_entry(child, &node->child_head, sibling_node) {
1744 len = strlen(child->func->name);
1745 if (pos + len + 1 >= maxlen) {
1746 fprintf(stderr, "String too short (%d chars)\n",
1750 strcpy(str + pos, child->func->name);
1751 if (output_tree(fout, child, str, maxlen, pos + len))
1759 * make_flamegraph() - Write out a flame graph
1761 * @fout: Output file
1762 * Returns 0 if OK, -1 on error
1764 static int make_flamegraph(FILE *fout)
1766 struct flame_node *tree;
1769 if (make_flame_tree(&tree))
1773 if (output_tree(fout, tree, str, sizeof(str), 0))
1780 * prof_tool() - Performs requested action
1782 * @argc: Number of arguments (used to obtain the command
1783 * @argv: List of arguments
1784 * @trace_fname: Filename of input file (trace data from U-Boot)
1785 * @map_fname: Filename of map file (System.map from U-Boot)
1786 * @trace_config_fname: Trace-configuration file, or NULL if none
1787 * @out_fname: Output filename
1789 static int prof_tool(int argc, char *const argv[],
1790 const char *trace_fname, const char *map_fname,
1791 const char *trace_config_fname, const char *out_fname,
1792 enum out_format_t out_format)
1796 if (read_map_file(map_fname))
1798 if (trace_fname && read_trace_file(trace_fname))
1800 if (trace_config_fname && read_trace_config_file(trace_config_fname))
1803 check_trace_config();
1805 for (; argc; argc--, argv++) {
1806 const char *cmd = *argv;
1808 if (!strcmp(cmd, "dump-ftrace")) {
1811 if (out_format != OUT_FMT_FUNCTION &&
1812 out_format != OUT_FMT_FUNCGRAPH)
1813 out_format = OUT_FMT_FUNCTION;
1814 fout = fopen(out_fname, "w");
1816 fprintf(stderr, "Cannot write file '%s'\n",
1820 err = make_ftrace(fout, out_format);
1822 } else if (!strcmp(cmd, "dump-flamegraph")) {
1825 fout = fopen(out_fname, "w");
1827 fprintf(stderr, "Cannot write file '%s'\n",
1831 err = make_flamegraph(fout);
1834 warn("Unknown command '%s'\n", cmd);
1841 int main(int argc, char *argv[])
1843 enum out_format_t out_format = OUT_FMT_DEFAULT;
1844 const char *map_fname = "System.map";
1845 const char *trace_fname = NULL;
1846 const char *config_fname = NULL;
1847 const char *out_fname = NULL;
1851 while ((opt = getopt(argc, argv, "c:f:m:o:t:v:")) != -1) {
1854 config_fname = optarg;
1857 if (!strcmp("function", optarg)) {
1858 out_format = OUT_FMT_FUNCTION;
1859 } else if (!strcmp("funcgraph", optarg)) {
1860 out_format = OUT_FMT_FUNCGRAPH;
1863 "Invalid format: use function, funcgraph, calls, timing\n");
1874 trace_fname = optarg;
1877 verbose = atoi(optarg);
1883 argc -= optind; argv += optind;
1887 if (!out_fname || !map_fname || !trace_fname) {
1889 "Must provide trace data, System.map file and output file\n");
1893 debug("Debug enabled\n");
1894 return prof_tool(argc, argv, trace_fname, map_fname, config_fname,
1895 out_fname, out_format);