1 /* Generate assembler source containing symbol information
3 * Copyright 2002 by Kai Germaschewski
5 * This software may be used and distributed according to the terms
6 * of the GNU General Public License, incorporated herein by reference.
8 * Usage: nm -n vmlinux | scripts/kallsyms [--all-symbols] > symbols.S
12 * (25/Aug/2004) Paulo Marques <pmarques@grupopie.com>
13 * Changed the compression method from stem compression to "table lookup"
16 * Table compression uses all the unused char codes on the symbols and
17 * maps these to the most used substrings (tokens). For instance, it might
18 * map char code 0xF7 to represent "write_" and then in every symbol where
19 * "write_" appears it can be replaced by 0xF7, saving 5 bytes.
20 * The used codes themselves are also placed in the table so that the
21 * decompresion can work without "special cases".
22 * Applied to kernel symbols, this usually produces a compression ratio
32 /* maximum token length used. It doesn't pay to increase it a lot, because
33 * very long substrings probably don't repeat themselves too often. */
34 #define MAX_TOK_SIZE 11
35 #define KSYM_NAME_LEN 127
37 /* we use only a subset of the complete symbol table to gather the token count,
38 * to speed up compression, at the expense of a little compression ratio */
39 #define WORKING_SET 1024
41 /* first find the best token only on the list of tokens that would profit more
42 * than GOOD_BAD_THRESHOLD. Only if this list is empty go to the "bad" list.
43 * Increasing this value will put less tokens on the "good" list, so the search
44 * is faster. However, if the good list runs out of tokens, we must painfully
45 * search the bad list. */
46 #define GOOD_BAD_THRESHOLD 10
48 /* token hash parameters */
50 #define HASH_TABLE_SIZE (1 << HASH_BITS)
51 #define HASH_MASK (HASH_TABLE_SIZE - 1)
52 #define HASH_BASE_OFFSET 2166136261U
53 #define HASH_FOLD(a) ((a)&(HASH_MASK))
55 /* flags to mark symbols */
56 #define SYM_FLAG_VALID 1
57 #define SYM_FLAG_SAMPLED 2
60 unsigned long long addr;
68 static struct sym_entry *table;
70 static unsigned long long _stext, _etext, _sinittext, _einittext;
71 static int all_symbols = 0;
72 static char symbol_prefix_char = '\0';
75 unsigned char data[MAX_TOK_SIZE];
77 /* profit: the number of bytes that could be saved by inserting this
78 * token into the table */
80 struct token *next; /* next token on the hash list */
81 struct token *right; /* next token on the good/bad list */
82 struct token *left; /* previous token on the good/bad list */
83 struct token *smaller; /* token that is less one letter than this one */
86 struct token bad_head, good_head;
87 struct token *hash_table[HASH_TABLE_SIZE];
89 /* the table that holds the result of the compression */
90 unsigned char best_table[256][MAX_TOK_SIZE+1];
91 unsigned char best_table_len[256];
97 fprintf(stderr, "Usage: kallsyms [--all-symbols] [--symbol-prefix=<prefix char>] < in.map > out.S\n");
102 * This ignores the intensely annoying "mapping symbols" found
103 * in ARM ELF files: $a, $t and $d.
106 is_arm_mapping_symbol(const char *str)
108 return str[0] == '$' && strchr("atd", str[1])
109 && (str[2] == '\0' || str[2] == '.');
113 read_symbol(FILE *in, struct sym_entry *s)
119 rc = fscanf(in, "%llx %c %499s\n", &s->addr, &s->type, str);
129 /* skip prefix char */
130 if (symbol_prefix_char && str[0] == symbol_prefix_char)
133 /* Ignore most absolute/undefined (?) symbols. */
134 if (strcmp(sym, "_stext") == 0)
136 else if (strcmp(sym, "_etext") == 0)
138 else if (strcmp(sym, "_sinittext") == 0)
139 _sinittext = s->addr;
140 else if (strcmp(sym, "_einittext") == 0)
141 _einittext = s->addr;
142 else if (toupper(s->type) == 'A')
144 /* Keep these useful absolute symbols */
145 if (strcmp(sym, "__kernel_syscall_via_break") &&
146 strcmp(sym, "__kernel_syscall_via_epc") &&
147 strcmp(sym, "__kernel_sigtramp") &&
152 else if (toupper(s->type) == 'U' ||
153 is_arm_mapping_symbol(sym))
156 /* include the type field in the symbol name, so that it gets
157 * compressed together */
158 s->len = strlen(str) + 1;
159 s->sym = (char *) malloc(s->len + 1);
160 strcpy(s->sym + 1, str);
167 symbol_valid(struct sym_entry *s)
169 /* Symbols which vary between passes. Passes 1 and 2 must have
170 * identical symbol lists. The kallsyms_* symbols below are only added
171 * after pass 1, they would be included in pass 2 when --all-symbols is
172 * specified so exclude them to get a stable symbol list.
174 static char *special_symbols[] = {
175 "kallsyms_addresses",
179 "kallsyms_token_table",
180 "kallsyms_token_index",
182 /* Exclude linker generated symbols which vary between passes */
183 "_SDA_BASE_", /* ppc */
184 "_SDA2_BASE_", /* ppc */
189 /* skip prefix char */
190 if (symbol_prefix_char && *(s->sym + 1) == symbol_prefix_char)
193 /* if --all-symbols is not specified, then symbols outside the text
194 * and inittext sections are discarded */
196 if ((s->addr < _stext || s->addr > _etext)
197 && (s->addr < _sinittext || s->addr > _einittext))
199 /* Corner case. Discard any symbols with the same value as
200 * _etext or _einittext, they can move between pass 1 and 2
201 * when the kallsyms data is added. If these symbols move then
202 * they may get dropped in pass 2, which breaks the kallsyms
205 if ((s->addr == _etext && strcmp(s->sym + offset, "_etext")) ||
206 (s->addr == _einittext && strcmp(s->sym + offset, "_einittext")))
210 /* Exclude symbols which vary between passes. */
211 if (strstr(s->sym + offset, "_compiled."))
214 for (i = 0; special_symbols[i]; i++)
215 if( strcmp(s->sym + offset, special_symbols[i]) == 0 )
227 table = realloc(table, sizeof(*table) * size);
229 fprintf(stderr, "out of memory\n");
233 if (read_symbol(in, &table[cnt]) == 0)
238 static void output_label(char *label)
240 if (symbol_prefix_char)
241 printf(".globl %c%s\n", symbol_prefix_char, label);
243 printf(".globl %s\n", label);
245 if (symbol_prefix_char)
246 printf("%c%s:\n", symbol_prefix_char, label);
248 printf("%s:\n", label);
251 /* uncompress a compressed symbol. When this function is called, the best table
252 * might still be compressed itself, so the function needs to be recursive */
253 static int expand_symbol(unsigned char *data, int len, char *result)
255 int c, rlen, total=0;
259 /* if the table holds a single char that is the same as the one
260 * we are looking for, then end the search */
261 if (best_table[c][0]==c && best_table_len[c]==1) {
265 /* if not, recurse and expand */
266 rlen = expand_symbol(best_table[c], best_table_len[c], result);
281 int i, k, off, valid;
282 unsigned int best_idx[256];
283 unsigned int *markers;
284 char buf[KSYM_NAME_LEN+1];
286 printf("#include <asm/types.h>\n");
287 printf("#if BITS_PER_LONG == 64\n");
288 printf("#define PTR .quad\n");
289 printf("#define ALGN .align 8\n");
291 printf("#define PTR .long\n");
292 printf("#define ALGN .align 4\n");
297 output_label("kallsyms_addresses");
299 for (i = 0; i < cnt; i++) {
300 if (table[i].flags & SYM_FLAG_VALID) {
301 printf("\tPTR\t%#llx\n", table[i].addr);
307 output_label("kallsyms_num_syms");
308 printf("\tPTR\t%d\n", valid);
311 /* table of offset markers, that give the offset in the compressed stream
312 * every 256 symbols */
313 markers = (unsigned int *) malloc(sizeof(unsigned int)*((valid + 255) / 256));
315 output_label("kallsyms_names");
318 for (i = 0; i < cnt; i++) {
320 if (!table[i].flags & SYM_FLAG_VALID)
323 if ((valid & 0xFF) == 0)
324 markers[valid >> 8] = off;
326 printf("\t.byte 0x%02x", table[i].len);
327 for (k = 0; k < table[i].len; k++)
328 printf(", 0x%02x", table[i].sym[k]);
331 off += table[i].len + 1;
336 output_label("kallsyms_markers");
337 for (i = 0; i < ((valid + 255) >> 8); i++)
338 printf("\tPTR\t%d\n", markers[i]);
343 output_label("kallsyms_token_table");
345 for (i = 0; i < 256; i++) {
347 expand_symbol(best_table[i],best_table_len[i],buf);
348 printf("\t.asciz\t\"%s\"\n", buf);
349 off += strlen(buf) + 1;
353 output_label("kallsyms_token_index");
354 for (i = 0; i < 256; i++)
355 printf("\t.short\t%d\n", best_idx[i]);
360 /* table lookup compression functions */
362 static inline unsigned int rehash_token(unsigned int hash, unsigned char data)
364 return ((hash * 16777619) ^ data);
367 static unsigned int hash_token(unsigned char *data, int len)
369 unsigned int hash=HASH_BASE_OFFSET;
372 for (i = 0; i < len; i++)
373 hash = rehash_token(hash, data[i]);
375 return HASH_FOLD(hash);
378 /* find a token given its data and hash value */
379 static struct token *find_token_hash(unsigned char *data, int len, unsigned int hash)
383 ptr = hash_table[hash];
386 if ((ptr->len == len) && (memcmp(ptr->data, data, len) == 0))
394 static inline void insert_token_in_group(struct token *head, struct token *ptr)
396 ptr->right = head->right;
397 ptr->right->left = ptr;
402 static inline void remove_token_from_group(struct token *ptr)
404 ptr->left->right = ptr->right;
405 ptr->right->left = ptr->left;
409 /* build the counts for all the tokens that start with "data", and have lenghts
411 static void learn_token(unsigned char *data, int len)
413 struct token *ptr,*last_ptr;
415 unsigned int hash = HASH_BASE_OFFSET;
416 unsigned int hashes[MAX_TOK_SIZE + 1];
418 if (len > MAX_TOK_SIZE)
421 /* calculate and store the hash values for all the sub-tokens */
422 hash = rehash_token(hash, data[0]);
423 for (i = 2; i <= len; i++) {
424 hash = rehash_token(hash, data[i-1]);
425 hashes[i] = HASH_FOLD(hash);
431 for (i = len; i >= 2; i--) {
434 if (!ptr) ptr = find_token_hash(data, i, hash);
437 /* create a new token entry */
438 ptr = (struct token *) malloc(sizeof(*ptr));
440 memcpy(ptr->data, data, i);
443 /* when we create an entry, it's profit is 0 because
444 * we also take into account the size of the token on
445 * the compressed table. We then subtract GOOD_BAD_THRESHOLD
446 * so that the test to see if this token belongs to
447 * the good or bad list, is a comparison to zero */
448 ptr->profit = -GOOD_BAD_THRESHOLD;
450 ptr->next = hash_table[hash];
451 hash_table[hash] = ptr;
453 insert_token_in_group(&bad_head, ptr);
457 newprofit = ptr->profit + (ptr->len - 1);
458 /* check to see if this token needs to be moved to a
460 if((ptr->profit < 0) && (newprofit >= 0)) {
461 remove_token_from_group(ptr);
462 insert_token_in_group(&good_head,ptr);
464 ptr->profit = newprofit;
467 if (last_ptr) last_ptr->smaller = ptr;
474 /* decrease the counts for all the tokens that start with "data", and have lenghts
475 * from 2 to "len". This function is much simpler than learn_token because we have
476 * more guarantees (tho tokens exist, the ->smaller pointer is set, etc.)
477 * The two separate functions exist only because of compression performance */
478 static void forget_token(unsigned char *data, int len)
484 if (len > MAX_TOK_SIZE) len = MAX_TOK_SIZE;
486 hash = hash_token(data, len);
487 ptr = find_token_hash(data, len, hash);
489 for (i = len; i >= 2; i--) {
491 newprofit = ptr->profit - (ptr->len - 1);
492 if ((ptr->profit >= 0) && (newprofit < 0)) {
493 remove_token_from_group(ptr);
494 insert_token_in_group(&bad_head, ptr);
496 ptr->profit=newprofit;
502 /* count all the possible tokens in a symbol */
503 static void learn_symbol(unsigned char *symbol, int len)
507 for (i = 0; i < len - 1; i++)
508 learn_token(symbol + i, len - i);
511 /* decrease the count for all the possible tokens in a symbol */
512 static void forget_symbol(unsigned char *symbol, int len)
516 for (i = 0; i < len - 1; i++)
517 forget_token(symbol + i, len - i);
520 /* set all the symbol flags and do the initial token count */
521 static void build_initial_tok_table(void)
523 int i, use_it, valid;
526 for (i = 0; i < cnt; i++) {
528 if ( symbol_valid(&table[i]) ) {
529 table[i].flags |= SYM_FLAG_VALID;
535 for (i = 0; i < cnt; i++) {
537 /* subsample the available symbols. This method is almost like
538 * a Bresenham's algorithm to get uniformly distributed samples
539 * across the symbol table */
540 if (table[i].flags & SYM_FLAG_VALID) {
542 use_it += WORKING_SET;
544 if (use_it >= valid) {
545 table[i].flags |= SYM_FLAG_SAMPLED;
549 if (table[i].flags & SYM_FLAG_SAMPLED)
550 learn_symbol(table[i].sym, table[i].len);
554 /* replace a given token in all the valid symbols. Use the sampled symbols
555 * to update the counts */
556 static void compress_symbols(unsigned char *str, int tlen, int idx)
558 int i, len, learn, size;
561 for (i = 0; i < cnt; i++) {
563 if (!(table[i].flags & SYM_FLAG_VALID)) continue;
570 /* find the token on the symbol */
571 p = (unsigned char *) strstr((char *) p, (char *) str);
575 /* if this symbol was used to count, decrease it */
576 if (table[i].flags & SYM_FLAG_SAMPLED)
577 forget_symbol(table[i].sym, len);
582 size = (len - (p - table[i].sym)) - tlen + 1;
583 memmove(p + 1, p + tlen, size);
587 } while (size >= tlen);
591 /* if this symbol was used to count, learn it again */
592 if(table[i].flags & SYM_FLAG_SAMPLED)
593 learn_symbol(table[i].sym, len);
598 /* search the token with the maximum profit */
599 static struct token *find_best_token(void)
601 struct token *ptr,*best,*head;
606 /* failsafe: if the "good" list is empty search from the "bad" list */
607 if(good_head.right == &good_head) head = &bad_head;
608 else head = &good_head;
612 while (ptr != head) {
613 if (ptr->profit > bestprofit) {
614 bestprofit = ptr->profit;
623 /* this is the core of the algorithm: calculate the "best" table */
624 static void optimize_result(void)
629 /* using the '\0' symbol last allows compress_symbols to use standard
630 * fast string functions */
631 for (i = 255; i >= 0; i--) {
633 /* if this table slot is empty (it is not used by an actual
634 * original char code */
635 if (!best_table_len[i]) {
637 /* find the token with the breates profit value */
638 best = find_best_token();
640 /* place it in the "best" table */
641 best_table_len[i] = best->len;
642 memcpy(best_table[i], best->data, best_table_len[i]);
643 /* zero terminate the token so that we can use strstr
644 in compress_symbols */
645 best_table[i][best_table_len[i]]='\0';
647 /* replace this token in all the valid symbols */
648 compress_symbols(best_table[i], best_table_len[i], i);
653 /* start by placing the symbols that are actually used on the table */
654 static void insert_real_symbols_in_table(void)
658 memset(best_table, 0, sizeof(best_table));
659 memset(best_table_len, 0, sizeof(best_table_len));
661 for (i = 0; i < cnt; i++) {
662 if (table[i].flags & SYM_FLAG_VALID) {
663 for (j = 0; j < table[i].len; j++) {
672 static void optimize_token_table(void)
674 memset(hash_table, 0, sizeof(hash_table));
676 good_head.left = &good_head;
677 good_head.right = &good_head;
679 bad_head.left = &bad_head;
680 bad_head.right = &bad_head;
682 build_initial_tok_table();
684 insert_real_symbols_in_table();
686 /* When valid symbol is not registered, exit to error */
687 if (good_head.left == good_head.right &&
688 bad_head.left == bad_head.right) {
689 fprintf(stderr, "No valid symbol.\n");
698 main(int argc, char **argv)
702 for (i = 1; i < argc; i++) {
703 if(strcmp(argv[i], "--all-symbols") == 0)
705 else if (strncmp(argv[i], "--symbol-prefix=", 16) == 0) {
706 char *p = &argv[i][16];
708 if ((*p == '"' && *(p+2) == '"') || (*p == '\'' && *(p+2) == '\''))
710 symbol_prefix_char = *p;
714 } else if (argc != 1)
718 optimize_token_table();