Re: ldso : dladdr support

[musl <b.brezillon.musl@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 3712 bytes --]

Same as before except I use bit mask to support multiple hash algorithms.

On 08/08/2012 13:52, Szabolcs Nagy wrote:
> * musl <b.brezillon.musl@gmail.com> [2012-08-08 11:55:45 +0200]:
>> Here is the new patch for dladdr + gnu hash support.
>>
>> On 08/08/2012 01:09, Rich Felker wrote:
>>> don't remember which) where I described some optimizations that should
>>> be present if gnu hash is to be supported. Basically, the dynamic
>>> linker should keep track of whether there's one of the two hashes
>>> that's supported by ALL loaded libs, and in that case, only ever use
>>> that hash, to avoid computing two different hashes.
>> The hash for given algo is only computed once (if needed).
>> That's the reason for the computed table.
>> If all libs uses the same hash algo, the other will never be computed.
> a lib can support both hash tables
> (so it's non-trivial to use the one that all libs support,
> eg instead of a single alg index a bit mask should be used)
>
See patch changes.
>> +#define SYSV_HASH_ALG_IDX   0
>> +#define GNU_HASH_ALG_IDX    1
>> +#define HASH_ALG_CNT		2
> if we go with this approach then use shorter names
> (thes are only used locally)
>
> eg. SYSV_HASH, GNU_HASH, NHASH
Corrected.
>> +	for (h1 &= (uint32_t)-2;; sym++) {
>> +		h2 = *hashval++;
>> +		if ((h1 == (h2 & ~1)) && !strcmp(s, strings + sym->st_name))
> these is still a ~1
>
> looking at it now probably writing out & 0xfffffffe
> is the cleanest
>
>> +	static uint32_t precomp[HASH_ALG_CNT][3] = {
>> +	  {0x6b366be, 0x6b3afd, 0x595a4cc},
>> +	  {0xf9040207, 0xf4dc4ae, 0x1f4039c9},
>> +	};
>> +	uint32_t h[HASH_ALG_CNT];
>> +	uint8_t computed[HASH_ALG_CNT] = {0, 0};
>> +	uint8_t alg = dso->hashalg;
>> +	if (alg == SYSV_HASH_ALG_IDX)
>> +		h[alg] = sysv_hash(s);
>> +	else
>> +		h[alg] = gnu_hash(s);
>> +
>> +	computed[alg] = 1;
>> +
>> +	if (h[alg] == precomp[alg][0] && !strcmp(s, "dlopen")) rtld_used = 1;
>> +	if (h[alg] == precomp[alg][1] && !strcmp(s, "dlsym")) rtld_used = 1;
>> +	if (h[alg] == precomp[alg][2] && !strcmp(s, "__stack_chk_fail")) ssp_used = 1;
>> +
>>  	for (; dso; dso=dso->next) {
>>  		Sym *sym;
>> +
>>  		if (!dso->global) continue;
>> -		sym = lookup(s, h, dso);
>> +
>> +		alg = dso->hashalg;
>> +		if (!computed[alg]) {
>> +			if (alg == SYSV_HASH_ALG_IDX) {
>> +				h[alg] = sysv_hash(s);
>> +				sym = sysv_lookup(s, h[alg], dso);
>> +			}
>> +			else {
>> +				h[alg] = gnu_hash(s);
>> +				sym = gnu_lookup(s, h[alg], dso);
>> +			}
>> +			computed[alg] = 1;
>> +		} else {
>> +			if (alg == SYSV_HASH_ALG_IDX)
>> +				sym = sysv_lookup(s, h[alg], dso);
>> +			else
>> +				sym = gnu_lookup(s, h[alg], dso);
>> +		}
> instead of arrays i'd write
>
> alg = dso->hashalg;
> if (alg == SYSV_HASH) {
> 	if (sysv_ok) {
> 		...
> 		sysv_ok = 1;
> 	}
> 	sym = sysv_lookup(s, sysv_h, dso);
> } else {
> }
Haven't changed it yet.
> since there are many ifs anyway
>
> the table approach is nicer when all data and functions
> are in a table:
>
> if (!ok[alg]) {
> 	h[alg] = hash[alg](s);
> 	ok[alg] = 1;
> }
> sym = lookup[alg](s, h[alg], dso);
>
>> +	p->hashalg = SYSV_HASH_ALG_IDX;
>>  	p->strings = (void *)(p->base + dyn[DT_STRTAB]);
>> +	for (; v[0]; v+=2) if (v[0] == DT_GNU_HASH) {
>> +		p->hashtab = (void *)(p->base + v[1]);
>> +		p->hashalg = GNU_HASH_ALG_IDX;
>> +	}
> so it seems gnu hash is used whenever it's present
> i'm not sure if that's the right default..
>
> another possibility is to have a plain find_sym function
> which is simple and only supports sysv hash
> and whenever it encounters a lib that has no sysv hash in
> it a find_sym_gnu is called that does the hard work
> (so using gnu only libs is penalized, i don't know how
> common that case is though)


[-- Attachment #2: ldso-add-dladdr-and-gnu-hash-support.patch --]
[-- Type: text/x-patch, Size: 10514 bytes --]

From 359843b57223db412847f69702e19511dfeb435d Mon Sep 17 00:00:00 2001
From: Boris BREZILLON <b.brezillon@overkiz.com>
Date: Wed, 8 Aug 2012 15:49:29 +0200
Subject: [PATCH] ldso : add dladdr and gnu hash support

---
 include/dlfcn.h    |   15 +++
 src/ldso/dynlink.c |  272 ++++++++++++++++++++++++++++++++++++++++++++++++----
 2 files changed, 269 insertions(+), 18 deletions(-)

diff --git a/include/dlfcn.h b/include/dlfcn.h
index dea74c7..8c45822 100644
--- a/include/dlfcn.h
+++ b/include/dlfcn.h
@@ -18,6 +18,21 @@ char  *dlerror(void);
 void  *dlopen(const char *, int);
 void  *dlsym(void *, const char *);
 
+#ifdef _GNU_SOURCE
+typedef struct {
+    const char *dli_fname;  /* Pathname of shared object that
+                               contains address */
+    void       *dli_fbase;  /* Address at which shared object
+                               is loaded */
+    const char *dli_sname;  /* Name of nearest symbol with address
+                               lower than addr */
+    void       *dli_saddr;  /* Exact address of symbol named
+                               in dli_sname */
+} Dl_info;
+
+int dladdr (void *addr, Dl_info *info);
+#endif
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/src/ldso/dynlink.c b/src/ldso/dynlink.c
index f55c6f1..4a236b2 100644
--- a/src/ldso/dynlink.c
+++ b/src/ldso/dynlink.c
@@ -1,3 +1,4 @@
+#define _GNU_SOURCE
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -28,14 +29,20 @@ typedef Elf32_Phdr Phdr;
 typedef Elf32_Sym Sym;
 #define R_TYPE(x) ((x)&255)
 #define R_SYM(x) ((x)>>8)
+#define ELF_ST_TYPE ELF32_ST_TYPE
 #else
 typedef Elf64_Ehdr Ehdr;
 typedef Elf64_Phdr Phdr;
 typedef Elf64_Sym Sym;
 #define R_TYPE(x) ((x)&0xffffffff)
 #define R_SYM(x) ((x)>>32)
+#define ELF_ST_TYPE ELF64_ST_TYPE
 #endif
 
+#define SYSV_HASH	0
+#define GNU_HASH	1
+#define NHASH		2
+
 struct debug {
 	int ver;
 	void *head;
@@ -52,7 +59,8 @@ struct dso {
 
 	int refcnt;
 	Sym *syms;
-	uint32_t *hashtab;
+	uint32_t hashalgs;
+	uint32_t *hashtabs[NHASH];
 	char *strings;
 	unsigned char *map;
 	size_t map_len;
@@ -94,7 +102,7 @@ static void decode_vec(size_t *v, size_t *a, size_t cnt)
 	}
 }
 
-static uint32_t hash(const char *s0)
+static uint32_t sysv_hash(const char *s0)
 {
 	const unsigned char *s = (void *)s0;
 	uint_fast32_t h = 0;
@@ -105,11 +113,20 @@ static uint32_t hash(const char *s0)
 	return h & 0xfffffff;
 }
 
-static Sym *lookup(const char *s, uint32_t h, struct dso *dso)
+static uint32_t gnu_hash (const char *s0)
+{
+	const unsigned char *s = (void *)s0;
+	uint_fast32_t h = 5381;
+	for (; *s; s++)
+		h = h*33 + *s;
+	return h & 0xffffffff;
+}
+
+static Sym *sysv_lookup(const char *s, uint32_t h, struct dso *dso)
 {
 	size_t i;
 	Sym *syms = dso->syms;
-	uint32_t *hashtab = dso->hashtab;
+	uint32_t *hashtab = dso->hashtabs[SYSV_HASH];
 	char *strings = dso->strings;
 	for (i=hashtab[2+h%hashtab[0]]; i; i=hashtab[2+hashtab[0]+i]) {
 		if (!strcmp(s, strings+syms[i].st_name))
@@ -118,20 +135,101 @@ static Sym *lookup(const char *s, uint32_t h, struct dso *dso)
 	return 0;
 }
 
+static Sym *gnu_lookup(const char *s, uint32_t h1, struct dso *dso)
+{
+	size_t i;
+	Sym *sym;
+	char *strings = dso->strings;
+	uint32_t *hashtab = dso->hashtabs[GNU_HASH];
+	uint32_t nbuckets = hashtab[0];
+	size_t *maskwords = (size_t *)(hashtab + 4);
+	uint32_t *buckets = hashtab + 4 + (hashtab[2]*(sizeof(size_t)/sizeof(uint32_t)));
+	uint32_t symndx = hashtab[1];
+	Sym *syms = dso->syms;
+	uint32_t shift2 = hashtab[3];
+	uint32_t h2 = h1 >> shift2;
+	uint32_t *hashvals = buckets + nbuckets;
+	uint32_t *hashval;
+	size_t c = sizeof(size_t) * 8;
+	size_t n = (h1/c) & (hashtab[2]-1);
+	size_t bitmask = (1 << (h1%c)) | (1 << (h2%c));
+
+	if ((maskwords[n] & bitmask) != bitmask)
+		return 0;
+
+	n = buckets[h1 % nbuckets];
+	if (!n)
+		return 0;
+
+	sym = syms + n;
+	hashval = hashvals + n - symndx;
+
+	for (h1 &= (uint32_t)-2;; sym++) {
+		h2 = *hashval++;
+		if ((h1 == (h2 & ~1)) && !strcmp(s, strings + sym->st_name))
+			return sym;
+
+		if (h2 & 1)
+			break;
+	}
+
+	return 0;
+}
+
 #define OK_TYPES (1<<STT_NOTYPE | 1<<STT_OBJECT | 1<<STT_FUNC | 1<<STT_COMMON)
 #define OK_BINDS (1<<STB_GLOBAL | 1<<STB_WEAK)
 
 static void *find_sym(struct dso *dso, const char *s, int need_def)
 {
-	uint32_t h = hash(s);
 	void *def = 0;
-	if (h==0x6b366be && !strcmp(s, "dlopen")) rtld_used = 1;
-	if (h==0x6b3afd && !strcmp(s, "dlsym")) rtld_used = 1;
-	if (h==0x595a4cc && !strcmp(s, "__stack_chk_fail")) ssp_used = 1;
+	static uint32_t precomp[NHASH][3] = {
+		{0x6b366be, 0x6b3afd, 0x595a4cc},
+		{0xf9040207, 0xf4dc4ae, 0x1f4039c9},
+	};
+	uint32_t h[NHASH];
+	uint8_t ok = 0;
+	uint8_t algs = dso->hashalgs;
+	uint8_t alg;
+	if (algs & (1 << SYSV_HASH)) {
+		h[SYSV_HASH] = sysv_hash(s);
+		alg = SYSV_HASH;
+	} else {
+		h[GNU_HASH] = gnu_hash(s);
+		alg = GNU_HASH;
+	}
+
+	ok |= (1 << alg);
+
+	if (h[alg] == precomp[alg][0] && !strcmp(s, "dlopen")) rtld_used = 1;
+	if (h[alg] == precomp[alg][1] && !strcmp(s, "dlsym")) rtld_used = 1;
+	if (h[alg] == precomp[alg][2] && !strcmp(s, "__stack_chk_fail")) ssp_used = 1;
+
 	for (; dso; dso=dso->next) {
 		Sym *sym;
+
 		if (!dso->global) continue;
-		sym = lookup(s, h, dso);
+
+		algs = dso->hashalgs;
+		if (!(algs & ok)) {
+			if (algs & (1 << SYSV_HASH)) {
+				alg = SYSV_HASH;
+				h[alg] = sysv_hash(s);
+				sym = sysv_lookup(s, h[alg], dso);
+			}
+			else {
+				alg = GNU_HASH;
+				h[alg] = gnu_hash(s);
+				sym = gnu_lookup(s, h[alg], dso);
+			}
+
+			ok |= (1 << alg);
+		} else {
+			if ((algs & ok) & (1 << SYSV_HASH))
+				sym = sysv_lookup(s, h[SYSV_HASH], dso);
+			else
+				sym = gnu_lookup(s, h[GNU_HASH], dso);
+		}
+
 		if (sym && (!need_def || sym->st_shndx) && sym->st_value
 		 && (1<<(sym->st_info&0xf) & OK_TYPES)
 		 && (1<<(sym->st_info>>4) & OK_BINDS)) {
@@ -320,11 +418,28 @@ static int path_open(const char *name, const char *search, char *buf, size_t buf
 
 static void decode_dyn(struct dso *p)
 {
-	size_t dyn[DYN_CNT] = {0};
-	decode_vec(p->dynv, dyn, DYN_CNT);
-	p->syms = (void *)(p->base + dyn[DT_SYMTAB]);
-	p->hashtab = (void *)(p->base + dyn[DT_HASH]);
-	p->strings = (void *)(p->base + dyn[DT_STRTAB]);
+	size_t *v;
+	p->hashalgs = 0;
+	for (v = p->dynv; v[0]; v+=2) {
+		switch (v[0]) {
+		case DT_SYMTAB:
+			p->syms = (void *)(p->base + v[1]);
+			break;
+		case DT_HASH:
+			p->hashtabs[SYSV_HASH] = (void *)(p->base + v[1]);
+			p->hashalgs |= (1 << SYSV_HASH);
+			break;
+		case DT_STRTAB:
+			p->strings = (void *)(p->base + v[1]);
+			break;
+		case DT_GNU_HASH:
+			p->hashtabs[GNU_HASH] = (void *)(p->base + v[1]);
+			p->hashalgs |= (1 << GNU_HASH);
+			break;
+		default:
+			break;
+		}
+	}
 }
 
 static struct dso *load_library(const char *name)
@@ -784,8 +899,11 @@ end:
 static void *do_dlsym(struct dso *p, const char *s, void *ra)
 {
 	size_t i;
-	uint32_t h;
 	Sym *sym;
+	uint32_t ok = 0;
+	uint8_t algs = p->hashalgs;
+	uint32_t h[NHASH];
+
 	if (p == RTLD_NEXT) {
 		for (p=head; p && (unsigned char *)ra-p->map>p->map_len; p=p->next);
 		if (!p) p=head;
@@ -798,12 +916,39 @@ static void *do_dlsym(struct dso *p, const char *s, void *ra)
 		if (!res) goto failed;
 		return res;
 	}
-	h = hash(s);
-	sym = lookup(s, h, p);
+
+	if (algs & (1 << SYSV_HASH)) {
+		h[SYSV_HASH] = sysv_hash(s);
+		sym = sysv_lookup(s, h[SYSV_HASH], p);
+		ok |= 1 << SYSV_HASH;
+	} else {
+		h[GNU_HASH] = gnu_hash(s);
+		sym = gnu_lookup(s, h[GNU_HASH], p);
+		ok |= 1 << GNU_HASH;
+	}
+
+
 	if (sym && sym->st_value && (1<<(sym->st_info&0xf) & OK_TYPES))
 		return p->base + sym->st_value;
 	if (p->deps) for (i=0; p->deps[i]; i++) {
-		sym = lookup(s, h, p->deps[i]);
+		algs = p->deps[i]->hashalgs;
+		if (!(algs & ok)) {
+			if (algs & SYSV_HASH) {
+				h[SYSV_HASH] = sysv_hash(s);
+				sym = sysv_lookup(s, h[SYSV_HASH], p->deps[i]);
+				ok |= SYSV_HASH;
+			} else {
+				h[GNU_HASH] = gnu_hash(s);
+				sym = gnu_lookup(s, h[GNU_HASH], p->deps[i]);
+				ok |= GNU_HASH;
+			}
+		} else {
+			if (algs & SYSV_HASH)
+				sym = sysv_lookup(s, h[SYSV_HASH], p->deps[i]);
+			else
+				sym = gnu_lookup(s, h[GNU_HASH], p->deps[i]);
+		}
+
 		if (sym && sym->st_value && (1<<(sym->st_info&0xf) & OK_TYPES))
 			return p->deps[i]->base + sym->st_value;
 	}
@@ -813,6 +958,97 @@ failed:
 	return 0;
 }
 
+struct sym_search {
+	void *addr;
+	Dl_info *info;
+};
+
+static int find_closest_sym (struct dso *dso, Sym *sym, struct sym_search *search)
+{
+	void *symaddr = dso->base + sym->st_value;
+	char *strings = dso->strings;
+	Dl_info *info = search->info;
+	void *addr = search->addr;
+	void *prevaddr = info->dli_saddr;
+
+	if (sym->st_value == 0 && sym->st_shndx == SHN_UNDEF)
+		return 1;
+
+	if (ELF_ST_TYPE(sym->st_info) == STT_TLS)
+		return 1;
+
+	if (addr < symaddr)
+		return 1;
+
+	if (prevaddr && (addr - symaddr) > (addr - prevaddr))
+		return 1;
+
+	info->dli_saddr = symaddr;
+	info->dli_sname = strings + sym->st_name;
+
+	if (addr == symaddr)
+		return 0;
+
+	return 1;
+
+}
+
+static int do_dladdr (void *addr, Dl_info *info)
+{
+	struct sym_search search;
+	struct dso *p;
+	memset (info, 0, sizeof (*info));
+	search.info = info;
+	search.addr = addr;
+	for (p=head; p; p=p->next) {
+		if ((unsigned char *)addr >= p->map && (unsigned char *)addr < p->map + p->map_len) {
+			Sym *syms = p->syms;
+			uint32_t *hashtab;
+			size_t i;
+			info->dli_fname = p->name;
+			info->dli_fbase = p->base;
+			if (p->hashalgs & (1 << SYSV_HASH)) {
+				hashtab = p->hashtabs[SYSV_HASH];
+				for (i = 0; i < hashtab[1]; i++) {
+					if (!find_closest_sym (p, syms + i, &search))
+						return 1;
+				}
+			} else if(p->hashalgs & (1 << GNU_HASH)) {
+				uint32_t *buckets;
+				uint32_t nbuckets;
+				uint32_t *hashvals;
+				uint32_t symndx;
+				hashtab = p->hashtabs[GNU_HASH];
+				buckets = hashtab + 4 + (hashtab[2] * (sizeof(size_t)/sizeof(uint32_t)));
+				nbuckets = hashtab[0];
+				hashvals = buckets + nbuckets;
+				symndx = hashtab[1];
+				for (i = 0; i < nbuckets; ++i) {
+					uint32_t n = buckets[i];
+					Sym *sym = syms + n;
+					uint32_t *hashval = hashvals + n - symndx;
+
+					do {
+						if (!find_closest_sym (p, sym, &search))
+							return 1;
+					}while (!(*hashval++ & 1));
+				}
+			}
+			return 1;
+		}
+	}
+	return 0;
+}
+
+int dladdr (void *addr, Dl_info *info)
+{
+	int res;
+	pthread_rwlock_rdlock(&lock);
+	res = do_dladdr (addr, info);
+	pthread_rwlock_unlock(&lock);
+	return res;
+}
+
 void *__dlsym(void *p, const char *s, void *ra)
 {
 	void *res;
-- 
1.7.9.5