memprof: extend symtab with C symbols

This commit provides an ability to enrich memprof's symbol table with
information about C symbols. The parser is updated correspondingly.

If there is .symtab section or at least .dynsym segment in a shared
library, then memprof dumps its contents in symtab. The <sym> was
extended the following way:
| sym := sym-lua | sym-cfunc | sym-trace | sym-final
| sym-cfunc := sym-header sym-addr sym-name

If none of those are present, then instead of a symbol name and its
address there will be name and address of a shared library containing
that symbol.

Resolves tarantool/tarantool#5813

Reviewed-by: Sergey Kaplun <skaplun@tarantool.org>
Reviewed-by: Igor Munkin <imun@tarantool.org>
Signed-off-by: Igor Munkin <imun@tarantool.org>

Author: mkokryashkin

Makefile.original

@@ -100,7 +100,7 @@ FILES_JITLIB= bc.lua bcsave.lua dump.lua p.lua v.lua zone.lua \
 	      dis_x86.lua dis_x64.lua dis_arm.lua dis_arm64.lua \
 	      dis_arm64be.lua dis_ppc.lua dis_mips.lua dis_mipsel.lua \
 	      dis_mips64.lua dis_mips64el.lua vmdef.lua
-FILES_UTILSLIB= bufread.lua symtab.lua
+FILES_UTILSLIB= avl.lua bufread.lua symtab.lua
 FILES_MEMPROFLIB= parse.lua humanize.lua
 FILES_TOOLSLIB= memprof.lua
 FILE_TMEMPROF= luajit-parse-memprof

src/lj_arch.h

@@ -583,6 +583,11 @@
 #define LJ_HASMEMPROF		0
 #else
 #define LJ_HASMEMPROF		1
+#if LJ_TARGET_LINUX
+#define LJ_HASRESOLVER		1
+#else
+#define LJ_HASRESOLVER		0
+#endif
 #endif
 
 #endif

src/lj_memprof.c

@@ -8,6 +8,8 @@
 #define lj_memprof_c
 #define LUA_CORE
 
+#define _GNU_SOURCE
+
 #include <errno.h>
 
 #include "lj_arch.h"
@@ -19,6 +21,14 @@
 #include "lj_frame.h"
 #include "lj_debug.h"
 
+#if LJ_HASRESOLVER
+#include <elf.h>
+#include <link.h>
+#include <stdio.h>
+#include <sys/auxv.h>
+#include "lj_gc.h"
+#endif
+
 #if LJ_HASJIT
 #include "lj_dispatch.h"
 #endif
@@ -71,12 +81,326 @@ static void dump_symtab_proto(struct lj_wbuf *out, const GCproto *pt)
   lj_wbuf_addu64(out, (uint64_t)pt->firstline);
 }
 
+#if LJ_HASRESOLVER
+
+struct ghashtab_header {
+  uint32_t nbuckets;
+  uint32_t symoffset;
+  uint32_t bloom_size;
+  uint32_t bloom_shift;
+};
+
+static uint32_t ghashtab_size(ElfW(Addr) ghashtab)
+{
+  /*
+  ** There is no easy way to get count of symbols in GNU hashtable, so the
+  ** only way to do this is to take highest possible non-empty bucket and
+  ** iterate through its symbols until the last chain is over.
+  */
+  uint32_t last_entry = 0;
+
+  const uint32_t *chain = NULL;
+  struct ghashtab_header *header = (struct ghashtab_header *)ghashtab;
+  /*
+  ** sizeof(size_t) returns 8, if compiled with 64-bit compiler, and 4 if
+  ** compiled with 32-bit compiler. It is the best option to determine which
+  ** kind of CPU we are running on.
+  */
+  const char *buckets = (char *)ghashtab + sizeof(struct ghashtab_header) +
+                        sizeof(size_t) * header->bloom_size;
+
+  uint32_t *cur_bucket = (uint32_t *)buckets;
+  uint32_t i;
+  for (i = 0; i < header->nbuckets; ++i) {
+    if (last_entry < *cur_bucket)
+      last_entry = *cur_bucket;
+    cur_bucket++;
+  }
+
+  if (last_entry < header->symoffset)
+    return header->symoffset;
+
+  chain = (uint32_t *)(buckets + sizeof(uint32_t) * header->nbuckets);
+  /* The chain ends with the lowest bit set to 1. */
+  while (!(chain[last_entry - header->symoffset] & 1))
+    last_entry++;
+
+  return ++last_entry;
+}
+
+static void write_c_symtab(ElfW(Sym *) sym, char *strtab, ElfW(Addr) so_addr,
+			   size_t sym_cnt, struct lj_wbuf *buf)
+{
+  /*
+  ** Index 0 in ELF symtab is used to represent undefined symbols. Hence, we
+  ** can just start with index 1.
+  **
+  ** For more information, see:
+  ** https://docs.oracle.com/cd/E23824_01/html/819-0690/chapter6-79797.html
+  */
+
+  ElfW(Word) sym_index;
+  for (sym_index = 1; sym_index < sym_cnt; sym_index++) {
+    /*
+    ** ELF32_ST_TYPE and ELF64_ST_TYPE are the same, so we can use
+    ** ELF32_ST_TYPE for both 64-bit and 32-bit ELFs.
+    **
+    ** For more, see https://github.com/torvalds/linux/blob/9137eda53752ef73148e42b0d7640a00f1bc96b1/include/uapi/linux/elf.h#L135
+    */
+    if (ELF32_ST_TYPE(sym[sym_index].st_info) == STT_FUNC &&
+        sym[sym_index].st_name != 0) {
+      char *sym_name = &strtab[sym[sym_index].st_name];
+      lj_wbuf_addbyte(buf, SYMTAB_CFUNC);
+      lj_wbuf_addu64(buf, sym[sym_index].st_value + so_addr);
+      lj_wbuf_addstring(buf, sym_name);
+    }
+  }
+}
+
+static int dump_sht_symtab(const char *elf_name, struct lj_wbuf *buf,
+			   lua_State *L, const ElfW(Addr) so_addr)
+{
+  int status = 0;
+
+  char *strtab = NULL;
+  ElfW(Shdr *) section_headers = NULL;
+  ElfW(Sym *) sym = NULL;
+  ElfW(Ehdr) elf_header = {};
+
+  ElfW(Off) sym_off = 0;
+  ElfW(Off) strtab_off = 0;
+
+  size_t sym_cnt = 0;
+  size_t strtab_size = 0;
+  size_t header_index = 0;
+
+  size_t shoff = 0; /* Section headers offset. */
+  size_t shnum = 0; /* Section headers number. */
+  size_t shentsize = 0; /* Section header entry size. */
+
+  FILE *elf_file = fopen(elf_name, "rb");
+
+  if (elf_file == NULL)
+    return -1;
+
+  if (fread(&elf_header, sizeof(elf_header), 1, elf_file) != sizeof(elf_header)
+      && ferror(elf_file) != 0)
+    goto error;
+  if (memcmp(elf_header.e_ident, ELFMAG, SELFMAG) != 0)
+    /* Not a valid ELF file. */
+    goto error;
+
+  shoff = elf_header.e_shoff;
+  shnum = elf_header.e_shnum;
+  shentsize = elf_header.e_shentsize;
+
+  if (shoff == 0 || shnum == 0 || shentsize == 0)
+    /* No sections in ELF. */
+    goto error;
+
+  /*
+  ** Memory occupied by section headers is unlikely to be more than 160B, but
+  ** 32-bit and 64-bit ELF files may have sections of different sizes and some
+  ** of the sections may duiplicate, so we need to take that into account.
+  */
+  section_headers = lj_mem_new(L, shnum * shentsize);
+  if (section_headers == NULL)
+    goto error;
+
+  if (fseek(elf_file, shoff, SEEK_SET) != 0)
+    goto error;
+
+  if (fread(section_headers, shentsize, shnum, elf_file) != shentsize * shnum
+      && ferror(elf_file) != 0)
+    goto error;
+
+  for (header_index = 0; header_index < shnum; ++header_index) {
+    if (section_headers[header_index].sh_type == SHT_SYMTAB) {
+      ElfW(Shdr) sym_hdr = section_headers[header_index];
+      ElfW(Shdr) strtab_hdr = section_headers[sym_hdr.sh_link];
+      size_t symtab_size = sym_hdr.sh_size;
+
+      sym_off = sym_hdr.sh_offset;
+      sym_cnt = symtab_size / sym_hdr.sh_entsize;
+
+      strtab_off = strtab_hdr.sh_offset;
+      strtab_size = strtab_hdr.sh_size;
+      break;
+    }
+  }
+
+  if (sym_off == 0 || strtab_off == 0 || sym_cnt == 0)
+    goto error;
+
+  /* Load symtab into memory. */
+  sym = lj_mem_new(L, sym_cnt * sizeof(ElfW(Sym)));
+  if (sym == NULL)
+    goto error;
+  if (fseek(elf_file, sym_off, SEEK_SET) != 0)
+    goto error;
+  if (fread(sym, sizeof(ElfW(Sym)), sym_cnt, elf_file) !=
+      sizeof(ElfW(Sym)) * sym_cnt && ferror(elf_file) != 0)
+    goto error;
+
+
+  /* Load strtab into memory. */
+  strtab = lj_mem_new(L, strtab_size * sizeof(char));
+  if (strtab == NULL)
+    goto error;
+  if (fseek(elf_file, strtab_off, SEEK_SET) != 0)
+    goto error;
+  if (fread(strtab, sizeof(char), strtab_size, elf_file) !=
+      sizeof(char) * strtab_size && ferror(elf_file) != 0)
+    goto error;
+
+  write_c_symtab(sym, strtab, so_addr, sym_cnt, buf);
+
+  goto end;
+
+error:
+  status = -1;
+
+end:
+  if (sym != NULL)
+    lj_mem_free(G(L), sym, sym_cnt * sizeof(ElfW(Sym)));
+  if(strtab != NULL)
+    lj_mem_free(G(L), strtab, strtab_size * sizeof(char));
+  if(section_headers != NULL)
+    lj_mem_free(G(L), section_headers, shnum * shentsize);
+
+  fclose(elf_file);
+
+  return status;
+}
+
+static int dump_dyn_symtab(struct dl_phdr_info *info, struct lj_wbuf *buf)
+{
+  size_t header_index;
+  for (header_index = 0; header_index < info->dlpi_phnum; ++header_index) {
+    if (info->dlpi_phdr[header_index].p_type == PT_DYNAMIC) {
+      ElfW(Dyn *) dyn =
+	(ElfW(Dyn) *)(info->dlpi_addr + info->dlpi_phdr[header_index].p_vaddr);
+      ElfW(Sym *) sym = NULL;
+      ElfW(Word *) hashtab = NULL;
+      ElfW(Addr) ghashtab = 0;
+      ElfW(Word) sym_cnt = 0;
+
+      char *strtab = 0;
+
+      for(; dyn->d_tag != DT_NULL; dyn++) {
+        switch(dyn->d_tag) {
+        case DT_HASH:
+          hashtab = (ElfW(Word *))dyn->d_un.d_ptr;
+          break;
+        case DT_GNU_HASH:
+          ghashtab = dyn->d_un.d_ptr;
+          break;
+        case DT_STRTAB:
+          strtab = (char *)dyn->d_un.d_ptr;
+          break;
+        case DT_SYMTAB:
+          sym = (ElfW(Sym *))dyn->d_un.d_ptr;
+          break;
+        default:
+          break;
+        }
+      }
+
+      if ((hashtab == NULL && ghashtab == 0) || strtab == NULL || sym == NULL)
+        /* Not enough data to resolve symbols. */
+        return 1;
+
+      /*
+      ** A hash table consists of Elf32_Word or Elf64_Word objects that provide
+      ** for symbol table access. Hash table has the following organization:
+      ** +-------------------+
+      ** |      nbucket      |
+      ** +-------------------+
+      ** |      nchain       |
+      ** +-------------------+
+      ** |     bucket[0]     |
+      ** |       ...         |
+      ** | bucket[nbucket-1] |
+      ** +-------------------+
+      ** |     chain[0]      |
+      ** |       ...         |
+      ** |  chain[nchain-1]  |
+      ** +-------------------+
+      ** Chain table entries parallel the symbol table. The number of symbol
+      ** table entries should equal nchain, so symbol table indexes also select
+      ** chain table entries. Since the chain array values are indexes for not
+      ** only the chain array itself, but also for the symbol table, the chain
+      ** array must be the same size as the symbol table. This makes nchain
+      ** equal to the length of the symbol table.
+      **
+      ** For more, see https://docs.oracle.com/cd/E23824_01/html/819-0690/chapter6-48031.html
+      */
+      sym_cnt = ghashtab == 0 ? hashtab[1] : ghashtab_size(ghashtab);
+      write_c_symtab(sym, strtab, info->dlpi_addr, sym_cnt, buf);
+      return 0;
+    }
+  }
+
+  return 1;
+}
+
+struct symbol_resolver_conf {
+  struct lj_wbuf *buf;
+  lua_State *L;
+};
+
+static int resolve_symbolnames(struct dl_phdr_info *info, size_t info_size,
+			       void *data)
+{
+  struct symbol_resolver_conf *conf = data;
+  struct lj_wbuf *buf = conf->buf;
+  lua_State *L = conf->L;
+
+  UNUSED(info_size);
+
+  /* Skip vDSO library. */
+  if (info->dlpi_addr == getauxval(AT_SYSINFO_EHDR))
+    return 0;
+
+  /*
+  ** Main way: try to open ELF and read SHT_SYMTAB, SHT_STRTAB and SHT_HASH
+  ** sections from it.
+  */
+  if (dump_sht_symtab(info->dlpi_name, buf, L, info->dlpi_addr) == 0) {
+    /* Empty body. */
+  }
+  /* First fallback: dump functions only from PT_DYNAMIC segment. */
+  else if(dump_dyn_symtab(info, buf) == 0) {
+    /* Empty body. */
+  }
+  /*
+  ** Last resort: dump ELF size and address to show .so name for its functions
+  ** in memprof output.
+  */
+  else {
+    lj_wbuf_addbyte(buf, SYMTAB_CFUNC);
+    lj_wbuf_addu64(buf, info->dlpi_addr);
+    lj_wbuf_addstring(buf, info->dlpi_name);
+  }
+
+  return 0;
+}
+
+#endif /* LJ_HASRESOLVER */
+
 static void dump_symtab(struct lj_wbuf *out, const struct global_State *g)
 {
   const GCRef *iter = &g->gc.root;
   const GCobj *o;
   const size_t ljs_header_len = sizeof(ljs_header) / sizeof(ljs_header[0]);
 
+#if LJ_HASRESOLVER
+  struct symbol_resolver_conf conf = {
+    .buf = out,
+    .L = gco2th(gcref(g->cur_L)),
+  };
+#endif
+
   /* Write prologue. */
   lj_wbuf_addn(out, ljs_header, ljs_header_len);
 
@@ -99,6 +423,10 @@ static void dump_symtab(struct lj_wbuf *out, const struct global_State *g)
     iter = &o->gch.nextgc;
   }
 
+#if LJ_HASRESOLVER
+  /* Write C symbols. */
+  dl_iterate_phdr(resolve_symbolnames, &conf);
+#endif
   lj_wbuf_addbyte(out, SYMTAB_FINAL);
 }