internal_function
__libdw_fde_by_offset (Dwarf_CFI *cache, Dwarf_Off offset)
{
Dwarf_CFI_Entry entry;
Dwarf_Off next_offset;
int result = INTUSE(dwarf_next_cfi) (cache->e_ident,
&cache->data->d, CFI_IS_EH (cache),
offset, &next_offset, &entry);
if (result != 0)
{
if (result > 0)
invalid:
__libdw_seterrno (DWARF_E_INVALID_DWARF);
return NULL;
}
if (unlikely (dwarf_cfi_cie_p (&entry)))
goto invalid;
/* We have a new FDE to consider. */
struct dwarf_fde *fde = intern_fde (cache, &entry.fde);
if (fde == (void *) -1l || fde == NULL)
return NULL;
/* If this happened to be what we would have read next, notice it. */
if (cache->next_offset == offset)
cache->next_offset = next_offset;
return fde;
}
/* Load ELF 'filename', parse the .eh_frame contents, and for each entry in the
* second argument check whether its address is contained in the range of some
* Frame Description Entry. If it does, fill in the function range of the
* entry. In other words, try to assign start address and length of function
* corresponding to each backtrace entry. We'll need that for the disassembly.
*
* Fails quietly - we should still be able to use the build ids.
*
* I wonder if this is really better than parsing eu-readelf text output.
*/
static GHashTable *elf_iterate_fdes(const char *filename, GList *entries, Elf *e)
{
const unsigned char *e_ident;
Elf_Data *scn_data;
GElf_Shdr shdr;
GElf_Phdr phdr;
size_t phnum;
GHashTable *retval = NULL; /* NULL = error */
e_ident = (unsigned char *)elf_getident(e, NULL);
if (e_ident == NULL)
{
VERB1 log_elf_error("elf_getident", filename);
return NULL;
}
/* Look up the .eh_frame section */
if (!xelf_section_by_name(e, ".eh_frame", filename, &scn_data, &shdr))
{
VERB1 log("Section .eh_frame not found in %s", filename);
return NULL;
}
/* Get the address at which the executable segment is loaded. If the
* .eh_frame addresses are absolute, this is used to convert them to
* relative to the beginning of executable segment. We are looking for the
* first LOAD segment that is executable, I hope this is sufficient.
*/
if (elf_getphdrnum(e, &phnum) != 0)
{
VERB1 log_elf_error("elf_getphdrnum", filename);
return NULL;
}
uintptr_t exec_base;
int i;
for (i = 0; i < phnum; i++)
{
if (gelf_getphdr(e, i, &phdr) != &phdr)
{
VERB1 log_elf_error("gelf_getphdr", filename);
return NULL;
}
if (phdr.p_type == PT_LOAD && phdr.p_flags & PF_X)
{
exec_base = (uintptr_t)phdr.p_vaddr;
goto base_found;
}
}
VERB1 log("Can't determine executable base for '%s'", filename);
return NULL;
base_found:
VERB2 log("Executable base: %jx", (uintmax_t)exec_base);
/* We now have a handle to .eh_frame data. We'll use dwarf_next_cfi to
* iterate through all FDEs looking for those matching the addresses we
* have.
* Some info on .eh_frame can be found at http://www.airs.com/blog/archives/460
* and in DWARF documentation for .debug_frame. The initial_location and
* address_range decoding is 'inspired' by elfutils source.
* XXX: If this linear scan is too slow, we can do binary search on
* .eh_frame_hdr -- see http://www.airs.com/blog/archives/462
*/
int ret;
Dwarf_Off cfi_offset;
Dwarf_Off cfi_offset_next = 0;
Dwarf_CFI_Entry cfi;
struct cie_encoding {
Dwarf_Off cie_offset;
int ptr_len;
bool pcrel;
} *cie;
GList *cie_list = NULL;
/* Init hash table
* keys are pointers to integers which we allocate with malloc
* values stored directly */
GHashTable *hash = g_hash_table_new_full(g_int64_hash, g_int64_equal, free, NULL);
while(1)
{
cfi_offset = cfi_offset_next;
ret = dwarf_next_cfi(e_ident, scn_data, 1, cfi_offset, &cfi_offset_next, &cfi);
if (ret > 0)
{
/* We're at the end. */
break;
}
if (ret < 0)
{
/* Error. If cfi_offset_next was updated, we may skip the
* erroneous cfi. */
if (cfi_offset_next > cfi_offset)
{
continue;
}
VERB1 log("dwarf_next_cfi failed for %s: %s", filename, dwarf_errmsg(-1));
goto ret_free;
}
if (dwarf_cfi_cie_p(&cfi))
{
/* Current CFI is a CIE. We store its offset and FDE encoding
* attributes to be used when reading FDEs.
*/
/* Default FDE encoding (i.e. no R in augmentation string) is
* DW_EH_PE_absptr.
*/
cie = btp_mallocz(sizeof(*cie));
cie->cie_offset = cfi_offset;
cie->ptr_len = encoded_size(DW_EH_PE_absptr, e_ident);
/* Search the augmentation data for FDE pointer encoding.
* Unfortunately, 'P' can come before 'R' (which we are looking
* for), so we may have to parse the whole thing. See the
* abovementioned blog post for details.
*/
const char *aug = cfi.cie.augmentation;
const uint8_t *augdata = cfi.cie.augmentation_data;
bool skip_cie = 0;
if (*aug == 'z')
{
aug++;
}
while (*aug != '\0')
{
if(*aug == 'R')
{
cie->ptr_len = encoded_size(*augdata, e_ident);
if (cie->ptr_len != 4 && cie->ptr_len != 8)
{
VERB1 log("Unknown FDE encoding (CIE %jx) in %s",
(uintmax_t)cfi_offset, filename);
skip_cie = 1;
}
if ((*augdata & 0x70) == DW_EH_PE_pcrel)
{
cie->pcrel = 1;
}
break;
}
else if (*aug == 'L')
{
augdata++;
}
else if (*aug == 'P')
{
unsigned size = encoded_size(*augdata, e_ident);
if (size == 0)
{
VERB1 log("Unknown size for personality encoding in %s",
filename);
skip_cie = 1;
break;
}
augdata += (size + 1);
}
else
{
VERB1 log("Unknown augmentation char in %s", filename);
skip_cie = 1;
break;
}
aug++;
}
if (skip_cie)
{
free(cie);
continue;
}
cie_list = g_list_append(cie_list, cie);
}
else
{
/* Current CFI is an FDE.
*/
GList *it = cie_list;
cie = NULL;
/* Find the CIE data that we should have saved earlier. XXX: We can
* use hash table/tree to speed up the search, the number of CIEs
* should usally be very low though. */
while (it != NULL)
{
cie = it->data;
/* In .eh_frame, CIE_pointer is relative, but libdw converts it
* to absolute offset. */
if(cfi.fde.CIE_pointer == cie->cie_offset)
{
break; /* Found. */
}
it = g_list_next(it);
}
if (it == NULL)
{
VERB1 log("CIE not found for FDE %jx in %s",
(uintmax_t)cfi_offset, filename);
continue;
}
/* Read the two numbers we need and if they are PC-relative,
* compute the offset from VMA base
*/
uintptr_t initial_location = fde_read_address(cfi.fde.start, cie->ptr_len);
uintptr_t address_range = fde_read_address(cfi.fde.start+cie->ptr_len, cie->ptr_len);
if (cie->pcrel)
{
/* We need to determine how long is the 'length' (and
* consequently CIE id) field of this FDE -- it can be either 4
* or 12 bytes long. */
uintptr_t length = fde_read_address(scn_data->d_buf + cfi_offset, 4);
uintptr_t skip = (length == 0xffffffffUL ? 12 : 4);
uintptr_t mask = (cie->ptr_len == 4 ? 0xffffffffUL : 0xffffffffffffffffUL);
initial_location += (uintptr_t)shdr.sh_offset + (uintptr_t)cfi_offset + 2*skip;
initial_location &= mask;
}
else
{
/* Assuming that not pcrel means absolute address (what if the file is a library?).
* Convert to text-section-start-relative.
*/
initial_location -= exec_base;
}
/* Insert the pair into hash */
uintptr_t *key = addr_alloc(initial_location + exec_base);
g_hash_table_insert(hash, key, (gpointer)address_range);
VERB3 log("FDE start: 0x%jx length: %u", (uintmax_t)*key, (unsigned)address_range);
/* Iterate through the backtrace entries and check each address
* member whether it belongs into the range given by current FDE.
*/
for (it = entries; it != NULL; it = g_list_next(it))
{
struct backtrace_entry *entry = it->data;
if (initial_location <= entry->build_id_offset
&& entry->build_id_offset < initial_location + address_range)
{
/* Convert to before-relocation absolute addresses, disassembler uses those. */
entry->function_initial_loc = exec_base + initial_location;
entry->function_length = address_range;
/*TODO: remove the entry from the list to save a bit of time in next iteration?*/
}
}
}
}
retval = hash; /* success */
ret_free:
list_free_with_free(cie_list);
if (retval == NULL)
g_hash_table_destroy(hash);
return retval;
}
struct sr_elf_fde *
sr_elf_get_eh_frame(const char *filename,
char **error_message)
{
#ifdef WITH_ELFUTILS
/* Open the input file. */
int fd = open(filename, O_RDONLY);
if (fd < 0)
{
*error_message = sr_asprintf("Failed to open file %s: %s",
filename,
strerror(errno));
return NULL;
}
/* Initialize libelf on the opened file. */
Elf *elf = elf_begin(fd, ELF_C_READ, NULL);
if (!elf)
{
*error_message = sr_asprintf("Failed to run elf_begin on file %s: %s",
filename,
elf_errmsg(-1));
close(fd);
return NULL;
}
unsigned char *e_ident = (unsigned char *)elf_getident(elf, NULL);
if (!e_ident)
{
*error_message = sr_asprintf("elf_getident failed for %s: %s",
filename,
elf_errmsg(-1));
elf_end(elf);
close(fd);
return NULL;
}
/* Look up the .eh_frame section */
GElf_Shdr shdr;
Elf_Data *section_data;
char *find_section_error_message;
if (!find_elf_section_by_name(elf,
".eh_frame",
§ion_data,
&shdr,
&find_section_error_message))
{
*error_message = sr_asprintf("Failed to find .eh_frame section for %s: %s",
filename,
find_section_error_message);
free(find_section_error_message);
elf_end(elf);
close(fd);
return NULL;
}
/* Get the address at which the executable segment is loaded. If
* the .eh_frame addresses are absolute, this is used to convert
* them to relative to the beginning of executable segment. We
* are looking for the first LOAD segment that is executable, I
* hope this is sufficient.
*/
size_t phnum;
if (elf_getphdrnum(elf, &phnum) != 0)
{
*error_message = sr_asprintf("elf_getphdrnum failed for %s: %s",
filename,
elf_errmsg(-1));
elf_end(elf);
close(fd);
return NULL;
}
uint64_t exec_base = -1;
int i;
for (i = 0; i < phnum; i++)
{
GElf_Phdr phdr;
if (gelf_getphdr(elf, i, &phdr) != &phdr)
{
*error_message = sr_asprintf("gelf_getphdr failed for %s: %s",
filename,
elf_errmsg(-1));
elf_end(elf);
close(fd);
return NULL;
}
if (phdr.p_type == PT_LOAD && phdr.p_flags & PF_X)
{
exec_base = phdr.p_vaddr;
break;
}
}
if (-1 == exec_base)
{
*error_message = sr_asprintf("Can't determine executable base for %s",
filename);
elf_end(elf);
close(fd);
return NULL;
}
/* We now have a handle to .eh_frame data. We'll use
* dwarf_next_cfi to iterate through all FDEs looking for those
* matching the addresses we have.
*
* Some info on .eh_frame can be found at
* http://www.airs.com/blog/archives/460 and in DWARF
* documentation for .debug_frame. The initial_location and
* address_range decoding is 'inspired' by elfutils source.
*
* @todo If this linear scan is too slow, we can do binary search
* on .eh_frame_hdr -- see http://www.airs.com/blog/archives/462
*/
struct sr_elf_fde *result = NULL, *last = NULL;
struct cie *cie_list = NULL, *cie_list_last = NULL;
Dwarf_Off cfi_offset_next = 0;
while (true)
{
Dwarf_CFI_Entry cfi;
Dwarf_Off cfi_offset = cfi_offset_next;
int ret = dwarf_next_cfi(e_ident,
section_data,
1,
cfi_offset,
&cfi_offset_next,
&cfi);
if (ret > 0)
{
/* We're at the end. */
break;
}
if (ret < 0)
{
/* Error. If cfi_offset_next was updated, we may skip the
* erroneous cfi. */
if (cfi_offset_next > cfi_offset)
continue;
*error_message = sr_asprintf("dwarf_next_cfi failed for %s: %s",
filename,
dwarf_errmsg(-1));
cie_free(cie_list);
sr_elf_eh_frame_free(result);
elf_end(elf);
close(fd);
return NULL;
}
if (dwarf_cfi_cie_p(&cfi))
{
/* Current CFI is a CIE. We store its offset and FDE encoding
* attributes to be used when reading FDEs.
*/
/* Default FDE encoding (i.e. no R in augmentation string) is
* DW_EH_PE_absptr.
*/
char *cie_error_message;
struct cie *cie = read_cie(&cfi,
cfi_offset,
e_ident,
&cie_error_message);
if (!cie)
{
*error_message = sr_asprintf("CIE reading failed for %s: %s",
filename,
cie_error_message);
free(cie_error_message);
cie_free(cie_list);
sr_elf_eh_frame_free(result);
elf_end(elf);
close(fd);
return NULL;
}
/* Append the newly parsed CIE to our list of CIEs. */
if (cie_list)
{
cie_list_last->next = cie;
cie_list_last = cie;
}
else
cie_list = cie_list_last = cie;
}
else
{
/* Current CFI is an FDE.
*/
struct cie *cie = cie_list;
/* Find the CIE data that we should have saved earlier. */
while (cie)
{
/* In .eh_frame, CIE_pointer is relative, but libdw converts it
* to absolute offset. */
if (cfi.fde.CIE_pointer == cie->cie_offset)
break; /* Found. */
cie = cie->next;
}
if (!cie)
{
*error_message = sr_asprintf("CIE not found for FDE %jx in %s",
(uintmax_t)cfi_offset,
filename);
cie_free(cie_list);
sr_elf_eh_frame_free(result);
elf_end(elf);
close(fd);
return NULL;
}
/* Read the two numbers we need and if they are PC-relative,
* compute the offset from VMA base
*/
uint64_t initial_location = fde_read_address(cfi.fde.start,
cie->ptr_len);
uint64_t address_range = fde_read_address(cfi.fde.start + cie->ptr_len,
cie->ptr_len);
if (cie->pcrel)
{
/* We need to determine how long is the 'length' (and
* consequently CIE id) field of this FDE -- it can be
* either 4 or 12 bytes long. */
uint64_t length = fde_read_address(section_data->d_buf + cfi_offset, 4);
uint64_t skip = (length == 0xffffffffUL ? 12 : 4);
uint64_t mask = (cie->ptr_len == 4 ? 0xffffffffUL : 0xffffffffffffffffUL);
initial_location += shdr.sh_offset + cfi_offset + 2 * skip;
initial_location &= mask;
}
else
{
/* Assuming that not pcrel means absolute address
* (what if the file is a library?). Convert to
* text-section-start-relative.
*/
initial_location -= exec_base;
}
struct sr_elf_fde *fde = sr_malloc(sizeof(struct sr_elf_fde));
fde->exec_base = exec_base;
fde->start_address = initial_location;
fde->length = address_range;
fde->next = NULL;
/* Append the newly parsed Frame Description Entry to our
* list of FDEs.
*/
if (result)
{
last->next = fde;
last = fde;
}
else
result = last = fde;
}
}
cie_free(cie_list);
elf_end(elf);
close(fd);
return result;
#else /* WITH_ELFUTILS */
*error_message = sr_asprintf("satyr compiled without elfutils");
return NULL;
#endif /* WITH_ELFUTILS */
}
/* Use a binary search table in .eh_frame_hdr format, yield an FDE offset. */
static Dwarf_Off
binary_search_fde (Dwarf_CFI *cache, Dwarf_Addr address)
{
const size_t size = 2 * encoded_value_size (&cache->data->d, cache->e_ident,
cache->search_table_encoding,
NULL);
/* Dummy used by read_encoded_value. */
Dwarf_CFI dummy_cfi =
{
.e_ident = cache->e_ident,
.datarel = cache->search_table_vaddr,
.frame_vaddr = cache->search_table_vaddr,
};
size_t l = 0, u = cache->search_table_entries;
while (l < u)
{
size_t idx = (l + u) / 2;
const uint8_t *p = &cache->search_table[idx * size];
Dwarf_Addr start;
if (unlikely (read_encoded_value (&dummy_cfi,
cache->search_table_encoding, &p,
&start)))
break;
if (address < start)
u = idx;
else
{
Dwarf_Addr fde;
if (unlikely (read_encoded_value (&dummy_cfi,
cache->search_table_encoding, &p,
&fde)))
break;
if (address >= start)
{
l = idx + 1;
/* If this is the last entry, its upper bound is assumed to be
the end of the module.
XXX really should be end of containing PT_LOAD segment */
if (l < cache->search_table_entries)
{
/* Look at the start address in the following entry. */
Dwarf_Addr end;
if (unlikely (read_encoded_value
(&dummy_cfi, cache->search_table_encoding, &p,
&end)))
break;
if (address >= end)
continue;
}
return fde - cache->frame_vaddr;
}
}
}
return (Dwarf_Off) -1l;
}
struct dwarf_fde *
internal_function
__libdw_find_fde (Dwarf_CFI *cache, Dwarf_Addr address)
{
/* Look for a cached FDE covering this address. */
const struct dwarf_fde fde_key = { .start = address, .end = 0 };
struct dwarf_fde **found = tfind (&fde_key, &cache->fde_tree, &compare_fde);
if (found != NULL)
return *found;
/* Use .eh_frame_hdr binary search table if possible. */
if (cache->search_table != NULL)
{
Dwarf_Off offset = binary_search_fde (cache, address);
if (offset == (Dwarf_Off) -1l)
goto no_match;
struct dwarf_fde *fde = __libdw_fde_by_offset (cache, offset);
if (unlikely (fde != NULL)
/* Sanity check the address range. */
&& unlikely (address < fde->start || address >= fde->end))
{
__libdw_seterrno (DWARF_E_INVALID_DWARF);
return NULL;
}
return fde;
}
/* It's not there. Read more CFI entries until we find it. */
while (1)
{
Dwarf_Off last_offset = cache->next_offset;
Dwarf_CFI_Entry entry;
int result = INTUSE(dwarf_next_cfi) (cache->e_ident,
&cache->data->d, CFI_IS_EH (cache),
last_offset, &cache->next_offset,
&entry);
if (result > 0)
break;
if (result < 0)
{
if (cache->next_offset == last_offset)
/* We couldn't progress past the bogus FDE. */
break;
/* Skip the loser and look at the next entry. */
continue;
}
if (dwarf_cfi_cie_p (&entry))
{
/* This is a CIE, not an FDE. We eagerly intern these
because the next FDE will usually refer to this CIE. */
__libdw_intern_cie (cache, last_offset, &entry.cie);
continue;
}
/* We have a new FDE to consider. */
struct dwarf_fde *fde = intern_fde (cache, &entry.fde);
if (fde == (void *) -1l) /* Bad FDE, but we can keep looking. */
continue;
if (fde == NULL) /* Bad data. */
return NULL;
/* Is this the one we're looking for? */
if (fde->start <= address && fde->end > address)
return fde;
}
no_match:
/* We found no FDE covering this address. */
__libdw_seterrno (DWARF_E_NO_MATCH);
return NULL;
}