/* view_size can be the size of the first mapping, to handle non-contiguous
* modules -- we'll update the module's size here
*/
void
os_module_area_init(module_area_t *ma, app_pc base, size_t view_size, bool at_map,
const char *filepath, uint64 inode HEAPACCT(which_heap_t which))
{
app_pc mod_base, mod_end;
ptr_int_t load_delta;
char *soname = NULL;
ASSERT(module_is_header(base, view_size));
/* i#1589: use privload data if it exists (for client lib) */
if (!privload_fill_os_module_info(base, &mod_base, &mod_end, &soname, &ma->os_data)) {
/* XXX i#1860: on Android we'll fail to fill in info from .dynamic, so
* we'll have incomplete data until the loader maps the segment with .dynamic.
* ma->os_data.have_dynamic_info indicates whether we have the info.
*/
module_walk_program_headers(base, view_size, at_map,
!at_map, /* i#1589: ld.so relocates .dynamic */
&mod_base, NULL, &mod_end, &soname, &ma->os_data);
}
if (ma->os_data.contiguous) {
app_pc map_end = ma->os_data.segments[ma->os_data.num_segments - 1].end;
module_list_add_mapping(ma, base, map_end);
/* update, since may just be 1st segment size */
ma->end = map_end;
} else {
/* Add the non-contiguous segments (i#160/PR 562667). We could just add
* them all separately but vmvectors are more efficient with fewer
* entries so we merge. We don't want general merging in our vector
* either.
*/
app_pc seg_base;
uint i;
ASSERT(ma->os_data.num_segments > 0 && ma->os_data.segments != NULL);
seg_base = ma->os_data.segments[0].start;
for (i = 1; i < ma->os_data.num_segments; i++) {
if (ma->os_data.segments[i].start > ma->os_data.segments[i - 1].end ||
/* XXX: for shared we just add the first one. But if the first
* module is unloaded we'll be missing an entry for the others.
* We assume this won't happen b/c our only use of this now is
* the MacOS dyld shared cache's shared __LINKEDIT segment. If
* it could happen we should switch to a refcount in the vector.
*/
ma->os_data.segments[i - 1].shared) {
if (!ma->os_data.segments[i - 1].shared ||
!vmvector_overlap(loaded_module_areas, seg_base,
ma->os_data.segments[i - 1].end)) {
module_list_add_mapping(ma, seg_base,
ma->os_data.segments[i - 1].end);
}
seg_base = ma->os_data.segments[i].start;
}
}
if (!ma->os_data.segments[i - 1].shared ||
!vmvector_overlap(loaded_module_areas, seg_base,
ma->os_data.segments[i - 1].end))
module_list_add_mapping(ma, seg_base, ma->os_data.segments[i - 1].end);
DOLOG(2, LOG_VMAREAS, {
LOG(GLOBAL, LOG_INTERP | LOG_VMAREAS, 2, "segment list\n");
for (i = 0; i < ma->os_data.num_segments; i++) {
LOG(GLOBAL, LOG_INTERP | LOG_VMAREAS, 2,
"\tsegment %d: [" PFX "," PFX ") prot=%x\n", i,
ma->os_data.segments[i].start, ma->os_data.segments[i].end,
ma->os_data.segments[i].prot);
}
});
static int
dl_iterate_get_path_cb(struct dl_phdr_info *info, size_t size, void *data)
{
dl_iterate_data_t *iter_data = (dl_iterate_data_t *)data;
/* info->dlpi_addr is offset from preferred so we need to calculate the
* absolute address of the base.
* we can calculate the absolute address of the first segment, but ELF
* doesn't seem to guarantee that either the elf header (base of
* file) or the program headers (info->dlpi_phdr) are later than
* the min_vaddr, so it's a little confusing as to what would be
* in the maps file or whatever and would thus be the base we're looking
* to match: for now we assume the page with min_vaddr is that base.
* If elf header, program header, and 1st segment could all be on
* separate pages, I don't see any way to find the elf header in
* such cases short of walking backward and looking for the magic #s.
*/
app_pc pref_start, pref_end;
app_pc min_vaddr = module_vaddr_from_prog_header((app_pc)info->dlpi_phdr,
info->dlpi_phnum, NULL, NULL);
app_pc base = info->dlpi_addr + min_vaddr;
/* Note that dl_iterate_phdr doesn't give a name for the executable or
* ld-linux.so presumably b/c those are mapped by the kernel so the
* user-space loader doesn't need to know their file paths.
*/
LOG(GLOBAL, LOG_VMAREAS, 2,
"dl_iterate_get_path_cb: addr=" PFX " hdrs=" PFX " base=" PFX " name=%s\n",
info->dlpi_addr, info->dlpi_phdr, base, info->dlpi_name);
/* all we have is an addr somewhere in the module, so we need the end */
if (module_walk_program_headers(base,
/* FIXME: don't have view size: but
* anything larger than header sizes works
*/
PAGE_SIZE, false,
true, /* i#1589: ld.so relocated .dynamic */
&pref_start, NULL, &pref_end, NULL, NULL)) {
/* we're passed back start,end of preferred base */
if ((iter_data->target_addr != NULL && iter_data->target_addr >= base &&
iter_data->target_addr < base + (pref_end - pref_start)) ||
(iter_data->target_path != NULL &&
/* if we're passed an ambiguous name, we return first hit.
* if passed full path, should normally be what was used to
* load, so should match.
*/
strstr(info->dlpi_name, iter_data->target_path) != NULL)) {
if (iter_data->path_size > 0) {
/* We want just the path, not the filename */
char *slash = strrchr(info->dlpi_name, '/');
ASSERT_CURIOSITY(slash != NULL);
ASSERT_CURIOSITY((slash - info->dlpi_name) < iter_data->path_size);
strncpy(iter_data->path_out, info->dlpi_name,
MIN(iter_data->path_size, (slash - info->dlpi_name)));
iter_data->path_out[iter_data->path_size] = '\0';
}
iter_data->mod_start = base;
iter_data->mod_end = base + (pref_end - pref_start);
return 1; /* done iterating */
}
} else {
ASSERT_NOT_REACHED();
}
return 0; /* keep looking */
}
/* callback for dl_iterate_phdr() for adding existing modules to our lists */
static int
dl_iterate_get_areas_cb(struct dl_phdr_info *info, size_t size, void *data)
{
int *count = (int *)data;
uint i;
/* see comments in dl_iterate_get_path_cb() */
app_pc modend;
app_pc min_vaddr = module_vaddr_from_prog_header((app_pc)info->dlpi_phdr,
info->dlpi_phnum, NULL, &modend);
app_pc modbase = info->dlpi_addr + min_vaddr;
size_t modsize = modend - min_vaddr;
LOG(GLOBAL, LOG_VMAREAS, 2,
"dl_iterate_get_areas_cb: addr=" PFX " hdrs=" PFX " base=" PFX " name=%s\n",
info->dlpi_addr, info->dlpi_phdr, modbase, info->dlpi_name);
ASSERT(info->dlpi_phnum == module_num_program_headers(modbase));
ASSERT(count != NULL);
if (*count == 0) {
/* since we don't get a name for the executable, for now we
* assume that the first iter is the executable itself.
* XXX: this seems to hold, but there's no guarantee: can we do better?
*/
executable_start = modbase;
}
#ifndef X64
if (modsize == PAGE_SIZE && info->dlpi_name[0] == '\0') {
/* Candidate for VDSO. Xref PR 289138 on using AT_SYSINFO to locate. */
/* Xref VSYSCALL_PAGE_START_HARDCODED but later linuxes randomize */
char *soname;
if (module_walk_program_headers(modbase, modsize, false,
true, /* i#1589: ld.so relocated .dynamic */
NULL, NULL, NULL, &soname, NULL) &&
strncmp(soname, VSYSCALL_PAGE_SO_NAME, strlen(VSYSCALL_PAGE_SO_NAME)) == 0) {
ASSERT(!dynamo_initialized); /* .data should be +w */
ASSERT(vsyscall_page_start == NULL);
vsyscall_page_start = modbase;
LOG(GLOBAL, LOG_VMAREAS, 1, "found vsyscall page @ " PFX "\n",
vsyscall_page_start);
}
}
#endif
if (modbase != vsyscall_page_start)
module_list_add(modbase, modsize, false, info->dlpi_name, 0 /*don't have inode*/);
for (i = 0; i < info->dlpi_phnum; i++) {
app_pc start, end;
uint prot;
size_t align;
if (module_read_program_header(modbase, i, &start, &end, &prot, &align)) {
start += info->dlpi_addr;
end += info->dlpi_addr;
LOG(GLOBAL, LOG_VMAREAS, 2, "\tsegment %d: " PFX "-" PFX " %s align=%d\n", i,
start, end, memprot_string(prot), align);
start = (app_pc)ALIGN_BACKWARD(start, PAGE_SIZE);
end = (app_pc)ALIGN_FORWARD(end, PAGE_SIZE);
LOG(GLOBAL, LOG_VMAREAS, 4,
"find_executable_vm_areas: adding: " PFX "-" PFX " prot=%d\n", start, end,
prot);
all_memory_areas_lock();
update_all_memory_areas(start, end, prot, DR_MEMTYPE_IMAGE);
all_memory_areas_unlock();
if (app_memory_allocation(NULL, start, end - start, prot,
true /*image*/
_IF_DEBUG("ELF SO")))
(*count)++;
}
}
return 0; /* keep iterating */
}
/* See memquery.h for full interface specs, which are identical to
* memquery_library_bounds().
*/
int
memquery_library_bounds_by_iterator(const char *name, app_pc *start/*IN/OUT*/,
app_pc *end/*OUT*/,
char *fullpath/*OPTIONAL OUT*/, size_t path_size)
{
int count = 0;
bool found_library = false;
char libname[MAXIMUM_PATH];
const char *name_cmp = name;
memquery_iter_t iter;
app_pc last_base = NULL;
app_pc last_end = NULL;
size_t image_size = 0;
app_pc cur_end = NULL;
app_pc mod_start = NULL;
ASSERT(name != NULL || start != NULL);
/* If name is non-NULL, start can be NULL, so we have to walk the whole
* address space even when we have syscalls for memquery (e.g., on Mac).
* Even if start is non-NULL, it could be in the middle of the library.
*/
memquery_iterator_start(&iter, NULL,
/* We're never called from a fragile place like a
* signal handler, so as long as it's not real early
* it's ok to alloc.
*/
dynamo_heap_initialized);
libname[0] = '\0';
while (memquery_iterator_next(&iter)) {
LOG(GLOBAL, LOG_VMAREAS, 5, "start="PFX" end="PFX" prot=%x comment=%s\n",
iter.vm_start, iter.vm_end, iter.prot, iter.comment);
/* Record the base of each differently-named set of entries up until
* we find our target, when we'll clobber libpath
*/
if (!found_library &&
strncmp(libname, iter.comment, BUFFER_SIZE_ELEMENTS(libname)) != 0) {
last_base = iter.vm_start;
/* last_end is used to know what's readable beyond last_base */
if (TEST(MEMPROT_READ, iter.prot))
last_end = iter.vm_end;
else
last_end = last_base;
/* remember name so we can find the base of a multiply-mapped so */
strncpy(libname, iter.comment, BUFFER_SIZE_ELEMENTS(libname));
NULL_TERMINATE_BUFFER(libname);
}
if ((name_cmp != NULL &&
(strstr(iter.comment, name_cmp) != NULL ||
/* For Linux, include mid-library (non-.bss) anonymous mappings.
* Our private loader
* fills mapping holes with anonymous memory instead of a
* MEMPROT_NONE mapping from the original file.
* For Mac, this includes mid-library .bss.
*/
(found_library && iter.comment[0] == '\0' && image_size != 0 &&
iter.vm_end - mod_start < image_size))) ||
(name == NULL && *start >= iter.vm_start && *start < iter.vm_end)) {
if (!found_library) {
size_t mod_readable_sz;
char *dst = (fullpath != NULL) ? fullpath : libname;
size_t dstsz = (fullpath != NULL) ? path_size :
BUFFER_SIZE_ELEMENTS(libname);
char *slash = strrchr(iter.comment, '/');
ASSERT_CURIOSITY(slash != NULL);
ASSERT_CURIOSITY((slash - iter.comment) < dstsz);
/* we keep the last '/' at end */
++slash;
strncpy(dst, iter.comment, MIN(dstsz, (slash - iter.comment)));
/* if max no null */
dst[dstsz - 1] = '\0';
if (name == NULL)
name_cmp = dst;
found_library = true;
/* Most library have multiple segments, and some have the
* ELF header repeated in a later mapping, so we can't rely
* on is_elf_so_header() and header walking.
* We use the name tracking to remember the first entry
* that had this name.
*/
if (last_base == NULL) {
mod_start = iter.vm_start;
mod_readable_sz = iter.vm_end - iter.vm_start;
} else {
mod_start = last_base;
mod_readable_sz = last_end - last_base;
}
if (module_is_header(mod_start, mod_readable_sz)) {
app_pc mod_base, mod_end;
if (module_walk_program_headers(mod_start, mod_readable_sz, false,
&mod_base, NULL, &mod_end, NULL,
NULL)) {
image_size = mod_end - mod_base;
LOG(GLOBAL, LOG_VMAREAS, 4, "%s: image size is "PIFX"\n",
__FUNCTION__, image_size);
ASSERT_CURIOSITY(image_size != 0);
} else {
ASSERT_NOT_REACHED();
}
} else {
ASSERT(false && "expected elf header");
}
}
count++;
cur_end = iter.vm_end;
} else if (found_library) {
/* hit non-matching, we expect module segments to be adjacent */
break;
}
}
/* Xref PR 208443: .bss sections are anonymous (no file name listed in
* maps file), but not every library has one. We have to parse the ELF
* header to know since we can't assume that a subsequent anonymous
* region is .bss. */
if (image_size != 0 && cur_end - mod_start < image_size) {
/* Found a .bss section. Check current mapping (note might only be
* part of the mapping (due to os region merging? FIXME investigate). */
ASSERT_CURIOSITY(iter.vm_start == cur_end /* no gaps, FIXME might there be
* a gap if the file has large
* alignment and no data section?
* curiosity for now*/);
ASSERT_CURIOSITY(iter.inode == 0); /* .bss is anonymous */
ASSERT_CURIOSITY(iter.vm_end - mod_start >= image_size);/* should be big enough */
count++;
cur_end = mod_start + image_size;
} else {
/* Shouldn't have more mapped then the size of the module, unless it's a
* second adjacent separate map of the same file. Curiosity for now. */
ASSERT_CURIOSITY(image_size == 0 || cur_end - mod_start == image_size);
}
memquery_iterator_stop(&iter);
if (start != NULL)
*start = mod_start;
if (end != NULL)
*end = cur_end;
return count;
}
