bool
pages_boot(void) {
os_page = os_page_detect();
if (os_page > PAGE) {
malloc_write("<jemalloc>: Unsupported system page size\n");
if (opt_abort) {
abort();
}
return true;
}
#ifndef _WIN32
mmap_flags = MAP_PRIVATE | MAP_ANON;
#endif
#ifdef __CHERI_PURE_CAPABILITY__
mmap_flags |= MAP_ALIGNED(LG_PAGE);
#endif
#ifdef JEMALLOC_SYSCTL_VM_OVERCOMMIT
os_overcommits = os_overcommits_sysctl();
#elif defined(JEMALLOC_PROC_SYS_VM_OVERCOMMIT_MEMORY)
os_overcommits = os_overcommits_proc();
# ifdef MAP_NORESERVE
if (os_overcommits) {
mmap_flags |= MAP_NORESERVE;
}
# endif
#else
os_overcommits = false;
#endif
init_thp_state();
#ifdef __FreeBSD__
/*
* FreeBSD doesn't need the check; madvise(2) is known to work.
*/
#else
/* Detect lazy purge runtime support. */
if (pages_can_purge_lazy) {
bool committed = false;
void *madv_free_page = os_pages_map(NULL, PAGE, PAGE, &committed);
if (madv_free_page == NULL) {
return true;
}
assert(pages_can_purge_lazy_runtime);
if (pages_purge_lazy(madv_free_page, PAGE)) {
pages_can_purge_lazy_runtime = false;
}
os_pages_unmap(madv_free_page, PAGE);
}
#endif
return false;
}
/*
* Map a shared object into memory. The argument is a file descriptor,
* which must be open on the object and positioned at its beginning.
*
* The return value is a pointer to a newly-allocated Obj_Entry structure
* for the shared object. Returns NULL on failure.
*/
Obj_Entry *
_rtld_map_object(const char *path, int fd, const struct stat *sb)
{
Obj_Entry *obj;
Elf_Ehdr *ehdr;
Elf_Phdr *phdr;
size_t phsize;
Elf_Phdr *phlimit;
Elf_Phdr *segs[2];
int nsegs;
caddr_t mapbase = MAP_FAILED;
size_t mapsize = 0;
size_t bsssize = 0;
int mapflags;
Elf_Off base_offset;
#ifdef MAP_ALIGNED
Elf_Addr base_alignment;
#endif
Elf_Addr base_vaddr;
Elf_Addr base_vlimit;
Elf_Addr text_vlimit;
int text_flags;
caddr_t base_addr;
Elf_Off data_offset;
Elf_Addr data_vaddr;
Elf_Addr data_vlimit;
int data_flags;
caddr_t data_addr;
Elf_Addr phdr_vaddr;
size_t phdr_memsz;
caddr_t gap_addr;
size_t gap_size;
int i;
#ifdef RTLD_LOADER
Elf_Addr clear_vaddr;
caddr_t clear_addr;
size_t nclear;
#endif
if (sb != NULL && sb->st_size < (off_t)sizeof (Elf_Ehdr)) {
_rtld_error("%s: unrecognized file format1", path);
return NULL;
}
obj = _rtld_obj_new();
obj->path = xstrdup(path);
obj->pathlen = strlen(path);
if (sb != NULL) {
obj->dev = sb->st_dev;
obj->ino = sb->st_ino;
}
ehdr = mmap(NULL, _rtld_pagesz, PROT_READ, MAP_FILE | MAP_SHARED, fd,
(off_t)0);
obj->ehdr = ehdr;
if (ehdr == MAP_FAILED) {
_rtld_error("%s: read error: %s", path, xstrerror(errno));
goto bad;
}
/* Make sure the file is valid */
if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0 ||
ehdr->e_ident[EI_CLASS] != ELFCLASS) {
_rtld_error("%s: unrecognized file format2 [%x != %x]", path,
ehdr->e_ident[EI_CLASS], ELFCLASS);
goto bad;
}
/* Elf_e_ident includes class */
if (ehdr->e_ident[EI_VERSION] != EV_CURRENT ||
ehdr->e_version != EV_CURRENT ||
ehdr->e_ident[EI_DATA] != ELFDEFNNAME(MACHDEP_ENDIANNESS)) {
_rtld_error("%s: unsupported file version", path);
goto bad;
}
if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN) {
_rtld_error("%s: unsupported file type", path);
goto bad;
}
switch (ehdr->e_machine) {
ELFDEFNNAME(MACHDEP_ID_CASES)
default:
_rtld_error("%s: unsupported machine", path);
goto bad;
}
/*
* We rely on the program header being in the first page. This is
* not strictly required by the ABI specification, but it seems to
* always true in practice. And, it simplifies things considerably.
*/
assert(ehdr->e_phentsize == sizeof(Elf_Phdr));
assert(ehdr->e_phoff + ehdr->e_phnum * sizeof(Elf_Phdr) <=
_rtld_pagesz);
/*
* Scan the program header entries, and save key information.
*
* We rely on there being exactly two load segments, text and data,
* in that order.
*/
phdr = (Elf_Phdr *) ((caddr_t)ehdr + ehdr->e_phoff);
phsize = ehdr->e_phnum * sizeof(phdr[0]);
obj->phdr = NULL;
phdr_vaddr = EA_UNDEF;
phdr_memsz = 0;
phlimit = phdr + ehdr->e_phnum;
nsegs = 0;
while (phdr < phlimit) {
switch (phdr->p_type) {
case PT_INTERP:
obj->interp = (void *)(uintptr_t)phdr->p_vaddr;
dbg(("%s: PT_INTERP %p", obj->path, obj->interp));
break;
case PT_LOAD:
if (nsegs < 2)
segs[nsegs] = phdr;
++nsegs;
dbg(("%s: PT_LOAD %p", obj->path, phdr));
break;
case PT_PHDR:
phdr_vaddr = phdr->p_vaddr;
phdr_memsz = phdr->p_memsz;
dbg(("%s: PT_PHDR %p phsize %zu", obj->path,
(void *)(uintptr_t)phdr_vaddr, phdr_memsz));
break;
case PT_DYNAMIC:
obj->dynamic = (void *)(uintptr_t)phdr->p_vaddr;
dbg(("%s: PT_DYNAMIC %p", obj->path, obj->dynamic));
break;
}
++phdr;
}
phdr = (Elf_Phdr *) ((caddr_t)ehdr + ehdr->e_phoff);
obj->entry = (void *)(uintptr_t)ehdr->e_entry;
if (!obj->dynamic) {
_rtld_error("%s: not dynamically linked", path);
goto bad;
}
if (nsegs != 2) {
_rtld_error("%s: wrong number of segments (%d != 2)", path,
nsegs);
goto bad;
}
/*
* Map the entire address space of the object as a file
* region to stake out our contiguous region and establish a
* base for relocation. We use a file mapping so that
* the kernel will give us whatever alignment is appropriate
* for the platform we're running on.
*
* We map it using the text protection, map the data segment
* into the right place, then map an anon segment for the bss
* and unmap the gaps left by padding to alignment.
*/
#ifdef MAP_ALIGNED
base_alignment = segs[0]->p_align;
#endif
base_offset = round_down(segs[0]->p_offset);
base_vaddr = round_down(segs[0]->p_vaddr);
base_vlimit = round_up(segs[1]->p_vaddr + segs[1]->p_memsz);
text_vlimit = round_up(segs[0]->p_vaddr + segs[0]->p_memsz);
text_flags = protflags(segs[0]->p_flags);
data_offset = round_down(segs[1]->p_offset);
data_vaddr = round_down(segs[1]->p_vaddr);
data_vlimit = round_up(segs[1]->p_vaddr + segs[1]->p_filesz);
data_flags = protflags(segs[1]->p_flags);
#ifdef RTLD_LOADER
clear_vaddr = segs[1]->p_vaddr + segs[1]->p_filesz;
#endif
obj->textsize = text_vlimit - base_vaddr;
obj->vaddrbase = base_vaddr;
obj->isdynamic = ehdr->e_type == ET_DYN;
obj->phdr_loaded = false;
for (i = 0; i < nsegs; i++) {
if (phdr_vaddr != EA_UNDEF &&
segs[i]->p_vaddr <= phdr_vaddr &&
segs[i]->p_memsz >= phdr_memsz) {
obj->phdr_loaded = true;
break;
}
if (segs[i]->p_offset <= ehdr->e_phoff &&
segs[i]->p_memsz >= phsize) {
phdr_vaddr = segs[i]->p_vaddr + ehdr->e_phoff;
phdr_memsz = phsize;
obj->phdr_loaded = true;
break;
}
}
if (obj->phdr_loaded) {
obj->phdr = (void *)(uintptr_t)phdr_vaddr;
obj->phsize = phdr_memsz;
} else {
Elf_Phdr *buf;
buf = xmalloc(phsize);
if (buf == NULL) {
_rtld_error("%s: cannot allocate program header", path);
goto bad;
}
memcpy(buf, phdr, phsize);
obj->phdr = buf;
obj->phsize = phsize;
}
dbg(("%s: phdr %p phsize %zu (%s)", obj->path, obj->phdr, obj->phsize,
obj->phdr_loaded ? "loaded" : "allocated"));
/* Unmap header if it overlaps the first load section. */
if (base_offset < _rtld_pagesz) {
munmap(ehdr, _rtld_pagesz);
obj->ehdr = MAP_FAILED;
}
/*
* Calculate log2 of the base section alignment.
*/
mapflags = 0;
#ifdef MAP_ALIGNED
if (base_alignment > _rtld_pagesz) {
unsigned int log2 = 0;
for (; base_alignment > 1; base_alignment >>= 1)
log2++;
mapflags = MAP_ALIGNED(log2);
}
void *
pages_map(void *addr, size_t size, size_t alignment, bool *commit) {
assert(alignment >= PAGE);
assert(ALIGNMENT_ADDR2BASE(addr, alignment) == (vaddr_t)addr);
#if defined(__FreeBSD__) && defined(MAP_EXCL)
/*
* FreeBSD has mechanisms both to mmap at specific address without
* touching existing mappings, and to mmap with specific alignment.
*/
{
#ifdef __CHERI_PURE_CAPABILITY__
if (size & CHERI_ALIGN_MASK(size))
abort();
#endif
if (os_overcommits) {
*commit = true;
}
int prot = *commit ? PAGES_PROT_COMMIT : PAGES_PROT_DECOMMIT;
int flags = mmap_flags;
if (addr != NULL) {
flags |= MAP_FIXED | MAP_EXCL;
} else {
unsigned alignment_bits = ffs_zu(alignment);
assert(alignment_bits > 1);
flags |= MAP_ALIGNED(alignment_bits - 1);
}
void *ret = mmap(addr, size, prot, flags, -1, 0);
if (ret == MAP_FAILED) {
ret = NULL;
}
return ret;
}
#endif
/*
* Ideally, there would be a way to specify alignment to mmap() (like
* NetBSD has), but in the absence of such a feature, we have to work
* hard to efficiently create aligned mappings. The reliable, but
* slow method is to create a mapping that is over-sized, then trim the
* excess. However, that always results in one or two calls to
* os_pages_unmap(), and it can leave holes in the process's virtual
* memory map if memory grows downward.
*
* Optimistically try mapping precisely the right amount before falling
* back to the slow method, with the expectation that the optimistic
* approach works most of the time.
*/
void *ret = os_pages_map(addr, size, os_page, commit);
if (ret == NULL || ret == addr) {
return ret;
}
assert(addr == NULL);
if (ALIGNMENT_ADDR2OFFSET((vaddr_t)ret, alignment) != 0) {
os_pages_unmap(ret, size);
return pages_map_slow(size, alignment, commit);
}
assert(PAGE_ADDR2BASE(ret) == ret);
return ret;
}