static void
pages_unmap(void *addr, size_t size)
{
if (munmap(addr, size) == -1) {
char buf[BUFERROR_BUF];
buferror(errno, buf, sizeof(buf));
malloc_write("<jemalloc>: Error in munmap(): ");
malloc_write(buf);
malloc_write("\n");
if (opt_abort)
abort();
}
}
void je_malloc_disable_init() {
if (pthread_atfork(je_malloc_disable_prefork,
je_malloc_disable_postfork_parent, je_malloc_disable_postfork_child) != 0) {
malloc_write("<jemalloc>: Error in pthread_atfork()\n");
if (opt_abort)
abort();
}
}
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;
}
static void
pthread_create_once(void)
{
pthread_create_fptr = dlsym(RTLD_NEXT, "pthread_create");
if (pthread_create_fptr == NULL) {
malloc_write("<jemalloc>: Error in dlsym(RTLD_NEXT, "
"\"pthread_create\")\n");
abort();
}
isthreaded = true;
}
static void *
pages_map(void *addr, size_t size, bool noreserve)
{
void *ret;
/*
* We don't use MAP_FIXED here, because it can cause the *replacement*
* of existing mappings, and we only want to create new mappings.
*/
int flags = MAP_PRIVATE | MAP_ANON;
#ifdef MAP_NORESERVE
if (noreserve)
flags |= MAP_NORESERVE;
#endif
ret = mmap(addr, size, PROT_READ | PROT_WRITE, flags, -1, 0);
assert(ret != NULL);
if (ret == MAP_FAILED)
ret = NULL;
else if (addr != NULL && ret != addr) {
/*
* We succeeded in mapping memory, but not in the right place.
*/
if (munmap(ret, size) == -1) {
char buf[BUFERROR_BUF];
buferror(errno, buf, sizeof(buf));
malloc_write("<jemalloc>: Error in munmap(): ");
malloc_write(buf);
malloc_write("\n");
if (opt_abort)
abort();
}
ret = NULL;
}
assert(ret == NULL || (addr == NULL && ret != addr)
|| (addr != NULL && ret == addr));
return (ret);
}
bool
chunk_mmap_boot(void)
{
#ifdef NO_TLS
if (pthread_key_create(&mmap_unaligned_tsd, NULL) != 0) {
malloc_write("<jemalloc>: Error in pthread_key_create()\n");
return (true);
}
#endif
return (false);
}
static void
init_thp_state(void) {
if (!have_madvise_huge) {
if (metadata_thp_enabled() && opt_abort) {
malloc_write("<jemalloc>: no MADV_HUGEPAGE support\n");
abort();
}
goto label_error;
}
static const char sys_state_madvise[] = "always [madvise] never\n";
static const char sys_state_always[] = "[always] madvise never\n";
static const char sys_state_never[] = "always madvise [never]\n";
char buf[sizeof(sys_state_madvise)];
#if defined(JEMALLOC_USE_SYSCALL) && defined(SYS_open)
int fd = (int)syscall(SYS_open,
"/sys/kernel/mm/transparent_hugepage/enabled", O_RDONLY);
#else
int fd = open("/sys/kernel/mm/transparent_hugepage/enabled", O_RDONLY);
#endif
if (fd == -1) {
goto label_error;
}
ssize_t nread = malloc_read_fd(fd, &buf, sizeof(buf));
#if defined(JEMALLOC_USE_SYSCALL) && defined(SYS_close)
syscall(SYS_close, fd);
#else
close(fd);
#endif
if (strncmp(buf, sys_state_madvise, (size_t)nread) == 0) {
init_system_thp_mode = thp_mode_default;
} else if (strncmp(buf, sys_state_always, (size_t)nread) == 0) {
init_system_thp_mode = thp_mode_always;
} else if (strncmp(buf, sys_state_never, (size_t)nread) == 0) {
init_system_thp_mode = thp_mode_never;
} else {
goto label_error;
}
return;
label_error:
opt_thp = init_system_thp_mode = thp_mode_not_supported;
}
static bool
base_pages_alloc(size_t minsize)
{
size_t csize;
bool zero;
if (base_pages != NULL) {
/* TODO: remove this implementation restriction */
malloc_write("<jemalloc>: Internal allocation limit reached\n");
abort();
}
assert(minsize != 0);
csize = CHUNK_CEILING(minsize);
zero = false;
base_pages = chunk_alloc(csize, true, &zero);
if (base_pages == NULL)
return (true);
base_next_addr = base_pages;
base_past_addr = (void *)((uintptr_t)base_pages + csize);
return (false);
}
void
stats_print(void (*write_cb)(void *, const char *), void *cbopaque,
const char *opts)
{
int err;
uint64_t epoch;
size_t u64sz;
bool general = true;
bool merged = true;
bool unmerged = true;
bool bins = true;
bool large = true;
/*
* Refresh stats, in case mallctl() was called by the application.
*
* Check for OOM here, since refreshing the ctl cache can trigger
* allocation. In practice, none of the subsequent mallctl()-related
* calls in this function will cause OOM if this one succeeds.
* */
epoch = 1;
u64sz = sizeof(uint64_t);
err = je_mallctl("epoch", &epoch, &u64sz, &epoch, sizeof(uint64_t));
if (err != 0) {
if (err == EAGAIN) {
malloc_write("<jemalloc>: Memory allocation failure in "
"mallctl(\"epoch\", ...)\n");
return;
}
malloc_write("<jemalloc>: Failure in mallctl(\"epoch\", "
"...)\n");
abort();
}
if (opts != NULL) {
unsigned i;
for (i = 0; opts[i] != '\0'; i++) {
switch (opts[i]) {
case 'g':
general = false;
break;
case 'm':
merged = false;
break;
case 'a':
unmerged = false;
break;
case 'b':
bins = false;
break;
case 'l':
large = false;
break;
default:;
}
}
}
malloc_cprintf(write_cb, cbopaque,
"___ Begin jemalloc statistics ___\n");
if (general) {
int err;
const char *cpv;
bool bv;
unsigned uv;
ssize_t ssv;
size_t sv, bsz, ssz, sssz, cpsz;
bsz = sizeof(bool);
ssz = sizeof(size_t);
sssz = sizeof(ssize_t);
cpsz = sizeof(const char *);
CTL_GET("version", &cpv, const char *);
malloc_cprintf(write_cb, cbopaque, "Version: %s\n", cpv);
CTL_GET("config.debug", &bv, bool);
malloc_cprintf(write_cb, cbopaque, "Assertions %s\n",
bv ? "enabled" : "disabled");
#define OPT_WRITE_BOOL(n) \
if ((err = je_mallctl("opt."#n, &bv, &bsz, NULL, 0)) \
== 0) { \
malloc_cprintf(write_cb, cbopaque, \
" opt."#n": %s\n", bv ? "true" : "false"); \
}
#define OPT_WRITE_SIZE_T(n) \
if ((err = je_mallctl("opt."#n, &sv, &ssz, NULL, 0)) \
== 0) { \
malloc_cprintf(write_cb, cbopaque, \
" opt."#n": %zu\n", sv); \
}
#define OPT_WRITE_SSIZE_T(n) \
if ((err = je_mallctl("opt."#n, &ssv, &sssz, NULL, 0)) \
== 0) { \
malloc_cprintf(write_cb, cbopaque, \
" opt."#n": %zd\n", ssv); \
}
#define OPT_WRITE_CHAR_P(n) \
if ((err = je_mallctl("opt."#n, &cpv, &cpsz, NULL, 0)) \
== 0) { \
malloc_cprintf(write_cb, cbopaque, \
" opt."#n": \"%s\"\n", cpv); \
}
malloc_cprintf(write_cb, cbopaque,
"Run-time option settings:\n");
OPT_WRITE_BOOL(abort)
OPT_WRITE_SIZE_T(lg_chunk)
OPT_WRITE_CHAR_P(dss)
OPT_WRITE_SIZE_T(narenas)
OPT_WRITE_SSIZE_T(lg_dirty_mult)
OPT_WRITE_BOOL(stats_print)
OPT_WRITE_BOOL(junk)
OPT_WRITE_SIZE_T(quarantine)
OPT_WRITE_BOOL(redzone)
OPT_WRITE_BOOL(zero)
OPT_WRITE_BOOL(utrace)
OPT_WRITE_BOOL(valgrind)
OPT_WRITE_BOOL(xmalloc)
OPT_WRITE_BOOL(tcache)
OPT_WRITE_SSIZE_T(lg_tcache_max)
OPT_WRITE_BOOL(prof)
OPT_WRITE_CHAR_P(prof_prefix)
OPT_WRITE_BOOL(prof_active)
OPT_WRITE_SSIZE_T(lg_prof_sample)
OPT_WRITE_BOOL(prof_accum)
OPT_WRITE_SSIZE_T(lg_prof_interval)
OPT_WRITE_BOOL(prof_gdump)
OPT_WRITE_BOOL(prof_final)
OPT_WRITE_BOOL(prof_leak)
#undef OPT_WRITE_BOOL
#undef OPT_WRITE_SIZE_T
#undef OPT_WRITE_SSIZE_T
#undef OPT_WRITE_CHAR_P
malloc_cprintf(write_cb, cbopaque, "CPUs: %u\n", ncpus);
CTL_GET("arenas.narenas", &uv, unsigned);
malloc_cprintf(write_cb, cbopaque, "Arenas: %u\n", uv);
malloc_cprintf(write_cb, cbopaque, "Pointer size: %zu\n",
sizeof(void *));
CTL_GET("arenas.quantum", &sv, size_t);
malloc_cprintf(write_cb, cbopaque, "Quantum size: %zu\n", sv);
CTL_GET("arenas.page", &sv, size_t);
malloc_cprintf(write_cb, cbopaque, "Page size: %zu\n", sv);
CTL_GET("opt.lg_dirty_mult", &ssv, ssize_t);
if (ssv >= 0) {
malloc_cprintf(write_cb, cbopaque,
"Min active:dirty page ratio per arena: %u:1\n",
(1U << ssv));
} else {
malloc_cprintf(write_cb, cbopaque,
"Min active:dirty page ratio per arena: N/A\n");
}
if ((err = je_mallctl("arenas.tcache_max", &sv, &ssz, NULL, 0))
== 0) {
malloc_cprintf(write_cb, cbopaque,
"Maximum thread-cached size class: %zu\n", sv);
}
if ((err = je_mallctl("opt.prof", &bv, &bsz, NULL, 0)) == 0 &&
bv) {
CTL_GET("opt.lg_prof_sample", &sv, size_t);
malloc_cprintf(write_cb, cbopaque,
"Average profile sample interval: %"PRIu64
" (2^%zu)\n", (((uint64_t)1U) << sv), sv);
CTL_GET("opt.lg_prof_interval", &ssv, ssize_t);
if (ssv >= 0) {
malloc_cprintf(write_cb, cbopaque,
"Average profile dump interval: %"PRIu64
" (2^%zd)\n",
(((uint64_t)1U) << ssv), ssv);
} else {
malloc_cprintf(write_cb, cbopaque,
"Average profile dump interval: N/A\n");
}
}
CTL_GET("opt.lg_chunk", &sv, size_t);
malloc_cprintf(write_cb, cbopaque, "Chunk size: %zu (2^%zu)\n",
(ZU(1) << sv), sv);
}
if (config_stats) {
size_t *cactive;
size_t allocated, active, mapped;
size_t chunks_current, chunks_high;
uint64_t chunks_total;
size_t huge_allocated;
uint64_t huge_nmalloc, huge_ndalloc;
CTL_GET("stats.cactive", &cactive, size_t *);
CTL_GET("stats.allocated", &allocated, size_t);
CTL_GET("stats.active", &active, size_t);
CTL_GET("stats.mapped", &mapped, size_t);
malloc_cprintf(write_cb, cbopaque,
"Allocated: %zu, active: %zu, mapped: %zu\n",
allocated, active, mapped);
malloc_cprintf(write_cb, cbopaque,
"Current active ceiling: %zu\n", atomic_read_z(cactive));
/* Print chunk stats. */
CTL_GET("stats.chunks.total", &chunks_total, uint64_t);
CTL_GET("stats.chunks.high", &chunks_high, size_t);
CTL_GET("stats.chunks.current", &chunks_current, size_t);
malloc_cprintf(write_cb, cbopaque, "chunks: nchunks "
"highchunks curchunks\n");
malloc_cprintf(write_cb, cbopaque,
" %13"PRIu64" %12zu %12zu\n",
chunks_total, chunks_high, chunks_current);
/* Print huge stats. */
CTL_GET("stats.huge.nmalloc", &huge_nmalloc, uint64_t);
CTL_GET("stats.huge.ndalloc", &huge_ndalloc, uint64_t);
CTL_GET("stats.huge.allocated", &huge_allocated, size_t);
malloc_cprintf(write_cb, cbopaque,
"huge: nmalloc ndalloc allocated\n");
malloc_cprintf(write_cb, cbopaque,
" %12"PRIu64" %12"PRIu64" %12zu\n",
huge_nmalloc, huge_ndalloc, huge_allocated);
if (merged) {
unsigned narenas;
CTL_GET("arenas.narenas", &narenas, unsigned);
{
VARIABLE_ARRAY(bool, initialized, narenas);
size_t isz;
unsigned i, ninitialized;
isz = sizeof(bool) * narenas;
xmallctl("arenas.initialized", initialized,
&isz, NULL, 0);
for (i = ninitialized = 0; i < narenas; i++) {
if (initialized[i])
ninitialized++;
}
if (ninitialized > 1 || unmerged == false) {
/* Print merged arena stats. */
malloc_cprintf(write_cb, cbopaque,
"\nMerged arenas stats:\n");
stats_arena_print(write_cb, cbopaque,
narenas, bins, large);
}
}
}
if (unmerged) {
unsigned narenas;
/* Print stats for each arena. */
CTL_GET("arenas.narenas", &narenas, unsigned);
{
VARIABLE_ARRAY(bool, initialized, narenas);
size_t isz;
unsigned i;
isz = sizeof(bool) * narenas;
xmallctl("arenas.initialized", initialized,
&isz, NULL, 0);
for (i = 0; i < narenas; i++) {
if (initialized[i]) {
malloc_cprintf(write_cb,
cbopaque,
"\narenas[%u]:\n", i);
stats_arena_print(write_cb,
cbopaque, i, bins, large);
}
}
}
}
}
malloc_cprintf(write_cb, cbopaque, "--- End jemalloc statistics ---\n");
}
bool
chunk_swap_enable(const int *fds, unsigned nfds, bool prezeroed)
{
bool ret;
unsigned i;
off_t off;
void *vaddr;
size_t cumsize, voff;
size_t sizes[nfds];
malloc_mutex_lock(&swap_mtx);
/* Get file sizes. */
for (i = 0, cumsize = 0; i < nfds; i++) {
off = lseek(fds[i], 0, SEEK_END);
if (off == ((off_t)-1)) {
ret = true;
goto RETURN;
}
if (PAGE_CEILING(off) != off) {
/* Truncate to a multiple of the page size. */
off &= ~PAGE_MASK;
if (ftruncate(fds[i], off) != 0) {
ret = true;
goto RETURN;
}
}
sizes[i] = off;
if (cumsize + off < cumsize) {
/*
* Cumulative file size is greater than the total
* address space. Bail out while it's still obvious
* what the problem is.
*/
ret = true;
goto RETURN;
}
cumsize += off;
}
/* Round down to a multiple of the chunk size. */
cumsize &= ~chunksize_mask;
if (cumsize == 0) {
ret = true;
goto RETURN;
}
/*
* Allocate a chunk-aligned region of anonymous memory, which will
* be the final location for the memory-mapped files.
*/
vaddr = chunk_alloc_mmap_noreserve(cumsize);
if (vaddr == NULL) {
ret = true;
goto RETURN;
}
/* Overlay the files onto the anonymous mapping. */
for (i = 0, voff = 0; i < nfds; i++) {
void *addr = mmap((void *)((uintptr_t)vaddr + voff), sizes[i],
PROT_READ | PROT_WRITE, MAP_SHARED | MAP_FIXED, fds[i], 0);
if (addr == MAP_FAILED) {
char buf[BUFERROR_BUF];
buferror(errno, buf, sizeof(buf));
malloc_write(
"<jemalloc>: Error in mmap(..., MAP_FIXED, ...): ");
malloc_write(buf);
malloc_write("\n");
if (opt_abort)
abort();
if (munmap(vaddr, voff) == -1) {
buferror(errno, buf, sizeof(buf));
malloc_write("<jemalloc>: Error in munmap(): ");
malloc_write(buf);
malloc_write("\n");
}
ret = true;
goto RETURN;
}
assert(addr == (void *)((uintptr_t)vaddr + voff));
/*
* Tell the kernel that the mapping will be accessed randomly,
* and that it should not gratuitously sync pages to the
* filesystem.
*/
#ifdef MADV_RANDOM
madvise(addr, sizes[i], MADV_RANDOM);
#endif
#ifdef MADV_NOSYNC
madvise(addr, sizes[i], MADV_NOSYNC);
#endif
voff += sizes[i];
}
swap_prezeroed = prezeroed;
swap_base = vaddr;
swap_end = swap_base;
swap_max = (void *)((uintptr_t)vaddr + cumsize);
/* Copy the fds array for mallctl purposes. */
swap_fds = (int *)base_alloc(nfds * sizeof(int));
if (swap_fds == NULL) {
ret = true;
goto RETURN;
}
memcpy(swap_fds, fds, nfds * sizeof(int));
swap_nfds = nfds;
#ifdef JEMALLOC_STATS
swap_avail = cumsize;
#endif
swap_enabled = true;
ret = false;
RETURN:
malloc_mutex_unlock(&swap_mtx);
return (ret);
}
void *
huge_ralloc(void *ptr, size_t oldsize, size_t size, size_t extra,
size_t alignment, bool zero)
{
void *ret;
size_t copysize;
/* Try to avoid moving the allocation. */
ret = huge_ralloc_no_move(ptr, oldsize, size, extra);
if (ret != NULL)
return (ret);
/*
* size and oldsize are different enough that we need to use a
* different size class. In that case, fall back to allocating new
* space and copying.
*/
if (alignment != 0)
ret = huge_palloc(size + extra, alignment, zero);
else
ret = huge_malloc(size + extra, zero);
if (ret == NULL) {
if (extra == 0)
return (NULL);
/* Try again, this time without extra. */
if (alignment != 0)
ret = huge_palloc(size, alignment, zero);
else
ret = huge_malloc(size, zero);
if (ret == NULL)
return (NULL);
}
/*
* Copy at most size bytes (not size+extra), since the caller has no
* expectation that the extra bytes will be reliably preserved.
*/
copysize = (size < oldsize) ? size : oldsize;
/*
* Use mremap(2) if this is a huge-->huge reallocation, and neither the
* source nor the destination are in swap or dss.
*/
#ifdef JEMALLOC_MREMAP_FIXED
if (oldsize >= chunksize
# ifdef JEMALLOC_SWAP
&& (swap_enabled == false || (chunk_in_swap(ptr) == false &&
chunk_in_swap(ret) == false))
# endif
# ifdef JEMALLOC_DSS
&& chunk_in_dss(ptr) == false && chunk_in_dss(ret) == false
# endif
) {
size_t newsize = huge_salloc(ret);
if (mremap(ptr, oldsize, newsize, MREMAP_MAYMOVE|MREMAP_FIXED,
ret) == MAP_FAILED) {
/*
* Assuming no chunk management bugs in the allocator,
* the only documented way an error can occur here is
* if the application changed the map type for a
* portion of the old allocation. This is firmly in
* undefined behavior territory, so write a diagnostic
* message, and optionally abort.
*/
char buf[BUFERROR_BUF];
buferror(errno, buf, sizeof(buf));
malloc_write("<jemalloc>: Error in mremap(): ");
malloc_write(buf);
malloc_write("\n");
if (opt_abort)
abort();
memcpy(ret, ptr, copysize);
idalloc(ptr);
} else
huge_dalloc(ptr, false);
} else
#endif
{
memcpy(ret, ptr, copysize);
idalloc(ptr);
}
return (ret);
}
