Rewrite ZipOS

This reduces the virtual memory usage of Emacs for me by 30%. We now
have a simpler implementation that uses read(), rather mmap()ing the
whole executable.
This commit is contained in:
Justine Tunney 2023-10-03 07:27:25 -07:00
parent ff77f2a6af
commit b01282e23e
No known key found for this signature in database
GPG key ID: BE714B4575D6E328
21 changed files with 408 additions and 421 deletions

View file

@ -16,68 +16,56 @@
TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
PERFORMANCE OF THIS SOFTWARE.
*/
#include "libc/assert.h"
#include "libc/calls/blocksigs.internal.h"
#include "libc/calls/calls.h"
#include "libc/calls/internal.h"
#include "libc/calls/state.internal.h"
#include "libc/calls/struct/sigset.h"
#include "libc/calls/syscall-sysv.internal.h"
#include "libc/calls/syscall_support-sysv.internal.h"
#include "libc/dce.h"
#include "libc/calls/struct/fd.internal.h"
#include "libc/errno.h"
#include "libc/intrin/asan.internal.h"
#include "libc/intrin/asancodes.h"
#include "libc/intrin/atomic.h"
#include "libc/intrin/cmpxchg.h"
#include "libc/intrin/directmap.internal.h"
#include "libc/intrin/extend.internal.h"
#include "libc/intrin/strace.internal.h"
#include "libc/intrin/weaken.h"
#include "libc/runtime/internal.h"
#include "libc/runtime/memtrack.internal.h"
#include "libc/runtime/zipos.internal.h"
#include "libc/sysv/consts/f.h"
#include "libc/sysv/consts/fd.h"
#include "libc/str/str.h"
#include "libc/sysv/consts/at.h"
#include "libc/sysv/consts/map.h"
#include "libc/sysv/consts/o.h"
#include "libc/sysv/consts/prot.h"
#include "libc/sysv/consts/s.h"
#include "libc/sysv/consts/sig.h"
#include "libc/sysv/errfuns.h"
#include "libc/thread/thread.h"
#include "libc/thread/tls.h"
#include "libc/zip.internal.h"
#include "third_party/zlib/zconf.h"
#include "third_party/zlib/zlib.h"
static char *__zipos_mapend;
static size_t __zipos_maptotal;
static pthread_mutex_t __zipos_lock_obj;
static void __zipos_wipe(void) {
pthread_mutex_init(&__zipos_lock_obj, 0);
static bool __zipos_method_is_supported(int method) {
switch (method) {
case kZipCompressionNone:
case kZipCompressionDeflate:
return true;
default:
return false;
}
}
static void __zipos_lock(void) {
pthread_mutex_lock(&__zipos_lock_obj);
}
static void __zipos_unlock(void) {
pthread_mutex_unlock(&__zipos_lock_obj);
}
static void *__zipos_mmap_space(size_t mapsize) {
static void *__zipos_mmap_space(struct Zipos *zipos, size_t mapsize) {
char *start;
size_t offset;
unassert(mapsize);
offset = __zipos_maptotal;
__zipos_maptotal += mapsize;
offset = zipos->maptotal;
zipos->maptotal += mapsize;
start = (char *)kMemtrackZiposStart;
if (!__zipos_mapend) __zipos_mapend = start;
__zipos_mapend = _extend(start, __zipos_maptotal, __zipos_mapend, MAP_PRIVATE,
kMemtrackZiposStart + kMemtrackZiposSize);
if (!zipos->mapend) zipos->mapend = start;
zipos->mapend = _extend(start, zipos->maptotal, zipos->mapend, MAP_PRIVATE,
kMemtrackZiposStart + kMemtrackZiposSize);
return start + offset;
}
void __zipos_free(struct ZiposHandle *h) {
if (!h) return;
if (IsAsan()) {
__asan_poison((char *)h + sizeof(struct ZiposHandle),
h->mapsize - sizeof(struct ZiposHandle), kAsanHeapFree);
@ -105,7 +93,7 @@ StartOver:
ph = &h->next;
}
if (!h) {
h = __zipos_mmap_space(mapsize);
h = __zipos_mmap_space(zipos, mapsize);
}
__zipos_unlock();
if (IsAsan()) {
@ -122,91 +110,85 @@ StartOver:
return h;
}
static int __zipos_mkfd(int minfd) {
int fd, e = errno;
if ((fd = __sys_fcntl(2, F_DUPFD_CLOEXEC, minfd)) != -1) {
return fd;
} else if (errno == EINVAL) {
errno = e;
return __fixupnewfd(__sys_fcntl(2, F_DUPFD, minfd), O_CLOEXEC);
} else {
return fd;
}
}
static int __zipos_setfd(int fd, struct ZiposHandle *h, unsigned flags) {
int want = fd;
atomic_compare_exchange_strong_explicit(
&g_fds.f, &want, fd + 1, memory_order_release, memory_order_relaxed);
g_fds.p[fd].kind = kFdZip;
g_fds.p[fd].handle = (intptr_t)h;
g_fds.p[fd].flags = flags | O_CLOEXEC;
g_fds.p[fd].extra = 0;
__fds_unlock();
return fd;
}
static int __zipos_load(struct Zipos *zipos, size_t cf, int flags,
struct ZiposUri *name) {
size_t lf;
size_t size;
int fd, minfd;
static int __zipos_load_fd(struct Zipos *zipos, int cf, int fd,
struct ZiposUri *name,
struct ZiposHandle **out_handle) {
uint64_t size;
struct ZiposHandle *h;
if (cf == ZIPOS_SYNTHETIC_DIRECTORY) {
size = name->len;
if (!(h = __zipos_alloc(zipos, size + 1))) return -1;
if (size) memcpy(h->data, name->path, size);
h->data[size] = 0;
h->mem = h->data;
} else {
lf = GetZipCfileOffset(zipos->map + cf);
npassert((ZIP_LFILE_MAGIC(zipos->map + lf) == kZipLfileHdrMagic));
size = GetZipLfileUncompressedSize(zipos->map + lf);
switch (ZIP_LFILE_COMPRESSIONMETHOD(zipos->map + lf)) {
case kZipCompressionNone:
if (!(h = __zipos_alloc(zipos, 0))) return -1;
h->mem = ZIP_LFILE_CONTENT(zipos->map + lf);
break;
case kZipCompressionDeflate:
if (!(h = __zipos_alloc(zipos, size))) return -1;
if (!__inflate(h->data, size, ZIP_LFILE_CONTENT(zipos->map + lf),
GetZipLfileCompressedSize(zipos->map + lf))) {
h->mem = h->data;
} else {
h->mem = 0;
eio();
}
break;
default:
uint8_t lfile[kZipLfileHdrMinSize];
uint64_t lf = GetZipCfileOffset(zipos->cdir + cf);
int compressed = ZIP_CFILE_COMPRESSIONMETHOD(zipos->cdir + cf);
if (!__zipos_method_is_supported(compressed) ||
pread(fd, lfile, kZipLfileHdrMinSize, lf) != kZipLfileHdrMinSize ||
ZIP_LFILE_MAGIC(lfile) != kZipLfileHdrMagic) {
return eio(); // this corruption
}
size = GetZipCfileUncompressedSize(zipos->cdir + cf);
if (!(h = __zipos_alloc(zipos, size))) return -1;
uint64_t off = lf + ZIP_LFILE_HDRSIZE(lfile);
if (!compressed) {
if (pread(fd, h->data, size, off) != size) {
__zipos_free(h);
return eio();
}
} else {
struct ZiposHandle *h2;
uint64_t compsize = GetZipCfileCompressedSize(zipos->cdir + cf);
if (!(h2 = __zipos_alloc(zipos, compsize))) {
__zipos_free(h);
return -1;
}
if (pread(fd, h2->data, compsize, off) != compsize ||
__inflate(h->data, size, h2->data, compsize)) {
__zipos_free(h2);
__zipos_free(h);
return eio();
}
__zipos_free(h2);
}
}
h->pos = 0;
h->cfile = cf;
h->size = size;
if (h->mem) {
minfd = 3;
__fds_lock();
TryAgain:
if (IsWindows() || IsMetal()) {
if ((fd = __reservefd_unlocked(-1)) != -1) {
return __zipos_setfd(fd, h, flags);
*out_handle = h;
return 0;
}
static int __zipos_load(struct Zipos *zipos, int cf, int flags,
struct ZiposUri *name) {
int fd;
if ((fd = openat(AT_FDCWD, zipos->progpath,
O_RDONLY | O_CLOEXEC |
(flags & (O_NONBLOCK | O_RANDOM | O_SEQUENTIAL)),
0)) != -1) {
struct ZiposHandle *h = 0;
if (__zipos_load_fd(zipos, cf, fd, name, &h) != -1) {
if (!IsWindows() && !IsMetal()) {
// unix doesn't use the g_fds table by default, so we need to
// explicitly ensure an entry exists for our new open()d file
__ensurefds(fd);
}
} else if ((fd = __zipos_mkfd(minfd)) != -1) {
if (__ensurefds_unlocked(fd) != -1) {
if (g_fds.p[fd].kind) {
sys_close(fd);
minfd = fd + 1;
goto TryAgain;
}
return __zipos_setfd(fd, h, flags);
}
sys_close(fd);
// turn it from a file fd to a zip fd
// layer the handle on windows so it can be restored on close
h->handle = g_fds.p[fd].handle;
g_fds.p[fd].kind = kFdZip;
g_fds.p[fd].handle = (intptr_t)h;
g_fds.p[fd].flags &= ~O_CLOEXEC;
g_fds.p[fd].flags |= flags & O_CLOEXEC;
return fd;
} else {
close(fd);
return -1;
}
__fds_unlock();
} else {
return -1;
}
__zipos_free(h);
return -1;
}
/**
@ -231,6 +213,10 @@ int __zipos_open(struct ZiposUri *name, int flags) {
(flags & O_ACCMODE) != O_RDONLY) {
return erofs();
}
if ((flags & ~(O_CLOEXEC | O_NONBLOCK | O_NOFOLLOW | O_NOCTTY | O_DIRECTORY |
O_NOATIME | O_RANDOM | O_SEQUENTIAL))) {
return einval();
}
// get the zipos global singleton
struct Zipos *zipos;
@ -242,15 +228,12 @@ int __zipos_open(struct ZiposUri *name, int flags) {
// majority of these calls will return ENOENT or ENOTDIR. we need to
// perform two extremely costly sigprocmask() calls below. thanks to
// zipos being a read-only filesystem, we can avoid it in many cases
ssize_t cf;
int cf;
if ((cf = __zipos_find(zipos, name)) == -1) {
return -1;
}
if (flags & O_EXCL) {
return eexist();
}
if (cf != ZIPOS_SYNTHETIC_DIRECTORY) {
int mode = GetZipCfileMode(zipos->map + cf);
int mode = GetZipCfileMode(zipos->cdir + cf);
if ((flags & O_DIRECTORY) && !S_ISDIR(mode)) {
return enotdir();
}
@ -265,8 +248,3 @@ int __zipos_open(struct ZiposUri *name, int flags) {
ALLOW_SIGNALS;
return rc;
}
__attribute__((__constructor__)) static void __zipos_ctor(void) {
__zipos_wipe();
pthread_atfork(__zipos_lock, __zipos_unlock, __zipos_wipe);
}