From c8e10eef30a421a22b24a5f18668e0ab1608aa89 Mon Sep 17 00:00:00 2001
From: Justine Tunney <jtunney@gmail.com>
Date: Mon, 23 Dec 2024 20:14:01 -0800
Subject: [PATCH] Make bulk_free() go faster

---
 .gitignore                         |  1 +
 libc/cosmo.h                       |  4 ++--
 libc/intrin/stack.c                | 15 ++++++++-------
 libc/thread/mapstack.c             |  4 ++--
 libc/thread/posixthread.internal.h |  4 +---
 libc/thread/pthread_atfork.c       |  2 ++
 libc/thread/pthread_create.c       | 20 +++++++++++---------
 libc/thread/pthread_exit.c         |  2 +-
 libc/thread/pthread_timedjoin_np.c |  2 +-
 libc/thread/thread.h               |  4 ++--
 test/libc/mem/malloc_test.c        |  3 +--
 third_party/dlmalloc/dlmalloc.c    | 10 +++++-----
 third_party/dlmalloc/threaded.inc  | 19 +++++++++++++++++--
 13 files changed, 54 insertions(+), 36 deletions(-)

diff --git a/.gitignore b/.gitignore
index 4c767cd51..0c6b21f03 100644
--- a/.gitignore
+++ b/.gitignore
@@ -15,3 +15,4 @@ __pycache__
 /tool/emacs/*.elc
 /perf.data
 /perf.data.old
+/qemu*core
diff --git a/libc/cosmo.h b/libc/cosmo.h
index e2691587a..d53c3045f 100644
--- a/libc/cosmo.h
+++ b/libc/cosmo.h
@@ -25,8 +25,8 @@ int cosmo_futex_wake(_COSMO_ATOMIC(int) *, int, char);
 int cosmo_futex_wait(_COSMO_ATOMIC(int) *, int, char, int,
                      const struct timespec *);
 
-errno_t cosmo_stack_alloc(size_t *, size_t *, void **) libcesque;
-errno_t cosmo_stack_free(void *, size_t, size_t) libcesque;
+errno_t cosmo_stack_alloc(unsigned *, unsigned *, void **) libcesque;
+errno_t cosmo_stack_free(void *, unsigned, unsigned) libcesque;
 void cosmo_stack_clear(void) libcesque;
 void cosmo_stack_setmaxstacks(int) libcesque;
 int cosmo_stack_getmaxstacks(void) libcesque;
diff --git a/libc/intrin/stack.c b/libc/intrin/stack.c
index e153b6ce8..c77e9a8d0 100644
--- a/libc/intrin/stack.c
+++ b/libc/intrin/stack.c
@@ -42,8 +42,8 @@
 struct CosmoStack {
   struct Dll elem;
   void *stackaddr;
-  size_t stacksize;
-  size_t guardsize;
+  unsigned stacksize;
+  unsigned guardsize;
 };
 
 struct CosmoStacks {
@@ -215,13 +215,13 @@ void cosmo_stack_setmaxstacks(int maxstacks) {
  * This function returns 0 on success, or an errno on error. See the
  * documentation of mmap() for a list possible errors that may occur.
  */
-errno_t cosmo_stack_alloc(size_t *inout_stacksize,  //
-                          size_t *inout_guardsize,  //
+errno_t cosmo_stack_alloc(unsigned *inout_stacksize,  //
+                          unsigned *inout_guardsize,  //
                           void **out_addr) {
 
   // validate arguments
-  size_t stacksize = *inout_stacksize;
-  size_t guardsize = *inout_guardsize;
+  unsigned stacksize = *inout_stacksize;
+  unsigned guardsize = *inout_guardsize;
   stacksize = (stacksize + __gransize - 1) & -__gransize;
   guardsize = (guardsize + __pagesize - 1) & -__pagesize;
   if (guardsize + __pagesize > stacksize)
@@ -283,7 +283,8 @@ static void cosmo_stack_setup(void) {
  * variable is never clobbered. You can only dependably count on this to
  * return an error on failure when you say `cosmo_stack_setmaxstacks(0)`
  */
-errno_t cosmo_stack_free(void *stackaddr, size_t stacksize, size_t guardsize) {
+errno_t cosmo_stack_free(void *stackaddr, unsigned stacksize,
+                         unsigned guardsize) {
   stacksize = (stacksize + __gransize - 1) & -__gransize;
   guardsize = (guardsize + __pagesize - 1) & -__pagesize;
   if (guardsize + __pagesize > stacksize)
diff --git a/libc/thread/mapstack.c b/libc/thread/mapstack.c
index 470ab58a6..28a3fd56e 100644
--- a/libc/thread/mapstack.c
+++ b/libc/thread/mapstack.c
@@ -35,8 +35,8 @@
  */
 void *NewCosmoStack(void) {
   void *stackaddr;
-  size_t stacksize = GetStackSize();
-  size_t guardsize = GetGuardSize();
+  unsigned stacksize = GetStackSize();
+  unsigned guardsize = GetGuardSize();
   errno_t err = cosmo_stack_alloc(&stacksize, &guardsize, &stackaddr);
   if (!err)
     return stackaddr;
diff --git a/libc/thread/posixthread.internal.h b/libc/thread/posixthread.internal.h
index 8fa216805..fe94dc066 100644
--- a/libc/thread/posixthread.internal.h
+++ b/libc/thread/posixthread.internal.h
@@ -78,8 +78,7 @@ struct PosixThread {
   atomic_int ptid;            // transitions 0 → tid
   atomic_int pt_refs;         // prevents decimation
   void *(*pt_start)(void *);  // creation callback
-  void *pt_arg;               // start's parameter
-  void *pt_rc;                // start's return value
+  void *pt_val;               // start param / return val
   char *pt_tls;               // bottom of tls allocation
   struct CosmoTib *tib;       // middle of tls allocation
   struct Dll list;            // list of threads
@@ -105,7 +104,6 @@ int _pthread_tid(struct PosixThread *) libcesque;
 intptr_t _pthread_syshand(struct PosixThread *) libcesque;
 long _pthread_cancel_ack(void) libcesque;
 void _pthread_decimate(void) libcesque;
-void _pthread_free(struct PosixThread *) libcesque;
 void _pthread_lock(void) libcesque;
 void _pthread_onfork_child(void) libcesque;
 void _pthread_onfork_parent(void) libcesque;
diff --git a/libc/thread/pthread_atfork.c b/libc/thread/pthread_atfork.c
index 5ef7a92c1..c7e32ed2c 100644
--- a/libc/thread/pthread_atfork.c
+++ b/libc/thread/pthread_atfork.c
@@ -63,11 +63,13 @@ static void _pthread_onfork(int i, const char *op) {
 }
 
 void _pthread_onfork_prepare(void) {
+  pthread_mutex_lock(&_atforks.lock);
   _pthread_onfork(0, "prepare");
 }
 
 void _pthread_onfork_parent(void) {
   _pthread_onfork(1, "parent");
+  pthread_mutex_unlock(&_atforks.lock);
 }
 
 void _pthread_onfork_child(void) {
diff --git a/libc/thread/pthread_create.c b/libc/thread/pthread_create.c
index ba5771a9e..351a18c8b 100644
--- a/libc/thread/pthread_create.c
+++ b/libc/thread/pthread_create.c
@@ -67,7 +67,7 @@ __static_yoink("_pthread_onfork_prepare");
 __static_yoink("_pthread_onfork_parent");
 __static_yoink("_pthread_onfork_child");
 
-void _pthread_free(struct PosixThread *pt) {
+static void _pthread_free(struct PosixThread *pt) {
 
   // thread must be removed from _pthread_list before calling
   unassert(dll_is_alone(&pt->list) && &pt->list != _pthread_list);
@@ -93,10 +93,13 @@ void _pthread_free(struct PosixThread *pt) {
   }
 
   // free heap memory associated with thread
-  if (pt->pt_flags & PT_OWNSIGALTSTACK)
-    free(pt->pt_attr.__sigaltstackaddr);
-  free(pt->pt_tls);
-  free(pt);
+  bulk_free(
+      (void *[]){
+          pt->pt_flags & PT_OWNSIGALTSTACK ? pt->pt_attr.__sigaltstackaddr : 0,
+          pt->pt_tls,
+          pt,
+      },
+      3);
 }
 
 void _pthread_decimate(void) {
@@ -137,7 +140,6 @@ void _pthread_decimate(void) {
 }
 
 static int PosixThread(void *arg, int tid) {
-  void *rc;
   struct PosixThread *pt = arg;
 
   // setup scheduling
@@ -167,11 +169,11 @@ static int PosixThread(void *arg, int tid) {
     } else {
       sys_sigprocmask(SIG_SETMASK, &pt->pt_attr.__sigmask, 0);
     }
-    rc = pt->pt_start(pt->pt_arg);
+    void *ret = pt->pt_start(pt->pt_val);
     // ensure pthread_cleanup_pop(), and pthread_exit() popped cleanup
     unassert(!pt->pt_cleanup);
     // calling pthread_exit() will either jump back here, or call exit
-    pthread_exit(rc);
+    pthread_exit(ret);
   }
 
   // avoid signal handler being triggered after we trash our own stack
@@ -196,7 +198,7 @@ static errno_t pthread_create_impl(pthread_t *thread,
   dll_init(&pt->list);
   pt->pt_locale = &__global_locale;
   pt->pt_start = start_routine;
-  pt->pt_arg = arg;
+  pt->pt_val = arg;
 
   // create thread local storage memory
   if (!(pt->pt_tls = _mktls(&pt->tib))) {
diff --git a/libc/thread/pthread_exit.c b/libc/thread/pthread_exit.c
index c50b867da..6c8d605bc 100644
--- a/libc/thread/pthread_exit.c
+++ b/libc/thread/pthread_exit.c
@@ -88,7 +88,7 @@ wontreturn void pthread_exit(void *rc) {
 
   // set state
   pt->pt_flags |= PT_NOCANCEL | PT_EXITING;
-  pt->pt_rc = rc;
+  pt->pt_val = rc;
 
   // free resources
   __cxa_thread_finalize();
diff --git a/libc/thread/pthread_timedjoin_np.c b/libc/thread/pthread_timedjoin_np.c
index 9022a9196..142ae4734 100644
--- a/libc/thread/pthread_timedjoin_np.c
+++ b/libc/thread/pthread_timedjoin_np.c
@@ -139,7 +139,7 @@ errno_t pthread_timedjoin_np(pthread_t thread, void **value_ptr,
                           memory_order_release);
     _pthread_zombify(pt);
     if (value_ptr)
-      *value_ptr = pt->pt_rc;
+      *value_ptr = pt->pt_val;
   }
 
   _pthread_unref(pt);
diff --git a/libc/thread/thread.h b/libc/thread/thread.h
index 4b469a209..e2827b7d4 100644
--- a/libc/thread/thread.h
+++ b/libc/thread/thread.h
@@ -130,8 +130,8 @@ typedef struct pthread_attr_s {
   int __contentionscope;
   int __sigaltstacksize;
   uint64_t __sigmask;
-  size_t __guardsize;
-  size_t __stacksize;
+  unsigned __guardsize;
+  unsigned __stacksize;
   void *__stackaddr;
   void *__sigaltstackaddr;
 } pthread_attr_t;
diff --git a/test/libc/mem/malloc_test.c b/test/libc/mem/malloc_test.c
index 5e69b98ca..b1b7d2609 100644
--- a/test/libc/mem/malloc_test.c
+++ b/test/libc/mem/malloc_test.c
@@ -22,7 +22,6 @@
 #include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/intrin/cxaatexit.h"
-#include "libc/intrin/kprintf.h"
 #include "libc/intrin/safemacros.h"
 #include "libc/macros.h"
 #include "libc/mem/gc.h"
@@ -162,7 +161,7 @@ void *bulk[1024];
 void BulkFreeBenchSetup(void) {
   size_t i;
   for (i = 0; i < ARRAYLEN(bulk); ++i) {
-    bulk[i] = malloc(rand() % 64);
+    bulk[i] = rand() % 64 ? malloc(rand() % 64) : 0;
   }
 }
 
diff --git a/third_party/dlmalloc/dlmalloc.c b/third_party/dlmalloc/dlmalloc.c
index 0adc13f4f..b20e28cd9 100644
--- a/third_party/dlmalloc/dlmalloc.c
+++ b/third_party/dlmalloc/dlmalloc.c
@@ -62,11 +62,6 @@
 #include "locks.inc"
 #include "chunks.inc"
 #include "headfoot.inc"
-
-#if ONLY_MSPACES
-#include "threaded.inc"
-#endif
-
 #include "global.inc"
 #include "system.inc"
 #include "hooks.inc"
@@ -74,6 +69,11 @@
 #include "indexing.inc"
 #include "binmaps.inc"
 #include "runtimechecks.inc"
+
+#if ONLY_MSPACES
+#include "threaded.inc"
+#endif
+
 #include "init.inc"
 #include "debuglib.inc"
 #include "statistics.inc"
diff --git a/third_party/dlmalloc/threaded.inc b/third_party/dlmalloc/threaded.inc
index e8768dbc3..3dbfb5b35 100644
--- a/third_party/dlmalloc/threaded.inc
+++ b/third_party/dlmalloc/threaded.inc
@@ -61,8 +61,23 @@ int dlmalloc_trim(size_t pad) {
 }
 
 size_t dlbulk_free(void *array[], size_t nelem) {
-  for (size_t i = 0; i < nelem; ++i)
-    mspace_free(0, array[i]);
+  size_t j = 0;
+  mstate msp = (mstate)-1;
+  for (size_t i = 0; i < nelem; ++i) {
+    mstate next;
+    if (array[i]) {
+      next = get_mstate_for(mem2chunk(array[i]));
+      if (next != msp) {
+        if (j)
+          mspace_bulk_free(msp, array, j);
+        msp = next;
+        j = 0;
+      }
+      array[j++] = array[i];
+    }
+  }
+  if (j)
+    mspace_bulk_free(msp, array, j);
   return 0;
 }