Add "hook" module.

The hook module allows a low-reader-overhead way of finding hooks to invoke and
calling them.

For now, none of the allocation pathways are tied into the hooks; this will come
later.

Makefile.in

@@ -102,6 +102,7 @@ C_SRCS := $(srcroot)src/jemalloc.c \
 	$(srcroot)src/extent_dss.c \
 	$(srcroot)src/extent_mmap.c \
 	$(srcroot)src/hash.c \
+	$(srcroot)src/hook.c \
 	$(srcroot)src/large.c \
 	$(srcroot)src/log.c \
 	$(srcroot)src/malloc_io.c \
@@ -172,6 +173,7 @@ TESTS_UNIT := \
 	$(srcroot)test/unit/extent_quantize.c \
 	$(srcroot)test/unit/fork.c \
 	$(srcroot)test/unit/hash.c \
+	$(srcroot)test/unit/hook.c \
 	$(srcroot)test/unit/junk.c \
 	$(srcroot)test/unit/junk_alloc.c \
 	$(srcroot)test/unit/junk_free.c \

include/jemalloc/internal/hook.h

@@ -0,0 +1,125 @@
+#ifndef JEMALLOC_INTERNAL_HOOK_H
+#define JEMALLOC_INTERNAL_HOOK_H
+
+#include "jemalloc/internal/tsd.h"
+
+/*
+ * This API is *extremely* experimental, and may get ripped out, changed in API-
+ * and ABI-incompatible ways, be insufficiently or incorrectly documented, etc.
+ *
+ * It allows hooking the stateful parts of the API to see changes as they
+ * happen.
+ *
+ * Allocation hooks are called after the allocation is done, free hooks are
+ * called before the free is done, and expand hooks are called after the
+ * allocation is expanded.
+ *
+ * For realloc and rallocx, if the expansion happens in place, the expansion
+ * hook is called.  If it is moved, then the alloc hook is called on the new
+ * location, and then the free hook is called on the old location.
+ *
+ * (We omit no-ops, like free(NULL), etc.).
+ *
+ * Reentrancy:
+ *   Is not protected against.  If your hooks allocate, then the hooks will be
+ *   called again.  Note that you can guard against this with a thread-local
+ *   "in_hook" bool.
+ * Threading:
+ *   The installation of a hook synchronizes with all its uses.  If you can
+ *   prove the installation of a hook happens-before a jemalloc entry point,
+ *   then the hook will get invoked (unless there's a racing removal).
+ *
+ *   Hook insertion appears to be atomic at a per-thread level (i.e. if a thread
+ *   allocates and has the alloc hook invoked, then a subsequent free on the
+ *   same thread will also have the free hook invoked).
+ *
+ *   The *removal* of a hook does *not* block until all threads are done with
+ *   the hook.  Hook authors have to be resilient to this, and need some
+ *   out-of-band mechanism for cleaning up any dynamically allocated memory
+ *   associated with their hook.
+ * Ordering:
+ *   Order of hook execution is unspecified, and may be different than insertion
+ *   order.
+ */
+
+enum hook_alloc_e {
+	hook_alloc_malloc,
+	hook_alloc_posix_memalign,
+	hook_alloc_aligned_alloc,
+	hook_alloc_calloc,
+	hook_alloc_memalign,
+	hook_alloc_valloc,
+	hook_alloc_mallocx,
+
+	/* The reallocating functions have both alloc and dalloc variants */
+	hook_alloc_realloc,
+	hook_alloc_rallocx,
+};
+/*
+ * We put the enum typedef after the enum, since this file may get included by
+ * jemalloc_cpp.cpp, and C++ disallows enum forward declarations.
+ */
+typedef enum hook_alloc_e hook_alloc_t;
+
+enum hook_dalloc_e {
+	hook_dalloc_free,
+	hook_dalloc_dallocx,
+	hook_dalloc_sdallocx,
+
+	/*
+	 * The dalloc halves of reallocation (not called if in-place expansion
+	 * happens).
+	 */
+	hook_dalloc_realloc,
+	hook_dalloc_rallocx,
+};
+typedef enum hook_dalloc_e hook_dalloc_t;
+
+
+enum hook_expand_e {
+	hook_expand_realloc,
+	hook_expand_rallocx,
+	hook_expand_xallocx,
+};
+typedef enum hook_expand_e hook_expand_t;
+
+typedef void (*hook_alloc)(
+    void *extra, hook_alloc_t type, void *result, uintptr_t result_raw,
+    uintptr_t args_raw[3]);
+
+typedef void (*hook_dalloc)(
+    void *extra, hook_dalloc_t type, void *address, uintptr_t args_raw[3]);
+
+typedef void (*hook_expand)(
+    void *extra, hook_expand_t type, void *address, size_t old_usize,
+    size_t new_usize, uintptr_t result_raw, uintptr_t args_raw[4]);
+
+typedef struct hooks_s hooks_t;
+struct hooks_s {
+	hook_alloc alloc_hook;
+	hook_dalloc dalloc_hook;
+	hook_expand expand_hook;
+};
+
+/*
+ * Returns an opaque handle to be used when removing the hook.  NULL means that
+ * we couldn't install the hook.
+ */
+bool hook_boot();
+
+void *hook_install(tsdn_t *tsdn, hooks_t *hooks, void *extra);
+/* Uninstalls the hook with the handle previously returned from hook_install. */
+void hook_remove(tsdn_t *tsdn, void *opaque);
+
+/* Hooks */
+
+void hook_invoke_alloc(hook_alloc_t type, void *result, uintptr_t result_raw,
+    uintptr_t args_raw[3]);
+
+void hook_invoke_dalloc(hook_dalloc_t type, void *address,
+    uintptr_t args_raw[3]);
+
+void hook_invoke_expand(hook_expand_t type, void *address, size_t old_usize,
+    size_t new_usize, uintptr_t result_raw, uintptr_t args_raw[4]);
+
+#endif /* JEMALLOC_INTERNAL_HOOK_H */

include/jemalloc/internal/witness.h

@@ -49,6 +49,7 @@
 #define WITNESS_RANK_RTREE		17U
 #define WITNESS_RANK_BASE		18U
 #define WITNESS_RANK_ARENA_LARGE	19U
+#define WITNESS_RANK_HOOK		20U
 
 #define WITNESS_RANK_LEAF		0xffffffffU
 #define WITNESS_RANK_BIN		WITNESS_RANK_LEAF

src/hook.c

@@ -0,0 +1,133 @@
+#include "jemalloc/internal/jemalloc_preamble.h"
+
+#include "jemalloc/internal/hook.h"
+
+#include "jemalloc/internal/atomic.h"
+#include "jemalloc/internal/mutex.h"
+#include "jemalloc/internal/seq.h"
+
+typedef struct hooks_internal_s hooks_internal_t;
+struct hooks_internal_s {
+	hooks_t hooks;
+	void *extra;
+	bool in_use;
+};
+
+seq_define(hooks_internal_t, hooks)
+
+#define HOOKS_MAX 4
+static seq_hooks_t hooks[HOOKS_MAX];
+static malloc_mutex_t hooks_mu;
+
+bool
+hook_boot() {
+	return malloc_mutex_init(&hooks_mu, "hooks", WITNESS_RANK_HOOK,
+	    malloc_mutex_rank_exclusive);
+}
+
+static void *
+hook_install_locked(hooks_t *to_install, void *extra) {
+	hooks_internal_t hooks_internal;
+	for (int i = 0; i < HOOKS_MAX; i++) {
+		bool success = seq_try_load_hooks(&hooks_internal, &hooks[i]);
+		/* We hold mu; no concurrent access. */
+		assert(success);
+		if (!hooks_internal.in_use) {
+			hooks_internal.hooks = *to_install;
+			hooks_internal.extra = extra;
+			hooks_internal.in_use = true;
+			seq_store_hooks(&hooks[i], &hooks_internal);
+			return &hooks[i];
+		}
+	}
+	return NULL;
+}
+
+void *
+hook_install(tsdn_t *tsdn, hooks_t *to_install, void *extra) {
+	malloc_mutex_lock(tsdn, &hooks_mu);
+	void *ret = hook_install_locked(to_install, extra);
+	if (ret != NULL) {
+		tsd_global_slow_inc(tsdn);
+	}
+	malloc_mutex_unlock(tsdn, &hooks_mu);
+	return ret;
+}
+
+static void
+hook_remove_locked(seq_hooks_t *to_remove) {
+	hooks_internal_t hooks_internal;
+	bool success = seq_try_load_hooks(&hooks_internal, to_remove);
+	/* We hold mu; no concurrent access. */
+	assert(success);
+	/* Should only remove hooks that were added. */
+	assert(hooks_internal.in_use);
+	hooks_internal.in_use = false;
+	seq_store_hooks(to_remove, &hooks_internal);
+}
+
+void
+hook_remove(tsdn_t *tsdn, void *opaque) {
+	if (config_debug) {
+		char *hooks_begin = (char *)&hooks[0];
+		char *hooks_end = (char *)&hooks[HOOKS_MAX];
+		char *hook = (char *)opaque;
+		assert(hooks_begin <= hook && hook < hooks_end
+		    && (hook - hooks_begin) % sizeof(seq_hooks_t) == 0);
+	}
+	malloc_mutex_lock(tsdn, &hooks_mu);
+	hook_remove_locked((seq_hooks_t *)opaque);
+	tsd_global_slow_dec(tsdn);
+	malloc_mutex_unlock(tsdn, &hooks_mu);
+}
+
+#define FOR_EACH_HOOK_BEGIN(hooks_internal_ptr)				\
+for (int for_each_hook_counter = 0;					\
+    for_each_hook_counter < HOOKS_MAX;					\
+    for_each_hook_counter++) {						\
+	bool for_each_hook_success = seq_try_load_hooks(		\
+	    (hooks_internal_ptr), &hooks[for_each_hook_counter]);	\
+	if (!for_each_hook_success) {					\
+		continue;						\
+	}								\
+	if (!(hooks_internal_ptr)->in_use) {				\
+		continue;						\
+	}
+#define FOR_EACH_HOOK_END						\
+}
+
+void
+hook_invoke_alloc(hook_alloc_t type, void *result, uintptr_t result_raw,
+    uintptr_t args_raw[3]) {
+	hooks_internal_t hook;
+	FOR_EACH_HOOK_BEGIN(&hook)
+		hook_alloc h = hook.hooks.alloc_hook;
+		if (h != NULL) {
+			h(hook.extra, type, result, result_raw, args_raw);
+		}
+	FOR_EACH_HOOK_END
+}
+
+void
+hook_invoke_dalloc(hook_dalloc_t type, void *address, uintptr_t args_raw[3]) {
+	hooks_internal_t hook;
+	FOR_EACH_HOOK_BEGIN(&hook)
+		hook_dalloc h = hook.hooks.dalloc_hook;
+		if (h != NULL) {
+			h(hook.extra, type, address, args_raw);
+		}
+	FOR_EACH_HOOK_END
+}
+
+void
+hook_invoke_expand(hook_expand_t type, void *address, size_t old_usize,
+    size_t new_usize, uintptr_t result_raw, uintptr_t args_raw[4]) {
+	hooks_internal_t hook;
+	FOR_EACH_HOOK_BEGIN(&hook)
+		hook_expand h = hook.hooks.expand_hook;
+		if (h != NULL) {
+			h(hook.extra, type, address, old_usize, new_usize,
+			    result_raw, args_raw);
+		}
+	FOR_EACH_HOOK_END
+}

src/jemalloc.c

@@ -7,6 +7,7 @@
 #include "jemalloc/internal/ctl.h"
 #include "jemalloc/internal/extent_dss.h"
 #include "jemalloc/internal/extent_mmap.h"
+#include "jemalloc/internal/hook.h"
 #include "jemalloc/internal/jemalloc_internal_types.h"
 #include "jemalloc/internal/log.h"
 #include "jemalloc/internal/malloc_io.h"
@@ -1311,6 +1312,7 @@ malloc_init_hard_a0_locked() {
 	    malloc_mutex_rank_exclusive)) {
 		return true;
 	}
+	hook_boot();
 	/*
 	 * Create enough scaffolding to allow recursive allocation in
 	 * malloc_ncpus().

test/unit/hook.c

@@ -0,0 +1,180 @@
+#include "test/jemalloc_test.h"
+
+#include "jemalloc/internal/hook.h"
+
+static void *arg_extra;
+static int arg_type;
+static void *arg_result;
+static void *arg_address;
+static size_t arg_old_usize;
+static size_t arg_new_usize;
+static uintptr_t arg_result_raw;
+static uintptr_t arg_args_raw[4];
+
+static int call_count = 0;
+
+static void
+reset_args() {
+	arg_extra = NULL;
+	arg_type = 12345;
+	arg_result = NULL;
+	arg_address = NULL;
+	arg_old_usize = 0;
+	arg_new_usize = 0;
+	arg_result_raw = 0;
+	memset(arg_args_raw, 77, sizeof(arg_args_raw));
+}
+
+static void
+set_args_raw(uintptr_t *args_raw, int nargs) {
+	memcpy(arg_args_raw, args_raw, sizeof(uintptr_t) * nargs);
+}
+
+static void
+assert_args_raw(uintptr_t *args_raw_expected, int nargs) {
+	int cmp = memcmp(args_raw_expected, arg_args_raw,
+	    sizeof(uintptr_t) * nargs);
+	assert_d_eq(cmp, 0, "Raw args mismatch");
+}
+
+static void
+test_alloc_hook(void *extra, hook_alloc_t type, void *result,
+    uintptr_t result_raw, uintptr_t args_raw[3]) {
+	call_count++;
+	arg_extra = extra;
+	arg_type = (int)type;
+	arg_result = result;
+	arg_result_raw = result_raw;
+	set_args_raw(args_raw, 3);
+}
+
+static void
+test_dalloc_hook(void *extra, hook_dalloc_t type, void *address,
+    uintptr_t args_raw[3]) {
+	call_count++;
+	arg_extra = extra;
+	arg_type = (int)type;
+	arg_address = address;
+	set_args_raw(args_raw, 3);
+}
+
+static void
+test_expand_hook(void *extra, hook_expand_t type, void *address,
+    size_t old_usize, size_t new_usize, uintptr_t result_raw,
+    uintptr_t args_raw[4]) {
+	call_count++;
+	arg_extra = extra;
+	arg_type = (int)type;
+	arg_address = address;
+	arg_old_usize = old_usize;
+	arg_new_usize = new_usize;
+	arg_result_raw = result_raw;
+	set_args_raw(args_raw, 4);
+}
+
+TEST_BEGIN(test_hooks_basic) {
+	/* Just verify that the record their arguments correctly. */
+	hooks_t hooks = {
+		&test_alloc_hook, &test_dalloc_hook, &test_expand_hook};
+	void *handle = hook_install(TSDN_NULL, &hooks, (void *)111);
+	uintptr_t args_raw[4] = {10, 20, 30, 40};
+
+	/* Alloc */
+	reset_args();
+	hook_invoke_alloc(hook_alloc_posix_memalign, (void *)222, 333,
+	    args_raw);
+	assert_ptr_eq(arg_extra, (void *)111, "Passed wrong user pointer");
+	assert_d_eq((int)hook_alloc_posix_memalign, arg_type,
+	    "Passed wrong alloc type");
+	assert_ptr_eq((void *)222, arg_result, "Passed wrong result address");
+	assert_u64_eq(333, arg_result_raw, "Passed wrong result");
+	assert_args_raw(args_raw, 3);
+
+	/* Dalloc */
+	reset_args();
+	hook_invoke_dalloc(hook_dalloc_sdallocx, (void *)222, args_raw);
+	assert_d_eq((int)hook_dalloc_sdallocx, arg_type,
+	    "Passed wrong dalloc type");
+	assert_ptr_eq((void *)111, arg_extra, "Passed wrong user pointer");
+	assert_ptr_eq((void *)222, arg_address, "Passed wrong address");
+	assert_args_raw(args_raw, 3);
+
+	/* Expand */
+	reset_args();
+	hook_invoke_expand(hook_expand_xallocx, (void *)222, 333, 444, 555,
+	    args_raw);
+	assert_d_eq((int)hook_expand_xallocx, arg_type,
+	    "Passed wrong expand type");
+	assert_ptr_eq((void *)111, arg_extra, "Passed wrong user pointer");
+	assert_ptr_eq((void *)222, arg_address, "Passed wrong address");
+	assert_zu_eq(333, arg_old_usize, "Passed wrong old usize");
+	assert_zu_eq(444, arg_new_usize, "Passed wrong new usize");
+	assert_zu_eq(555, arg_result_raw, "Passed wrong result");
+	assert_args_raw(args_raw, 4);
+
+	hook_remove(TSDN_NULL, handle);
+}
+TEST_END
+
+TEST_BEGIN(test_hooks_null) {
+	/* Null hooks should be ignored, not crash. */
+	hooks_t hooks1 = {NULL, NULL, NULL};
+	hooks_t hooks2 = {&test_alloc_hook, NULL, NULL};
+	hooks_t hooks3 = {NULL, &test_dalloc_hook, NULL};
+	hooks_t hooks4 = {NULL, NULL, &test_expand_hook};
+
+	void *handle1 = hook_install(TSDN_NULL, &hooks1, NULL);
+	void *handle2 = hook_install(TSDN_NULL, &hooks2, NULL);
+	void *handle3 = hook_install(TSDN_NULL, &hooks3, NULL);
+	void *handle4 = hook_install(TSDN_NULL, &hooks4, NULL);
+
+	assert_ptr_ne(handle1, NULL, "Hook installation failed");
+	assert_ptr_ne(handle2, NULL, "Hook installation failed");
+	assert_ptr_ne(handle3, NULL, "Hook installation failed");
+	assert_ptr_ne(handle4, NULL, "Hook installation failed");
+
+	uintptr_t args_raw[4] = {10, 20, 30, 40};
+
+	call_count = 0;
+	hook_invoke_alloc(hook_alloc_malloc, NULL, 0, args_raw);
+	assert_d_eq(call_count, 1, "Called wrong number of times");
+
+	call_count = 0;
+	hook_invoke_dalloc(hook_dalloc_free, NULL, args_raw);
+	assert_d_eq(call_count, 1, "Called wrong number of times");
+
+	call_count = 0;
+	hook_invoke_expand(hook_expand_realloc, NULL, 0, 0, 0, args_raw);
+	assert_d_eq(call_count, 1, "Called wrong number of times");
+
+	hook_remove(TSDN_NULL, handle1);
+	hook_remove(TSDN_NULL, handle2);
+	hook_remove(TSDN_NULL, handle3);
+	hook_remove(TSDN_NULL, handle4);
+}
+TEST_END
+
+TEST_BEGIN(test_hooks_remove) {
+	hooks_t hooks = {&test_alloc_hook, NULL, NULL};
+	void *handle = hook_install(TSDN_NULL, &hooks, NULL);
+	assert_ptr_ne(handle, NULL, "Hook installation failed");
+	call_count = 0;
+	uintptr_t args_raw[4] = {10, 20, 30, 40};
+	hook_invoke_alloc(hook_alloc_malloc, NULL, 0, args_raw);
+	assert_d_eq(call_count, 1, "Hook not invoked");
+
+	call_count = 0;
+	hook_remove(TSDN_NULL, handle);
+	hook_invoke_alloc(hook_alloc_malloc, NULL, 0, NULL);
+	assert_d_eq(call_count, 0, "Hook invoked after removal");
+
+}
+TEST_END
+
+int
+main(void) {
+	return test(
+	    test_hooks_basic,
+	    test_hooks_null,
+	    test_hooks_remove);
+}