void
sgen_gray_object_enqueue_section (SgenGrayQueue *queue, GrayQueueSection *section, gboolean is_parallel)
{
STATE_TRANSITION (section, GRAY_QUEUE_SECTION_STATE_FLOATING, GRAY_QUEUE_SECTION_STATE_ENQUEUED);
if (queue->first)
queue->first->size = queue->cursor - queue->first->entries + 1;
section->next = queue->first;
section->prev = NULL;
if (queue->first)
queue->first->prev = section;
else
queue->last = section;
queue->first = section;
queue->cursor = queue->first->entries + queue->first->size - 1;
#ifdef SGEN_CHECK_GRAY_OBJECT_ENQUEUE
if (queue->enqueue_check_func) {
int i;
for (i = 0; i < section->size; ++i)
queue->enqueue_check_func (section->entries [i].obj);
}
#endif
if (is_parallel) {
mono_memory_write_barrier ();
mono_atomic_inc_i32 (&queue->num_sections);
} else {
queue->num_sections++;
}
}
MONO_SIG_HANDLER_FUNC (static, profiler_signal_handler)
{
int old_errno = errno;
MONO_SIG_HANDLER_GET_CONTEXT;
/* See the comment in mono_runtime_shutdown_stat_profiler (). */
if (mono_native_thread_id_get () == sampling_thread) {
mono_atomic_inc_i32 (&profiler_interrupt_signals_received);
return;
}
mono_atomic_inc_i32 (&profiler_signals_received);
// Did a non-attached or detaching thread get the signal?
if (mono_thread_info_get_small_id () == -1 ||
!mono_domain_get () ||
!mono_tls_get_jit_tls ()) {
errno = old_errno;
return;
}
// See the comment in sampling_thread_func ().
mono_atomic_store_i32 (&mono_thread_info_current ()->profiler_signal_ack, 1);
mono_atomic_inc_i32 (&profiler_signals_accepted);
int hp_save_index = mono_hazard_pointer_save_for_signal_handler ();
mono_thread_info_set_is_async_context (TRUE);
MONO_PROFILER_RAISE (sample_hit, (mono_arch_ip_from_context (ctx), ctx));
mono_thread_info_set_is_async_context (FALSE);
mono_hazard_pointer_restore_for_signal_handler (hp_save_index);
errno = old_errno;
mono_chain_signal (MONO_SIG_HANDLER_PARAMS);
}
/**
* mono_thread_hazardous_queue_free:
* \param p the pointer to free
* \param free_func the function that can free the pointer
* Queue \p p to be freed later. \p p will be freed once the hazard free queue is pumped.
*
* This function doesn't pump the free queue so try to accommodate a call at an appropriate time.
* See \c mono_thread_hazardous_try_free_some for when it's appropriate.
*/
void
mono_thread_hazardous_queue_free (gpointer p, MonoHazardousFreeFunc free_func)
{
DelayedFreeItem item = { p, free_func };
mono_atomic_inc_i32 (&hazardous_pointer_count);
mono_lock_free_array_queue_push (&delayed_free_queue, &item);
guint32 queue_size = delayed_free_queue.num_used_entries;
if (queue_size && queue_size_cb)
queue_size_cb (queue_size);
}
GrayQueueSection*
sgen_gray_object_steal_section (SgenGrayQueue *queue)
{
gint32 sections_remaining;
GrayQueueSection *section = NULL;
/*
* With each push/pop into the queue we increment the number of sections.
* There is only one thread accessing the top (the owner) and potentially
* multiple workers trying to steal sections from the bottom, so we need
* to lock. A num sections decrement from the owner means that the first
* section is reserved, while a decrement by the stealer means that the
* last section is reserved. If after we decrement the num sections, we
* have at least one more section present, it means we can't race with
* the other thread. If this is not the case the steal end abandons the
* pop, setting back the num_sections, while the owner end will take a
* lock to make sure we are not racing with the stealer (since the stealer
* might have popped an entry and be in the process of updating the entry
* that the owner is trying to pop.
*/
if (queue->num_sections <= 1)
return NULL;
/* Give up if there is contention on the last section */
if (mono_os_mutex_trylock (&queue->steal_mutex) != 0)
return NULL;
sections_remaining = mono_atomic_dec_i32 (&queue->num_sections);
if (sections_remaining <= 0) {
/* The section that we tried to steal might be the head of the queue. */
mono_atomic_inc_i32 (&queue->num_sections);
} else {
/* We have reserved for us the tail section of the queue */
section = queue->last;
SGEN_ASSERT (0, section, "Why we don't have any sections to steal?");
SGEN_ASSERT (0, !section->next, "Why aren't we stealing the tail?");
queue->last = section->prev;
section->prev = NULL;
SGEN_ASSERT (0, queue->last, "Why are we stealing the last section?");
queue->last->next = NULL;
STATE_TRANSITION (section, GRAY_QUEUE_SECTION_STATE_ENQUEUED, GRAY_QUEUE_SECTION_STATE_FLOATING);
}
mono_os_mutex_unlock (&queue->steal_mutex);
return section;
}
void
sgen_gray_object_alloc_queue_section (SgenGrayQueue *queue, gboolean is_parallel)
{
GrayQueueSection *section;
if (queue->free_list) {
/* Use the previously allocated queue sections if possible */
section = queue->free_list;
queue->free_list = section->next;
STATE_TRANSITION (section, GRAY_QUEUE_SECTION_STATE_FREE_LIST, GRAY_QUEUE_SECTION_STATE_FLOATING);
} else {
HEAVY_STAT (stat_gray_queue_section_alloc ++);
/* Allocate a new section */
section = (GrayQueueSection *)sgen_alloc_internal (INTERNAL_MEM_GRAY_QUEUE);
STATE_SET (section, GRAY_QUEUE_SECTION_STATE_FLOATING);
}
/* Section is empty */
section->size = 0;
STATE_TRANSITION (section, GRAY_QUEUE_SECTION_STATE_FLOATING, GRAY_QUEUE_SECTION_STATE_ENQUEUED);
/* Link it with the others */
section->next = queue->first;
section->prev = NULL;
if (queue->first)
queue->first->prev = section;
else
queue->last = section;
queue->first = section;
queue->cursor = section->entries - 1;
if (is_parallel) {
mono_memory_write_barrier ();
/*
* FIXME
* we could probably optimize the code to only rely on the write barrier
* for synchronization with the stealer thread. Additionally we could also
* do a write barrier once every other gray queue change, and request
* to have a minimum of sections before stealing, to keep consistency.
*/
mono_atomic_inc_i32 (&queue->num_sections);
} else {
queue->num_sections++;
}
}
GCObject*
sgen_alloc_obj (GCVTable vtable, size_t size)
{
GCObject *res;
TLAB_ACCESS_INIT;
if (!SGEN_CAN_ALIGN_UP (size))
return NULL;
if (G_UNLIKELY (sgen_has_per_allocation_action)) {
static int alloc_count;
int current_alloc = mono_atomic_inc_i32 (&alloc_count);
if (sgen_verify_before_allocs) {
if ((current_alloc % sgen_verify_before_allocs) == 0) {
LOCK_GC;
sgen_check_whole_heap_stw ();
UNLOCK_GC;
}
}
if (sgen_collect_before_allocs) {
if (((current_alloc % sgen_collect_before_allocs) == 0) && sgen_nursery_section) {
LOCK_GC;
sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE, TRUE);
UNLOCK_GC;
}
}
}
ENTER_CRITICAL_REGION;
res = sgen_try_alloc_obj_nolock (vtable, size);
if (res) {
EXIT_CRITICAL_REGION;
return res;
}
EXIT_CRITICAL_REGION;
LOCK_GC;
res = sgen_alloc_obj_nolock (vtable, size);
UNLOCK_GC;
return res;
}
gboolean
sgen_cement_lookup_or_register (GCObject *obj)
{
guint hv;
int i;
CementHashEntry *hash = cement_hash;
if (!cement_enabled)
return FALSE;
hv = sgen_aligned_addr_hash (obj);
i = SGEN_CEMENT_HASH (hv);
SGEN_ASSERT (5, sgen_ptr_in_nursery (obj), "Can only cement pointers to nursery objects");
if (!hash [i].obj) {
GCObject *old_obj;
old_obj = (GCObject*)mono_atomic_cas_ptr ((gpointer*)&hash [i].obj, obj, NULL);
/* Check if the slot was occupied by some other object */
if (old_obj != NULL && old_obj != obj)
return FALSE;
} else if (hash [i].obj != obj) {
return FALSE;
}
if (hash [i].count >= SGEN_CEMENT_THRESHOLD)
return TRUE;
if (mono_atomic_inc_i32 ((gint32*)&hash [i].count) == SGEN_CEMENT_THRESHOLD) {
SGEN_ASSERT (9, sgen_get_current_collection_generation () >= 0, "We can only cement objects when we're in a collection pause.");
SGEN_ASSERT (9, SGEN_OBJECT_IS_PINNED (obj), "Can only cement pinned objects");
SGEN_CEMENT_OBJECT (obj);
sgen_binary_protocol_cement (obj, (gpointer)SGEN_LOAD_VTABLE (obj),
(int)sgen_safe_object_get_size (obj));
}
return FALSE;
}
/*
* Provide a variant that takes just the vtable for small fixed-size objects.
* The aligned size is already computed and stored in vt->gc_descr.
* Note: every SGEN_SCAN_START_SIZE or so we are given the chance to do some special
* processing. We can keep track of where objects start, for example,
* so when we scan the thread stacks for pinned objects, we can start
* a search for the pinned object in SGEN_SCAN_START_SIZE chunks.
*/
GCObject*
sgen_alloc_obj_nolock (GCVTable vtable, size_t size)
{
/* FIXME: handle OOM */
void **p;
char *new_next;
size_t real_size = size;
TLAB_ACCESS_INIT;
CANARIFY_SIZE(size);
HEAVY_STAT (++stat_objects_alloced);
if (real_size <= SGEN_MAX_SMALL_OBJ_SIZE)
HEAVY_STAT (stat_bytes_alloced += size);
else
HEAVY_STAT (stat_bytes_alloced_los += size);
size = ALIGN_UP (size);
SGEN_ASSERT (6, sgen_vtable_get_descriptor (vtable), "VTable without descriptor");
if (G_UNLIKELY (sgen_has_per_allocation_action)) {
static int alloc_count;
int current_alloc = mono_atomic_inc_i32 (&alloc_count);
if (sgen_collect_before_allocs) {
if (((current_alloc % sgen_collect_before_allocs) == 0) && sgen_nursery_section) {
sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE, TRUE);
if (!sgen_degraded_mode && sgen_can_alloc_size (size) && real_size <= SGEN_MAX_SMALL_OBJ_SIZE) {
// FIXME:
g_assert_not_reached ();
}
}
} else if (sgen_verify_before_allocs) {
if ((current_alloc % sgen_verify_before_allocs) == 0)
sgen_check_whole_heap_stw ();
}
}
/*
* We must already have the lock here instead of after the
* fast path because we might be interrupted in the fast path
* (after confirming that new_next < TLAB_TEMP_END) by the GC,
* and we'll end up allocating an object in a fragment which
* no longer belongs to us.
*
* The managed allocator does not do this, but it's treated
* specially by the world-stopping code.
*/
if (real_size > SGEN_MAX_SMALL_OBJ_SIZE) {
p = (void **)sgen_los_alloc_large_inner (vtable, ALIGN_UP (real_size));
} else {
/* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
p = (void**)TLAB_NEXT;
/* FIXME: handle overflow */
new_next = (char*)p + size;
TLAB_NEXT = new_next;
if (G_LIKELY (new_next < TLAB_TEMP_END)) {
/* Fast path */
CANARIFY_ALLOC(p,real_size);
SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
sgen_binary_protocol_alloc (p , vtable, size, sgen_client_get_provenance ());
g_assert (*p == NULL);
mono_atomic_store_seq (p, vtable);
return (GCObject*)p;
}
/* Slow path */
/* there are two cases: the object is too big or we run out of space in the TLAB */
/* we also reach here when the thread does its first allocation after a minor
* collection, since the tlab_ variables are initialized to NULL.
* there can be another case (from ORP), if we cooperate with the runtime a bit:
* objects that need finalizers can have the high bit set in their size
* so the above check fails and we can readily add the object to the queue.
* This avoids taking again the GC lock when registering, but this is moot when
* doing thread-local allocation, so it may not be a good idea.
*/
if (TLAB_NEXT >= TLAB_REAL_END) {
int available_in_tlab;
/*
* Run out of space in the TLAB. When this happens, some amount of space
* remains in the TLAB, but not enough to satisfy the current allocation
* request. Currently, we retire the TLAB in all cases, later we could
* keep it if the remaining space is above a treshold, and satisfy the
* allocation directly from the nursery.
*/
TLAB_NEXT -= size;
/* when running in degraded mode, we continue allocing that way
* for a while, to decrease the number of useless nursery collections.
*/
if (sgen_degraded_mode && sgen_degraded_mode < sgen_nursery_size)
return alloc_degraded (vtable, size, FALSE);
available_in_tlab = (int)(TLAB_REAL_END - TLAB_NEXT);//We'll never have tlabs > 2Gb
if (size > sgen_tlab_size || available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
/* Allocate directly from the nursery */
p = (void **)sgen_nursery_alloc (size);
if (!p) {
/*
* We couldn't allocate from the nursery, so we try
* collecting. Even after the collection, we might
* still not have enough memory to allocate the
* object. The reason will most likely be that we've
* run out of memory, but there is the theoretical
* possibility that other threads might have consumed
* the freed up memory ahead of us.
*
* What we do in this case is allocate degraded, i.e.,
* from the major heap.
*
* Ideally we'd like to detect the case of other
* threads allocating ahead of us and loop (if we
* always loop we will loop endlessly in the case of
* OOM).
*/
sgen_ensure_free_space (real_size, GENERATION_NURSERY);
if (!sgen_degraded_mode)
p = (void **)sgen_nursery_alloc (size);
}
if (!p)
return alloc_degraded (vtable, size, TRUE);
zero_tlab_if_necessary (p, size);
} else {
size_t alloc_size = 0;
if (TLAB_START)
SGEN_LOG (3, "Retire TLAB: %p-%p [%ld]", TLAB_START, TLAB_REAL_END, (long)(TLAB_REAL_END - TLAB_NEXT - size));
sgen_nursery_retire_region (p, available_in_tlab);
p = (void **)sgen_nursery_alloc_range (sgen_tlab_size, size, &alloc_size);
if (!p) {
/* See comment above in similar case. */
sgen_ensure_free_space (sgen_tlab_size, GENERATION_NURSERY);
if (!sgen_degraded_mode)
p = (void **)sgen_nursery_alloc_range (sgen_tlab_size, size, &alloc_size);
}
if (!p)
return alloc_degraded (vtable, size, TRUE);
/* Allocate a new TLAB from the current nursery fragment */
TLAB_START = (char*)p;
TLAB_NEXT = TLAB_START;
TLAB_REAL_END = TLAB_START + alloc_size;
TLAB_TEMP_END = TLAB_START + MIN (SGEN_SCAN_START_SIZE, alloc_size);
zero_tlab_if_necessary (TLAB_START, alloc_size);
/* Allocate from the TLAB */
p = (void **)TLAB_NEXT;
TLAB_NEXT += size;
sgen_set_nursery_scan_start ((char*)p);
}
} else {
/* Reached tlab_temp_end */
/* record the scan start so we can find pinned objects more easily */
sgen_set_nursery_scan_start ((char*)p);
/* we just bump tlab_temp_end as well */
TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
}
CANARIFY_ALLOC(p,real_size);
}
if (G_LIKELY (p)) {
SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, sgen_client_vtable_get_name (vtable), size);
sgen_binary_protocol_alloc (p, vtable, size, sgen_client_get_provenance ());
mono_atomic_store_seq (p, vtable);
}
return (GCObject*)p;
}
static gsize
sampling_thread_func (gpointer unused)
{
MonoInternalThread *thread = mono_thread_internal_current ();
thread->flags |= MONO_THREAD_FLAG_DONT_MANAGE;
ERROR_DECL (error);
MonoString *name = mono_string_new_checked (mono_get_root_domain (), "Profiler Sampler", error);
mono_error_assert_ok (error);
mono_thread_set_name_internal (thread, name, FALSE, FALSE, error);
mono_error_assert_ok (error);
mono_thread_info_set_flags (MONO_THREAD_INFO_FLAGS_NO_GC | MONO_THREAD_INFO_FLAGS_NO_SAMPLE);
int old_policy;
struct sched_param old_sched;
pthread_getschedparam (pthread_self (), &old_policy, &old_sched);
/*
* Attempt to switch the thread to real time scheduling. This will not
* necessarily work on all OSs; for example, most Linux systems will give
* us EPERM here unless configured to allow this.
*
* TODO: This does not work on Mac (and maybe some other OSs). On Mac, we
* have to use the Mach thread policy routines to switch to real-time
* scheduling. This is quite tricky as we need to specify how often we'll
* be doing work (easy), the normal processing time needed (also easy),
* and the maximum amount of processing time needed (hard). This is
* further complicated by the fact that if we misbehave and take too long
* to do our work, the kernel may knock us back down to the normal thread
* scheduling policy without telling us.
*/
struct sched_param sched = { .sched_priority = sched_get_priority_max (SCHED_FIFO) };
pthread_setschedparam (pthread_self (), SCHED_FIFO, &sched);
MonoProfilerSampleMode mode;
init:
mono_profiler_get_sample_mode (NULL, &mode, NULL);
if (mode == MONO_PROFILER_SAMPLE_MODE_NONE) {
mono_profiler_sampling_thread_wait ();
if (!mono_atomic_load_i32 (&sampling_thread_running))
goto done;
goto init;
}
clock_init (mode);
for (guint64 sleep = clock_get_time_ns (); mono_atomic_load_i32 (&sampling_thread_running); clock_sleep_ns_abs (sleep)) {
uint32_t freq;
MonoProfilerSampleMode new_mode;
mono_profiler_get_sample_mode (NULL, &new_mode, &freq);
if (new_mode != mode) {
clock_cleanup ();
goto init;
}
sleep += 1000000000 / freq;
FOREACH_THREAD_SAFE_EXCLUDE (info, MONO_THREAD_INFO_FLAGS_NO_SAMPLE) {
g_assert (mono_thread_info_get_tid (info) != sampling_thread);
/*
* Require an ack for the last sampling signal sent to the thread
* so that we don't overflow the signal queue, leading to all sorts
* of problems (e.g. GC STW failing).
*/
if (profiler_signal != SIGPROF && !mono_atomic_cas_i32 (&info->profiler_signal_ack, 0, 1))
continue;
mono_threads_pthread_kill (info, profiler_signal);
mono_atomic_inc_i32 (&profiler_signals_sent);
} FOREACH_THREAD_SAFE_END
}
