[profiler] Do the signal ack after checking that the thread is detached.
-Subproject commit 58bf3183031479747adcceceb622aea3aa076e38
+Subproject commit dab16db1ba42b153098d4e3fe28f2a00861b6863
void
mono_gc_base_cleanup (void)
{
- sgen_thread_pool_shutdown (major_collector.get_sweep_pool ());
-
- sgen_workers_shutdown ();
+ sgen_thread_pool_shutdown ();
// We should have consumed any outstanding moves.
g_assert (sgen_pointer_queue_is_empty (&moved_objects_queue));
We can't suspend the workers that will do all the heavy lifting.
FIXME Use some state bit in SgenThreadInfo for this.
*/
- if (sgen_thread_pool_is_thread_pool_thread (major_collector.get_sweep_pool (), mono_thread_info_get_tid (info)) ||
- sgen_workers_is_worker_thread (mono_thread_info_get_tid (info))) {
+ if (sgen_thread_pool_is_thread_pool_thread (mono_thread_info_get_tid (info))) {
if (reason)
*reason = 4;
return FALSE;
acfg->tmpfname = g_strdup_printf ("%s.s", acfg->image->name);
acfg->fp = fopen (acfg->tmpfname, "w+");
} else {
- int i = g_file_open_tmp ("mono_aot_XXXXXX", &acfg->tmpfname, NULL);
- acfg->fp = fdopen (i, "w+");
+ if (strcmp (acfg->aot_opts.temp_path, "") == 0) {
+ int i = g_file_open_tmp ("mono_aot_XXXXXX", &acfg->tmpfname, NULL);
+ acfg->fp = fdopen (i, "w+");
+ } else {
+ acfg->tmpbasename = g_build_filename (acfg->aot_opts.temp_path, "temp", NULL);
+ acfg->tmpfname = g_strdup_printf ("%s.s", acfg->tmpbasename);
+ acfg->fp = fopen (acfg->tmpfname, "w+");
+ }
}
if (acfg->fp == 0 && !acfg->aot_opts.llvm_only) {
aot_printerrf (acfg, "Unable to open file '%s': %s\n", acfg->tmpfname, strerror (errno));
globals = (void **)info->globals;
g_assert (globals);
}
+ found_aot_name = g_strdup (aot_name);
} else {
char *err;
static int i8_align;
static gpointer single_step_tramp, breakpoint_tramp;
-static gpointer get_tls_tramp;
/*
* The code generated for sequence points reads from this location, which is
return code;
}
+static guint8*
+emit_ldr_imm (guint8 *code, int dreg, int sreg, int imm)
+{
+ if (!arm_is_imm12 (imm)) {
+ g_assert (dreg != sreg);
+ code = emit_big_add (code, dreg, sreg, imm);
+ ARM_LDR_IMM (code, dreg, dreg, 0);
+ } else {
+ ARM_LDR_IMM (code, dreg, sreg, imm);
+ }
+ return code;
+}
+
/* If dreg == sreg, this clobbers IP */
static guint8*
emit_sub_imm (guint8 *code, int dreg, int sreg, int imm)
void
mono_arch_init (void)
{
- const char *cpu_arch;
+ char *cpu_arch;
#ifdef TARGET_WATCHOS
mini_get_debug_options ()->soft_breakpoints = TRUE;
* works. Most ARM devices have VFP units these days, so
* normally soft float code would not be exercised much.
*/
- const char *soft = g_getenv ("MONO_ARM_FORCE_SOFT_FLOAT");
+ char *soft = g_getenv ("MONO_ARM_FORCE_SOFT_FLOAT");
if (soft && !strncmp (soft, "1", 1))
arm_fpu = MONO_ARM_FPU_NONE;
if (cfg->compile_aot) {
g_assert (info_var);
g_assert (info_var->opcode == OP_REGOFFSET);
- g_assert (arm_is_imm12 (info_var->inst_offset));
}
if (!cfg->soft_breakpoints && !cfg->compile_aot) {
var = ss_method_var;
g_assert (var);
g_assert (var->opcode == OP_REGOFFSET);
- g_assert (arm_is_imm12 (var->inst_offset));
- ARM_LDR_IMM (code, dreg, var->inst_basereg, var->inst_offset);
-
+ code = emit_ldr_imm (code, dreg, var->inst_basereg, var->inst_offset);
/* Read the value and check whether it is non-zero. */
ARM_LDR_IMM (code, dreg, dreg, 0);
ARM_CMP_REG_IMM (code, dreg, 0, 0);
var = ss_trigger_page_var;
g_assert (var);
g_assert (var->opcode == OP_REGOFFSET);
- g_assert (arm_is_imm12 (var->inst_offset));
- ARM_LDR_IMM (code, dreg, var->inst_basereg, var->inst_offset);
+ code = emit_ldr_imm (code, dreg, var->inst_basereg, var->inst_offset);
} else {
ARM_LDR_IMM (code, dreg, ARMREG_PC, 0);
ARM_B (code, 0);
guint32 offset = code - cfg->native_code;
guint32 val;
- ARM_LDR_IMM (code, dreg, info_var->inst_basereg, info_var->inst_offset);
+ var = info_var;
+ code = emit_ldr_imm (code, dreg, var->inst_basereg, var->inst_offset);
/* Add the offset */
val = ((offset / 4) * sizeof (guint8*)) + MONO_STRUCT_OFFSET (SeqPointInfo, bp_addrs);
/* Load the info->bp_addrs [offset], which is either 0 or the address of a trigger page */
if (info_var) {
g_assert (info_var->opcode == OP_REGOFFSET);
- g_assert (arm_is_imm12 (info_var->inst_offset));
- ARM_LDR_IMM (code, dreg, info_var->inst_basereg, info_var->inst_offset);
+ code = emit_ldr_imm (code, dreg, info_var->inst_basereg, info_var->inst_offset);
/* Load the trigger page addr */
ARM_LDR_IMM (code, dreg, dreg, MONO_STRUCT_OFFSET (SeqPointInfo, ss_trigger_page));
ARM_STR_IMM (code, dreg, ss_trigger_page_var->inst_basereg, ss_trigger_page_var->inst_offset);
register_icall (mono_gsharedvt_constrained_call, "mono_gsharedvt_constrained_call", "object ptr ptr ptr ptr ptr", FALSE);
register_icall (mono_gsharedvt_value_copy, "mono_gsharedvt_value_copy", "void ptr ptr ptr", TRUE);
- register_icall_no_wrapper (mono_gc_get_range_copy_func (), "mono_gc_range_copy", "void ptr ptr int");
+ //WARNING We do runtime selection here but the string *MUST* be to a fallback function that has same signature and behavior
+ register_icall_no_wrapper (mono_gc_get_range_copy_func (), "mono_gc_wbarrier_range_copy", "void ptr ptr int");
register_icall (mono_object_castclass_with_cache, "mono_object_castclass_with_cache", "object object ptr ptr", FALSE);
register_icall (mono_object_isinst_with_cache, "mono_object_isinst_with_cache", "object object ptr ptr", FALSE);
#define REDZONE_SIZE 224
-/* MS_BLOCK_SIZE must be a multiple of the system pagesize, which for some
- architectures is 64k. */
-#if defined(TARGET_POWERPC) || defined(TARGET_POWERPC64)
-#define ARCH_MIN_MS_BLOCK_SIZE (64*1024)
-#define ARCH_MIN_MS_BLOCK_SIZE_SHIFT 16
-#endif
-
#elif defined(TARGET_ARM)
#define REDZONE_SIZE 0
static void
gray_queue_redirect (SgenGrayQueue *queue)
{
- sgen_workers_take_from_queue (queue);
+ sgen_workers_take_from_queue (current_collection_generation, queue);
}
void
typedef struct {
ScanJob scan_job;
- int job_index;
+ int job_index, job_split_count;
} ParallelScanJob;
static ScanCopyContext
* object ops changes, like forced concurrent finish.
*/
SGEN_ASSERT (0, sgen_workers_is_worker_thread (mono_native_thread_id_get ()), "We need a context for the scan job");
- job->ops = sgen_workers_get_idle_func_object_ops ();
+ job->ops = sgen_workers_get_idle_func_object_ops (worker_data);
}
return CONTEXT_FROM_OBJECT_OPERATIONS (job->ops, sgen_workers_get_job_gray_queue (worker_data, job->gc_thread_gray_queue));
ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
SGEN_TV_GETTIME (atv);
- major_collector.scan_card_table (CARDTABLE_SCAN_GLOBAL, ctx, job_data->job_index, sgen_workers_get_job_split_count ());
+ major_collector.scan_card_table (CARDTABLE_SCAN_GLOBAL, ctx, job_data->job_index, job_data->job_split_count);
SGEN_TV_GETTIME (btv);
time_minor_scan_major_blocks += SGEN_TV_ELAPSED (atv, btv);
}
ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
SGEN_TV_GETTIME (atv);
- sgen_los_scan_card_table (CARDTABLE_SCAN_GLOBAL, ctx, job_data->job_index, sgen_workers_get_job_split_count ());
+ sgen_los_scan_card_table (CARDTABLE_SCAN_GLOBAL, ctx, job_data->job_index, job_data->job_split_count);
SGEN_TV_GETTIME (btv);
time_minor_scan_los += SGEN_TV_ELAPSED (atv, btv);
}
ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
g_assert (concurrent_collection_in_progress);
- major_collector.scan_card_table (CARDTABLE_SCAN_MOD_UNION, ctx, job_data->job_index, sgen_workers_get_job_split_count ());
+ major_collector.scan_card_table (CARDTABLE_SCAN_MOD_UNION, ctx, job_data->job_index, job_data->job_split_count);
}
static void
ScanCopyContext ctx = scan_copy_context_for_scan_job (worker_data_untyped, (ScanJob*)job_data);
g_assert (concurrent_collection_in_progress);
- sgen_los_scan_card_table (CARDTABLE_SCAN_MOD_UNION, ctx, job_data->job_index, sgen_workers_get_job_split_count ());
+ sgen_los_scan_card_table (CARDTABLE_SCAN_MOD_UNION, ctx, job_data->job_index, job_data->job_split_count);
}
static void
g_assert (concurrent_collection_in_progress);
- major_collector.scan_card_table (CARDTABLE_SCAN_MOD_UNION_PRECLEAN, ctx, job_data->job_index, sgen_workers_get_job_split_count ());
+ major_collector.scan_card_table (CARDTABLE_SCAN_MOD_UNION_PRECLEAN, ctx, job_data->job_index, job_data->job_split_count);
}
static void
g_assert (concurrent_collection_in_progress);
- sgen_los_scan_card_table (CARDTABLE_SCAN_MOD_UNION_PRECLEAN, ctx, job_data->job_index, sgen_workers_get_job_split_count ());
+ sgen_los_scan_card_table (CARDTABLE_SCAN_MOD_UNION_PRECLEAN, ctx, job_data->job_index, job_data->job_split_count);
}
static void
{
ParallelScanJob *psj;
ScanJob *sj;
- int split_count = sgen_workers_get_job_split_count ();
+ int split_count = sgen_workers_get_job_split_count (GENERATION_OLD);
int i;
/* Mod union preclean jobs */
for (i = 0; i < split_count; i++) {
psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("preclean major mod union cardtable", job_major_mod_union_preclean, sizeof (ParallelScanJob));
psj->scan_job.gc_thread_gray_queue = NULL;
psj->job_index = i;
- sgen_workers_enqueue_job (&psj->scan_job.job, TRUE);
+ psj->job_split_count = split_count;
+ sgen_workers_enqueue_job (GENERATION_OLD, &psj->scan_job.job, TRUE);
}
for (i = 0; i < split_count; i++) {
psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("preclean los mod union cardtable", job_los_mod_union_preclean, sizeof (ParallelScanJob));
psj->scan_job.gc_thread_gray_queue = NULL;
psj->job_index = i;
- sgen_workers_enqueue_job (&psj->scan_job.job, TRUE);
+ psj->job_split_count = split_count;
+ sgen_workers_enqueue_job (GENERATION_OLD, &psj->scan_job.job, TRUE);
}
sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan last pinned", job_scan_last_pinned, sizeof (ScanJob));
sj->gc_thread_gray_queue = NULL;
- sgen_workers_enqueue_job (&sj->job, TRUE);
+ sgen_workers_enqueue_job (GENERATION_OLD, &sj->job, TRUE);
}
static void
-init_gray_queue (SgenGrayQueue *gc_thread_gray_queue, gboolean use_workers)
+init_gray_queue (SgenGrayQueue *gc_thread_gray_queue)
{
- if (use_workers)
- sgen_workers_init_distribute_gray_queue ();
sgen_gray_object_queue_init (gc_thread_gray_queue, NULL, TRUE);
}
static void
enqueue_scan_remembered_set_jobs (SgenGrayQueue *gc_thread_gray_queue, SgenObjectOperations *ops, gboolean enqueue)
{
- int i, split_count = sgen_workers_get_job_split_count ();
+ int i, split_count = sgen_workers_get_job_split_count (GENERATION_NURSERY);
ScanJob *sj;
sj = (ScanJob*)sgen_thread_pool_job_alloc ("scan wbroots", job_scan_wbroots, sizeof (ScanJob));
sj->ops = ops;
sj->gc_thread_gray_queue = gc_thread_gray_queue;
- sgen_workers_enqueue_job (&sj->job, enqueue);
+ sgen_workers_enqueue_job (GENERATION_NURSERY, &sj->job, enqueue);
for (i = 0; i < split_count; i++) {
ParallelScanJob *psj;
psj->scan_job.ops = ops;
psj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
psj->job_index = i;
- sgen_workers_enqueue_job (&psj->scan_job.job, enqueue);
+ psj->job_split_count = split_count;
+ sgen_workers_enqueue_job (GENERATION_NURSERY, &psj->scan_job.job, enqueue);
psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("scan LOS remsets", job_scan_los_card_table, sizeof (ParallelScanJob));
psj->scan_job.ops = ops;
psj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
psj->job_index = i;
- sgen_workers_enqueue_job (&psj->scan_job.job, enqueue);
+ psj->job_split_count = split_count;
+ sgen_workers_enqueue_job (GENERATION_NURSERY, &psj->scan_job.job, enqueue);
}
}
scrrj->heap_start = heap_start;
scrrj->heap_end = heap_end;
scrrj->root_type = ROOT_TYPE_NORMAL;
- sgen_workers_enqueue_job (&scrrj->scan_job.job, enqueue);
+ sgen_workers_enqueue_job (current_collection_generation, &scrrj->scan_job.job, enqueue);
if (current_collection_generation == GENERATION_OLD) {
/* During minors we scan the cardtable for these roots instead */
scrrj->heap_start = heap_start;
scrrj->heap_end = heap_end;
scrrj->root_type = ROOT_TYPE_WBARRIER;
- sgen_workers_enqueue_job (&scrrj->scan_job.job, enqueue);
+ sgen_workers_enqueue_job (current_collection_generation, &scrrj->scan_job.job, enqueue);
}
/* Threads */
stdj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
stdj->heap_start = heap_start;
stdj->heap_end = heap_end;
- sgen_workers_enqueue_job (&stdj->scan_job.job, enqueue);
+ sgen_workers_enqueue_job (current_collection_generation, &stdj->scan_job.job, enqueue);
/* Scan the list of objects ready for finalization. */
sfej->scan_job.ops = ops;
sfej->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
sfej->queue = &fin_ready_queue;
- sgen_workers_enqueue_job (&sfej->scan_job.job, enqueue);
+ sgen_workers_enqueue_job (current_collection_generation, &sfej->scan_job.job, enqueue);
sfej = (ScanFinalizerEntriesJob*)sgen_thread_pool_job_alloc ("scan critical finalizer entries", job_scan_finalizer_entries, sizeof (ScanFinalizerEntriesJob));
sfej->scan_job.ops = ops;
sfej->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
sfej->queue = &critical_fin_queue;
- sgen_workers_enqueue_job (&sfej->scan_job.job, enqueue);
+ sgen_workers_enqueue_job (current_collection_generation, &sfej->scan_job.job, enqueue);
}
/*
binary_protocol_collection_begin (gc_stats.minor_gc_count, GENERATION_NURSERY);
- if (sgen_concurrent_collection_in_progress ()) {
- /* FIXME Support parallel nursery collections with concurrent major */
- object_ops_nopar = &sgen_minor_collector.serial_ops_with_concurrent_major;
- } else {
- object_ops_nopar = &sgen_minor_collector.serial_ops;
- if (sgen_minor_collector.is_parallel && sgen_nursery_size >= SGEN_PARALLEL_MINOR_MIN_NURSERY_SIZE) {
- object_ops_par = &sgen_minor_collector.parallel_ops;
- is_parallel = TRUE;
- }
+ object_ops_nopar = sgen_concurrent_collection_in_progress ()
+ ? &sgen_minor_collector.serial_ops_with_concurrent_major
+ : &sgen_minor_collector.serial_ops;
+ if (sgen_minor_collector.is_parallel && sgen_nursery_size >= SGEN_PARALLEL_MINOR_MIN_NURSERY_SIZE) {
+ object_ops_par = sgen_concurrent_collection_in_progress ()
+ ? &sgen_minor_collector.parallel_ops_with_concurrent_major
+ : &sgen_minor_collector.parallel_ops;
+ is_parallel = TRUE;
}
if (do_verify_nursery || do_dump_nursery_content)
sgen_memgov_minor_collection_start ();
- init_gray_queue (&gc_thread_gray_queue, is_parallel);
+ init_gray_queue (&gc_thread_gray_queue);
ctx = CONTEXT_FROM_OBJECT_OPERATIONS (object_ops_nopar, &gc_thread_gray_queue);
gc_stats.minor_gc_count ++;
/* pin from pinned handles */
sgen_init_pinning ();
+ if (concurrent_collection_in_progress)
+ sgen_init_pinning_for_conc ();
sgen_client_binary_protocol_mark_start (GENERATION_NURSERY);
pin_from_roots (nursery_section->data, nursery_section->end_data, ctx);
/* pin cemented objects */
pin_objects_in_nursery (FALSE, ctx);
sgen_pinning_trim_queue_to_section (nursery_section);
+ if (concurrent_collection_in_progress)
+ sgen_finish_pinning_for_conc ();
if (remset_consistency_checks)
sgen_check_remset_consistency ();
if (is_parallel) {
gray_queue_redirect (&gc_thread_gray_queue);
- sgen_workers_start_all_workers (object_ops_nopar, object_ops_par, NULL);
- sgen_workers_join ();
+ sgen_workers_start_all_workers (GENERATION_NURSERY, object_ops_nopar, object_ops_par, NULL);
+ sgen_workers_join (GENERATION_NURSERY);
}
TV_GETTIME (btv);
TV_GETTIME (atv);
sgen_init_pinning ();
+ if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT)
+ sgen_init_pinning_for_conc ();
SGEN_LOG (6, "Collecting pinned addresses");
pin_from_roots ((void*)lowest_heap_address, (void*)highest_heap_address, ctx);
if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
SGEN_LOG (2, "Finding pinned pointers: %zd in %lld usecs", sgen_get_pinned_count (), (long long)TV_ELAPSED (atv, btv));
SGEN_LOG (4, "Start scan with %zd pinned objects", sgen_get_pinned_count ());
+ if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT)
+ sgen_finish_pinning_for_conc ();
+
major_collector.init_to_space ();
SGEN_ASSERT (0, sgen_workers_all_done (), "Why are the workers not done when we start or finish a major collection?");
if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
if (object_ops_par != NULL)
- sgen_workers_set_num_active_workers (0);
- if (sgen_workers_have_idle_work ()) {
+ sgen_workers_set_num_active_workers (GENERATION_OLD, 0);
+ if (sgen_workers_have_idle_work (GENERATION_OLD)) {
/*
* We force the finish of the worker with the new object ops context
* which can also do copying. We need to have finished pinning.
*/
- sgen_workers_start_all_workers (object_ops_nopar, object_ops_par, NULL);
+ sgen_workers_start_all_workers (GENERATION_OLD, object_ops_nopar, object_ops_par, NULL);
- sgen_workers_join ();
+ sgen_workers_join (GENERATION_OLD);
}
}
* the roots.
*/
if (mode == COPY_OR_MARK_FROM_ROOTS_START_CONCURRENT) {
- sgen_workers_set_num_active_workers (1);
+ sgen_workers_set_num_active_workers (GENERATION_OLD, 1);
gray_queue_redirect (gc_thread_gray_queue);
if (precleaning_enabled) {
- sgen_workers_start_all_workers (object_ops_nopar, object_ops_par, workers_finish_callback);
+ sgen_workers_start_all_workers (GENERATION_OLD, object_ops_nopar, object_ops_par, workers_finish_callback);
} else {
- sgen_workers_start_all_workers (object_ops_nopar, object_ops_par, NULL);
+ sgen_workers_start_all_workers (GENERATION_OLD, object_ops_nopar, object_ops_par, NULL);
}
}
if (mode == COPY_OR_MARK_FROM_ROOTS_FINISH_CONCURRENT) {
- int i, split_count = sgen_workers_get_job_split_count ();
+ int i, split_count = sgen_workers_get_job_split_count (GENERATION_OLD);
gboolean parallel = object_ops_par != NULL;
/* If we're not parallel we finish the collection on the gc thread */
psj->scan_job.ops = parallel ? NULL : object_ops_nopar;
psj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
psj->job_index = i;
- sgen_workers_enqueue_job (&psj->scan_job.job, parallel);
+ psj->job_split_count = split_count;
+ sgen_workers_enqueue_job (GENERATION_OLD, &psj->scan_job.job, parallel);
psj = (ParallelScanJob*)sgen_thread_pool_job_alloc ("scan LOS mod union cardtable", job_scan_los_mod_union_card_table, sizeof (ParallelScanJob));
psj->scan_job.ops = parallel ? NULL : object_ops_nopar;
psj->scan_job.gc_thread_gray_queue = gc_thread_gray_queue;
psj->job_index = i;
- sgen_workers_enqueue_job (&psj->scan_job.job, parallel);
+ psj->job_split_count = split_count;
+ sgen_workers_enqueue_job (GENERATION_OLD, &psj->scan_job.job, parallel);
}
if (parallel) {
* stack that contained roots and pinned objects and also scan the mod union card
* table.
*/
- sgen_workers_start_all_workers (object_ops_nopar, object_ops_par, NULL);
- sgen_workers_join ();
+ sgen_workers_start_all_workers (GENERATION_OLD, object_ops_nopar, object_ops_par, NULL);
+ sgen_workers_join (GENERATION_OLD);
}
}
current_collection_generation = GENERATION_OLD;
- sgen_workers_assert_gray_queue_is_empty ();
+ sgen_workers_assert_gray_queue_is_empty (GENERATION_OLD);
if (!concurrent)
sgen_cement_reset ();
object_ops_nopar = &major_collector.major_ops_serial;
}
- sgen_workers_assert_gray_queue_is_empty ();
+ sgen_workers_assert_gray_queue_is_empty (GENERATION_OLD);
finish_gray_stack (GENERATION_OLD, CONTEXT_FROM_OBJECT_OPERATIONS (object_ops_nopar, gc_thread_gray_queue));
TV_GETTIME (atv);
memset (&counts, 0, sizeof (ScannedObjectCounts));
major_collector.finish_major_collection (&counts);
- sgen_workers_assert_gray_queue_is_empty ();
+ sgen_workers_assert_gray_queue_is_empty (GENERATION_OLD);
SGEN_ASSERT (0, sgen_workers_all_done (), "Can't have workers working after major collection has finished");
if (concurrent_collection_in_progress)
/* world must be stopped already */
TV_GETTIME (time_start);
- init_gray_queue (&gc_thread_gray_queue, FALSE);
+ init_gray_queue (&gc_thread_gray_queue);
major_start_collection (&gc_thread_gray_queue, reason, FALSE, &old_next_pin_slot);
major_finish_collection (&gc_thread_gray_queue, reason, is_overflow, old_next_pin_slot, forced);
sgen_gray_object_queue_dispose (&gc_thread_gray_queue);
binary_protocol_concurrent_start ();
- init_gray_queue (&gc_thread_gray_queue, TRUE);
+ init_gray_queue (&gc_thread_gray_queue);
// FIXME: store reason and pass it when finishing
major_start_collection (&gc_thread_gray_queue, reason, TRUE, NULL);
sgen_gray_object_queue_dispose (&gc_thread_gray_queue);
* The workers will be resumed with a finishing pause context to avoid
* additional cardtable and object scanning.
*/
- sgen_workers_stop_all_workers ();
+ sgen_workers_stop_all_workers (GENERATION_OLD);
SGEN_TV_GETTIME (time_major_conc_collection_end);
gc_stats.major_gc_time_concurrent += SGEN_TV_ELAPSED (time_major_conc_collection_start, time_major_conc_collection_end);
current_collection_generation = GENERATION_OLD;
sgen_cement_reset ();
- init_gray_queue (&gc_thread_gray_queue, FALSE);
+ init_gray_queue (&gc_thread_gray_queue);
major_finish_collection (&gc_thread_gray_queue, "finishing", FALSE, -1, forced);
sgen_gray_object_queue_dispose (&gc_thread_gray_queue);
if (major_collector.post_param_init)
major_collector.post_param_init (&major_collector);
- if (major_collector.is_concurrent || sgen_minor_collector.is_parallel) {
- int num_workers = 1;
- if (major_collector.is_parallel || sgen_minor_collector.is_parallel) {
- num_workers = mono_cpu_count ();
- if (num_workers <= 1) {
- num_workers = 1;
- major_collector.is_parallel = FALSE;
- sgen_minor_collector.is_parallel = FALSE;
- }
- }
- if (major_collector.is_concurrent || sgen_minor_collector.is_parallel)
- sgen_workers_init (num_workers, (SgenWorkerCallback) major_collector.worker_init_cb);
- }
+ sgen_thread_pool_start ();
sgen_memgov_init (max_heap, soft_limit, debug_print_allowance, allowance_ratio, save_target);
SgenObjectOperations serial_ops;
SgenObjectOperations serial_ops_with_concurrent_major;
SgenObjectOperations parallel_ops;
+ SgenObjectOperations parallel_ops_with_concurrent_major;
void (*prepare_to_space) (char *to_space_bitmap, size_t space_bitmap_size);
void (*clear_fragments) (void);
guint8* (*get_cardtable_mod_union_for_reference) (char *object);
long long (*get_and_reset_num_major_objects_marked) (void);
void (*count_cards) (long long *num_total_cards, long long *num_marked_cards);
- SgenThreadPool* (*get_sweep_pool) (void);
-
- void (*worker_init_cb) (gpointer worker);
+ void (*init_block_free_lists) (gpointer *list_p);
};
extern SgenMajorCollector major_collector;
#include "mono/sgen/sgen-thread-pool.h"
#include "mono/sgen/sgen-client.h"
#include "mono/utils/mono-memory-model.h"
+#include "mono/utils/mono-proclib.h"
-#if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
-#define MS_BLOCK_SIZE ARCH_MIN_MS_BLOCK_SIZE
-#define MS_BLOCK_SIZE_SHIFT ARCH_MIN_MS_BLOCK_SIZE_SHIFT
-#else
-#define MS_BLOCK_SIZE_SHIFT 14 /* INT FASTENABLE */
-#define MS_BLOCK_SIZE (1 << MS_BLOCK_SIZE_SHIFT)
-#endif
-#define MAJOR_SECTION_SIZE MS_BLOCK_SIZE
-#define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
+static int ms_block_size;
+
+/*
+ * Blocks must be at least this size, meaning that if we detect a
+ * page size lower than this, we'll use this instead.
+ */
+#define MS_BLOCK_SIZE_MIN (1024 * 16)
+
+#define CARDS_PER_BLOCK (ms_block_size / CARD_SIZE_IN_BYTES)
/*
* Don't allocate single blocks, but alloc a contingent of this many
*/
#define MS_BLOCK_ALLOC_NUM 32
+#define MS_NUM_MARK_WORDS ((ms_block_size / SGEN_ALLOC_ALIGN + sizeof (guint32) * 8 - 1) / (sizeof (guint32) * 8))
+
+/*
+ * Use this instead of sizeof (MSBlockInfo) since the mark_words
+ * array size depends on page size at runtime.
+ */
+#define SIZEOF_MS_BLOCK_INFO (sizeof (MSBlockInfo) + sizeof (guint32) * (MS_NUM_MARK_WORDS - MONO_ZERO_LEN_ARRAY))
+
/*
* Number of bytes before the first object in a block. At the start
* of a block is the MSBlockHeader, then opional padding, then come
- * the objects, so this must be >= sizeof (MSBlockHeader).
+ * the objects, so this must be >= SIZEOF_MS_BLOCK_INFO.
*/
-#define MS_BLOCK_SKIP ((sizeof (MSBlockHeader) + 15) & ~15)
-
-#define MS_BLOCK_FREE (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
+#define MS_BLOCK_SKIP ((SIZEOF_MS_BLOCK_INFO + 15) & ~15)
-#define MS_NUM_MARK_WORDS (MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (guint32) * 8 - 1) / (sizeof (guint32) * 8)
+#define MS_BLOCK_FREE (ms_block_size - MS_BLOCK_SKIP)
/*
* Blocks progress from one state to the next:
void ** volatile free_list;
MSBlockInfo * volatile next_free;
guint8 * volatile cardtable_mod_union;
- guint32 mark_words [MS_NUM_MARK_WORDS];
+ guint32 mark_words [MONO_ZERO_LEN_ARRAY];
};
#define MS_BLOCK_FOR_BLOCK_INFO(b) ((char*)(b))
#define MS_BLOCK_OBJ(b,i) ((GCObject *)(MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i)))
#define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size) (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
-#define MS_BLOCK_DATA_FOR_OBJ(o) ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
+#define MS_BLOCK_DATA_FOR_OBJ(o) ((char*)((mword)(o) & ~(mword)(ms_block_size - 1)))
typedef struct {
MSBlockInfo info;
} while (0)
-#define MS_OBJ_ALLOCED(o,b) (*(void**)(o) && (*(char**)(o) < MS_BLOCK_FOR_BLOCK_INFO (b) || *(char**)(o) >= MS_BLOCK_FOR_BLOCK_INFO (b) + MS_BLOCK_SIZE))
+#define MS_OBJ_ALLOCED(o,b) (*(void**)(o) && (*(char**)(o) < MS_BLOCK_FOR_BLOCK_INFO (b) || *(char**)(o) >= MS_BLOCK_FOR_BLOCK_INFO (b) + ms_block_size))
#define MS_BLOCK_OBJ_SIZE_FACTOR (pow (2.0, 1.0 / 3))
static gboolean concurrent_mark;
static gboolean concurrent_sweep = TRUE;
-SgenThreadPool sweep_pool_inst;
-SgenThreadPool *sweep_pool;
+int sweep_pool_context = -1;
#define BLOCK_IS_TAGGED_HAS_REFERENCES(bl) SGEN_POINTER_IS_TAGGED_1 ((bl))
#define BLOCK_TAG_HAS_REFERENCES(bl) SGEN_POINTER_TAG_1 ((bl))
static void
update_heap_boundaries_for_block (MSBlockInfo *block)
{
- sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
+ sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size);
}
/*
*/
int alloc_num = MS_BLOCK_ALLOC_NUM;
for (;;) {
- p = (char *)sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE,
+ p = (char *)sgen_alloc_os_memory_aligned (ms_block_size * alloc_num, ms_block_size,
(SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE),
alloc_num == 1 ? "major heap section" : NULL, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
if (p)
empty = empty_blocks;
*(void**)block = empty;
} while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
- p += MS_BLOCK_SIZE;
+ p += ms_block_size;
}
SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
*(void**)block = NULL;
- g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
+ g_assert (!((mword)block & (ms_block_size - 1)));
return block;
}
void *empty;
char *block = MS_BLOCK_FOR_BLOCK_INFO (info);
- sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
+ sgen_memgov_release_space (ms_block_size, SPACE_MAJOR);
if (info->cardtable_mod_union)
- sgen_card_table_free_mod_union (info->cardtable_mod_union, block, MS_BLOCK_SIZE);
- memset (block, 0, MS_BLOCK_SIZE);
+ sgen_card_table_free_mod_union (info->cardtable_mod_union, block, ms_block_size);
+ memset (block, 0, ms_block_size);
do {
empty = empty_blocks;
SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
- binary_protocol_block_free (block, MS_BLOCK_SIZE);
+ binary_protocol_block_free (block, ms_block_size);
}
static gboolean
char *obj_start;
int i;
- if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
+ if (!sgen_memgov_try_alloc_space (ms_block_size, SPACE_MAJOR))
return FALSE;
info = (MSBlockInfo*)ms_get_empty_block ();
update_heap_boundaries_for_block (info);
- binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
+ binary_protocol_block_alloc (info, ms_block_size);
/* build free list */
obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
MSBlockInfo *block;
FOREACH_BLOCK_NO_LOCK (block) {
- if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
+ if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size) {
int count = MS_BLOCK_FREE / block->obj_size;
int i;
wait:
job = sweep_job;
if (job)
- sgen_thread_pool_job_wait (sweep_pool, job);
+ sgen_thread_pool_job_wait (sweep_pool_context, job);
SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
}
int idx;
char *obj;
- if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
+ if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size) <= object))
continue;
idx = MS_BLOCK_OBJ_INDEX (object, block);
int w, b;
gboolean marked;
- if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
+ if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size) <= ptr))
continue;
SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
return mod_union;
else if (!allocate)
return NULL;
- mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
+ mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size);
other = (guint8 *)SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
if (!other) {
SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
return mod_union;
}
- sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
+ sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size);
return other;
}
sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
{
int obj_index;
+ void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, 0, obj_size);
- for (obj_index = 0; obj_index < count; ++obj_index) {
+ for (obj_index = 0; obj_index < count; ++obj_index, obj = (void*)((mword)obj + obj_size)) {
int word, bit;
- void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
MS_CALC_MARK_BIT (word, bit, obj);
if (MS_MARK_BIT (block, word, bit)) {
gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
gboolean success = old_state == expected_state;
if (success)
- binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
+ binary_protocol_block_set_state (block, ms_block_size, old_state, new_state);
return success;
}
{
SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set");
block->state = new_state;
- binary_protocol_block_set_state (block, MS_BLOCK_SIZE, expected_state, new_state);
+ binary_protocol_block_set_state (block, ms_block_size, expected_state, new_state);
}
/*
free_blocks [j] = NULL;
}
- sgen_workers_foreach (sgen_worker_clear_free_block_lists);
+ sgen_workers_foreach (GENERATION_NURSERY, sgen_worker_clear_free_block_lists);
+ sgen_workers_foreach (GENERATION_OLD, sgen_worker_clear_free_block_lists);
}
static void sweep_finish (void);
*/
if (concurrent_sweep && lazy_sweep) {
sweep_blocks_job = sgen_thread_pool_job_alloc ("sweep_blocks", sweep_blocks_job_func, sizeof (SgenThreadPoolJob));
- sgen_thread_pool_job_enqueue (sweep_pool, sweep_blocks_job);
+ sgen_thread_pool_job_enqueue (sweep_pool_context, sweep_blocks_job);
}
sweep_finish ();
SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
if (concurrent_sweep) {
sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
- sgen_thread_pool_job_enqueue (sweep_pool, sweep_job);
+ sgen_thread_pool_job_enqueue (sweep_pool_context, sweep_job);
} else {
sweep_job_func (NULL, NULL);
}
}
/* We expect workers to have very few blocks on the freelist, just evacuate them */
- sgen_workers_foreach (sgen_worker_clear_free_block_lists_evac);
+ sgen_workers_foreach (GENERATION_NURSERY, sgen_worker_clear_free_block_lists_evac);
+ sgen_workers_foreach (GENERATION_OLD, sgen_worker_clear_free_block_lists_evac);
if (lazy_sweep && concurrent_sweep) {
/*
*/
SgenThreadPoolJob *job = sweep_blocks_job;
if (job)
- sgen_thread_pool_job_wait (sweep_pool, job);
+ sgen_thread_pool_job_wait (sweep_pool_context, job);
}
if (lazy_sweep && !concurrent_sweep)
#endif
}
-static SgenThreadPool*
-major_get_sweep_pool (void)
-{
- return sweep_pool;
-}
-
static int
compare_pointers (const void *va, const void *vb) {
char *a = *(char**)va, *b = *(char**)vb;
SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
- if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
+ if ((char*)block != ((char*)empty_block_arr [d-1]) + ms_block_size) {
first = d;
continue;
}
* we're iterating.
*/
int j;
- sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
+ sgen_free_os_memory (empty_block_arr [first], ms_block_size * num_blocks, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
for (j = first; j <= d; ++j)
empty_block_arr [j] = NULL;
dest = first;
while (num_empty_blocks > section_reserve) {
void *next = *(void**)empty_blocks;
- sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
+ sgen_free_os_memory (empty_blocks, ms_block_size, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
empty_blocks = next;
/*
* Needs not be atomic because this is running
FOREACH_BLOCK_NO_LOCK (block) {
size_t first_entry, last_entry;
SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
- sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
+ sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size,
&first_entry, &last_entry);
mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
} END_FOREACH_BLOCK_NO_LOCK;
get_bytes_survived_last_sweep (void)
{
SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
- return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
+ return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * ms_block_size;
}
static gboolean
FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
if (has_references)
- callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
+ callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size);
} END_FOREACH_BLOCK_NO_LOCK;
}
major_finish_sweep_checking ();
FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
if (has_references)
- callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
+ callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size);
} END_FOREACH_BLOCK_NO_LOCK;
}
initial_skip_card (guint8 *card_data)
{
mword *cards = (mword*)card_data;
- mword card;
+ mword card = 0;
int i;
for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
card = cards [i];
#define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
#define MS_BLOCK_OBJ_FAST(b,os,i) ((b) + MS_BLOCK_SKIP + (os) * (i))
-#define MS_OBJ_ALLOCED_FAST(o,b) (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
+#define MS_OBJ_ALLOCED_FAST(o,b) (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + ms_block_size))
static void
scan_card_table_for_block (MSBlockInfo *block, CardTableScanType scan_type, ScanCopyContext ctx)
{
SgenGrayQueue *queue = ctx.queue;
ScanObjectFunc scan_func = ctx.ops->scan_object;
+ /*
+ * FIXME: On systems with very large pages, we allocate fairly large
+ * arrays on the stack here. This shouldn't be a problem once block
+ * size is no longer required to be a multiple of the system page size.
+ */
#ifndef SGEN_HAVE_OVERLAPPING_CARDS
- guint8 cards_copy [CARDS_PER_BLOCK];
+ guint8 *cards_copy = alloca (sizeof (guint8) * CARDS_PER_BLOCK);
#endif
- guint8 cards_preclean [CARDS_PER_BLOCK];
+ guint8 *cards_preclean = alloca (sizeof (guint8) * CARDS_PER_BLOCK);
gboolean small_objects;
int block_obj_size;
char *block_start;
if (has_dirty_cards) {
size_t num_cards;
guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
- sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
+ sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size, &num_cards);
SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
}
} END_FOREACH_BLOCK_NO_LOCK;
collector->sweeps_lazily = lazy_sweep;
}
-/* We are guaranteed to be called by the worker in question */
+/*
+ * We are guaranteed to be called by the worker in question.
+ * This provides initialization for threads that plan to do
+ * parallel object allocation. We need to store these lists
+ * in additional data structures so we can traverse them
+ * at major/sweep start.
+ */
static void
-sgen_worker_init_callback (gpointer worker_untyped)
+sgen_init_block_free_lists (gpointer *list_p)
{
int i;
- WorkerData *worker = (WorkerData*) worker_untyped;
- MSBlockInfo ***worker_free_blocks = (MSBlockInfo ***) sgen_alloc_internal_dynamic (sizeof (MSBlockInfo**) * MS_BLOCK_TYPE_MAX, INTERNAL_MEM_MS_TABLES, TRUE);
+ MSBlockInfo ***worker_free_blocks = (MSBlockInfo ***) mono_native_tls_get_value (worker_block_free_list_key);
+
+ /*
+ * For simplification, a worker thread uses the same free block lists,
+ * regardless of the context it is part of (major/minor).
+ */
+ if (worker_free_blocks) {
+ *list_p = (gpointer)worker_free_blocks;
+ return;
+ }
+
+ worker_free_blocks = (MSBlockInfo ***) sgen_alloc_internal_dynamic (sizeof (MSBlockInfo**) * MS_BLOCK_TYPE_MAX, INTERNAL_MEM_MS_TABLES, TRUE);
for (i = 0; i < MS_BLOCK_TYPE_MAX; i++)
worker_free_blocks [i] = (MSBlockInfo **) sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
- worker->free_block_lists = worker_free_blocks;
+ *list_p = (gpointer)worker_free_blocks;
mono_native_tls_set_value (worker_block_free_list_key, worker_free_blocks);
}
-static void
-thread_pool_init_func (void *data_untyped)
-{
- sgen_client_thread_register_worker ();
-}
-
static void
sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent, gboolean is_parallel)
{
int i;
- sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
+ ms_block_size = mono_pagesize ();
+
+ if (ms_block_size < MS_BLOCK_SIZE_MIN)
+ ms_block_size = MS_BLOCK_SIZE_MIN;
+
+ sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, SIZEOF_MS_BLOCK_INFO);
+
+ if (mono_cpu_count () <= 1)
+ is_parallel = FALSE;
num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
block_obj_sizes = (int *)sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
/* We can do this because we always init the minor before the major */
- if (is_parallel || sgen_get_minor_collector ()->is_parallel) {
+ if (is_parallel || sgen_get_minor_collector ()->is_parallel)
mono_native_tls_alloc (&worker_block_free_list_key, NULL);
- collector->worker_init_cb = sgen_worker_init_callback;
- }
mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
- collector->section_size = MAJOR_SECTION_SIZE;
+ collector->section_size = ms_block_size;
concurrent_mark = is_concurrent;
collector->is_concurrent = is_concurrent;
collector->is_valid_object = major_is_valid_object;
collector->describe_pointer = major_describe_pointer;
collector->count_cards = major_count_cards;
- collector->get_sweep_pool = major_get_sweep_pool;
+ collector->init_block_free_lists = sgen_init_block_free_lists;
collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
/*cardtable requires major pages to be 8 cards aligned*/
- g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
+ g_assert ((ms_block_size % (8 * CARD_SIZE_IN_BYTES)) == 0);
- if (concurrent_sweep) {
- SgenThreadPool **thread_datas = &sweep_pool;
- sweep_pool = &sweep_pool_inst;
- sgen_thread_pool_init (sweep_pool, 1, thread_pool_init_func, NULL, NULL, NULL, (SgenThreadPoolData**)&thread_datas);
- }
+ if (is_concurrent && is_parallel)
+ sgen_workers_create_context (GENERATION_OLD, mono_cpu_count ());
+ else if (is_concurrent)
+ sgen_workers_create_context (GENERATION_OLD, 1);
+
+ if (concurrent_sweep)
+ sweep_pool_context = sgen_thread_pool_create_context (1, NULL, NULL, NULL, NULL, NULL);
}
void
#if defined(SGEN_SIMPLE_NURSERY)
#ifdef SGEN_SIMPLE_PAR_NURSERY
-/* Not supported with concurrent major yet */
+
+#ifdef SGEN_CONCURRENT_MAJOR
+#define SERIAL_COPY_OBJECT simple_par_nursery_with_concurrent_major_copy_object
+#define SERIAL_COPY_OBJECT_FROM_OBJ simple_par_nursery_with_concurrent_major_copy_object_from_obj
+#else
#define SERIAL_COPY_OBJECT simple_par_nursery_copy_object
#define SERIAL_COPY_OBJECT_FROM_OBJ simple_par_nursery_copy_object_from_obj
+#endif
+
#else
+
#ifdef SGEN_CONCURRENT_MAJOR
#define SERIAL_COPY_OBJECT simple_nursery_serial_with_concurrent_major_copy_object
#define SERIAL_COPY_OBJECT_FROM_OBJ simple_nursery_serial_with_concurrent_major_copy_object_from_obj
#define SERIAL_COPY_OBJECT simple_nursery_serial_copy_object
#define SERIAL_COPY_OBJECT_FROM_OBJ simple_nursery_serial_copy_object_from_obj
#endif
+
#endif
#elif defined (SGEN_SPLIT_NURSERY)
#if defined(SGEN_SIMPLE_NURSERY)
#ifdef SGEN_SIMPLE_PAR_NURSERY
+#ifdef SGEN_CONCURRENT_MAJOR
+#define SERIAL_SCAN_OBJECT simple_par_nursery_serial_with_concurrent_major_scan_object
+#define SERIAL_SCAN_VTYPE simple_par_nursery_serial_with_concurrent_major_scan_vtype
+#define SERIAL_SCAN_PTR_FIELD simple_par_nursery_serial_with_concurrent_major_scan_ptr_field
+#define SERIAL_DRAIN_GRAY_STACK simple_par_nursery_serial_with_concurrent_major_drain_gray_stack
+#else
#define SERIAL_SCAN_OBJECT simple_par_nursery_serial_scan_object
#define SERIAL_SCAN_VTYPE simple_par_nursery_serial_scan_vtype
#define SERIAL_SCAN_PTR_FIELD simple_par_nursery_serial_scan_ptr_field
#define SERIAL_DRAIN_GRAY_STACK simple_par_nursery_serial_drain_gray_stack
+#endif
#else
#ifdef SGEN_CONCURRENT_MAJOR
#define SERIAL_SCAN_OBJECT simple_nursery_serial_with_concurrent_major_scan_object
void
sgen_init_pinning (void)
{
- mono_os_mutex_lock (&pin_queue_mutex);
memset (pin_hash_filter, 0, sizeof (pin_hash_filter));
pin_queue.mem_type = INTERNAL_MEM_PIN_QUEUE;
+}
+
+void
+sgen_init_pinning_for_conc (void)
+{
+ mono_os_mutex_lock (&pin_queue_mutex);
sgen_pointer_queue_clear (&pin_queue_objs);
}
{
last_num_pinned = pin_queue.next_slot;
sgen_pointer_queue_clear (&pin_queue);
+}
+
+void
+sgen_finish_pinning_for_conc (void)
+{
mono_os_mutex_unlock (&pin_queue_mutex);
}
void sgen_pin_stage_ptr (void *ptr);
void sgen_optimize_pin_queue (void);
void sgen_init_pinning (void);
+void sgen_init_pinning_for_conc (void);
+void sgen_finish_pinning_for_conc (void);
void sgen_finish_pinning (void);
void sgen_pinning_register_pinned_in_nursery (GCObject *obj);
void sgen_scan_pin_queue_objects (ScanCopyContext ctx);
* If the thread is not a worker thread we insert 0, which is interpreted
* as gc thread. Worker indexes are 1 based.
*/
- worker_index = sgen_workers_is_worker_thread (tid);
- if (!worker_index)
- worker_index = sgen_thread_pool_is_thread_pool_thread (major_collector.get_sweep_pool (), tid);
+ worker_index = sgen_thread_pool_is_thread_pool_thread (tid);
/* FIXME Consider using different index bases for different thread pools */
buffer->buffer [index++] = (unsigned char) worker_index;
}
#include "mono/sgen/sgen-protocol.h"
#include "mono/sgen/sgen-layout-stats.h"
#include "mono/sgen/sgen-client.h"
+#include "mono/sgen/sgen-workers.h"
#include "mono/utils/mono-memory-model.h"
+#include "mono/utils/mono-proclib.h"
static inline GCObject*
alloc_for_promotion (GCVTable vtable, GCObject *obj, size_t objsize, gboolean has_references)
FILL_MINOR_COLLECTOR_SCAN_OBJECT (ops);
}
+#define SGEN_SIMPLE_PAR_NURSERY
+
+#include "sgen-minor-copy-object.h"
+#include "sgen-minor-scan-object.h"
+
+static void
+fill_parallel_with_concurrent_major_ops (SgenObjectOperations *ops)
+{
+ ops->copy_or_mark_object = SERIAL_COPY_OBJECT;
+ FILL_MINOR_COLLECTOR_SCAN_OBJECT (ops);
+}
+
void
sgen_simple_nursery_init (SgenMinorCollector *collector, gboolean parallel)
{
+ if (mono_cpu_count () <= 1)
+ parallel = FALSE;
+
collector->is_split = FALSE;
collector->is_parallel = parallel;
fill_serial_ops (&collector->serial_ops);
fill_serial_with_concurrent_major_ops (&collector->serial_ops_with_concurrent_major);
fill_parallel_ops (&collector->parallel_ops);
+ fill_parallel_with_concurrent_major_ops (&collector->parallel_ops_with_concurrent_major);
+
+ /*
+ * The nursery worker context is created first so it will have priority over
+ * concurrent mark and concurrent sweep.
+ */
+ if (parallel)
+ sgen_workers_create_context (GENERATION_NURSERY, mono_cpu_count ());
}
#include "mono/sgen/sgen-gc.h"
#include "mono/sgen/sgen-thread-pool.h"
+#include "mono/sgen/sgen-client.h"
#include "mono/utils/mono-os-mutex.h"
+static mono_mutex_t lock;
+static mono_cond_t work_cond;
+static mono_cond_t done_cond;
+
+static int threads_num;
+static MonoNativeThreadId threads [SGEN_THREADPOOL_MAX_NUM_THREADS];
+
+static volatile gboolean threadpool_shutdown;
+static volatile int threads_finished;
+
+static int contexts_num;
+static SgenThreadPoolContext pool_contexts [SGEN_THREADPOOL_MAX_NUM_CONTEXTS];
+
enum {
STATE_WAITING,
STATE_IN_PROGRESS,
/* Assumes that the lock is held. */
static SgenThreadPoolJob*
-get_job_and_set_in_progress (SgenThreadPool *pool)
+get_job_and_set_in_progress (SgenThreadPoolContext *context)
{
- for (size_t i = 0; i < pool->job_queue.next_slot; ++i) {
- SgenThreadPoolJob *job = (SgenThreadPoolJob *)pool->job_queue.data [i];
+ for (size_t i = 0; i < context->job_queue.next_slot; ++i) {
+ SgenThreadPoolJob *job = (SgenThreadPoolJob *)context->job_queue.data [i];
if (job->state == STATE_WAITING) {
job->state = STATE_IN_PROGRESS;
return job;
/* Assumes that the lock is held. */
static ssize_t
-find_job_in_queue (SgenThreadPool *pool, SgenThreadPoolJob *job)
+find_job_in_queue (SgenThreadPoolContext *context, SgenThreadPoolJob *job)
{
- for (ssize_t i = 0; i < pool->job_queue.next_slot; ++i) {
- if (pool->job_queue.data [i] == job)
+ for (ssize_t i = 0; i < context->job_queue.next_slot; ++i) {
+ if (context->job_queue.data [i] == job)
return i;
}
return -1;
/* Assumes that the lock is held. */
static void
-remove_job (SgenThreadPool *pool, SgenThreadPoolJob *job)
+remove_job (SgenThreadPoolContext *context, SgenThreadPoolJob *job)
{
ssize_t index;
SGEN_ASSERT (0, job->state == STATE_DONE, "Why are we removing a job that's not done?");
- index = find_job_in_queue (pool, job);
+ index = find_job_in_queue (context, job);
SGEN_ASSERT (0, index >= 0, "Why is the job we're trying to remove not in the queue?");
- pool->job_queue.data [index] = NULL;
- sgen_pointer_queue_remove_nulls (&pool->job_queue);
+ context->job_queue.data [index] = NULL;
+ sgen_pointer_queue_remove_nulls (&context->job_queue);
sgen_thread_pool_job_free (job);
}
static gboolean
-continue_idle_job (SgenThreadPool *pool, void *thread_data)
+continue_idle_job (SgenThreadPoolContext *context, void *thread_data)
{
- if (!pool->continue_idle_job_func)
+ if (!context->continue_idle_job_func)
return FALSE;
- return pool->continue_idle_job_func (thread_data);
+ return context->continue_idle_job_func (thread_data, context - pool_contexts);
}
static gboolean
-should_work (SgenThreadPool *pool, void *thread_data)
+should_work (SgenThreadPoolContext *context, void *thread_data)
{
- if (!pool->should_work_func)
+ if (!context->should_work_func)
return TRUE;
- return pool->should_work_func (thread_data);
+ return context->should_work_func (thread_data);
}
-static mono_native_thread_return_t
-thread_func (SgenThreadPoolData *thread_data)
+/*
+ * Tells whether we should lock and attempt to get work from
+ * a higher priority context.
+ */
+static gboolean
+has_priority_work (int worker_index, int current_context)
{
- SgenThreadPool *pool = thread_data->pool;
-
- pool->thread_init_func (thread_data);
+ int i;
- mono_os_mutex_lock (&pool->lock);
- for (;;) {
- gboolean do_idle;
- SgenThreadPoolJob *job;
+ for (i = 0; i < current_context; i++) {
+ SgenThreadPoolContext *context = &pool_contexts [i];
+ void *thread_data;
- if (!should_work (pool, thread_data) && !pool->threadpool_shutdown) {
- mono_os_cond_wait (&pool->work_cond, &pool->lock);
+ if (worker_index >= context->num_threads)
continue;
+ thread_data = (context->thread_datas) ? context->thread_datas [worker_index] : NULL;
+ if (!should_work (context, thread_data))
+ continue;
+ if (context->job_queue.next_slot > 0)
+ return TRUE;
+ if (continue_idle_job (context, thread_data))
+ return TRUE;
+ }
+
+ /* Return if job enqueued on current context. Jobs have priority over idle work */
+ if (pool_contexts [current_context].job_queue.next_slot > 0)
+ return TRUE;
+
+ return FALSE;
+}
+
+/*
+ * Gets the highest priority work. If there is none, it waits
+ * for work_cond. Should always be called with lock held.
+ */
+static void
+get_work (int worker_index, int *work_context, int *do_idle, SgenThreadPoolJob **job)
+{
+ while (!threadpool_shutdown) {
+ int i;
+
+ for (i = 0; i < contexts_num; i++) {
+ SgenThreadPoolContext *context = &pool_contexts [i];
+ void *thread_data;
+
+ if (worker_index >= context->num_threads)
+ continue;
+ thread_data = (context->thread_datas) ? context->thread_datas [worker_index] : NULL;
+
+ if (!should_work (context, thread_data))
+ continue;
+
+ /*
+ * It's important that we check the continue idle flag with the lock held.
+ * Suppose we didn't check with the lock held, and the result is FALSE. The
+ * main thread might then set continue idle and signal us before we can take
+ * the lock, and we'd lose the signal.
+ */
+ *do_idle = continue_idle_job (context, thread_data);
+ *job = get_job_and_set_in_progress (context);
+
+ if (*job || *do_idle) {
+ *work_context = i;
+ return;
+ }
}
+
/*
- * It's important that we check the continue idle flag with the lock held.
- * Suppose we didn't check with the lock held, and the result is FALSE. The
- * main thread might then set continue idle and signal us before we can take
- * the lock, and we'd lose the signal.
+ * Nothing to do on any context
+ * pthread_cond_wait() can return successfully despite the condition
+ * not being signalled, so we have to run this in a loop until we
+ * really have work to do.
*/
- do_idle = continue_idle_job (pool, thread_data);
- job = get_job_and_set_in_progress (pool);
+ mono_os_cond_wait (&work_cond, &lock);
+ }
+}
- if (!job && !do_idle && !pool->threadpool_shutdown) {
- /*
- * pthread_cond_wait() can return successfully despite the condition
- * not being signalled, so we have to run this in a loop until we
- * really have work to do.
- */
- mono_os_cond_wait (&pool->work_cond, &pool->lock);
- continue;
+static mono_native_thread_return_t
+thread_func (int worker_index)
+{
+ int current_context;
+ void *thread_data = NULL;
+
+ sgen_client_thread_register_worker ();
+
+ for (current_context = 0; current_context < contexts_num; current_context++) {
+ if (worker_index >= pool_contexts [current_context].num_threads ||
+ !pool_contexts [current_context].thread_init_func)
+ break;
+
+ thread_data = (pool_contexts [current_context].thread_datas) ? pool_contexts [current_context].thread_datas [worker_index] : NULL;
+ pool_contexts [current_context].thread_init_func (thread_data);
+ }
+
+ current_context = 0;
+
+ mono_os_mutex_lock (&lock);
+ for (;;) {
+ gboolean do_idle = FALSE;
+ SgenThreadPoolJob *job = NULL;
+ SgenThreadPoolContext *context = NULL;
+
+ get_work (worker_index, ¤t_context, &do_idle, &job);
+
+ if (!threadpool_shutdown) {
+ context = &pool_contexts [current_context];
+ thread_data = (context->thread_datas) ? context->thread_datas [worker_index] : NULL;
}
- mono_os_mutex_unlock (&pool->lock);
+ mono_os_mutex_unlock (&lock);
if (job) {
job->func (thread_data, job);
- mono_os_mutex_lock (&pool->lock);
+ mono_os_mutex_lock (&lock);
SGEN_ASSERT (0, job->state == STATE_IN_PROGRESS, "The job should still be in progress.");
job->state = STATE_DONE;
- remove_job (pool, job);
+ remove_job (context, job);
/*
* Only the main GC thread will ever wait on the done condition, so we don't
* have to broadcast.
*/
- mono_os_cond_signal (&pool->done_cond);
+ mono_os_cond_signal (&done_cond);
} else if (do_idle) {
- SGEN_ASSERT (0, pool->idle_job_func, "Why do we have idle work when there's no idle job function?");
+ SGEN_ASSERT (0, context->idle_job_func, "Why do we have idle work when there's no idle job function?");
do {
- pool->idle_job_func (thread_data);
- do_idle = continue_idle_job (pool, thread_data);
- } while (do_idle && !pool->job_queue.next_slot);
+ context->idle_job_func (thread_data);
+ do_idle = continue_idle_job (context, thread_data);
+ } while (do_idle && !has_priority_work (worker_index, current_context));
- mono_os_mutex_lock (&pool->lock);
+ mono_os_mutex_lock (&lock);
if (!do_idle)
- mono_os_cond_signal (&pool->done_cond);
+ mono_os_cond_signal (&done_cond);
} else {
- SGEN_ASSERT (0, pool->threadpool_shutdown, "Why did we unlock if no jobs and not shutting down?");
- mono_os_mutex_lock (&pool->lock);
- pool->threads_finished++;
- mono_os_cond_signal (&pool->done_cond);
- mono_os_mutex_unlock (&pool->lock);
+ SGEN_ASSERT (0, threadpool_shutdown, "Why did we unlock if no jobs and not shutting down?");
+ mono_os_mutex_lock (&lock);
+ threads_finished++;
+ mono_os_cond_signal (&done_cond);
+ mono_os_mutex_unlock (&lock);
return 0;
}
}
return (mono_native_thread_return_t)0;
}
+int
+sgen_thread_pool_create_context (int num_threads, SgenThreadPoolThreadInitFunc init_func, SgenThreadPoolIdleJobFunc idle_func, SgenThreadPoolContinueIdleJobFunc continue_idle_func, SgenThreadPoolShouldWorkFunc should_work_func, void **thread_datas)
+{
+ int context_id = contexts_num;
+
+ SGEN_ASSERT (0, contexts_num < SGEN_THREADPOOL_MAX_NUM_CONTEXTS, "Maximum sgen thread pool contexts reached");
+
+ pool_contexts [context_id].thread_init_func = init_func;
+ pool_contexts [context_id].idle_job_func = idle_func;
+ pool_contexts [context_id].continue_idle_job_func = continue_idle_func;
+ pool_contexts [context_id].should_work_func = should_work_func;
+ pool_contexts [context_id].thread_datas = thread_datas;
+
+ SGEN_ASSERT (0, num_threads <= SGEN_THREADPOOL_MAX_NUM_THREADS, "Maximum sgen thread pool threads exceeded");
+
+ pool_contexts [context_id].num_threads = num_threads;
+
+ sgen_pointer_queue_init (&pool_contexts [contexts_num].job_queue, 0);
+
+ contexts_num++;
+
+ return context_id;
+}
+
void
-sgen_thread_pool_init (SgenThreadPool *pool, int num_threads, SgenThreadPoolThreadInitFunc init_func, SgenThreadPoolIdleJobFunc idle_func, SgenThreadPoolContinueIdleJobFunc continue_idle_func, SgenThreadPoolShouldWorkFunc should_work_func_p, SgenThreadPoolData **thread_datas)
+sgen_thread_pool_start (void)
{
int i;
- SGEN_ASSERT (0, num_threads > 0, "Why are we creating a threadpool with no threads?");
-
- pool->threads_num = (num_threads < MAX_NUM_THREADS) ? num_threads : MAX_NUM_THREADS;
+ for (i = 0; i < contexts_num; i++) {
+ if (threads_num < pool_contexts [i].num_threads)
+ threads_num = pool_contexts [i].num_threads;
+ }
- mono_os_mutex_init (&pool->lock);
- mono_os_cond_init (&pool->work_cond);
- mono_os_cond_init (&pool->done_cond);
+ if (!threads_num)
+ return;
- pool->thread_init_func = init_func;
- pool->idle_job_func = idle_func;
- pool->continue_idle_job_func = continue_idle_func;
- pool->should_work_func = should_work_func_p;
+ mono_os_mutex_init (&lock);
+ mono_os_cond_init (&work_cond);
+ mono_os_cond_init (&done_cond);
- sgen_pointer_queue_init (&pool->job_queue, 0);
- pool->threads_finished = 0;
- pool->threadpool_shutdown = FALSE;
+ threads_finished = 0;
+ threadpool_shutdown = FALSE;
- for (i = 0; i < pool->threads_num; i++) {
- thread_datas [i]->pool = pool;
- mono_native_thread_create (&pool->threads [i], thread_func, thread_datas [i]);
+ for (i = 0; i < threads_num; i++) {
+ mono_native_thread_create (&threads [i], thread_func, (void*)(gsize)i);
}
}
void
-sgen_thread_pool_shutdown (SgenThreadPool *pool)
+sgen_thread_pool_shutdown (void)
{
- if (!pool)
+ if (!threads_num)
return;
- mono_os_mutex_lock (&pool->lock);
- pool->threadpool_shutdown = TRUE;
- mono_os_cond_broadcast (&pool->work_cond);
- while (pool->threads_finished < pool->threads_num)
- mono_os_cond_wait (&pool->done_cond, &pool->lock);
- mono_os_mutex_unlock (&pool->lock);
+ mono_os_mutex_lock (&lock);
+ threadpool_shutdown = TRUE;
+ mono_os_cond_broadcast (&work_cond);
+ while (threads_finished < threads_num)
+ mono_os_cond_wait (&done_cond, &lock);
+ mono_os_mutex_unlock (&lock);
- mono_os_mutex_destroy (&pool->lock);
- mono_os_cond_destroy (&pool->work_cond);
- mono_os_cond_destroy (&pool->done_cond);
+ mono_os_mutex_destroy (&lock);
+ mono_os_cond_destroy (&work_cond);
+ mono_os_cond_destroy (&done_cond);
}
SgenThreadPoolJob*
}
void
-sgen_thread_pool_job_enqueue (SgenThreadPool *pool, SgenThreadPoolJob *job)
+sgen_thread_pool_job_enqueue (int context_id, SgenThreadPoolJob *job)
{
- mono_os_mutex_lock (&pool->lock);
+ mono_os_mutex_lock (&lock);
- sgen_pointer_queue_add (&pool->job_queue, job);
- mono_os_cond_signal (&pool->work_cond);
+ sgen_pointer_queue_add (&pool_contexts [context_id].job_queue, job);
+ mono_os_cond_broadcast (&work_cond);
- mono_os_mutex_unlock (&pool->lock);
+ mono_os_mutex_unlock (&lock);
}
void
-sgen_thread_pool_job_wait (SgenThreadPool *pool, SgenThreadPoolJob *job)
+sgen_thread_pool_job_wait (int context_id, SgenThreadPoolJob *job)
{
SGEN_ASSERT (0, job, "Where's the job?");
- mono_os_mutex_lock (&pool->lock);
+ mono_os_mutex_lock (&lock);
- while (find_job_in_queue (pool, job) >= 0)
- mono_os_cond_wait (&pool->done_cond, &pool->lock);
+ while (find_job_in_queue (&pool_contexts [context_id], job) >= 0)
+ mono_os_cond_wait (&done_cond, &lock);
- mono_os_mutex_unlock (&pool->lock);
+ mono_os_mutex_unlock (&lock);
}
void
-sgen_thread_pool_idle_signal (SgenThreadPool *pool)
+sgen_thread_pool_idle_signal (int context_id)
{
- SGEN_ASSERT (0, pool->idle_job_func, "Why are we signaling idle without an idle function?");
+ SGEN_ASSERT (0, pool_contexts [context_id].idle_job_func, "Why are we signaling idle without an idle function?");
- mono_os_mutex_lock (&pool->lock);
+ mono_os_mutex_lock (&lock);
- if (pool->continue_idle_job_func (NULL))
- mono_os_cond_broadcast (&pool->work_cond);
+ if (pool_contexts [context_id].continue_idle_job_func (NULL, context_id))
+ mono_os_cond_broadcast (&work_cond);
- mono_os_mutex_unlock (&pool->lock);
+ mono_os_mutex_unlock (&lock);
}
void
-sgen_thread_pool_idle_wait (SgenThreadPool *pool)
+sgen_thread_pool_idle_wait (int context_id)
{
- SGEN_ASSERT (0, pool->idle_job_func, "Why are we waiting for idle without an idle function?");
+ SGEN_ASSERT (0, pool_contexts [context_id].idle_job_func, "Why are we waiting for idle without an idle function?");
- mono_os_mutex_lock (&pool->lock);
+ mono_os_mutex_lock (&lock);
- while (pool->continue_idle_job_func (NULL))
- mono_os_cond_wait (&pool->done_cond, &pool->lock);
+ while (pool_contexts [context_id].continue_idle_job_func (NULL, context_id))
+ mono_os_cond_wait (&done_cond, &lock);
- mono_os_mutex_unlock (&pool->lock);
+ mono_os_mutex_unlock (&lock);
}
void
-sgen_thread_pool_wait_for_all_jobs (SgenThreadPool *pool)
+sgen_thread_pool_wait_for_all_jobs (int context_id)
{
- mono_os_mutex_lock (&pool->lock);
+ mono_os_mutex_lock (&lock);
- while (!sgen_pointer_queue_is_empty (&pool->job_queue))
- mono_os_cond_wait (&pool->done_cond, &pool->lock);
+ while (!sgen_pointer_queue_is_empty (&pool_contexts [context_id].job_queue))
+ mono_os_cond_wait (&done_cond, &lock);
- mono_os_mutex_unlock (&pool->lock);
+ mono_os_mutex_unlock (&lock);
}
/* Return 0 if is not a thread pool thread or the thread number otherwise */
int
-sgen_thread_pool_is_thread_pool_thread (SgenThreadPool *pool, MonoNativeThreadId some_thread)
+sgen_thread_pool_is_thread_pool_thread (MonoNativeThreadId some_thread)
{
int i;
- if (!pool)
- return 0;
-
- for (i = 0; i < pool->threads_num; i++) {
- if (some_thread == pool->threads [i])
+ for (i = 0; i < threads_num; i++) {
+ if (some_thread == threads [i])
return i + 1;
}
#include "mono/sgen/sgen-pointer-queue.h"
#include "mono/utils/mono-threads.h"
+#define SGEN_THREADPOOL_MAX_NUM_THREADS 8
+#define SGEN_THREADPOOL_MAX_NUM_CONTEXTS 3
+
typedef struct _SgenThreadPoolJob SgenThreadPoolJob;
-typedef struct _SgenThreadPool SgenThreadPool;
-typedef struct _SgenThreadPoolData SgenThreadPoolData;
+typedef struct _SgenThreadPoolContext SgenThreadPoolContext;
typedef void (*SgenThreadPoolJobFunc) (void *thread_data, SgenThreadPoolJob *job);
typedef void (*SgenThreadPoolThreadInitFunc) (void*);
typedef void (*SgenThreadPoolIdleJobFunc) (void*);
-typedef gboolean (*SgenThreadPoolContinueIdleJobFunc) (void*);
+typedef gboolean (*SgenThreadPoolContinueIdleJobFunc) (void*, int);
typedef gboolean (*SgenThreadPoolShouldWorkFunc) (void*);
struct _SgenThreadPoolJob {
volatile gint32 state;
};
-#define MAX_NUM_THREADS 8
-
-struct _SgenThreadPool {
- mono_mutex_t lock;
- mono_cond_t work_cond;
- mono_cond_t done_cond;
-
- int threads_num;
- MonoNativeThreadId threads [MAX_NUM_THREADS];
-
+struct _SgenThreadPoolContext {
/* Only accessed with the lock held. */
SgenPointerQueue job_queue;
SgenThreadPoolContinueIdleJobFunc continue_idle_job_func;
SgenThreadPoolShouldWorkFunc should_work_func;
- volatile gboolean threadpool_shutdown;
- volatile int threads_finished;
+ void **thread_datas;
+ int num_threads;
};
-struct _SgenThreadPoolData {
- SgenThreadPool *pool;
-};
-void sgen_thread_pool_init (SgenThreadPool *pool, int num_threads, SgenThreadPoolThreadInitFunc init_func, SgenThreadPoolIdleJobFunc idle_func, SgenThreadPoolContinueIdleJobFunc continue_idle_func, SgenThreadPoolShouldWorkFunc should_work_func, SgenThreadPoolData **thread_datas);
+int sgen_thread_pool_create_context (int num_threads, SgenThreadPoolThreadInitFunc init_func, SgenThreadPoolIdleJobFunc idle_func, SgenThreadPoolContinueIdleJobFunc continue_idle_func, SgenThreadPoolShouldWorkFunc should_work_func, void **thread_datas);
+void sgen_thread_pool_start (void);
-void sgen_thread_pool_shutdown (SgenThreadPool *pool);
+void sgen_thread_pool_shutdown (void);
SgenThreadPoolJob* sgen_thread_pool_job_alloc (const char *name, SgenThreadPoolJobFunc func, size_t size);
/* This only needs to be called on jobs that are not enqueued. */
void sgen_thread_pool_job_free (SgenThreadPoolJob *job);
-void sgen_thread_pool_job_enqueue (SgenThreadPool *pool, SgenThreadPoolJob *job);
+void sgen_thread_pool_job_enqueue (int context_id, SgenThreadPoolJob *job);
/* This must only be called after the job has been enqueued. */
-void sgen_thread_pool_job_wait (SgenThreadPool *pool, SgenThreadPoolJob *job);
+void sgen_thread_pool_job_wait (int context_id, SgenThreadPoolJob *job);
-void sgen_thread_pool_idle_signal (SgenThreadPool *pool);
-void sgen_thread_pool_idle_wait (SgenThreadPool *pool);
+void sgen_thread_pool_idle_signal (int context_id);
+void sgen_thread_pool_idle_wait (int context_id);
-void sgen_thread_pool_wait_for_all_jobs (SgenThreadPool *pool);
+void sgen_thread_pool_wait_for_all_jobs (int context_id);
-int sgen_thread_pool_is_thread_pool_thread (SgenThreadPool *pool, MonoNativeThreadId thread);
+int sgen_thread_pool_is_thread_pool_thread (MonoNativeThreadId thread);
#endif
#include "mono/utils/mono-membar.h"
#include "mono/sgen/sgen-client.h"
-static int workers_num;
-static int active_workers_num;
-static volatile gboolean started;
-static volatile gboolean forced_stop;
-static WorkerData *workers_data;
-static SgenWorkerCallback worker_init_cb;
-
-static SgenThreadPool pool_inst;
-static SgenThreadPool *pool; /* null if we're not using workers */
-
-/*
- * When using multiple workers, we need to have the last worker
- * enqueue the preclean jobs (if there are any). This lock ensures
- * that when the last worker takes it, all the other workers have
- * gracefully finished, so it can restart them.
- */
-static mono_mutex_t finished_lock;
-static volatile gboolean workers_finished;
-static int worker_awakenings;
-
-static SgenSectionGrayQueue workers_distribute_gray_queue;
-static gboolean workers_distribute_gray_queue_inited;
+static WorkerContext worker_contexts [GENERATION_MAX];
/*
* Allowed transitions:
#define SGEN_WORKER_MIN_SECTIONS_SIGNAL 4
-typedef gint32 State;
-
-static SgenObjectOperations * volatile idle_func_object_ops;
-static SgenObjectOperations *idle_func_object_ops_par, *idle_func_object_ops_nopar;
-/*
- * finished_callback is called only when the workers finish work normally (when they
- * are not forced to finish). The callback is used to enqueue preclean jobs.
- */
-static volatile SgenWorkersFinishCallback finish_callback;
-
static guint64 stat_workers_num_finished;
static gboolean
else if (new_state == STATE_WORKING)
SGEN_ASSERT (0, old_state == STATE_WORK_ENQUEUED, "We can only transition to WORKING from WORK ENQUEUED");
if (new_state == STATE_NOT_WORKING || new_state == STATE_WORKING)
- SGEN_ASSERT (6, sgen_thread_pool_is_thread_pool_thread (pool, mono_native_thread_id_get ()), "Only the worker thread is allowed to transition to NOT_WORKING or WORKING");
+ SGEN_ASSERT (6, sgen_thread_pool_is_thread_pool_thread (mono_native_thread_id_get ()), "Only the worker thread is allowed to transition to NOT_WORKING or WORKING");
return InterlockedCompareExchange (&data->state, new_state, old_state) == old_state;
}
}
static void
-sgen_workers_ensure_awake (void)
+sgen_workers_ensure_awake (WorkerContext *context)
{
int i;
gboolean need_signal = FALSE;
* or when the last worker is enqueuing preclean work. In both cases we can't
* have a worker working using a nopar context, which means it is safe.
*/
- idle_func_object_ops = (active_workers_num > 1) ? idle_func_object_ops_par : idle_func_object_ops_nopar;
- workers_finished = FALSE;
+ context->idle_func_object_ops = (context->active_workers_num > 1) ? context->idle_func_object_ops_par : context->idle_func_object_ops_nopar;
+ context->workers_finished = FALSE;
- for (i = 0; i < active_workers_num; i++) {
+ for (i = 0; i < context->active_workers_num; i++) {
State old_state;
gboolean did_set_state;
do {
- old_state = workers_data [i].state;
+ old_state = context->workers_data [i].state;
if (old_state == STATE_WORK_ENQUEUED)
break;
- did_set_state = set_state (&workers_data [i], old_state, STATE_WORK_ENQUEUED);
+ did_set_state = set_state (&context->workers_data [i], old_state, STATE_WORK_ENQUEUED);
} while (!did_set_state);
if (!state_is_working_or_enqueued (old_state))
}
if (need_signal)
- sgen_thread_pool_idle_signal (pool);
+ sgen_thread_pool_idle_signal (context->thread_pool_context);
}
static void
{
State old_state;
int i, working = 0;
+ WorkerContext *context = data->context;
++stat_workers_num_finished;
- mono_os_mutex_lock (&finished_lock);
+ mono_os_mutex_lock (&context->finished_lock);
- for (i = 0; i < active_workers_num; i++) {
- if (state_is_working_or_enqueued (workers_data [i].state))
+ for (i = 0; i < context->active_workers_num; i++) {
+ if (state_is_working_or_enqueued (context->workers_data [i].state))
working++;
}
if (working == 1) {
- SgenWorkersFinishCallback callback = finish_callback;
- SGEN_ASSERT (0, idle_func_object_ops == idle_func_object_ops_nopar, "Why are we finishing with parallel context");
+ SgenWorkersFinishCallback callback = context->finish_callback;
+ SGEN_ASSERT (0, context->idle_func_object_ops == context->idle_func_object_ops_nopar, "Why are we finishing with parallel context");
/* We are the last one left. Enqueue preclean job if we have one and awake everybody */
SGEN_ASSERT (0, data->state != STATE_NOT_WORKING, "How did we get from doing idle work to NOT WORKING without setting it ourselves?");
if (callback) {
- finish_callback = NULL;
+ context->finish_callback = NULL;
callback ();
- worker_awakenings = 0;
+ context->worker_awakenings = 0;
/* Make sure each worker has a chance of seeing the enqueued jobs */
- sgen_workers_ensure_awake ();
+ sgen_workers_ensure_awake (context);
SGEN_ASSERT (0, data->state == STATE_WORK_ENQUEUED, "Why did we fail to set our own state to ENQUEUED");
goto work_available;
}
* of performance as non-parallel mode even if we fail to distribute work properly.
*/
if (working == 2)
- idle_func_object_ops = idle_func_object_ops_nopar;
+ context->idle_func_object_ops = context->idle_func_object_ops_nopar;
- workers_finished = TRUE;
- mono_os_mutex_unlock (&finished_lock);
+ context->workers_finished = TRUE;
+ mono_os_mutex_unlock (&context->finished_lock);
- binary_protocol_worker_finish (sgen_timestamp (), forced_stop);
+ binary_protocol_worker_finish (sgen_timestamp (), context->forced_stop);
sgen_gray_object_queue_trim_free_list (&data->private_gray_queue);
return;
work_available:
- mono_os_mutex_unlock (&finished_lock);
+ mono_os_mutex_unlock (&context->finished_lock);
}
void
-sgen_workers_enqueue_job (SgenThreadPoolJob *job, gboolean enqueue)
+sgen_workers_enqueue_job (int generation, SgenThreadPoolJob *job, gboolean enqueue)
{
if (!enqueue) {
job->func (NULL, job);
return;
}
- sgen_thread_pool_job_enqueue (pool, job);
+ sgen_thread_pool_job_enqueue (worker_contexts [generation].thread_pool_context, job);
}
static gboolean
g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
g_assert (major->is_concurrent || minor->is_parallel);
- section = sgen_section_gray_queue_dequeue (&workers_distribute_gray_queue);
+ section = sgen_section_gray_queue_dequeue (&data->context->workers_distribute_gray_queue);
if (section) {
sgen_gray_object_enqueue_section (&data->private_gray_queue, section, major->is_parallel);
return TRUE;
SgenMinorCollector *minor = sgen_get_minor_collector ();
int generation = sgen_get_current_collection_generation ();
GrayQueueSection *section = NULL;
+ WorkerContext *context = data->context;
int i, current_worker;
if ((generation == GENERATION_OLD && !major->is_parallel) ||
/* If we're parallel, steal from other workers' private gray queues */
g_assert (sgen_gray_object_queue_is_empty (&data->private_gray_queue));
- current_worker = (int) (data - workers_data);
+ current_worker = (int) (data - context->workers_data);
- for (i = 1; i < active_workers_num && !section; i++) {
- int steal_worker = (current_worker + i) % active_workers_num;
- if (state_is_working_or_enqueued (workers_data [steal_worker].state))
- section = sgen_gray_object_steal_section (&workers_data [steal_worker].private_gray_queue);
+ for (i = 1; i < context->active_workers_num && !section; i++) {
+ int steal_worker = (current_worker + i) % context->active_workers_num;
+ if (state_is_working_or_enqueued (context->workers_data [steal_worker].state))
+ section = sgen_gray_object_steal_section (&context->workers_data [steal_worker].private_gray_queue);
}
if (section) {
SgenMajorCollector *major = sgen_get_major_collector ();
SgenMinorCollector *minor = sgen_get_minor_collector ();
- sgen_client_thread_register_worker ();
-
if (!major->is_concurrent && !minor->is_parallel)
return;
init_private_gray_queue (data);
- if (worker_init_cb)
- worker_init_cb (data);
+ /* Separate WorkerData for same thread share free_block_lists */
+ if (major->is_parallel || minor->is_parallel)
+ major->init_block_free_lists (&data->free_block_lists);
}
static gboolean
-continue_idle_func (void *data_untyped)
+sgen_workers_are_working (WorkerContext *context)
{
- if (data_untyped) {
- WorkerData *data = (WorkerData *)data_untyped;
- return state_is_working_or_enqueued (data->state);
- } else {
- /* Return if any of the threads is working */
- return !sgen_workers_all_done ();
+ int i;
+
+ for (i = 0; i < context->active_workers_num; i++) {
+ if (state_is_working_or_enqueued (context->workers_data [i].state))
+ return TRUE;
}
+ return FALSE;
+}
+
+static gboolean
+continue_idle_func (void *data_untyped, int thread_pool_context)
+{
+ if (data_untyped)
+ return state_is_working_or_enqueued (((WorkerData*)data_untyped)->state);
+
+ /* Return if any of the threads is working in the context */
+ if (worker_contexts [GENERATION_NURSERY].workers_num && worker_contexts [GENERATION_NURSERY].thread_pool_context == thread_pool_context)
+ return sgen_workers_are_working (&worker_contexts [GENERATION_NURSERY]);
+ if (worker_contexts [GENERATION_OLD].workers_num && worker_contexts [GENERATION_OLD].thread_pool_context == thread_pool_context)
+ return sgen_workers_are_working (&worker_contexts [GENERATION_OLD]);
+
+ g_assert_not_reached ();
+ return FALSE;
}
static gboolean
should_work_func (void *data_untyped)
{
WorkerData *data = (WorkerData*)data_untyped;
- int current_worker = (int) (data - workers_data);
+ WorkerContext *context = data->context;
+ int current_worker = (int) (data - context->workers_data);
- return started && current_worker < active_workers_num && state_is_working_or_enqueued (data->state);
+ return context->started && current_worker < context->active_workers_num && state_is_working_or_enqueued (data->state);
}
static void
marker_idle_func (void *data_untyped)
{
WorkerData *data = (WorkerData *)data_untyped;
+ WorkerContext *context = data->context;
- SGEN_ASSERT (0, continue_idle_func (data_untyped), "Why are we called when we're not supposed to work?");
+ SGEN_ASSERT (0, continue_idle_func (data_untyped, context->thread_pool_context), "Why are we called when we're not supposed to work?");
if (data->state == STATE_WORK_ENQUEUED) {
set_state (data, STATE_WORK_ENQUEUED, STATE_WORKING);
SGEN_ASSERT (0, data->state != STATE_NOT_WORKING, "How did we get from WORK ENQUEUED to NOT WORKING?");
}
- if (!forced_stop && (!sgen_gray_object_queue_is_empty (&data->private_gray_queue) || workers_get_work (data) || workers_steal_work (data))) {
- ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (idle_func_object_ops, &data->private_gray_queue);
+ if (!context->forced_stop && (!sgen_gray_object_queue_is_empty (&data->private_gray_queue) || workers_get_work (data) || workers_steal_work (data))) {
+ ScanCopyContext ctx = CONTEXT_FROM_OBJECT_OPERATIONS (context->idle_func_object_ops, &data->private_gray_queue);
SGEN_ASSERT (0, !sgen_gray_object_queue_is_empty (&data->private_gray_queue), "How is our gray queue empty if we just got work?");
sgen_drain_gray_stack (ctx);
if (data->private_gray_queue.num_sections >= SGEN_WORKER_MIN_SECTIONS_SIGNAL
- && workers_finished && worker_awakenings < active_workers_num) {
+ && context->workers_finished && context->worker_awakenings < context->active_workers_num) {
/* We bound the number of worker awakenings just to be sure */
- worker_awakenings++;
- mono_os_mutex_lock (&finished_lock);
- sgen_workers_ensure_awake ();
- mono_os_mutex_unlock (&finished_lock);
+ context->worker_awakenings++;
+ mono_os_mutex_lock (&context->finished_lock);
+ sgen_workers_ensure_awake (context);
+ mono_os_mutex_unlock (&context->finished_lock);
}
} else {
worker_try_finish (data);
}
static void
-init_distribute_gray_queue (void)
+init_distribute_gray_queue (WorkerContext *context)
{
- if (workers_distribute_gray_queue_inited) {
- g_assert (sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue));
- g_assert (workers_distribute_gray_queue.locked);
- return;
- }
-
- sgen_section_gray_queue_init (&workers_distribute_gray_queue, TRUE,
+ sgen_section_gray_queue_init (&context->workers_distribute_gray_queue, TRUE,
sgen_get_major_collector ()->is_concurrent ? concurrent_enqueue_check : NULL);
- workers_distribute_gray_queue_inited = TRUE;
-}
-
-void
-sgen_workers_init_distribute_gray_queue (void)
-{
- SGEN_ASSERT (0, sgen_get_major_collector ()->is_concurrent || sgen_get_minor_collector ()->is_parallel,
- "Why should we init the distribute gray queue if we don't need it?");
- init_distribute_gray_queue ();
}
void
-sgen_workers_init (int num_workers, SgenWorkerCallback callback)
+sgen_workers_create_context (int generation, int num_workers)
{
+ static gboolean stat_inited = FALSE;
int i;
- WorkerData **workers_data_ptrs = (WorkerData**)alloca(num_workers * sizeof(WorkerData*));
+ WorkerData **workers_data_ptrs = (WorkerData**)sgen_alloc_internal_dynamic (num_workers * sizeof(WorkerData*), INTERNAL_MEM_WORKER_DATA, TRUE);
+ WorkerContext *context = &worker_contexts [generation];
- mono_os_mutex_init (&finished_lock);
- //g_print ("initing %d workers\n", num_workers);
+ SGEN_ASSERT (0, !context->workers_num, "We can't init the worker context for a generation twice");
- workers_num = num_workers;
- active_workers_num = num_workers;
+ mono_os_mutex_init (&context->finished_lock);
- workers_data = (WorkerData *)sgen_alloc_internal_dynamic (sizeof (WorkerData) * num_workers, INTERNAL_MEM_WORKER_DATA, TRUE);
- memset (workers_data, 0, sizeof (WorkerData) * num_workers);
+ context->generation = generation;
+ context->workers_num = num_workers;
+ context->active_workers_num = num_workers;
- init_distribute_gray_queue ();
+ context->workers_data = (WorkerData *)sgen_alloc_internal_dynamic (sizeof (WorkerData) * num_workers, INTERNAL_MEM_WORKER_DATA, TRUE);
+ memset (context->workers_data, 0, sizeof (WorkerData) * num_workers);
- for (i = 0; i < num_workers; ++i)
- workers_data_ptrs [i] = &workers_data [i];
+ init_distribute_gray_queue (context);
- worker_init_cb = callback;
+ for (i = 0; i < num_workers; ++i) {
+ workers_data_ptrs [i] = &context->workers_data [i];
+ context->workers_data [i].context = context;
+ }
- pool = &pool_inst;
- sgen_thread_pool_init (pool, num_workers, thread_pool_init_func, marker_idle_func, continue_idle_func, should_work_func, (SgenThreadPoolData**)workers_data_ptrs);
+ context->thread_pool_context = sgen_thread_pool_create_context (num_workers, thread_pool_init_func, marker_idle_func, continue_idle_func, should_work_func, (void**)workers_data_ptrs);
- mono_counters_register ("# workers finished", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_workers_num_finished);
+ if (!stat_inited) {
+ mono_counters_register ("# workers finished", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_workers_num_finished);
+ stat_inited = TRUE;
+ }
}
void
-sgen_workers_shutdown (void)
+sgen_workers_stop_all_workers (int generation)
{
- if (pool)
- sgen_thread_pool_shutdown (pool);
-}
+ WorkerContext *context = &worker_contexts [generation];
-void
-sgen_workers_stop_all_workers (void)
-{
- finish_callback = NULL;
+ context->finish_callback = NULL;
mono_memory_write_barrier ();
- forced_stop = TRUE;
+ context->forced_stop = TRUE;
- sgen_thread_pool_wait_for_all_jobs (pool);
- sgen_thread_pool_idle_wait (pool);
- SGEN_ASSERT (0, sgen_workers_all_done (), "Can only signal enqueue work when in no work state");
+ sgen_thread_pool_wait_for_all_jobs (context->thread_pool_context);
+ sgen_thread_pool_idle_wait (context->thread_pool_context);
+ SGEN_ASSERT (0, !sgen_workers_are_working (context), "Can only signal enqueue work when in no work state");
- started = FALSE;
+ context->started = FALSE;
}
void
-sgen_workers_set_num_active_workers (int num_workers)
+sgen_workers_set_num_active_workers (int generation, int num_workers)
{
+ WorkerContext *context = &worker_contexts [generation];
if (num_workers) {
- SGEN_ASSERT (0, active_workers_num <= workers_num, "We can't start more workers than we initialized");
- active_workers_num = num_workers;
+ SGEN_ASSERT (0, num_workers <= context->workers_num, "We can't start more workers than we initialized");
+ context->active_workers_num = num_workers;
} else {
- active_workers_num = workers_num;
+ context->active_workers_num = context->workers_num;
}
}
void
-sgen_workers_start_all_workers (SgenObjectOperations *object_ops_nopar, SgenObjectOperations *object_ops_par, SgenWorkersFinishCallback callback)
+sgen_workers_start_all_workers (int generation, SgenObjectOperations *object_ops_nopar, SgenObjectOperations *object_ops_par, SgenWorkersFinishCallback callback)
{
- SGEN_ASSERT (0, !started, "Why are we starting to work without finishing previous cycle");
-
- idle_func_object_ops_par = object_ops_par;
- idle_func_object_ops_nopar = object_ops_nopar;
- forced_stop = FALSE;
- finish_callback = callback;
- worker_awakenings = 0;
- started = TRUE;
+ WorkerContext *context = &worker_contexts [generation];
+ SGEN_ASSERT (0, !context->started, "Why are we starting to work without finishing previous cycle");
+
+ context->idle_func_object_ops_par = object_ops_par;
+ context->idle_func_object_ops_nopar = object_ops_nopar;
+ context->forced_stop = FALSE;
+ context->finish_callback = callback;
+ context->worker_awakenings = 0;
+ context->started = TRUE;
mono_memory_write_barrier ();
/*
* We expect workers to start finishing only after all of them were awaken.
* Otherwise we might think that we have fewer workers and use wrong context.
*/
- mono_os_mutex_lock (&finished_lock);
- sgen_workers_ensure_awake ();
- mono_os_mutex_unlock (&finished_lock);
+ mono_os_mutex_lock (&context->finished_lock);
+ sgen_workers_ensure_awake (context);
+ mono_os_mutex_unlock (&context->finished_lock);
}
void
-sgen_workers_join (void)
+sgen_workers_join (int generation)
{
+ WorkerContext *context = &worker_contexts [generation];
int i;
- sgen_thread_pool_wait_for_all_jobs (pool);
- sgen_thread_pool_idle_wait (pool);
- SGEN_ASSERT (0, sgen_workers_all_done (), "Can only signal enqueue work when in no work state");
+ /*
+ * It might be the case that a worker didn't get to run anything
+ * in this context, because it was stuck working on a long job
+ * in another context. In this case its state is active (WORK_ENQUEUED)
+ * and we need to wait for it to finish itself.
+ * FIXME Avoid having to wait for the worker to report its own finish.
+ */
+
+ sgen_thread_pool_wait_for_all_jobs (context->thread_pool_context);
+ sgen_thread_pool_idle_wait (context->thread_pool_context);
+ SGEN_ASSERT (0, !sgen_workers_are_working (context), "Can only signal enqueue work when in no work state");
/* At this point all the workers have stopped. */
- SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue), "Why is there still work left to do?");
- for (i = 0; i < active_workers_num; ++i)
- SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&workers_data [i].private_gray_queue), "Why is there still work left to do?");
+ SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&context->workers_distribute_gray_queue), "Why is there still work left to do?");
+ for (i = 0; i < context->active_workers_num; ++i)
+ SGEN_ASSERT (0, sgen_gray_object_queue_is_empty (&context->workers_data [i].private_gray_queue), "Why is there still work left to do?");
- started = FALSE;
+ context->started = FALSE;
}
/*
* If we're stopped, there are also no pending jobs.
*/
gboolean
-sgen_workers_have_idle_work (void)
+sgen_workers_have_idle_work (int generation)
{
+ WorkerContext *context = &worker_contexts [generation];
int i;
- SGEN_ASSERT (0, forced_stop && sgen_workers_all_done (), "Checking for idle work should only happen if the workers are stopped.");
+ SGEN_ASSERT (0, context->forced_stop && !sgen_workers_are_working (context), "Checking for idle work should only happen if the workers are stopped.");
- if (!sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue))
+ if (!sgen_section_gray_queue_is_empty (&context->workers_distribute_gray_queue))
return TRUE;
- for (i = 0; i < active_workers_num; ++i) {
- if (!sgen_gray_object_queue_is_empty (&workers_data [i].private_gray_queue))
+ for (i = 0; i < context->active_workers_num; ++i) {
+ if (!sgen_gray_object_queue_is_empty (&context->workers_data [i].private_gray_queue))
return TRUE;
}
gboolean
sgen_workers_all_done (void)
{
- int i;
+ if (worker_contexts [GENERATION_NURSERY].workers_num && sgen_workers_are_working (&worker_contexts [GENERATION_NURSERY]))
+ return FALSE;
+ if (worker_contexts [GENERATION_OLD].workers_num && sgen_workers_are_working (&worker_contexts [GENERATION_OLD]))
+ return FALSE;
- for (i = 0; i < active_workers_num; i++) {
- if (state_is_working_or_enqueued (workers_data [i].state))
- return FALSE;
- }
return TRUE;
}
-/* Must only be used for debugging */
-gboolean
-sgen_workers_are_working (void)
-{
- return !sgen_workers_all_done ();
-}
-
void
-sgen_workers_assert_gray_queue_is_empty (void)
+sgen_workers_assert_gray_queue_is_empty (int generation)
{
- SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&workers_distribute_gray_queue), "Why is the workers gray queue not empty?");
+ SGEN_ASSERT (0, sgen_section_gray_queue_is_empty (&worker_contexts [generation].workers_distribute_gray_queue), "Why is the workers gray queue not empty?");
}
void
-sgen_workers_take_from_queue (SgenGrayQueue *queue)
+sgen_workers_take_from_queue (int generation, SgenGrayQueue *queue)
{
- sgen_gray_object_spread (queue, sgen_workers_get_job_split_count ());
+ WorkerContext *context = &worker_contexts [generation];
+
+ sgen_gray_object_spread (queue, sgen_workers_get_job_split_count (generation));
for (;;) {
GrayQueueSection *section = sgen_gray_object_dequeue_section (queue);
if (!section)
break;
- sgen_section_gray_queue_enqueue (&workers_distribute_gray_queue, section);
+ sgen_section_gray_queue_enqueue (&context->workers_distribute_gray_queue, section);
}
- SGEN_ASSERT (0, !sgen_workers_are_working (), "We should fully populate the distribute gray queue before we start the workers");
+ SGEN_ASSERT (0, !sgen_workers_are_working (context), "We should fully populate the distribute gray queue before we start the workers");
}
SgenObjectOperations*
-sgen_workers_get_idle_func_object_ops (void)
+sgen_workers_get_idle_func_object_ops (WorkerData *worker)
{
- g_assert (idle_func_object_ops);
- return idle_func_object_ops;
+ g_assert (worker->context->idle_func_object_ops);
+ return worker->context->idle_func_object_ops;
}
/*
* more than one worker, we split into a larger number of jobs so that, in case
* the work load is uneven, a worker that finished quickly can take up more jobs
* than another one.
+ *
+ * We also return 1 if there is no worker context for that generation.
*/
int
-sgen_workers_get_job_split_count (void)
+sgen_workers_get_job_split_count (int generation)
{
- return (active_workers_num > 1) ? active_workers_num * 4 : 1;
+ return (worker_contexts [generation].active_workers_num > 1) ? worker_contexts [generation].active_workers_num * 4 : 1;
}
void
-sgen_workers_foreach (SgenWorkerCallback callback)
+sgen_workers_foreach (int generation, SgenWorkerCallback callback)
{
+ WorkerContext *context = &worker_contexts [generation];
int i;
- for (i = 0; i < workers_num; i++)
- callback (&workers_data [i]);
+ for (i = 0; i < context->workers_num; i++)
+ callback (&context->workers_data [i]);
}
gboolean
sgen_workers_is_worker_thread (MonoNativeThreadId id)
{
- if (!pool)
- return FALSE;
- return sgen_thread_pool_is_thread_pool_thread (pool, id);
+ return sgen_thread_pool_is_thread_pool_thread (id);
}
#endif
#include "mono/sgen/sgen-thread-pool.h"
typedef struct _WorkerData WorkerData;
+typedef struct _WorkerContext WorkerContext;
+
+typedef gint32 State;
+
+typedef void (*SgenWorkersFinishCallback) (void);
+typedef void (*SgenWorkerCallback) (WorkerData *data);
+
struct _WorkerData {
- /*
- * Threadpool threads receive as their starting argument a WorkerData.
- * tp_data is meant for use inside the sgen thread pool and must be first.
- */
- SgenThreadPoolData tp_data;
gint32 state;
SgenGrayQueue private_gray_queue; /* only read/written by worker thread */
/*
* starts.
*/
gpointer free_block_lists;
+ WorkerContext *context;
};
-typedef void (*SgenWorkersFinishCallback) (void);
-typedef void (*SgenWorkerCallback) (WorkerData *data);
+struct _WorkerContext {
+ int workers_num;
+ int active_workers_num;
+ volatile gboolean started;
+ volatile gboolean forced_stop;
+ WorkerData *workers_data;
+
+ /*
+ * When using multiple workers, we need to have the last worker
+ * enqueue the preclean jobs (if there are any). This lock ensures
+ * that when the last worker takes it, all the other workers have
+ * gracefully finished, so it can restart them.
+ */
+ mono_mutex_t finished_lock;
+ volatile gboolean workers_finished;
+ int worker_awakenings;
+
+ SgenSectionGrayQueue workers_distribute_gray_queue;
+
+ SgenObjectOperations * volatile idle_func_object_ops;
+ SgenObjectOperations *idle_func_object_ops_par, *idle_func_object_ops_nopar;
+
+ /*
+ * finished_callback is called only when the workers finish work normally (when they
+ * are not forced to finish). The callback is used to enqueue preclean jobs.
+ */
+ volatile SgenWorkersFinishCallback finish_callback;
+
+ int generation;
+ int thread_pool_context;
+};
-void sgen_workers_init (int num_workers, SgenWorkerCallback callback);
-void sgen_workers_shutdown (void);
-void sgen_workers_stop_all_workers (void);
-void sgen_workers_set_num_active_workers (int num_workers);
-void sgen_workers_start_all_workers (SgenObjectOperations *object_ops_nopar, SgenObjectOperations *object_ops_par, SgenWorkersFinishCallback finish_job);
-void sgen_workers_init_distribute_gray_queue (void);
-void sgen_workers_enqueue_job (SgenThreadPoolJob *job, gboolean enqueue);
-void sgen_workers_distribute_gray_queue_sections (void);
-void sgen_workers_reset_data (void);
-void sgen_workers_join (void);
-gboolean sgen_workers_have_idle_work (void);
+void sgen_workers_create_context (int generation, int num_workers);
+void sgen_workers_stop_all_workers (int generation);
+void sgen_workers_set_num_active_workers (int generation, int num_workers);
+void sgen_workers_start_all_workers (int generation, SgenObjectOperations *object_ops_nopar, SgenObjectOperations *object_ops_par, SgenWorkersFinishCallback finish_job);
+void sgen_workers_enqueue_job (int generation, SgenThreadPoolJob *job, gboolean enqueue);
+void sgen_workers_join (int generation);
+gboolean sgen_workers_have_idle_work (int generation);
gboolean sgen_workers_all_done (void);
-gboolean sgen_workers_are_working (void);
-void sgen_workers_assert_gray_queue_is_empty (void);
-void sgen_workers_take_from_queue (SgenGrayQueue *queue);
-SgenObjectOperations* sgen_workers_get_idle_func_object_ops (void);
-int sgen_workers_get_job_split_count (void);
-void sgen_workers_foreach (SgenWorkerCallback callback);
+void sgen_workers_assert_gray_queue_is_empty (int generation);
+void sgen_workers_take_from_queue (int generation, SgenGrayQueue *queue);
+SgenObjectOperations* sgen_workers_get_idle_func_object_ops (WorkerData *worker);
+int sgen_workers_get_job_split_count (int generation);
+void sgen_workers_foreach (int generation, SgenWorkerCallback callback);
gboolean sgen_workers_is_worker_thread (MonoNativeThreadId id);
#endif
mono_pagesize (void)
{
static int saved_pagesize = 0;
+
if (saved_pagesize)
return saved_pagesize;
+
+ // Prefer sysconf () as it's signal safe.
+#if defined (HAVE_SYSCONF) && defined (_SC_PAGESIZE)
+ saved_pagesize = sysconf (_SC_PAGESIZE);
+#else
saved_pagesize = getpagesize ();
+#endif
+
return saved_pagesize;
}
class MSBuild (GitHubPackage):
def __init__ (self):
GitHubPackage.__init__ (self, 'mono', 'msbuild', '15.3',
- revision = '622cece0cf47460bf83f2931c5788134bac4c48d',
+ revision = 'bce6977da8fdb05f30d80bfc5649d9edd543b026',
git_branch = 'xplat-master')
def build (self):