dasm, ctrl memstream, & more moving to async layer, rather than per-layer thread pools

This commit is contained in:
Ryan Fleury
2024-11-03 12:11:24 -08:00
parent c2855a9a46
commit a40edbba51
9 changed files with 408 additions and 737 deletions
+3 -12
View File
@@ -1299,12 +1299,6 @@ ctrl_init(void)
ctrl_state->u2ms_ring_cv = os_condition_variable_alloc(); ctrl_state->u2ms_ring_cv = os_condition_variable_alloc();
ctrl_state->ctrl_thread_log = log_alloc(); ctrl_state->ctrl_thread_log = log_alloc();
ctrl_state->ctrl_thread = os_thread_launch(ctrl_thread__entry_point, 0, 0); ctrl_state->ctrl_thread = os_thread_launch(ctrl_thread__entry_point, 0, 0);
ctrl_state->ms_thread_count = Clamp(1, os_get_system_info()->logical_processor_count-1, 4);
ctrl_state->ms_threads = push_array(arena, OS_Handle, ctrl_state->ms_thread_count);
for(U64 idx = 0; idx < ctrl_state->ms_thread_count; idx += 1)
{
ctrl_state->ms_threads[idx] = os_thread_launch(ctrl_mem_stream_thread__entry_point, (void *)idx, 0);
}
} }
//////////////////////////////// ////////////////////////////////
@@ -1468,6 +1462,7 @@ ctrl_stored_hash_from_process_vaddr_range(CTRL_Handle process, Rng1U64 range, B3
} }
} }
} }
async_push_work(ctrl_mem_stream_work);
} }
} }
@@ -5854,13 +5849,9 @@ ctrl_u2ms_dequeue_req(CTRL_Handle *out_process, Rng1U64 *out_vaddr_range, B32 *o
//- rjf: entry point //- rjf: entry point
internal void ASYNC_WORK_DEF(ctrl_mem_stream_work)
ctrl_mem_stream_thread__entry_point(void *p)
{ {
ThreadNameF("[ctrl] mem stream thread #%I64u", (U64)p);
CTRL_ProcessMemoryCache *cache = &ctrl_state->process_memory_cache; CTRL_ProcessMemoryCache *cache = &ctrl_state->process_memory_cache;
for(;;)
{
//- rjf: unpack next request //- rjf: unpack next request
CTRL_Handle process = {0}; CTRL_Handle process = {0};
Rng1U64 vaddr_range = {0}; Rng1U64 vaddr_range = {0};
@@ -6019,5 +6010,5 @@ ctrl_mem_stream_thread__entry_point(void *p)
//- rjf: broadcast changes //- rjf: broadcast changes
os_condition_variable_broadcast(process_stripe->cv); os_condition_variable_broadcast(process_stripe->cv);
ProfEnd(); ProfEnd();
} return 0;
} }
+1 -4
View File
@@ -888,10 +888,6 @@ struct CTRL_State
U64 u2ms_ring_read_pos; U64 u2ms_ring_read_pos;
OS_Handle u2ms_ring_mutex; OS_Handle u2ms_ring_mutex;
OS_Handle u2ms_ring_cv; OS_Handle u2ms_ring_cv;
// rjf: memory stream threads
U64 ms_thread_count;
OS_Handle *ms_threads;
}; };
//////////////////////////////// ////////////////////////////////
@@ -1156,6 +1152,7 @@ internal B32 ctrl_u2ms_enqueue_req(CTRL_Handle process, Rng1U64 vaddr_range, B32
internal void ctrl_u2ms_dequeue_req(CTRL_Handle *out_process, Rng1U64 *out_vaddr_range, B32 *out_zero_terminated); internal void ctrl_u2ms_dequeue_req(CTRL_Handle *out_process, Rng1U64 *out_vaddr_range, B32 *out_zero_terminated);
//- rjf: entry point //- rjf: entry point
ASYNC_WORK_DEF(ctrl_mem_stream_work);
internal void ctrl_mem_stream_thread__entry_point(void *p); internal void ctrl_mem_stream_thread__entry_point(void *p);
#endif // CTRL_CORE_H #endif // CTRL_CORE_H
+4 -12
View File
@@ -274,12 +274,6 @@ dasm_init(void)
dasm_shared->u2p_ring_base = push_array_no_zero(arena, U8, dasm_shared->u2p_ring_size); dasm_shared->u2p_ring_base = push_array_no_zero(arena, U8, dasm_shared->u2p_ring_size);
dasm_shared->u2p_ring_cv = os_condition_variable_alloc(); dasm_shared->u2p_ring_cv = os_condition_variable_alloc();
dasm_shared->u2p_ring_mutex = os_mutex_alloc(); dasm_shared->u2p_ring_mutex = os_mutex_alloc();
dasm_shared->parse_thread_count = 1;
dasm_shared->parse_threads = push_array(arena, OS_Handle, dasm_shared->parse_thread_count);
for(U64 idx = 0; idx < dasm_shared->parse_thread_count; idx += 1)
{
dasm_shared->parse_threads[idx] = os_thread_launch(dasm_parse_thread__entry_point, (void *)idx, 0);
}
dasm_shared->evictor_detector_thread = os_thread_launch(dasm_evictor_detector_thread__entry_point, 0, 0); dasm_shared->evictor_detector_thread = os_thread_launch(dasm_evictor_detector_thread__entry_point, 0, 0);
} }
@@ -430,6 +424,7 @@ dasm_info_from_hash_params(DASM_Scope *scope, U128 hash, DASM_Params *params)
if(node_is_new) if(node_is_new)
{ {
dasm_u2p_enqueue_req(hash, params, max_U64); dasm_u2p_enqueue_req(hash, params, max_U64);
async_push_work(dasm_parse_work);
} }
} }
return info; return info;
@@ -522,12 +517,8 @@ dasm_u2p_dequeue_req(Arena *arena, U128 *hash_out, DASM_Params *params_out)
os_condition_variable_broadcast(dasm_shared->u2p_ring_cv); os_condition_variable_broadcast(dasm_shared->u2p_ring_cv);
} }
internal void ASYNC_WORK_DEF(dasm_parse_work)
dasm_parse_thread__entry_point(void *p)
{ {
ThreadNameF("[dasm] parse thread #%I64u", (U64)p);
for(;;)
{
Temp scratch = scratch_begin(0, 0); Temp scratch = scratch_begin(0, 0);
HS_Scope *hs_scope = hs_scope_open(); HS_Scope *hs_scope = hs_scope_open();
DI_Scope *di_scope = di_scope_open(); DI_Scope *di_scope = di_scope_open();
@@ -799,7 +790,7 @@ dasm_parse_thread__entry_point(void *p)
di_scope_close(di_scope); di_scope_close(di_scope);
hs_scope_close(hs_scope); hs_scope_close(hs_scope);
scratch_end(scratch); scratch_end(scratch);
} return 0;
} }
//////////////////////////////// ////////////////////////////////
@@ -870,6 +861,7 @@ dasm_evictor_detector_thread__entry_point(void *p)
{ {
if(dasm_u2p_enqueue_req(n->hash, &n->params, max_U64)) if(dasm_u2p_enqueue_req(n->hash, &n->params, max_U64))
{ {
async_push_work(dasm_parse_work);
n->last_time_requested_us = os_now_microseconds(); n->last_time_requested_us = os_now_microseconds();
n->last_user_clock_idx_requested = check_time_user_clocks; n->last_user_clock_idx_requested = check_time_user_clocks;
} }
+1 -5
View File
@@ -263,10 +263,6 @@ struct DASM_Shared
OS_Handle u2p_ring_cv; OS_Handle u2p_ring_cv;
OS_Handle u2p_ring_mutex; OS_Handle u2p_ring_mutex;
// rjf: parse threads
U64 parse_thread_count;
OS_Handle *parse_threads;
// rjf: evictor/detector thread // rjf: evictor/detector thread
OS_Handle evictor_detector_thread; OS_Handle evictor_detector_thread;
}; };
@@ -329,7 +325,7 @@ internal DASM_Info dasm_info_from_key_params(DASM_Scope *scope, U128 key, DASM_P
internal B32 dasm_u2p_enqueue_req(U128 hash, DASM_Params *params, U64 endt_us); internal B32 dasm_u2p_enqueue_req(U128 hash, DASM_Params *params, U64 endt_us);
internal void dasm_u2p_dequeue_req(Arena *arena, U128 *hash_out, DASM_Params *params_out); internal void dasm_u2p_dequeue_req(Arena *arena, U128 *hash_out, DASM_Params *params_out);
internal void dasm_parse_thread__entry_point(void *p); ASYNC_WORK_DEF(dasm_parse_work);
//////////////////////////////// ////////////////////////////////
//~ rjf: Evictor/Detector Thread //~ rjf: Evictor/Detector Thread
-306
View File
@@ -878,309 +878,3 @@ ASYNC_WORK_DEF(di_parse_work)
scratch_end(scratch); scratch_end(scratch);
return 0; return 0;
} }
internal void
di_parse_thread__entry_point(void *p)
{
ThreadNameF("[di] parse #%I64u", (U64)p);
for(;;)
{
Temp scratch = scratch_begin(0, 0);
////////////////////////////
//- rjf: grab next key
//
DI_Key key = {0};
di_u2p_dequeue_key(scratch.arena, &key);
String8 og_path = key.path;
U64 min_timestamp = key.min_timestamp;
////////////////////////////
//- rjf: unpack key
//
U64 hash = di_hash_from_string(og_path, StringMatchFlag_CaseInsensitive);
U64 slot_idx = hash%di_shared->slots_count;
U64 stripe_idx = slot_idx%di_shared->stripes_count;
DI_Slot *slot = &di_shared->slots[slot_idx];
DI_Stripe *stripe = &di_shared->stripes[stripe_idx];
////////////////////////////
//- rjf: take task
//
B32 got_task = 0;
OS_MutexScopeR(stripe->rw_mutex)
{
DI_Node *node = di_node_from_key_slot__stripe_mutex_r_guarded(slot, &key);
if(node != 0)
{
got_task = !ins_atomic_u64_eval_cond_assign(&node->is_working, 1, 0);
}
}
////////////////////////////
//- rjf: got task -> open O.G. file (may or may not be RDI)
//
B32 og_format_is_known = 0;
B32 og_is_pe = 0;
B32 og_is_pdb = 0;
B32 og_is_elf = 0;
B32 og_is_rdi = 0;
FileProperties og_props = {0};
if(got_task) ProfScope("analyze %.*s", str8_varg(og_path))
{
OS_Handle file = os_file_open(OS_AccessFlag_Read|OS_AccessFlag_ShareRead, og_path);
OS_Handle file_map = os_file_map_open(OS_AccessFlag_Read, file);
FileProperties props = og_props = os_properties_from_file(file);
void *base = os_file_map_view_open(file_map, OS_AccessFlag_Read, r1u64(0, props.size));
String8 data = str8((U8 *)base, props.size);
if(!og_format_is_known)
{
String8 msf20_magic = str8_lit("Microsoft C/C++ program database 2.00\r\n\x1aJG\0\0");
String8 msf70_magic = str8_lit("Microsoft C/C++ MSF 7.00\r\n\032DS\0\0");
String8 msfxx_magic = str8_lit("Microsoft C/C++");
if((data.size >= msf20_magic.size && str8_match(data, msf20_magic, StringMatchFlag_RightSideSloppy)) ||
(data.size >= msf70_magic.size && str8_match(data, msf70_magic, StringMatchFlag_RightSideSloppy)) ||
(data.size >= msfxx_magic.size && str8_match(data, msfxx_magic, StringMatchFlag_RightSideSloppy)))
{
og_format_is_known = 1;
og_is_pdb = 1;
}
}
if(!og_format_is_known)
{
if(data.size >= 8 && *(U64 *)data.str == RDI_MAGIC_CONSTANT)
{
og_format_is_known = 1;
og_is_rdi = 1;
}
}
if(!og_format_is_known)
{
if(data.size >= 4 &&
data.str[0] == 0x7f &&
data.str[1] == 'E' &&
data.str[2] == 'L' &&
data.str[3] == 'F')
{
og_format_is_known = 1;
og_is_elf = 1;
}
}
if(!og_format_is_known)
{
if(data.size >= 2 && *(U16 *)data.str == 0x5a4d)
{
og_format_is_known = 1;
og_is_pe = 1;
}
}
os_file_map_view_close(file_map, base, r1u64(0, props.size));
os_file_map_close(file_map);
os_file_close(file);
}
////////////////////////////
//- rjf: given O.G. path & analysis, determine RDI path
//
String8 rdi_path = {0};
if(got_task)
{
if(og_is_rdi)
{
rdi_path = og_path;
}
else if(og_format_is_known && og_is_pdb)
{
rdi_path = push_str8f(scratch.arena, "%S.rdi", str8_chop_last_dot(og_path));
}
}
////////////////////////////
//- rjf: check if rdi file is up-to-date
//
B32 rdi_file_is_up_to_date = 0;
if(got_task)
{
if(rdi_path.size != 0) ProfScope("check %.*s is up-to-date", str8_varg(rdi_path))
{
FileProperties props = os_properties_from_file_path(rdi_path);
rdi_file_is_up_to_date = (props.modified > og_props.modified);
}
}
////////////////////////////
//- rjf: if raddbg file is up to date based on timestamp, check the
// encoding generation number & size, to see if we need to regenerate it
// regardless
//
if(got_task && rdi_file_is_up_to_date) ProfScope("check %.*s version matches our's", str8_varg(rdi_path))
{
OS_Handle file = {0};
OS_Handle file_map = {0};
FileProperties file_props = {0};
void *file_base = 0;
file = os_file_open(OS_AccessFlag_Read|OS_AccessFlag_ShareRead, rdi_path);
file_map = os_file_map_open(OS_AccessFlag_Read, file);
file_props = os_properties_from_file(file);
file_base = os_file_map_view_open(file_map, OS_AccessFlag_Read, r1u64(0, file_props.size));
if(sizeof(RDI_Header) <= file_props.size)
{
RDI_Header *header = (RDI_Header*)file_base;
if(header->encoding_version != RDI_ENCODING_VERSION)
{
rdi_file_is_up_to_date = 0;
}
}
else
{
rdi_file_is_up_to_date = 0;
}
os_file_map_view_close(file_map, file_base, r1u64(0, file_props.size));
os_file_map_close(file_map);
os_file_close(file);
}
////////////////////////////
//- rjf: heuristically choose compression settings
//
B32 should_compress = 0;
#if 0
if(og_dbg_props.size > MB(64))
{
should_compress = 1;
}
#endif
////////////////////////////
//- rjf: rdi file not up-to-date? we need to generate it
//
if(got_task && !rdi_file_is_up_to_date) ProfScope("generate %.*s", str8_varg(rdi_path))
{
if(og_is_pdb)
{
//- rjf: push conversion task begin event
{
DI_Event event = {DI_EventKind_ConversionStarted};
event.string = rdi_path;
di_p2u_push_event(&event);
}
//- rjf: kick off process
OS_Handle process = {0};
{
OS_ProcessLaunchParams params = {0};
params.path = os_get_process_info()->binary_path;
params.inherit_env = 1;
params.consoleless = 1;
str8_list_pushf(scratch.arena, &params.cmd_line, "raddbg");
str8_list_pushf(scratch.arena, &params.cmd_line, "--convert");
str8_list_pushf(scratch.arena, &params.cmd_line, "--quiet");
if(should_compress)
{
str8_list_pushf(scratch.arena, &params.cmd_line, "--compress");
}
// str8_list_pushf(scratch.arena, &params.cmd_line, "--capture");
str8_list_pushf(scratch.arena, &params.cmd_line, "--pdb:%S", og_path);
str8_list_pushf(scratch.arena, &params.cmd_line, "--out:%S", rdi_path);
process = os_process_launch(&params);
}
//- rjf: wait for process to complete
{
U64 start_wait_t = os_now_microseconds();
for(;;)
{
B32 wait_done = os_process_join(process, os_now_microseconds()+1000);
if(wait_done)
{
rdi_file_is_up_to_date = 1;
break;
}
}
}
//- rjf: push conversion task end event
{
DI_Event event = {DI_EventKind_ConversionEnded};
event.string = rdi_path;
di_p2u_push_event(&event);
}
}
else
{
// NOTE(rjf): we cannot convert from this O.G. debug info format right now.
//- rjf: push conversion task failure event
{
DI_Event event = {DI_EventKind_ConversionFailureUnsupportedFormat};
event.string = rdi_path;
di_p2u_push_event(&event);
}
}
}
////////////////////////////
//- rjf: got task -> open file
//
OS_Handle file = {0};
OS_Handle file_map = {0};
FileProperties file_props = {0};
void *file_base = 0;
if(got_task)
{
file = os_file_open(OS_AccessFlag_Read|OS_AccessFlag_ShareRead|OS_AccessFlag_ShareWrite, rdi_path);
file_map = os_file_map_open(OS_AccessFlag_Read, file);
file_props = os_properties_from_file(file);
file_base = os_file_map_view_open(file_map, OS_AccessFlag_Read, r1u64(0, file_props.size));
}
////////////////////////////
//- rjf: do initial parse of rdi
//
RDI_Parsed rdi_parsed_maybe_compressed = di_rdi_parsed_nil;
if(got_task)
{
RDI_ParseStatus parse_status = rdi_parse((U8 *)file_base, file_props.size, &rdi_parsed_maybe_compressed);
(void)parse_status;
}
////////////////////////////
//- rjf: decompress & re-parse, if necessary
//
Arena *rdi_parsed_arena = 0;
RDI_Parsed rdi_parsed = rdi_parsed_maybe_compressed;
if(got_task)
{
U64 decompressed_size = rdi_decompressed_size_from_parsed(&rdi_parsed_maybe_compressed);
if(decompressed_size > file_props.size)
{
rdi_parsed_arena = arena_alloc();
U8 *decompressed_data = push_array_no_zero(rdi_parsed_arena, U8, decompressed_size);
rdi_decompress_parsed(decompressed_data, decompressed_size, &rdi_parsed_maybe_compressed);
RDI_ParseStatus parse_status = rdi_parse(decompressed_data, decompressed_size, &rdi_parsed);
(void)parse_status;
}
}
////////////////////////////
//- rjf: commit parsed info to cache
//
if(got_task) OS_MutexScopeW(stripe->rw_mutex)
{
DI_Node *node = di_node_from_key_slot__stripe_mutex_r_guarded(slot, &key);
if(node != 0)
{
node->is_working = 0;
node->file = file;
node->file_map = file_map;
node->file_base = file_base;
node->file_props = file_props;
node->arena = rdi_parsed_arena;
node->rdi = rdi_parsed;
node->parse_done = 1;
}
}
os_condition_variable_broadcast(stripe->cv);
scratch_end(scratch);
}
}
-1
View File
@@ -248,6 +248,5 @@ internal void di_p2u_push_event(DI_Event *event);
internal DI_EventList di_p2u_pop_events(Arena *arena, U64 endt_us); internal DI_EventList di_p2u_pop_events(Arena *arena, U64 endt_us);
ASYNC_WORK_DEF(di_parse_work); ASYNC_WORK_DEF(di_parse_work);
internal void di_parse_thread__entry_point(void *p);
#endif // DBGI_H #endif // DBGI_H
+1 -1
View File
@@ -423,7 +423,7 @@ ASYNC_WORK_DEF(fs_stream_work)
internal void internal void
fs_detector_thread__entry_point(void *p) fs_detector_thread__entry_point(void *p)
{ {
ThreadNameF("[fs] detector"); ThreadNameF("[fs] detector thread");
for(;;) for(;;)
{ {
U64 slots_per_stripe = fs_shared->slots_count/fs_shared->stripes_count; U64 slots_per_stripe = fs_shared->slots_count/fs_shared->stripes_count;
+1
View File
@@ -296,6 +296,7 @@ hs_data_from_hash(HS_Scope *scope, U128 hash)
internal void internal void
hs_evictor_thread__entry_point(void *p) hs_evictor_thread__entry_point(void *p)
{ {
ThreadNameF("[hs] evictor thread");
for(;;) for(;;)
{ {
for(U64 slot_idx = 0; slot_idx < hs_shared->slots_count; slot_idx += 1) for(U64 slot_idx = 0; slot_idx < hs_shared->slots_count; slot_idx += 1)
+1
View File
@@ -2347,6 +2347,7 @@ ASYNC_WORK_DEF(txt_parse_work)
internal void internal void
txt_evictor_thread__entry_point(void *p) txt_evictor_thread__entry_point(void *p)
{ {
ThreadNameF("[txt] evictor thread");
for(;;) for(;;)
{ {
U64 check_time_us = os_now_microseconds(); U64 check_time_us = os_now_microseconds();