2014-05-03 18:27:38 -04:00
/*
2017-06-30 03:22:17 -04:00
* This file is part of the MicroPython project , http : //micropython.org/
2014-05-03 18:27:38 -04:00
*
* The MIT License ( MIT )
*
2020-06-03 18:40:05 -04:00
* SPDX - FileCopyrightText : Copyright ( c ) 2013 , 2014 Damien P . George
2021-04-22 20:55:39 -04:00
* SPDX - FileCopyrightText : Copyright ( c ) 2014 Paul Sokolovsky
2014-05-03 18:27:38 -04:00
*
* Permission is hereby granted , free of charge , to any person obtaining a copy
* of this software and associated documentation files ( the " Software " ) , to deal
* in the Software without restriction , including without limitation the rights
* to use , copy , modify , merge , publish , distribute , sublicense , and / or sell
* copies of the Software , and to permit persons to whom the Software is
* furnished to do so , subject to the following conditions :
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software .
*
* THE SOFTWARE IS PROVIDED " AS IS " , WITHOUT WARRANTY OF ANY KIND , EXPRESS OR
* IMPLIED , INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY ,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT . IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM , DAMAGES OR OTHER
* LIABILITY , WHETHER IN AN ACTION OF CONTRACT , TORT OR OTHERWISE , ARISING FROM ,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE .
*/
2014-04-20 04:45:16 -04:00
# include <assert.h>
2013-10-21 18:45:08 -04:00
# include <stdio.h>
# include <string.h>
2015-01-01 15:27:54 -05:00
# include "py/gc.h"
# include "py/runtime.h"
2014-04-03 17:55:12 -04:00
2021-08-27 10:02:03 -04:00
# if MICROPY_DEBUG_VALGRIND
# include <valgrind/memcheck.h>
# endif
2019-03-08 18:29:14 -05:00
# include "supervisor/shared/safe_mode.h"
2020-07-15 20:58:38 -04:00
# if CIRCUITPY_MEMORYMONITOR
2020-07-16 22:01:43 -04:00
# include "shared-module/memorymonitor/__init__.h"
2020-07-15 20:58:38 -04:00
# endif
2014-01-07 10:20:33 -05:00
# if MICROPY_ENABLE_GC
2017-07-24 12:55:14 -04:00
# if MICROPY_DEBUG_VERBOSE // print debugging info
2014-02-10 14:45:54 -05:00
# define DEBUG_PRINT (1)
2014-02-16 11:11:42 -05:00
# define DEBUG_printf DEBUG_printf
2014-02-10 14:45:54 -05:00
# else // don't print debugging info
2014-11-05 16:16:41 -05:00
# define DEBUG_PRINT (0)
2014-02-26 17:40:35 -05:00
# define DEBUG_printf(...) (void)0
2014-02-10 14:45:54 -05:00
# endif
2018-01-23 19:22:05 -05:00
// Uncomment this if you want to use a debugger to capture state at every allocation and free.
// #define LOG_HEAP_ACTIVITY 1
2014-08-28 18:06:38 -04:00
// make this 1 to dump the heap each time it changes
# define EXTENSIVE_HEAP_PROFILING (0)
2017-12-07 17:10:44 -05:00
// make this 1 to zero out swept memory to more eagerly
// detect untraced object still in use
# define CLEAR_ON_SWEEP (0)
2013-10-21 18:45:08 -04:00
// ATB = allocation table byte
// 0b00 = FREE -- free block
// 0b01 = HEAD -- head of a chain of blocks
// 0b10 = TAIL -- in the tail of a chain of blocks
// 0b11 = MARK -- marked head block
# define AT_FREE (0)
# define AT_HEAD (1)
# define AT_TAIL (2)
# define AT_MARK (3)
# define BLOCKS_PER_ATB (4)
# define BLOCK_SHIFT(block) (2 * ((block) & (BLOCKS_PER_ATB - 1)))
2015-01-01 18:30:53 -05:00
# define ATB_GET_KIND(block) ((MP_STATE_MEM(gc_alloc_table_start)[(block) / BLOCKS_PER_ATB] >> BLOCK_SHIFT(block)) & 3)
# define ATB_ANY_TO_FREE(block) do { MP_STATE_MEM(gc_alloc_table_start)[(block) / BLOCKS_PER_ATB] &= (~(AT_MARK << BLOCK_SHIFT(block))); } while (0)
# define ATB_FREE_TO_HEAD(block) do { MP_STATE_MEM(gc_alloc_table_start)[(block) / BLOCKS_PER_ATB] |= (AT_HEAD << BLOCK_SHIFT(block)); } while (0)
# define ATB_FREE_TO_TAIL(block) do { MP_STATE_MEM(gc_alloc_table_start)[(block) / BLOCKS_PER_ATB] |= (AT_TAIL << BLOCK_SHIFT(block)); } while (0)
# define ATB_HEAD_TO_MARK(block) do { MP_STATE_MEM(gc_alloc_table_start)[(block) / BLOCKS_PER_ATB] |= (AT_MARK << BLOCK_SHIFT(block)); } while (0)
# define ATB_MARK_TO_HEAD(block) do { MP_STATE_MEM(gc_alloc_table_start)[(block) / BLOCKS_PER_ATB] &= (~(AT_TAIL << BLOCK_SHIFT(block))); } while (0)
2020-02-26 23:36:53 -05:00
# define BLOCK_FROM_PTR(ptr) (((byte *)(ptr) - MP_STATE_MEM(gc_pool_start)) / BYTES_PER_BLOCK)
2015-11-27 08:07:48 -05:00
# define PTR_FROM_BLOCK(block) (((block) * BYTES_PER_BLOCK + (uintptr_t)MP_STATE_MEM(gc_pool_start)))
2013-10-21 18:45:08 -04:00
# define ATB_FROM_BLOCK(bl) ((bl) / BLOCKS_PER_ATB)
2014-04-05 15:35:48 -04:00
# if MICROPY_ENABLE_FINALISER
// FTB = finaliser table byte
// if set, then the corresponding block may have a finaliser
# define BLOCKS_PER_FTB (8)
2015-01-01 18:30:53 -05:00
# define FTB_GET(block) ((MP_STATE_MEM(gc_finaliser_table_start)[(block) / BLOCKS_PER_FTB] >> ((block) & 7)) & 1)
# define FTB_SET(block) do { MP_STATE_MEM(gc_finaliser_table_start)[(block) / BLOCKS_PER_FTB] |= (1 << ((block) & 7)); } while (0)
# define FTB_CLEAR(block) do { MP_STATE_MEM(gc_finaliser_table_start)[(block) / BLOCKS_PER_FTB] &= (~(1 << ((block) & 7))); } while (0)
2014-04-05 15:35:48 -04:00
# endif
2016-05-26 06:53:34 -04:00
# if MICROPY_PY_THREAD && !MICROPY_PY_THREAD_GIL
2016-04-25 11:28:57 -04:00
# define GC_ENTER() mp_thread_mutex_lock(&MP_STATE_MEM(gc_mutex), 1)
# define GC_EXIT() mp_thread_mutex_unlock(&MP_STATE_MEM(gc_mutex))
# else
# define GC_ENTER()
# define GC_EXIT()
# endif
2016-11-23 20:31:53 -05:00
# ifdef LOG_HEAP_ACTIVITY
volatile uint32_t change_me ;
2017-05-05 15:04:20 -04:00
# pragma GCC push_options
# pragma GCC optimize ("O0")
2016-11-23 20:31:53 -05:00
void __attribute__ ( ( noinline ) ) gc_log_change ( uint32_t start_block , uint32_t length ) {
change_me + = start_block ;
change_me + = length ; // Break on this line.
}
2017-05-05 15:04:20 -04:00
# pragma GCC pop_options
2016-11-23 20:31:53 -05:00
# endif
2013-10-22 16:12:29 -04:00
// TODO waste less memory; currently requires that all entries in alloc_table have a corresponding block in pool
void gc_init ( void * start , void * end ) {
// align end pointer on block boundary
2020-02-26 23:36:53 -05:00
end = ( void * ) ( ( uintptr_t ) end & ( ~ ( BYTES_PER_BLOCK - 1 ) ) ) ;
DEBUG_printf ( " Initializing GC heap: %p..%p = " UINT_FMT " bytes \n " , start , end , ( byte * ) end - ( byte * ) start ) ;
2014-04-05 15:35:48 -04:00
// calculate parameters for GC (T=total, A=alloc table, F=finaliser table, P=pool; all in bytes):
// T = A + F + P
// F = A * BLOCKS_PER_ATB / BLOCKS_PER_FTB
// P = A * BLOCKS_PER_ATB * BYTES_PER_BLOCK
// => T = A * (1 + BLOCKS_PER_ATB / BLOCKS_PER_FTB + BLOCKS_PER_ATB * BYTES_PER_BLOCK)
2020-02-26 23:36:53 -05:00
size_t total_byte_len = ( byte * ) end - ( byte * ) start ;
# if MICROPY_ENABLE_FINALISER
2021-08-27 10:04:32 -04:00
MP_STATE_MEM ( gc_alloc_table_byte_len ) = ( total_byte_len - 1 ) * MP_BITS_PER_BYTE / ( MP_BITS_PER_BYTE + MP_BITS_PER_BYTE * BLOCKS_PER_ATB / BLOCKS_PER_FTB + MP_BITS_PER_BYTE * BLOCKS_PER_ATB * BYTES_PER_BLOCK ) ;
2020-02-26 23:36:53 -05:00
# else
2021-02-03 23:32:59 -05:00
MP_STATE_MEM ( gc_alloc_table_byte_len ) = total_byte_len / ( 1 + MP_BITS_PER_BYTE / 2 * BYTES_PER_BLOCK ) ;
2020-02-26 23:36:53 -05:00
# endif
2014-04-05 15:35:48 -04:00
2020-02-26 23:36:53 -05:00
MP_STATE_MEM ( gc_alloc_table_start ) = ( byte * ) start ;
2014-04-03 17:55:12 -04:00
2020-02-26 23:36:53 -05:00
# if MICROPY_ENABLE_FINALISER
2015-12-16 20:09:11 -05:00
size_t gc_finaliser_table_byte_len = ( MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB + BLOCKS_PER_FTB - 1 ) / BLOCKS_PER_FTB ;
2021-08-27 10:04:32 -04:00
MP_STATE_MEM ( gc_finaliser_table_start ) = MP_STATE_MEM ( gc_alloc_table_start ) + MP_STATE_MEM ( gc_alloc_table_byte_len ) + 1 ;
2020-02-26 23:36:53 -05:00
# endif
2014-04-03 17:55:12 -04:00
2015-12-16 20:09:11 -05:00
size_t gc_pool_block_len = MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB ;
2020-02-26 23:36:53 -05:00
MP_STATE_MEM ( gc_pool_start ) = ( byte * ) end - gc_pool_block_len * BYTES_PER_BLOCK ;
2015-11-27 08:07:48 -05:00
MP_STATE_MEM ( gc_pool_end ) = end ;
2013-10-22 16:12:29 -04:00
2020-02-26 23:36:53 -05:00
# if MICROPY_ENABLE_FINALISER
2015-11-27 08:07:48 -05:00
assert ( MP_STATE_MEM ( gc_pool_start ) > = MP_STATE_MEM ( gc_finaliser_table_start ) + gc_finaliser_table_byte_len ) ;
2020-02-26 23:36:53 -05:00
# endif
2014-08-08 07:33:49 -04:00
2021-09-14 19:24:37 -04:00
// Clear ATBs plus one more byte. The extra byte might be read when we read the final ATB and
// then try to count its tail. Clearing the byte ensures it is 0 and ends the chain. Without an
// FTB, it'll just clear the pool byte early.
memset ( MP_STATE_MEM ( gc_alloc_table_start ) , 0 , MP_STATE_MEM ( gc_alloc_table_byte_len ) + 1 ) ;
2013-10-22 16:12:29 -04:00
2020-02-26 23:36:53 -05:00
# if MICROPY_ENABLE_FINALISER
2014-04-05 15:35:48 -04:00
// clear FTBs
2015-01-01 18:30:53 -05:00
memset ( MP_STATE_MEM ( gc_finaliser_table_start ) , 0 , gc_finaliser_table_byte_len ) ;
2020-02-26 23:36:53 -05:00
# endif
2014-04-03 17:55:12 -04:00
2018-01-23 19:22:05 -05:00
// Set first free ATB index to the start of the heap.
2020-02-11 20:06:24 -05:00
for ( size_t i = 0 ; i < MICROPY_ATB_INDICES ; i + + ) {
MP_STATE_MEM ( gc_first_free_atb_index ) [ i ] = 0 ;
}
2018-01-23 19:22:05 -05:00
// Set last free ATB index to the end of the heap.
MP_STATE_MEM ( gc_last_free_atb_index ) = MP_STATE_MEM ( gc_alloc_table_byte_len ) - 1 ;
2020-02-11 20:06:24 -05:00
2018-01-23 19:22:05 -05:00
// Set the lowest long lived ptr to the end of the heap to start. This will be lowered as long
// lived objects are allocated.
2021-03-15 09:57:36 -04:00
MP_STATE_MEM ( gc_lowest_long_lived_ptr ) = ( void * ) PTR_FROM_BLOCK ( MP_STATE_MEM ( gc_alloc_table_byte_len * BLOCKS_PER_ATB ) ) ;
2014-08-22 13:17:02 -04:00
2014-04-05 15:35:48 -04:00
// unlock the GC
2021-05-04 09:56:43 -04:00
MP_STATE_THREAD ( gc_lock_depth ) = 0 ;
2014-04-05 15:35:48 -04:00
2014-10-31 17:30:46 -04:00
// allow auto collection
2018-01-23 19:22:05 -05:00
MP_STATE_MEM ( gc_auto_collect_enabled ) = true ;
2014-10-31 17:30:46 -04:00
py/gc: Implement GC running by allocation threshold.
Currently, MicroPython runs GC when it could not allocate a block of memory,
which happens when heap is exhausted. However, that policy can't work well
with "inifinity" heaps, e.g. backed by a virtual memory - there will be a
lot of swap thrashing long before VM will be exhausted. Instead, in such
cases "allocation threshold" policy is used: a GC is run after some number of
allocations have been made. Details vary, for example, number or total amount
of allocations can be used, threshold may be self-adjusting based on GC
outcome, etc.
This change implements a simple variant of such policy for MicroPython. Amount
of allocated memory so far is used for threshold, to make it useful to typical
finite-size, and small, heaps as used with MicroPython ports. And such GC policy
is indeed useful for such types of heaps too, as it allows to better control
fragmentation. For example, if a threshold is set to half size of heap, then
for an application which usually makes big number of small allocations, that
will (try to) keep half of heap memory in a nice defragmented state for an
occasional large allocation.
For an application which doesn't exhibit such behavior, there won't be any
visible effects, except for GC running more frequently, which however may
affect performance. To address this, the GC threshold is configurable, and
by default is off so far. It's configured with gc.threshold(amount_in_bytes)
call (can be queries without an argument).
2016-07-20 17:37:30 -04:00
# if MICROPY_GC_ALLOC_THRESHOLD
// by default, maxuint for gc threshold, effectively turning gc-by-threshold off
MP_STATE_MEM ( gc_alloc_threshold ) = ( size_t ) - 1 ;
MP_STATE_MEM ( gc_alloc_amount ) = 0 ;
# endif
2020-01-22 12:19:37 -05:00
# if MICROPY_PY_THREAD && !MICROPY_PY_THREAD_GIL
2016-04-25 11:28:57 -04:00
mp_thread_mutex_init ( & MP_STATE_MEM ( gc_mutex ) ) ;
# endif
2019-04-08 19:58:50 -04:00
MP_STATE_MEM ( permanent_pointers ) = NULL ;
2014-02-10 14:45:54 -05:00
DEBUG_printf ( " GC layout: \n " ) ;
2015-01-01 18:30:53 -05:00
DEBUG_printf ( " alloc table at %p, length " UINT_FMT " bytes, " UINT_FMT " blocks \n " , MP_STATE_MEM ( gc_alloc_table_start ) , MP_STATE_MEM ( gc_alloc_table_byte_len ) , MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB ) ;
2020-02-26 23:36:53 -05:00
# if MICROPY_ENABLE_FINALISER
2015-01-01 18:30:53 -05:00
DEBUG_printf ( " finaliser table at %p, length " UINT_FMT " bytes, " UINT_FMT " blocks \n " , MP_STATE_MEM ( gc_finaliser_table_start ) , gc_finaliser_table_byte_len , gc_finaliser_table_byte_len * BLOCKS_PER_FTB ) ;
2020-02-26 23:36:53 -05:00
# endif
2015-01-01 18:30:53 -05:00
DEBUG_printf ( " pool at %p, length " UINT_FMT " bytes, " UINT_FMT " blocks \n " , MP_STATE_MEM ( gc_pool_start ) , gc_pool_block_len * BYTES_PER_BLOCK , gc_pool_block_len ) ;
2013-10-22 16:12:29 -04:00
}
2019-03-12 17:05:31 -04:00
void gc_deinit ( void ) {
2021-02-25 19:50:57 -05:00
// Run any finalisers before we stop using the heap.
2019-03-08 18:29:14 -05:00
gc_sweep_all ( ) ;
MP_STATE_MEM ( gc_pool_start ) = 0 ;
}
2014-04-08 07:31:21 -04:00
void gc_lock ( void ) {
2021-05-04 09:56:43 -04:00
// This does not need to be atomic or have the GC mutex because:
// - each thread has its own gc_lock_depth so there are no races between threads;
// - a hard interrupt will only change gc_lock_depth during its execution, and
// upon return will restore the value of gc_lock_depth.
MP_STATE_THREAD ( gc_lock_depth ) + + ;
2014-04-08 07:31:21 -04:00
}
void gc_unlock ( void ) {
2021-05-04 09:56:43 -04:00
// This does not need to be atomic, See comment above in gc_lock.
MP_STATE_THREAD ( gc_lock_depth ) - - ;
2014-04-08 07:31:21 -04:00
}
2014-07-01 01:49:21 -04:00
bool gc_is_locked ( void ) {
2021-05-04 09:56:43 -04:00
return MP_STATE_THREAD ( gc_lock_depth ) ! = 0 ;
2014-07-01 01:49:21 -04:00
}
2017-12-08 18:48:26 -05:00
# ifndef TRACE_MARK
# if DEBUG_PRINT
# define TRACE_MARK(block, ptr) DEBUG_printf("gc_mark(%p)\n", ptr)
# else
# define TRACE_MARK(block, ptr)
# endif
# endif
2018-02-15 11:10:13 -05:00
// Take the given block as the topmost block on the stack. Check all it's
// children: mark the unmarked child blocks and put those newly marked
// blocks on the stack. When all children have been checked, pop off the
// topmost block on the stack and repeat with that one.
STATIC void gc_mark_subtree ( size_t block ) {
// Start with the block passed in the argument.
2018-02-15 11:22:34 -05:00
size_t sp = 0 ;
2018-02-15 11:10:13 -05:00
for ( ; ; ) {
2021-10-26 04:47:04 -04:00
MICROPY_GC_HOOK_LOOP
2013-10-22 10:25:25 -04:00
// work out number of consecutive blocks in the chain starting with this one
2015-12-16 20:09:11 -05:00
size_t n_blocks = 0 ;
2013-10-21 18:45:08 -04:00
do {
n_blocks + = 1 ;
} while ( ATB_GET_KIND ( block + n_blocks ) = = AT_TAIL ) ;
// check this block's children
2020-02-26 23:36:53 -05:00
void * * ptrs = ( void * * ) PTR_FROM_BLOCK ( block ) ;
for ( size_t i = n_blocks * BYTES_PER_BLOCK / sizeof ( void * ) ; i > 0 ; i - - , ptrs + + ) {
2021-10-26 04:47:04 -04:00
MICROPY_GC_HOOK_LOOP
2015-12-16 19:41:37 -05:00
void * ptr = * ptrs ;
2018-01-27 12:37:38 -05:00
if ( VERIFY_PTR ( ptr ) ) {
// Mark and push this pointer
size_t childblock = BLOCK_FROM_PTR ( ptr ) ;
if ( ATB_GET_KIND ( childblock ) = = AT_HEAD ) {
// an unmarked head, mark it, and push it on gc stack
TRACE_MARK ( childblock , ptr ) ;
ATB_HEAD_TO_MARK ( childblock ) ;
2018-02-15 11:22:34 -05:00
if ( sp < MICROPY_ALLOC_GC_STACK_SIZE ) {
MP_STATE_MEM ( gc_stack ) [ sp + + ] = childblock ;
2018-01-27 12:37:38 -05:00
} else {
MP_STATE_MEM ( gc_stack_overflow ) = 1 ;
}
}
}
2013-10-21 18:45:08 -04:00
}
2018-02-15 11:10:13 -05:00
// Are there any blocks on the stack?
2018-02-15 11:22:34 -05:00
if ( sp = = 0 ) {
2018-02-15 11:10:13 -05:00
break ; // No, stack is empty, we're done.
}
// pop the next block off the stack
2018-02-15 11:22:34 -05:00
block = MP_STATE_MEM ( gc_stack ) [ - - sp ] ;
2013-10-21 18:45:08 -04:00
}
}
2014-02-12 11:31:30 -05:00
STATIC void gc_deal_with_stack_overflow ( void ) {
2015-01-01 18:30:53 -05:00
while ( MP_STATE_MEM ( gc_stack_overflow ) ) {
MP_STATE_MEM ( gc_stack_overflow ) = 0 ;
2013-10-21 18:45:08 -04:00
// scan entire memory looking for blocks which have been marked but not their children
2015-12-16 20:09:11 -05:00
for ( size_t block = 0 ; block < MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB ; block + + ) {
2021-10-26 04:47:04 -04:00
MICROPY_GC_HOOK_LOOP
2013-10-21 18:45:08 -04:00
// trace (again) if mark bit set
if ( ATB_GET_KIND ( block ) = = AT_MARK ) {
2018-02-15 11:10:13 -05:00
gc_mark_subtree ( block ) ;
2013-10-21 18:45:08 -04:00
}
}
}
}
2014-02-12 11:31:30 -05:00
STATIC void gc_sweep ( void ) {
2014-06-05 15:48:02 -04:00
# if MICROPY_PY_GC_COLLECT_RETVAL
2015-02-07 12:24:10 -05:00
MP_STATE_MEM ( gc_collected ) = 0 ;
2014-06-05 15:48:02 -04:00
# endif
2013-10-21 18:45:08 -04:00
// free unmarked heads and their tails
int free_tail = 0 ;
2015-12-16 19:45:42 -05:00
for ( size_t block = 0 ; block < MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB ; block + + ) {
2021-10-26 04:47:04 -04:00
MICROPY_GC_HOOK_LOOP
2013-10-21 18:45:08 -04:00
switch ( ATB_GET_KIND ( block ) ) {
case AT_HEAD :
2020-02-26 23:36:53 -05:00
# if MICROPY_ENABLE_FINALISER
2014-04-05 15:35:48 -04:00
if ( FTB_GET ( block ) ) {
2020-02-26 23:36:53 -05:00
mp_obj_base_t * obj = ( mp_obj_base_t * ) PTR_FROM_BLOCK ( block ) ;
2015-12-16 19:45:42 -05:00
if ( obj - > type ! = NULL ) {
2014-04-05 15:35:48 -04:00
// if the object has a type then see if it has a __del__ method
mp_obj_t dest [ 2 ] ;
2015-12-16 19:45:42 -05:00
mp_load_method_maybe ( MP_OBJ_FROM_PTR ( obj ) , MP_QSTR___del__ , dest ) ;
2014-04-05 15:35:48 -04:00
if ( dest [ 0 ] ! = MP_OBJ_NULL ) {
2017-04-11 23:52:04 -04:00
// load_method returned a method, execute it in a protected environment
# if MICROPY_ENABLE_SCHEDULER
mp_sched_lock ( ) ;
# endif
mp_call_function_1_protected ( dest [ 0 ] , dest [ 1 ] ) ;
# if MICROPY_ENABLE_SCHEDULER
mp_sched_unlock ( ) ;
# endif
2014-04-05 15:35:48 -04:00
}
2014-04-03 17:55:12 -04:00
}
2014-04-05 15:35:48 -04:00
// clear finaliser flag
FTB_CLEAR ( block ) ;
2014-04-03 17:55:12 -04:00
}
2020-02-26 23:36:53 -05:00
# endif
2013-10-21 18:45:08 -04:00
free_tail = 1 ;
2021-01-27 09:44:53 -05:00
DEBUG_printf ( " gc_sweep(%p) \n " , ( void * ) PTR_FROM_BLOCK ( block ) ) ;
2016-11-23 20:31:53 -05:00
# ifdef LOG_HEAP_ACTIVITY
gc_log_change ( block , 0 ) ;
# endif
2014-06-05 15:48:02 -04:00
# if MICROPY_PY_GC_COLLECT_RETVAL
2015-02-07 12:24:10 -05:00
MP_STATE_MEM ( gc_collected ) + + ;
2014-06-05 15:48:02 -04:00
# endif
2019-11-28 06:47:21 -05:00
// fall through to free the head
MP_FALLTHROUGH
2013-10-21 18:45:08 -04:00
case AT_TAIL :
if ( free_tail ) {
ATB_ANY_TO_FREE ( block ) ;
2017-12-07 17:10:44 -05:00
# if CLEAR_ON_SWEEP
2020-02-26 23:36:53 -05:00
memset ( ( void * ) PTR_FROM_BLOCK ( block ) , 0 , BYTES_PER_BLOCK ) ;
2017-12-07 17:10:44 -05:00
# endif
2013-10-21 18:45:08 -04:00
}
break ;
case AT_MARK :
ATB_MARK_TO_HEAD ( block ) ;
free_tail = 0 ;
break ;
}
}
}
2019-04-18 18:44:44 -04:00
// Mark can handle NULL pointers because it verifies the pointer is within the heap bounds.
2021-03-15 09:57:36 -04:00
STATIC void gc_mark ( void * ptr ) {
2019-04-18 18:44:44 -04:00
if ( VERIFY_PTR ( ptr ) ) {
size_t block = BLOCK_FROM_PTR ( ptr ) ;
if ( ATB_GET_KIND ( block ) = = AT_HEAD ) {
// An unmarked head: mark it, and mark all its children
TRACE_MARK ( block , ptr ) ;
ATB_HEAD_TO_MARK ( block ) ;
gc_mark_subtree ( block ) ;
}
}
}
2013-10-23 15:20:17 -04:00
void gc_collect_start ( void ) {
2016-04-25 11:28:57 -04:00
GC_ENTER ( ) ;
2021-05-04 09:56:43 -04:00
MP_STATE_THREAD ( gc_lock_depth ) + + ;
py/gc: Implement GC running by allocation threshold.
Currently, MicroPython runs GC when it could not allocate a block of memory,
which happens when heap is exhausted. However, that policy can't work well
with "inifinity" heaps, e.g. backed by a virtual memory - there will be a
lot of swap thrashing long before VM will be exhausted. Instead, in such
cases "allocation threshold" policy is used: a GC is run after some number of
allocations have been made. Details vary, for example, number or total amount
of allocations can be used, threshold may be self-adjusting based on GC
outcome, etc.
This change implements a simple variant of such policy for MicroPython. Amount
of allocated memory so far is used for threshold, to make it useful to typical
finite-size, and small, heaps as used with MicroPython ports. And such GC policy
is indeed useful for such types of heaps too, as it allows to better control
fragmentation. For example, if a threshold is set to half size of heap, then
for an application which usually makes big number of small allocations, that
will (try to) keep half of heap memory in a nice defragmented state for an
occasional large allocation.
For an application which doesn't exhibit such behavior, there won't be any
visible effects, except for GC running more frequently, which however may
affect performance. To address this, the GC threshold is configurable, and
by default is off so far. It's configured with gc.threshold(amount_in_bytes)
call (can be queries without an argument).
2016-07-20 17:37:30 -04:00
# if MICROPY_GC_ALLOC_THRESHOLD
MP_STATE_MEM ( gc_alloc_amount ) = 0 ;
# endif
2015-01-01 18:30:53 -05:00
MP_STATE_MEM ( gc_stack_overflow ) = 0 ;
py: Introduce a Python stack for scoped allocation.
This patch introduces the MICROPY_ENABLE_PYSTACK option (disabled by
default) which enables a "Python stack" that allows to allocate and free
memory in a scoped, or Last-In-First-Out (LIFO) way, similar to alloca().
A new memory allocation API is introduced along with this Py-stack. It
includes both "local" and "nonlocal" LIFO allocation. Local allocation is
intended to be equivalent to using alloca(), whereby the same function must
free the memory. Nonlocal allocation is where another function may free
the memory, so long as it's still LIFO.
Follow-up patches will convert all uses of alloca() and VLA to the new
scoped allocation API. The old behaviour (using alloca()) will still be
available, but when MICROPY_ENABLE_PYSTACK is enabled then alloca() is no
longer required or used.
The benefits of enabling this option are (or will be once subsequent
patches are made to convert alloca()/VLA):
- Toolchains without alloca() can use this feature to obtain correct and
efficient scoped memory allocation (compared to using the heap instead
of alloca(), which is slower).
- Even if alloca() is available, enabling the Py-stack gives slightly more
efficient use of stack space when calling nested Python functions, due to
the way that compilers implement alloca().
- Enabling the Py-stack with the stackless mode allows for even more
efficient stack usage, as well as retaining high performance (because the
heap is no longer used to build and destroy stackless code states).
- With Py-stack and stackless enabled, Python-calling-Python is no longer
recursive in the C mp_execute_bytecode function.
The micropython.pystack_use() function is included to measure usage of the
Python stack.
2017-11-26 07:28:40 -05:00
2015-01-01 18:30:53 -05:00
// Trace root pointers. This relies on the root pointers being organised
// correctly in the mp_state_ctx structure. We scan nlr_top, dict_locals,
// dict_globals, then the root pointer section of mp_state_vm.
2020-02-26 23:36:53 -05:00
void * * ptrs = ( void * * ) ( void * ) & mp_state_ctx ;
2018-05-12 08:09:34 -04:00
size_t root_start = offsetof ( mp_state_ctx_t , thread . dict_locals ) ;
size_t root_end = offsetof ( mp_state_ctx_t , vm . qstr_last_chunk ) ;
2020-02-26 23:36:53 -05:00
gc_collect_root ( ptrs + root_start / sizeof ( void * ) , ( root_end - root_start ) / sizeof ( void * ) ) ;
py: Introduce a Python stack for scoped allocation.
This patch introduces the MICROPY_ENABLE_PYSTACK option (disabled by
default) which enables a "Python stack" that allows to allocate and free
memory in a scoped, or Last-In-First-Out (LIFO) way, similar to alloca().
A new memory allocation API is introduced along with this Py-stack. It
includes both "local" and "nonlocal" LIFO allocation. Local allocation is
intended to be equivalent to using alloca(), whereby the same function must
free the memory. Nonlocal allocation is where another function may free
the memory, so long as it's still LIFO.
Follow-up patches will convert all uses of alloca() and VLA to the new
scoped allocation API. The old behaviour (using alloca()) will still be
available, but when MICROPY_ENABLE_PYSTACK is enabled then alloca() is no
longer required or used.
The benefits of enabling this option are (or will be once subsequent
patches are made to convert alloca()/VLA):
- Toolchains without alloca() can use this feature to obtain correct and
efficient scoped memory allocation (compared to using the heap instead
of alloca(), which is slower).
- Even if alloca() is available, enabling the Py-stack gives slightly more
efficient use of stack space when calling nested Python functions, due to
the way that compilers implement alloca().
- Enabling the Py-stack with the stackless mode allows for even more
efficient stack usage, as well as retaining high performance (because the
heap is no longer used to build and destroy stackless code states).
- With Py-stack and stackless enabled, Python-calling-Python is no longer
recursive in the C mp_execute_bytecode function.
The micropython.pystack_use() function is included to measure usage of the
Python stack.
2017-11-26 07:28:40 -05:00
2019-04-18 18:44:44 -04:00
gc_mark ( MP_STATE_MEM ( permanent_pointers ) ) ;
2019-04-08 19:58:50 -04:00
py: Introduce a Python stack for scoped allocation.
This patch introduces the MICROPY_ENABLE_PYSTACK option (disabled by
default) which enables a "Python stack" that allows to allocate and free
memory in a scoped, or Last-In-First-Out (LIFO) way, similar to alloca().
A new memory allocation API is introduced along with this Py-stack. It
includes both "local" and "nonlocal" LIFO allocation. Local allocation is
intended to be equivalent to using alloca(), whereby the same function must
free the memory. Nonlocal allocation is where another function may free
the memory, so long as it's still LIFO.
Follow-up patches will convert all uses of alloca() and VLA to the new
scoped allocation API. The old behaviour (using alloca()) will still be
available, but when MICROPY_ENABLE_PYSTACK is enabled then alloca() is no
longer required or used.
The benefits of enabling this option are (or will be once subsequent
patches are made to convert alloca()/VLA):
- Toolchains without alloca() can use this feature to obtain correct and
efficient scoped memory allocation (compared to using the heap instead
of alloca(), which is slower).
- Even if alloca() is available, enabling the Py-stack gives slightly more
efficient use of stack space when calling nested Python functions, due to
the way that compilers implement alloca().
- Enabling the Py-stack with the stackless mode allows for even more
efficient stack usage, as well as retaining high performance (because the
heap is no longer used to build and destroy stackless code states).
- With Py-stack and stackless enabled, Python-calling-Python is no longer
recursive in the C mp_execute_bytecode function.
The micropython.pystack_use() function is included to measure usage of the
Python stack.
2017-11-26 07:28:40 -05:00
# if MICROPY_ENABLE_PYSTACK
// Trace root pointers from the Python stack.
2020-02-26 23:36:53 -05:00
ptrs = ( void * * ) ( void * ) MP_STATE_THREAD ( pystack_start ) ;
gc_collect_root ( ptrs , ( MP_STATE_THREAD ( pystack_cur ) - MP_STATE_THREAD ( pystack_start ) ) / sizeof ( void * ) ) ;
py: Introduce a Python stack for scoped allocation.
This patch introduces the MICROPY_ENABLE_PYSTACK option (disabled by
default) which enables a "Python stack" that allows to allocate and free
memory in a scoped, or Last-In-First-Out (LIFO) way, similar to alloca().
A new memory allocation API is introduced along with this Py-stack. It
includes both "local" and "nonlocal" LIFO allocation. Local allocation is
intended to be equivalent to using alloca(), whereby the same function must
free the memory. Nonlocal allocation is where another function may free
the memory, so long as it's still LIFO.
Follow-up patches will convert all uses of alloca() and VLA to the new
scoped allocation API. The old behaviour (using alloca()) will still be
available, but when MICROPY_ENABLE_PYSTACK is enabled then alloca() is no
longer required or used.
The benefits of enabling this option are (or will be once subsequent
patches are made to convert alloca()/VLA):
- Toolchains without alloca() can use this feature to obtain correct and
efficient scoped memory allocation (compared to using the heap instead
of alloca(), which is slower).
- Even if alloca() is available, enabling the Py-stack gives slightly more
efficient use of stack space when calling nested Python functions, due to
the way that compilers implement alloca().
- Enabling the Py-stack with the stackless mode allows for even more
efficient stack usage, as well as retaining high performance (because the
heap is no longer used to build and destroy stackless code states).
- With Py-stack and stackless enabled, Python-calling-Python is no longer
recursive in the C mp_execute_bytecode function.
The micropython.pystack_use() function is included to measure usage of the
Python stack.
2017-11-26 07:28:40 -05:00
# endif
2013-10-21 18:45:08 -04:00
}
2019-06-06 17:49:32 -04:00
void gc_collect_ptr ( void * ptr ) {
gc_mark ( ptr ) ;
}
2021-05-10 17:16:50 -04:00
// Address sanitizer needs to know that the access to ptrs[i] must always be
// considered OK, even if it's a load from an address that would normally be
// prohibited (due to being undefined, in a red zone, etc).
2021-06-18 00:15:37 -04:00
# if defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))
2021-05-10 17:16:50 -04:00
__attribute__ ( ( no_sanitize_address ) )
# endif
static void * gc_get_ptr ( void * * ptrs , int i ) {
2021-08-27 10:02:03 -04:00
# if MICROPY_DEBUG_VALGRIND
if ( ! VALGRIND_CHECK_MEM_IS_ADDRESSABLE ( & ptrs [ i ] , sizeof ( * ptrs ) ) ) {
return NULL ;
}
# endif
2021-05-10 17:16:50 -04:00
return ptrs [ i ] ;
}
2015-11-27 08:07:48 -05:00
void gc_collect_root ( void * * ptrs , size_t len ) {
for ( size_t i = 0 ; i < len ; i + + ) {
2021-10-26 04:47:04 -04:00
MICROPY_GC_HOOK_LOOP
2021-05-10 17:16:50 -04:00
void * ptr = gc_get_ptr ( ptrs , i ) ;
2019-04-18 18:44:44 -04:00
gc_mark ( ptr ) ;
2013-10-21 18:45:08 -04:00
}
}
2013-10-23 15:20:17 -04:00
void gc_collect_end ( void ) {
2013-10-21 18:45:08 -04:00
gc_deal_with_stack_overflow ( ) ;
gc_sweep ( ) ;
2020-02-11 20:06:24 -05:00
for ( size_t i = 0 ; i < MICROPY_ATB_INDICES ; i + + ) {
MP_STATE_MEM ( gc_first_free_atb_index ) [ i ] = 0 ;
}
2018-01-23 19:22:05 -05:00
MP_STATE_MEM ( gc_last_free_atb_index ) = MP_STATE_MEM ( gc_alloc_table_byte_len ) - 1 ;
2021-05-04 09:56:43 -04:00
MP_STATE_THREAD ( gc_lock_depth ) - - ;
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2013-10-22 10:25:25 -04:00
}
2013-10-21 18:45:08 -04:00
2018-05-31 09:10:48 -04:00
void gc_sweep_all ( void ) {
GC_ENTER ( ) ;
2021-05-04 09:56:43 -04:00
MP_STATE_THREAD ( gc_lock_depth ) + + ;
2018-05-31 09:10:48 -04:00
MP_STATE_MEM ( gc_stack_overflow ) = 0 ;
gc_collect_end ( ) ;
}
2013-10-22 10:25:25 -04:00
void gc_info ( gc_info_t * info ) {
2016-04-25 11:28:57 -04:00
GC_ENTER ( ) ;
2015-11-27 08:07:48 -05:00
info - > total = MP_STATE_MEM ( gc_pool_end ) - MP_STATE_MEM ( gc_pool_start ) ;
2013-10-22 10:25:25 -04:00
info - > used = 0 ;
info - > free = 0 ;
2016-06-30 17:09:55 -04:00
info - > max_free = 0 ;
2013-10-22 10:25:25 -04:00
info - > num_1block = 0 ;
info - > num_2block = 0 ;
info - > max_block = 0 ;
2016-06-29 18:59:24 -04:00
bool finish = false ;
2016-06-30 17:09:55 -04:00
for ( size_t block = 0 , len = 0 , len_free = 0 ; ! finish ; ) {
2015-12-16 20:09:11 -05:00
size_t kind = ATB_GET_KIND ( block ) ;
2013-10-22 10:25:25 -04:00
switch ( kind ) {
2013-10-21 18:45:08 -04:00
case AT_FREE :
2013-10-22 10:25:25 -04:00
info - > free + = 1 ;
2016-06-30 17:09:55 -04:00
len_free + = 1 ;
2013-10-22 10:25:25 -04:00
len = 0 ;
2013-10-21 18:45:08 -04:00
break ;
case AT_HEAD :
2013-10-22 10:25:25 -04:00
info - > used + = 1 ;
len = 1 ;
break ;
2013-10-21 18:45:08 -04:00
case AT_TAIL :
2013-10-22 10:25:25 -04:00
info - > used + = 1 ;
len + = 1 ;
2013-10-21 18:45:08 -04:00
break ;
case AT_MARK :
2013-10-22 10:25:25 -04:00
// shouldn't happen
2013-10-21 18:45:08 -04:00
break ;
}
2016-06-29 18:59:24 -04:00
block + + ;
finish = ( block = = MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB ) ;
// Get next block type if possible
if ( ! finish ) {
kind = ATB_GET_KIND ( block ) ;
}
if ( finish | | kind = = AT_FREE | | kind = = AT_HEAD ) {
if ( len = = 1 ) {
info - > num_1block + = 1 ;
} else if ( len = = 2 ) {
info - > num_2block + = 1 ;
}
if ( len > info - > max_block ) {
info - > max_block = len ;
}
2016-06-30 17:09:55 -04:00
if ( finish | | kind = = AT_HEAD ) {
if ( len_free > info - > max_free ) {
info - > max_free = len_free ;
}
len_free = 0 ;
}
2016-06-29 18:59:24 -04:00
}
2013-10-21 18:45:08 -04:00
}
2013-10-22 10:25:25 -04:00
info - > used * = BYTES_PER_BLOCK ;
info - > free * = BYTES_PER_BLOCK ;
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2013-10-21 18:45:08 -04:00
}
2020-01-12 23:32:51 -05:00
bool gc_alloc_possible ( void ) {
return MP_STATE_MEM ( gc_pool_start ) ! = 0 ;
}
2018-01-23 19:22:05 -05:00
// We place long lived objects at the end of the heap rather than the start. This reduces
// fragmentation by localizing the heap churn to one portion of memory (the start of the heap.)
2021-04-20 21:04:23 -04:00
void * gc_alloc ( size_t n_bytes , unsigned int alloc_flags , bool long_lived ) {
2018-10-26 15:27:44 -04:00
bool has_finaliser = alloc_flags & GC_ALLOC_FLAG_HAS_FINALISER ;
2015-12-16 20:09:11 -05:00
size_t n_blocks = ( ( n_bytes + BYTES_PER_BLOCK - 1 ) & ( ~ ( BYTES_PER_BLOCK - 1 ) ) ) / BYTES_PER_BLOCK ;
2014-06-18 04:20:41 -04:00
DEBUG_printf ( " gc_alloc( " UINT_FMT " bytes -> " UINT_FMT " blocks) \n " , n_bytes , n_blocks ) ;
2013-10-21 18:45:08 -04:00
2016-04-25 11:28:57 -04:00
// check for 0 allocation
if ( n_blocks = = 0 ) {
2014-04-08 07:31:21 -04:00
return NULL ;
2014-04-05 15:35:48 -04:00
}
2016-04-25 11:28:57 -04:00
// check if GC is locked
2021-05-04 09:56:43 -04:00
if ( MP_STATE_THREAD ( gc_lock_depth ) > 0 ) {
2013-10-21 18:45:08 -04:00
return NULL ;
}
2019-03-08 18:29:14 -05:00
if ( MP_STATE_MEM ( gc_pool_start ) = = 0 ) {
reset_into_safe_mode ( GC_ALLOC_OUTSIDE_VM ) ;
}
2021-05-04 09:56:43 -04:00
GC_ENTER ( ) ;
2018-01-23 19:22:05 -05:00
size_t found_block = 0xffffffff ;
2015-12-16 20:09:11 -05:00
size_t end_block ;
size_t start_block ;
py/gc: In gc_alloc, reset n_free var right before search for free mem.
Otherwise there is the possibility that n_free starts out non-zero from the
previous iteration, which may have found a few (but not enough) free blocks
at the end of the heap. If this is the case, and if the very first blocks
that are scanned the second time around (starting at
gc_last_free_atb_index) are found to give enough memory (including the
blocks at the end of the heap from the previous iteration that left n_free
non-zero) then memory will be allocated starting before the location that
gc_last_free_atb_index points to, most likely leading to corruption.
This serious bug did not manifest itself in the past because a gc_collect
always resets gc_last_free_atb_index to point to the start of the GC heap,
and the first block there is almost always allocated to a long-lived
object (eg entries from sys.path, or mounted filesystem objects), which
means that n_free would be reset at the start of the search loop.
But with threading enabled with the GIL disabled it is possible to trigger
the bug via the following sequence of events:
1. Thread A runs gc_alloc, fails to find enough memory, and has a non-zero
n_free at the end of the search.
2. Thread A calls gc_collect and frees a bunch of blocks on the GC heap.
3. Just after gc_collect finishes in thread A, thread B takes gc_mutex and
does an allocation, moving gc_last_free_atb_index to point to the
interior of the heap, to a place where there is most likely a run of
available blocks.
4. Thread A regains gc_mutex and does its second search for free memory,
starting with a non-zero n_free. Since it's likely that the first block
it searches is available it will allocate memory which overlaps with the
memory before gc_last_free_atb_index.
2018-08-10 01:46:45 -04:00
size_t n_free ;
2018-01-23 19:22:05 -05:00
bool collected = ! MP_STATE_MEM ( gc_auto_collect_enabled ) ;
py/gc: Implement GC running by allocation threshold.
Currently, MicroPython runs GC when it could not allocate a block of memory,
which happens when heap is exhausted. However, that policy can't work well
with "inifinity" heaps, e.g. backed by a virtual memory - there will be a
lot of swap thrashing long before VM will be exhausted. Instead, in such
cases "allocation threshold" policy is used: a GC is run after some number of
allocations have been made. Details vary, for example, number or total amount
of allocations can be used, threshold may be self-adjusting based on GC
outcome, etc.
This change implements a simple variant of such policy for MicroPython. Amount
of allocated memory so far is used for threshold, to make it useful to typical
finite-size, and small, heaps as used with MicroPython ports. And such GC policy
is indeed useful for such types of heaps too, as it allows to better control
fragmentation. For example, if a threshold is set to half size of heap, then
for an application which usually makes big number of small allocations, that
will (try to) keep half of heap memory in a nice defragmented state for an
occasional large allocation.
For an application which doesn't exhibit such behavior, there won't be any
visible effects, except for GC running more frequently, which however may
affect performance. To address this, the GC threshold is configurable, and
by default is off so far. It's configured with gc.threshold(amount_in_bytes)
call (can be queries without an argument).
2016-07-20 17:37:30 -04:00
# if MICROPY_GC_ALLOC_THRESHOLD
if ( ! collected & & MP_STATE_MEM ( gc_alloc_amount ) > = MP_STATE_MEM ( gc_alloc_threshold ) ) {
GC_EXIT ( ) ;
gc_collect ( ) ;
2018-05-20 23:36:21 -04:00
collected = 1 ;
py/gc: Implement GC running by allocation threshold.
Currently, MicroPython runs GC when it could not allocate a block of memory,
which happens when heap is exhausted. However, that policy can't work well
with "inifinity" heaps, e.g. backed by a virtual memory - there will be a
lot of swap thrashing long before VM will be exhausted. Instead, in such
cases "allocation threshold" policy is used: a GC is run after some number of
allocations have been made. Details vary, for example, number or total amount
of allocations can be used, threshold may be self-adjusting based on GC
outcome, etc.
This change implements a simple variant of such policy for MicroPython. Amount
of allocated memory so far is used for threshold, to make it useful to typical
finite-size, and small, heaps as used with MicroPython ports. And such GC policy
is indeed useful for such types of heaps too, as it allows to better control
fragmentation. For example, if a threshold is set to half size of heap, then
for an application which usually makes big number of small allocations, that
will (try to) keep half of heap memory in a nice defragmented state for an
occasional large allocation.
For an application which doesn't exhibit such behavior, there won't be any
visible effects, except for GC running more frequently, which however may
affect performance. To address this, the GC threshold is configurable, and
by default is off so far. It's configured with gc.threshold(amount_in_bytes)
call (can be queries without an argument).
2016-07-20 17:37:30 -04:00
GC_ENTER ( ) ;
}
# endif
2018-01-23 19:22:05 -05:00
bool keep_looking = true ;
// When we start searching on the other side of the crossover block we make sure to
// perform a collect. That way we'll get the closest free block in our section.
size_t crossover_block = BLOCK_FROM_PTR ( MP_STATE_MEM ( gc_lowest_long_lived_ptr ) ) ;
while ( keep_looking ) {
int8_t direction = 1 ;
2020-02-11 20:06:24 -05:00
size_t bucket = MIN ( n_blocks , MICROPY_ATB_INDICES ) - 1 ;
size_t first_free = MP_STATE_MEM ( gc_first_free_atb_index ) [ bucket ] ;
size_t start = first_free ;
2018-01-23 19:22:05 -05:00
if ( long_lived ) {
direction = - 1 ;
start = MP_STATE_MEM ( gc_last_free_atb_index ) ;
}
py/gc: In gc_alloc, reset n_free var right before search for free mem.
Otherwise there is the possibility that n_free starts out non-zero from the
previous iteration, which may have found a few (but not enough) free blocks
at the end of the heap. If this is the case, and if the very first blocks
that are scanned the second time around (starting at
gc_last_free_atb_index) are found to give enough memory (including the
blocks at the end of the heap from the previous iteration that left n_free
non-zero) then memory will be allocated starting before the location that
gc_last_free_atb_index points to, most likely leading to corruption.
This serious bug did not manifest itself in the past because a gc_collect
always resets gc_last_free_atb_index to point to the start of the GC heap,
and the first block there is almost always allocated to a long-lived
object (eg entries from sys.path, or mounted filesystem objects), which
means that n_free would be reset at the start of the search loop.
But with threading enabled with the GIL disabled it is possible to trigger
the bug via the following sequence of events:
1. Thread A runs gc_alloc, fails to find enough memory, and has a non-zero
n_free at the end of the search.
2. Thread A calls gc_collect and frees a bunch of blocks on the GC heap.
3. Just after gc_collect finishes in thread A, thread B takes gc_mutex and
does an allocation, moving gc_last_free_atb_index to point to the
interior of the heap, to a place where there is most likely a run of
available blocks.
4. Thread A regains gc_mutex and does its second search for free memory,
starting with a non-zero n_free. Since it's likely that the first block
it searches is available it will allocate memory which overlaps with the
memory before gc_last_free_atb_index.
2018-08-10 01:46:45 -04:00
n_free = 0 ;
2013-10-21 18:45:08 -04:00
// look for a run of n_blocks available blocks
2020-02-11 20:06:24 -05:00
for ( size_t i = start ; keep_looking & & first_free < = i & & i < = MP_STATE_MEM ( gc_last_free_atb_index ) ; i + = direction ) {
2015-01-01 18:30:53 -05:00
byte a = MP_STATE_MEM ( gc_alloc_table_start ) [ i ] ;
2018-01-23 19:22:05 -05:00
// Four ATB states are packed into a single byte.
int j = 0 ;
if ( direction = = - 1 ) {
j = 3 ;
}
for ( ; keep_looking & & 0 < = j & & j < = 3 ; j + = direction ) {
if ( ( a & ( 0x3 < < ( j * 2 ) ) ) = = 0 ) {
if ( + + n_free > = n_blocks ) {
found_block = i * BLOCKS_PER_ATB + j ;
keep_looking = false ;
}
} else {
if ( ! collected ) {
size_t block = i * BLOCKS_PER_ATB + j ;
if ( ( direction = = 1 & & block > = crossover_block ) | |
2021-03-15 09:57:36 -04:00
( direction = = - 1 & & block < crossover_block ) ) {
2018-01-23 19:22:05 -05:00
keep_looking = false ;
}
}
n_free = 0 ;
}
}
}
if ( n_free > = n_blocks ) {
break ;
2014-08-22 13:17:02 -04:00
}
2013-10-21 18:45:08 -04:00
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2013-10-21 18:45:08 -04:00
// nothing found!
if ( collected ) {
return NULL ;
}
2014-02-11 11:01:38 -05:00
DEBUG_printf ( " gc_alloc( " UINT_FMT " ): no free mem, triggering GC \n " , n_bytes ) ;
2013-10-21 18:45:08 -04:00
gc_collect ( ) ;
2018-01-23 19:22:05 -05:00
collected = true ;
// Try again since we've hopefully freed up space.
keep_looking = true ;
2016-04-25 11:28:57 -04:00
GC_ENTER ( ) ;
2013-10-21 18:45:08 -04:00
}
2018-01-23 19:22:05 -05:00
assert ( found_block ! = 0xffffffff ) ;
2013-10-21 18:45:08 -04:00
2018-01-23 19:22:05 -05:00
// Found free space ending at found_block inclusive.
// Also, set last free ATB index to block after last block we found, for start of
2020-02-11 20:06:24 -05:00
// next scan. Also, whenever we free or shrink a block we must check if this index needs
// adjusting (see gc_realloc and gc_free).
2018-01-23 19:22:05 -05:00
if ( ! long_lived ) {
end_block = found_block ;
start_block = found_block - n_free + 1 ;
2020-02-11 20:06:24 -05:00
if ( n_blocks < MICROPY_ATB_INDICES ) {
size_t next_free_atb = ( found_block + n_blocks ) / BLOCKS_PER_ATB ;
// Update all atb indices for larger blocks too.
for ( size_t i = n_blocks - 1 ; i < MICROPY_ATB_INDICES ; i + + ) {
MP_STATE_MEM ( gc_first_free_atb_index ) [ i ] = next_free_atb ;
}
2018-01-23 19:22:05 -05:00
}
} else {
start_block = found_block ;
end_block = found_block + n_free - 1 ;
2020-02-11 20:06:24 -05:00
// Always update the bounds of the long lived area because we assume it is contiguous. (It
// can still be reset by a sweep.)
MP_STATE_MEM ( gc_last_free_atb_index ) = ( found_block - 1 ) / BLOCKS_PER_ATB ;
2014-08-28 05:18:40 -04:00
}
2014-08-22 13:17:02 -04:00
2016-11-23 20:31:53 -05:00
# ifdef LOG_HEAP_ACTIVITY
gc_log_change ( start_block , end_block - start_block + 1 ) ;
# endif
2013-10-21 18:45:08 -04:00
// mark first block as used head
ATB_FREE_TO_HEAD ( start_block ) ;
// mark rest of blocks as used tail
// TODO for a run of many blocks can make this more efficient
2015-12-16 20:09:11 -05:00
for ( size_t bl = start_block + 1 ; bl < = end_block ; bl + + ) {
2013-10-21 18:45:08 -04:00
ATB_FREE_TO_TAIL ( bl ) ;
}
2014-04-05 15:35:48 -04:00
// get pointer to first block
2016-05-05 06:25:08 -04:00
// we must create this pointer before unlocking the GC so a collection can find it
2020-02-26 23:36:53 -05:00
void * ret_ptr = ( void * ) ( MP_STATE_MEM ( gc_pool_start ) + start_block * BYTES_PER_BLOCK ) ;
2014-06-18 04:20:41 -04:00
DEBUG_printf ( " gc_alloc(%p) \n " , ret_ptr ) ;
2014-04-05 15:35:48 -04:00
2018-01-23 19:22:05 -05:00
// If the allocation was long live then update the lowest value. Its used to trigger early
// collects when allocations fail in their respective section. Its also used to ignore calls to
// gc_make_long_lived where the pointer is already in the long lived section.
if ( long_lived & & ret_ptr < MP_STATE_MEM ( gc_lowest_long_lived_ptr ) ) {
MP_STATE_MEM ( gc_lowest_long_lived_ptr ) = ret_ptr ;
}
py/gc: Implement GC running by allocation threshold.
Currently, MicroPython runs GC when it could not allocate a block of memory,
which happens when heap is exhausted. However, that policy can't work well
with "inifinity" heaps, e.g. backed by a virtual memory - there will be a
lot of swap thrashing long before VM will be exhausted. Instead, in such
cases "allocation threshold" policy is used: a GC is run after some number of
allocations have been made. Details vary, for example, number or total amount
of allocations can be used, threshold may be self-adjusting based on GC
outcome, etc.
This change implements a simple variant of such policy for MicroPython. Amount
of allocated memory so far is used for threshold, to make it useful to typical
finite-size, and small, heaps as used with MicroPython ports. And such GC policy
is indeed useful for such types of heaps too, as it allows to better control
fragmentation. For example, if a threshold is set to half size of heap, then
for an application which usually makes big number of small allocations, that
will (try to) keep half of heap memory in a nice defragmented state for an
occasional large allocation.
For an application which doesn't exhibit such behavior, there won't be any
visible effects, except for GC running more frequently, which however may
affect performance. To address this, the GC threshold is configurable, and
by default is off so far. It's configured with gc.threshold(amount_in_bytes)
call (can be queries without an argument).
2016-07-20 17:37:30 -04:00
# if MICROPY_GC_ALLOC_THRESHOLD
MP_STATE_MEM ( gc_alloc_amount ) + = n_blocks ;
# endif
2016-05-05 06:25:08 -04:00
GC_EXIT ( ) ;
2016-08-26 01:35:26 -04:00
# if MICROPY_GC_CONSERVATIVE_CLEAR
// be conservative and zero out all the newly allocated blocks
2020-02-26 23:36:53 -05:00
memset ( ( byte * ) ret_ptr , 0 , ( end_block - start_block + 1 ) * BYTES_PER_BLOCK ) ;
2016-08-26 01:35:26 -04:00
# else
2014-04-26 17:23:42 -04:00
// zero out the additional bytes of the newly allocated blocks
py, gc: Zero out newly allocated blocks in the GC.
Also add some more debugging output to gc_dump_alloc_table().
Now that newly allocated heap is always zero'd, maybe we just make this
a policy for the uPy API to keep it simple (ie any new implementation of
memory allocation must zero all allocations). This follows the D
language philosophy.
Before this patch, a previously used memory block which had pointers in
it may still retain those pointers if the new user of that block does
not actually use the entire block. Eg, if I want 5 blocks worth of
heap, I actually get 8 (round up to nearest 4). Then I never use the
last 3, so they keep their old values, which may be pointers pointing to
the heap, hence preventing GC.
In rare (or maybe not that rare) cases, this leads to long, unintentional
"linked lists" within the GC'd heap, filling it up completely. It's
pretty rare, because you have to reuse exactly that memory which is part
of this "linked list", and reuse it in just the right way.
This should fix issue #522, and might have something to do with
issue #510.
2014-04-25 18:37:55 -04:00
// This is needed because the blocks may have previously held pointers
// to the heap and will not be set to something else if the caller
// doesn't actually use the entire block. As such they will continue
// to point to the heap and may prevent other blocks from being reclaimed.
2020-02-26 23:36:53 -05:00
memset ( ( byte * ) ret_ptr + n_bytes , 0 , ( end_block - start_block + 1 ) * BYTES_PER_BLOCK - n_bytes ) ;
2016-08-26 01:35:26 -04:00
# endif
py, gc: Zero out newly allocated blocks in the GC.
Also add some more debugging output to gc_dump_alloc_table().
Now that newly allocated heap is always zero'd, maybe we just make this
a policy for the uPy API to keep it simple (ie any new implementation of
memory allocation must zero all allocations). This follows the D
language philosophy.
Before this patch, a previously used memory block which had pointers in
it may still retain those pointers if the new user of that block does
not actually use the entire block. Eg, if I want 5 blocks worth of
heap, I actually get 8 (round up to nearest 4). Then I never use the
last 3, so they keep their old values, which may be pointers pointing to
the heap, hence preventing GC.
In rare (or maybe not that rare) cases, this leads to long, unintentional
"linked lists" within the GC'd heap, filling it up completely. It's
pretty rare, because you have to reuse exactly that memory which is part
of this "linked list", and reuse it in just the right way.
This should fix issue #522, and might have something to do with
issue #510.
2014-04-25 18:37:55 -04:00
2015-09-04 11:53:46 -04:00
# if MICROPY_ENABLE_FINALISER
2014-04-05 15:35:48 -04:00
if ( has_finaliser ) {
2014-04-26 17:23:42 -04:00
// clear type pointer in case it is never set
2020-02-26 23:36:53 -05:00
( ( mp_obj_base_t * ) ret_ptr ) - > type = NULL ;
2014-04-05 15:35:48 -04:00
// set mp_obj flag only if it has a finaliser
2016-04-25 11:28:57 -04:00
GC_ENTER ( ) ;
2014-04-05 15:35:48 -04:00
FTB_SET ( start_block ) ;
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2014-04-05 09:49:03 -04:00
}
2015-09-04 11:53:46 -04:00
# else
( void ) has_finaliser ;
# endif
2014-04-05 09:49:03 -04:00
2014-08-28 18:06:38 -04:00
# if EXTENSIVE_HEAP_PROFILING
gc_dump_alloc_table ( ) ;
# endif
2020-07-15 20:58:38 -04:00
# if CIRCUITPY_MEMORYMONITOR
2020-07-16 22:01:43 -04:00
memorymonitor_track_allocation ( end_block - start_block + 1 ) ;
2020-07-15 20:58:38 -04:00
# endif
2014-04-05 15:35:48 -04:00
return ret_ptr ;
2013-10-21 18:45:08 -04:00
}
2014-04-05 15:35:48 -04:00
/*
2014-07-03 08:25:24 -04:00
void * gc_alloc ( mp_uint_t n_bytes ) {
2014-04-03 17:55:12 -04:00
return _gc_alloc ( n_bytes , false ) ;
}
2014-07-03 08:25:24 -04:00
void * gc_alloc_with_finaliser ( mp_uint_t n_bytes ) {
2014-04-03 17:55:12 -04:00
return _gc_alloc ( n_bytes , true ) ;
}
2014-04-05 15:35:48 -04:00
*/
2014-04-03 17:55:12 -04:00
2013-10-22 15:26:36 -04:00
// force the freeing of a piece of memory
2015-11-06 17:32:47 -05:00
// TODO: freeing here does not call finaliser
2015-11-27 08:07:48 -05:00
void gc_free ( void * ptr ) {
2021-05-04 09:56:43 -04:00
if ( MP_STATE_THREAD ( gc_lock_depth ) > 0 ) {
2014-04-08 07:31:21 -04:00
// TODO how to deal with this error?
return ;
2014-04-05 15:35:48 -04:00
}
2021-05-04 09:56:43 -04:00
GC_ENTER ( ) ;
2014-06-16 04:44:29 -04:00
DEBUG_printf ( " gc_free(%p) \n " , ptr ) ;
2013-10-22 15:26:36 -04:00
2017-07-11 22:17:38 -04:00
if ( ptr = = NULL ) {
GC_EXIT ( ) ;
} else {
2019-12-10 18:05:47 -05:00
if ( MP_STATE_MEM ( gc_pool_start ) = = 0 ) {
reset_into_safe_mode ( GC_ALLOC_OUTSIDE_VM ) ;
}
2017-07-11 22:17:38 -04:00
// get the GC block number corresponding to this pointer
assert ( VERIFY_PTR ( ptr ) ) ;
2020-02-11 20:06:24 -05:00
size_t start_block = BLOCK_FROM_PTR ( ptr ) ;
assert ( ATB_GET_KIND ( start_block ) = = AT_HEAD ) ;
2014-08-28 18:06:38 -04:00
2017-07-11 22:17:38 -04:00
# if MICROPY_ENABLE_FINALISER
2020-02-11 20:06:24 -05:00
FTB_CLEAR ( start_block ) ;
2017-07-11 22:17:38 -04:00
# endif
2016-04-25 11:28:57 -04:00
2017-07-11 22:17:38 -04:00
// free head and all of its tail blocks
2020-02-11 20:06:24 -05:00
# ifdef LOG_HEAP_ACTIVITY
gc_log_change ( start_block , 0 ) ;
# endif
size_t block = start_block ;
2017-07-11 22:17:38 -04:00
do {
ATB_ANY_TO_FREE ( block ) ;
block + = 1 ;
} while ( ATB_GET_KIND ( block ) = = AT_TAIL ) ;
2020-02-11 20:06:24 -05:00
// Update the first free pointer for our size only. Not much calls gc_free directly so there
// is decent chance we'll want to allocate this size again. By only updating the specific
// size we don't risk something smaller fitting in.
size_t n_blocks = block - start_block ;
size_t bucket = MIN ( n_blocks , MICROPY_ATB_INDICES ) - 1 ;
size_t new_free_atb = start_block / BLOCKS_PER_ATB ;
if ( new_free_atb < MP_STATE_MEM ( gc_first_free_atb_index ) [ bucket ] ) {
MP_STATE_MEM ( gc_first_free_atb_index ) [ bucket ] = new_free_atb ;
}
// set the last_free pointer to this block if it's earlier in the heap
if ( new_free_atb > MP_STATE_MEM ( gc_last_free_atb_index ) ) {
MP_STATE_MEM ( gc_last_free_atb_index ) = new_free_atb ;
}
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2017-07-11 22:17:38 -04:00
# if EXTENSIVE_HEAP_PROFILING
gc_dump_alloc_table ( ) ;
# endif
2013-10-22 15:26:36 -04:00
}
}
2015-11-27 08:07:48 -05:00
size_t gc_nbytes ( const void * ptr ) {
2016-04-25 11:28:57 -04:00
GC_ENTER ( ) ;
2013-10-22 15:26:36 -04:00
if ( VERIFY_PTR ( ptr ) ) {
2015-12-16 20:09:11 -05:00
size_t block = BLOCK_FROM_PTR ( ptr ) ;
2013-10-21 18:45:08 -04:00
if ( ATB_GET_KIND ( block ) = = AT_HEAD ) {
// work out number of consecutive blocks in the chain starting with this on
2015-12-16 20:09:11 -05:00
size_t n_blocks = 0 ;
2013-10-21 18:45:08 -04:00
do {
n_blocks + = 1 ;
} while ( ATB_GET_KIND ( block + n_blocks ) = = AT_TAIL ) ;
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2013-10-21 18:45:08 -04:00
return n_blocks * BYTES_PER_BLOCK ;
}
}
// invalid pointer
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2013-10-21 18:45:08 -04:00
return 0 ;
}
2018-01-23 19:22:05 -05:00
bool gc_has_finaliser ( const void * ptr ) {
2021-03-15 09:57:36 -04:00
# if MICROPY_ENABLE_FINALISER
2018-01-23 19:22:05 -05:00
GC_ENTER ( ) ;
if ( VERIFY_PTR ( ptr ) ) {
bool has_finaliser = FTB_GET ( BLOCK_FROM_PTR ( ptr ) ) ;
GC_EXIT ( ) ;
return has_finaliser ;
}
// invalid pointer
GC_EXIT ( ) ;
2021-03-15 09:57:36 -04:00
# else
( void ) ptr ;
# endif
2018-01-23 19:22:05 -05:00
return false ;
}
void * gc_make_long_lived ( void * old_ptr ) {
// If its already in the long lived section then don't bother moving it.
if ( old_ptr > = MP_STATE_MEM ( gc_lowest_long_lived_ptr ) ) {
return old_ptr ;
}
size_t n_bytes = gc_nbytes ( old_ptr ) ;
if ( n_bytes = = 0 ) {
return old_ptr ;
}
bool has_finaliser = gc_has_finaliser ( old_ptr ) ;
// Try and find a new area in the long lived section to copy the memory to.
2021-03-15 09:57:36 -04:00
void * new_ptr = gc_alloc ( n_bytes , has_finaliser , true ) ;
2018-01-23 19:22:05 -05:00
if ( new_ptr = = NULL ) {
return old_ptr ;
} else if ( old_ptr > new_ptr ) {
// Return the old pointer if the new one is lower in the heap and free the new space.
gc_free ( new_ptr ) ;
return old_ptr ;
}
// We copy everything over and let the garbage collection process delete the old copy. That way
// we ensure we don't delete memory that has a second reference. (Though if there is we may
// confuse things when its mutable.)
memcpy ( new_ptr , old_ptr , n_bytes ) ;
return new_ptr ;
}
2014-03-12 15:00:23 -04:00
#if 0
2014-04-08 07:31:21 -04:00
// old, simple realloc that didn't expand memory in place
2014-07-03 08:25:24 -04:00
void * gc_realloc ( void * ptr , mp_uint_t n_bytes ) {
mp_uint_t n_existing = gc_nbytes ( ptr ) ;
2014-03-06 19:21:51 -05:00
if ( n_bytes < = n_existing ) {
return ptr ;
} else {
2014-04-25 07:44:53 -04:00
bool has_finaliser ;
if ( ptr = = NULL ) {
has_finaliser = false ;
} else {
2020-02-26 23:36:53 -05:00
# if MICROPY_ENABLE_FINALISER
2014-07-03 08:25:24 -04:00
has_finaliser = FTB_GET ( BLOCK_FROM_PTR ( ( mp_uint_t ) ptr ) ) ;
2020-02-26 23:36:53 -05:00
# else
2014-04-25 07:44:53 -04:00
has_finaliser = false ;
2020-02-26 23:36:53 -05:00
# endif
2014-04-25 07:44:53 -04:00
}
void * ptr2 = gc_alloc ( n_bytes , has_finaliser ) ;
2014-03-06 19:21:51 -05:00
if ( ptr2 = = NULL ) {
return ptr2 ;
}
memcpy ( ptr2 , ptr , n_existing ) ;
gc_free ( ptr ) ;
return ptr2 ;
}
}
2014-04-20 04:43:38 -04:00
# else // Alternative gc_realloc impl
2014-04-08 07:31:21 -04:00
2015-11-27 08:07:48 -05:00
void * gc_realloc ( void * ptr_in , size_t n_bytes , bool allow_move ) {
2014-04-20 13:16:25 -04:00
// check for pure allocation
2014-03-05 16:23:04 -05:00
if ( ptr_in = = NULL ) {
2018-01-23 19:22:05 -05:00
return gc_alloc ( n_bytes , false , false ) ;
2014-03-05 16:23:04 -05:00
}
2014-10-23 07:02:00 -04:00
// check for pure free
if ( n_bytes = = 0 ) {
gc_free ( ptr_in ) ;
return NULL ;
}
2021-05-04 09:56:43 -04:00
if ( MP_STATE_THREAD ( gc_lock_depth ) > 0 ) {
return NULL ;
}
2015-11-27 08:07:48 -05:00
void * ptr = ptr_in ;
2014-04-20 13:16:25 -04:00
2016-05-04 05:14:43 -04:00
GC_ENTER ( ) ;
2017-11-29 01:17:08 -05:00
// get the GC block number corresponding to this pointer
assert ( VERIFY_PTR ( ptr ) ) ;
size_t block = BLOCK_FROM_PTR ( ptr ) ;
assert ( ATB_GET_KIND ( block ) = = AT_HEAD ) ;
2014-04-20 13:16:25 -04:00
// compute number of new blocks that are requested
2015-12-16 20:09:11 -05:00
size_t new_blocks = ( n_bytes + BYTES_PER_BLOCK - 1 ) / BYTES_PER_BLOCK ;
2014-03-05 16:23:04 -05:00
2014-10-15 14:24:47 -04:00
// Get the total number of consecutive blocks that are already allocated to
// this chunk of memory, and then count the number of free blocks following
// it. Stop if we reach the end of the heap, or if we find enough extra
// free blocks to satisfy the realloc. Note that we need to compute the
// total size of the existing memory chunk so we can correctly and
// efficiently shrink it (see below for shrinking code).
2020-02-26 23:36:53 -05:00
size_t n_free = 0 ;
2015-12-16 20:09:11 -05:00
size_t n_blocks = 1 ; // counting HEAD block
size_t max_block = MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB ;
for ( size_t bl = block + n_blocks ; bl < max_block ; bl + + ) {
2014-10-15 14:24:47 -04:00
byte block_type = ATB_GET_KIND ( bl ) ;
if ( block_type = = AT_TAIL ) {
n_blocks + + ;
continue ;
2014-03-12 15:00:23 -04:00
}
2014-10-15 14:24:47 -04:00
if ( block_type = = AT_FREE ) {
n_free + + ;
if ( n_blocks + n_free > = new_blocks ) {
// stop as soon as we find enough blocks for n_bytes
break ;
}
continue ;
2014-04-20 13:16:25 -04:00
}
break ;
}
// return original ptr if it already has the requested number of blocks
if ( new_blocks = = n_blocks ) {
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2014-04-20 13:16:25 -04:00
return ptr_in ;
}
2014-03-05 16:23:04 -05:00
2014-04-20 13:16:25 -04:00
// check if we can shrink the allocated area
if ( new_blocks < n_blocks ) {
// free unneeded tail blocks
2015-12-16 20:09:11 -05:00
for ( size_t bl = block + new_blocks , count = n_blocks - new_blocks ; count > 0 ; bl + + , count - - ) {
2014-04-20 13:16:25 -04:00
ATB_ANY_TO_FREE ( bl ) ;
2014-04-20 04:45:16 -04:00
}
2014-08-28 18:06:38 -04:00
// set the last_free pointer to end of this block if it's earlier in the heap
2020-02-11 20:06:24 -05:00
size_t new_free_atb = ( block + new_blocks ) / BLOCKS_PER_ATB ;
size_t bucket = MIN ( n_blocks - new_blocks , MICROPY_ATB_INDICES ) - 1 ;
if ( new_free_atb < MP_STATE_MEM ( gc_first_free_atb_index ) [ bucket ] ) {
MP_STATE_MEM ( gc_first_free_atb_index ) [ bucket ] = new_free_atb ;
2018-01-23 19:22:05 -05:00
}
2020-02-11 20:06:24 -05:00
if ( new_free_atb > MP_STATE_MEM ( gc_last_free_atb_index ) ) {
MP_STATE_MEM ( gc_last_free_atb_index ) = new_free_atb ;
2014-08-28 18:06:38 -04:00
}
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2014-08-28 18:06:38 -04:00
# if EXTENSIVE_HEAP_PROFILING
gc_dump_alloc_table ( ) ;
# endif
2016-11-23 20:31:53 -05:00
# ifdef LOG_HEAP_ACTIVITY
gc_log_change ( block , new_blocks ) ;
# endif
2020-07-15 20:58:38 -04:00
# if CIRCUITPY_MEMORYMONITOR
2020-07-16 22:01:43 -04:00
memorymonitor_track_allocation ( new_blocks ) ;
2020-07-15 20:58:38 -04:00
# endif
2014-04-20 13:16:25 -04:00
return ptr_in ;
}
2014-04-20 04:45:16 -04:00
2014-04-20 13:16:25 -04:00
// check if we can expand in place
if ( new_blocks < = n_blocks + n_free ) {
// mark few more blocks as used tail
2015-12-16 20:09:11 -05:00
for ( size_t bl = block + n_blocks ; bl < block + new_blocks ; bl + + ) {
2014-04-20 13:16:25 -04:00
assert ( ATB_GET_KIND ( bl ) = = AT_FREE ) ;
ATB_FREE_TO_TAIL ( bl ) ;
}
py, gc: Zero out newly allocated blocks in the GC.
Also add some more debugging output to gc_dump_alloc_table().
Now that newly allocated heap is always zero'd, maybe we just make this
a policy for the uPy API to keep it simple (ie any new implementation of
memory allocation must zero all allocations). This follows the D
language philosophy.
Before this patch, a previously used memory block which had pointers in
it may still retain those pointers if the new user of that block does
not actually use the entire block. Eg, if I want 5 blocks worth of
heap, I actually get 8 (round up to nearest 4). Then I never use the
last 3, so they keep their old values, which may be pointers pointing to
the heap, hence preventing GC.
In rare (or maybe not that rare) cases, this leads to long, unintentional
"linked lists" within the GC'd heap, filling it up completely. It's
pretty rare, because you have to reuse exactly that memory which is part
of this "linked list", and reuse it in just the right way.
This should fix issue #522, and might have something to do with
issue #510.
2014-04-25 18:37:55 -04:00
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2016-08-26 01:35:26 -04:00
# if MICROPY_GC_CONSERVATIVE_CLEAR
// be conservative and zero out all the newly allocated blocks
2020-02-26 23:36:53 -05:00
memset ( ( byte * ) ptr_in + n_blocks * BYTES_PER_BLOCK , 0 , ( new_blocks - n_blocks ) * BYTES_PER_BLOCK ) ;
2016-08-26 01:35:26 -04:00
# else
2014-04-26 17:23:42 -04:00
// zero out the additional bytes of the newly allocated blocks (see comment above in gc_alloc)
2020-02-26 23:36:53 -05:00
memset ( ( byte * ) ptr_in + n_bytes , 0 , new_blocks * BYTES_PER_BLOCK - n_bytes ) ;
2016-08-26 01:35:26 -04:00
# endif
py, gc: Zero out newly allocated blocks in the GC.
Also add some more debugging output to gc_dump_alloc_table().
Now that newly allocated heap is always zero'd, maybe we just make this
a policy for the uPy API to keep it simple (ie any new implementation of
memory allocation must zero all allocations). This follows the D
language philosophy.
Before this patch, a previously used memory block which had pointers in
it may still retain those pointers if the new user of that block does
not actually use the entire block. Eg, if I want 5 blocks worth of
heap, I actually get 8 (round up to nearest 4). Then I never use the
last 3, so they keep their old values, which may be pointers pointing to
the heap, hence preventing GC.
In rare (or maybe not that rare) cases, this leads to long, unintentional
"linked lists" within the GC'd heap, filling it up completely. It's
pretty rare, because you have to reuse exactly that memory which is part
of this "linked list", and reuse it in just the right way.
This should fix issue #522, and might have something to do with
issue #510.
2014-04-25 18:37:55 -04:00
2014-08-28 18:06:38 -04:00
# if EXTENSIVE_HEAP_PROFILING
gc_dump_alloc_table ( ) ;
# endif
2016-11-23 20:31:53 -05:00
# ifdef LOG_HEAP_ACTIVITY
gc_log_change ( block , new_blocks ) ;
# endif
2020-07-15 20:58:38 -04:00
# if CIRCUITPY_MEMORYMONITOR
2020-07-16 22:01:43 -04:00
memorymonitor_track_allocation ( new_blocks ) ;
2020-07-15 20:58:38 -04:00
# endif
2014-04-20 13:16:25 -04:00
return ptr_in ;
}
2014-03-05 16:23:04 -05:00
2016-05-04 05:14:43 -04:00
# if MICROPY_ENABLE_FINALISER
bool ftb_state = FTB_GET ( block ) ;
# else
bool ftb_state = false ;
# endif
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2015-06-13 16:53:22 -04:00
if ( ! allow_move ) {
// not allowed to move memory block so return failure
return NULL ;
}
2014-04-20 13:16:25 -04:00
// can't resize inplace; try to find a new contiguous chain
2018-01-23 19:22:05 -05:00
void * ptr_out = gc_alloc ( n_bytes , ftb_state , false ) ;
2014-04-20 13:16:25 -04:00
// check that the alloc succeeded
if ( ptr_out = = NULL ) {
return NULL ;
2013-10-21 18:45:08 -04:00
}
2014-03-05 16:23:04 -05:00
2014-06-16 04:44:29 -04:00
DEBUG_printf ( " gc_realloc(%p -> %p) \n " , ptr_in , ptr_out ) ;
2014-04-20 13:16:25 -04:00
memcpy ( ptr_out , ptr_in , n_blocks * BYTES_PER_BLOCK ) ;
gc_free ( ptr_in ) ;
return ptr_out ;
2013-10-21 18:45:08 -04:00
}
2014-04-20 04:43:38 -04:00
# endif // Alternative gc_realloc impl
2014-03-12 15:00:23 -04:00
2019-04-08 19:58:50 -04:00
bool gc_never_free ( void * ptr ) {
2019-04-09 18:12:28 -04:00
// Check to make sure the pointer is on the heap in the first place.
if ( gc_nbytes ( ptr ) = = 0 ) {
return false ;
}
2019-04-08 19:58:50 -04:00
// Pointers are stored in a linked list where each block is BYTES_PER_BLOCK long and the first
// pointer is the next block of pointers.
2021-03-15 09:57:36 -04:00
void * * current_reference_block = MP_STATE_MEM ( permanent_pointers ) ;
2019-04-08 19:58:50 -04:00
while ( current_reference_block ! = NULL ) {
2021-03-15 09:57:36 -04:00
for ( size_t i = 1 ; i < BYTES_PER_BLOCK / sizeof ( void * ) ; i + + ) {
2019-04-08 19:58:50 -04:00
if ( current_reference_block [ i ] = = NULL ) {
current_reference_block [ i ] = ptr ;
return true ;
}
}
current_reference_block = current_reference_block [ 0 ] ;
}
2021-03-15 09:57:36 -04:00
void * * next_block = gc_alloc ( BYTES_PER_BLOCK , false , true ) ;
2019-04-08 19:58:50 -04:00
if ( next_block = = NULL ) {
return false ;
}
if ( MP_STATE_MEM ( permanent_pointers ) = = NULL ) {
MP_STATE_MEM ( permanent_pointers ) = next_block ;
} else {
current_reference_block [ 0 ] = next_block ;
}
next_block [ 1 ] = ptr ;
return true ;
}
2015-01-12 17:34:38 -05:00
void gc_dump_info ( void ) {
2014-02-11 11:01:38 -05:00
gc_info_t info ;
gc_info ( & info ) ;
2015-12-18 07:52:45 -05:00
mp_printf ( & mp_plat_print , " GC: total: %u, used: %u, free: %u \n " ,
( uint ) info . total , ( uint ) info . used , ( uint ) info . free ) ;
2016-06-30 17:09:55 -04:00
mp_printf ( & mp_plat_print , " No. of 1-blocks: %u, 2-blocks: %u, max blk sz: %u, max free sz: %u \n " ,
2020-02-26 23:36:53 -05:00
( uint ) info . num_1block , ( uint ) info . num_2block , ( uint ) info . max_block , ( uint ) info . max_free ) ;
2014-02-11 11:01:38 -05:00
}
2014-02-26 17:55:59 -05:00
void gc_dump_alloc_table ( void ) {
2016-04-25 11:28:57 -04:00
GC_ENTER ( ) ;
2015-12-16 20:09:11 -05:00
static const size_t DUMP_BYTES_PER_LINE = 64 ;
2014-08-28 18:06:38 -04:00
# if !EXTENSIVE_HEAP_PROFILING
// When comparing heap output we don't want to print the starting
// pointer of the heap because it changes from run to run.
2015-04-11 07:15:47 -04:00
mp_printf ( & mp_plat_print , " GC memory layout; from %p: " , MP_STATE_MEM ( gc_pool_start ) ) ;
2014-08-28 18:06:38 -04:00
# endif
2015-12-16 20:09:11 -05:00
for ( size_t bl = 0 ; bl < MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB ; bl + + ) {
2014-08-28 18:06:38 -04:00
if ( bl % DUMP_BYTES_PER_LINE = = 0 ) {
// a new line of blocks
{
// check if this line contains only free blocks
2015-12-16 20:09:11 -05:00
size_t bl2 = bl ;
2015-01-01 18:30:53 -05:00
while ( bl2 < MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB & & ATB_GET_KIND ( bl2 ) = = AT_FREE ) {
2014-10-24 18:12:25 -04:00
bl2 + + ;
}
if ( bl2 - bl > = 2 * DUMP_BYTES_PER_LINE ) {
// there are at least 2 lines containing only free blocks, so abbreviate their printing
2015-12-18 07:52:45 -05:00
mp_printf ( & mp_plat_print , " \n (%u lines all free) " , ( uint ) ( bl2 - bl ) / DUMP_BYTES_PER_LINE ) ;
2014-10-24 18:12:25 -04:00
bl = bl2 & ( ~ ( DUMP_BYTES_PER_LINE - 1 ) ) ;
2015-01-01 18:30:53 -05:00
if ( bl > = MP_STATE_MEM ( gc_alloc_table_byte_len ) * BLOCKS_PER_ATB ) {
2014-10-24 18:12:25 -04:00
// got to end of heap
2014-08-28 18:06:38 -04:00
break ;
}
}
}
// print header for new line of blocks
2015-03-31 18:07:02 -04:00
// (the cast to uint32_t is for 16-bit ports)
2020-04-16 03:13:57 -04:00
// mp_printf(&mp_plat_print, "\n%05x: ", (uint)(PTR_FROM_BLOCK(bl) & (uint32_t)0xfffff));
2015-04-11 07:15:47 -04:00
mp_printf ( & mp_plat_print , " \n %05x: " , ( uint ) ( ( bl * BYTES_PER_BLOCK ) & ( uint32_t ) 0xfffff ) ) ;
2014-02-26 17:55:59 -05:00
}
int c = ' ' ;
2013-10-22 10:25:25 -04:00
switch ( ATB_GET_KIND ( bl ) ) {
2020-02-26 23:36:53 -05:00
case AT_FREE :
c = ' . ' ;
break ;
2014-10-17 10:12:57 -04:00
/* this prints out if the object is reachable from BSS or STACK (for unix only)
case AT_HEAD : {
c = ' h ' ;
2015-01-08 04:32:45 -05:00
void * * ptrs = ( void * * ) ( void * ) & mp_state_ctx ;
mp_uint_t len = offsetof ( mp_state_ctx_t , vm . stack_top ) / sizeof ( mp_uint_t ) ;
2014-10-17 10:12:57 -04:00
for ( mp_uint_t i = 0 ; i < len ; i + + ) {
mp_uint_t ptr = ( mp_uint_t ) ptrs [ i ] ;
if ( VERIFY_PTR ( ptr ) & & BLOCK_FROM_PTR ( ptr ) = = bl ) {
c = ' B ' ;
break ;
}
}
if ( c = = ' h ' ) {
ptrs = ( void * * ) & c ;
2016-04-22 18:44:56 -04:00
len = ( ( mp_uint_t ) MP_STATE_THREAD ( stack_top ) - ( mp_uint_t ) & c ) / sizeof ( mp_uint_t ) ;
2014-10-17 10:12:57 -04:00
for ( mp_uint_t i = 0 ; i < len ; i + + ) {
mp_uint_t ptr = ( mp_uint_t ) ptrs [ i ] ;
if ( VERIFY_PTR ( ptr ) & & BLOCK_FROM_PTR ( ptr ) = = bl ) {
c = ' S ' ;
break ;
}
}
}
break ;
}
*/
2014-10-24 18:12:25 -04:00
/* this prints the uPy object type of the head block */
py, gc: Zero out newly allocated blocks in the GC.
Also add some more debugging output to gc_dump_alloc_table().
Now that newly allocated heap is always zero'd, maybe we just make this
a policy for the uPy API to keep it simple (ie any new implementation of
memory allocation must zero all allocations). This follows the D
language philosophy.
Before this patch, a previously used memory block which had pointers in
it may still retain those pointers if the new user of that block does
not actually use the entire block. Eg, if I want 5 blocks worth of
heap, I actually get 8 (round up to nearest 4). Then I never use the
last 3, so they keep their old values, which may be pointers pointing to
the heap, hence preventing GC.
In rare (or maybe not that rare) cases, this leads to long, unintentional
"linked lists" within the GC'd heap, filling it up completely. It's
pretty rare, because you have to reuse exactly that memory which is part
of this "linked list", and reuse it in just the right way.
This should fix issue #522, and might have something to do with
issue #510.
2014-04-25 18:37:55 -04:00
case AT_HEAD : {
2021-03-15 09:57:36 -04:00
# pragma GCC diagnostic push
# pragma GCC diagnostic ignored "-Wcast-align"
2020-02-26 23:36:53 -05:00
void * * ptr = ( void * * ) ( MP_STATE_MEM ( gc_pool_start ) + bl * BYTES_PER_BLOCK ) ;
2021-03-15 09:57:36 -04:00
# pragma GCC diagnostic pop
2020-02-26 23:36:53 -05:00
if ( * ptr = = & mp_type_tuple ) {
c = ' T ' ;
} else if ( * ptr = = & mp_type_list ) {
c = ' L ' ;
} else if ( * ptr = = & mp_type_dict ) {
c = ' D ' ;
} else if ( * ptr = = & mp_type_str | | * ptr = = & mp_type_bytes ) {
c = ' S ' ;
}
2016-05-11 12:21:53 -04:00
# if MICROPY_PY_BUILTINS_BYTEARRAY
2020-02-26 23:36:53 -05:00
else if ( * ptr = = & mp_type_bytearray ) {
c = ' A ' ;
}
2016-05-11 12:21:53 -04:00
# endif
# if MICROPY_PY_ARRAY
2020-02-26 23:36:53 -05:00
else if ( * ptr = = & mp_type_array ) {
c = ' A ' ;
}
2016-05-11 12:21:53 -04:00
# endif
2014-11-05 16:16:41 -05:00
# if MICROPY_PY_BUILTINS_FLOAT
2020-02-26 23:36:53 -05:00
else if ( * ptr = = & mp_type_float ) {
c = ' F ' ;
}
2014-11-05 16:16:41 -05:00
# endif
2020-02-26 23:36:53 -05:00
else if ( * ptr = = & mp_type_fun_bc ) {
c = ' B ' ;
} else if ( * ptr = = & mp_type_module ) {
c = ' M ' ;
} else {
2015-01-11 09:37:06 -05:00
c = ' h ' ;
#if 0
// This code prints "Q" for qstr-pool data, and "q" for qstr-str
// data. It can be useful to see how qstrs are being allocated,
// but is disabled by default because it is very slow.
for ( qstr_pool_t * pool = MP_STATE_VM ( last_pool ) ; c = = ' h ' & & pool ! = NULL ; pool = pool - > prev ) {
2020-02-26 23:36:53 -05:00
if ( ( qstr_pool_t * ) ptr = = pool ) {
2015-01-11 09:37:06 -05:00
c = ' Q ' ;
break ;
}
for ( const byte * * q = pool - > qstrs , * * q_top = pool - > qstrs + pool - > len ; q < q_top ; q + + ) {
2020-02-26 23:36:53 -05:00
if ( ( const byte * ) ptr = = * q ) {
2015-01-11 09:37:06 -05:00
c = ' q ' ;
break ;
}
}
}
# endif
}
py, gc: Zero out newly allocated blocks in the GC.
Also add some more debugging output to gc_dump_alloc_table().
Now that newly allocated heap is always zero'd, maybe we just make this
a policy for the uPy API to keep it simple (ie any new implementation of
memory allocation must zero all allocations). This follows the D
language philosophy.
Before this patch, a previously used memory block which had pointers in
it may still retain those pointers if the new user of that block does
not actually use the entire block. Eg, if I want 5 blocks worth of
heap, I actually get 8 (round up to nearest 4). Then I never use the
last 3, so they keep their old values, which may be pointers pointing to
the heap, hence preventing GC.
In rare (or maybe not that rare) cases, this leads to long, unintentional
"linked lists" within the GC'd heap, filling it up completely. It's
pretty rare, because you have to reuse exactly that memory which is part
of this "linked list", and reuse it in just the right way.
This should fix issue #522, and might have something to do with
issue #510.
2014-04-25 18:37:55 -04:00
break ;
}
2020-02-26 23:36:53 -05:00
case AT_TAIL :
c = ' = ' ;
break ;
case AT_MARK :
c = ' m ' ;
break ;
2013-10-22 10:25:25 -04:00
}
2015-04-11 07:15:47 -04:00
mp_printf ( & mp_plat_print , " %c " , c ) ;
2013-10-22 10:25:25 -04:00
}
2015-04-11 07:15:47 -04:00
mp_print_str ( & mp_plat_print , " \n " ) ;
2016-04-25 11:28:57 -04:00
GC_EXIT ( ) ;
2013-10-22 10:25:25 -04:00
}
2018-08-02 00:17:24 -04:00
#if 0
// For testing the GC functions
2014-02-10 14:45:54 -05:00
void gc_test ( void ) {
2014-07-03 08:25:24 -04:00
mp_uint_t len = 500 ;
mp_uint_t * heap = malloc ( len ) ;
gc_init ( heap , heap + len / sizeof ( mp_uint_t ) ) ;
2013-10-21 18:45:08 -04:00
void * ptrs [ 100 ] ;
{
2014-07-03 08:25:24 -04:00
mp_uint_t * * p = gc_alloc ( 16 , false ) ;
2014-04-05 15:35:48 -04:00
p [ 0 ] = gc_alloc ( 64 , false ) ;
p [ 1 ] = gc_alloc ( 1 , false ) ;
p [ 2 ] = gc_alloc ( 1 , false ) ;
p [ 3 ] = gc_alloc ( 1 , false ) ;
2014-07-03 08:25:24 -04:00
mp_uint_t * * * p2 = gc_alloc ( 16 , false ) ;
2013-10-21 18:45:08 -04:00
p2 [ 0 ] = p ;
p2 [ 1 ] = p ;
ptrs [ 0 ] = p2 ;
}
2020-02-26 23:36:53 -05:00
for ( int i = 0 ; i < 25 ; i + = 2 ) {
2014-07-03 08:25:24 -04:00
mp_uint_t * p = gc_alloc ( i , false ) ;
2013-10-21 18:45:08 -04:00
printf ( " p=%p \n " , p ) ;
if ( i & 3 ) {
2020-04-16 03:13:57 -04:00
// ptrs[i] = p;
2013-10-21 18:45:08 -04:00
}
}
2014-02-10 14:45:54 -05:00
printf ( " Before GC: \n " ) ;
2014-02-26 17:55:59 -05:00
gc_dump_alloc_table ( ) ;
2014-02-10 14:45:54 -05:00
printf ( " Starting GC... \n " ) ;
gc_collect_start ( ) ;
2020-02-26 23:36:53 -05:00
gc_collect_root ( ptrs , sizeof ( ptrs ) / sizeof ( void * ) ) ;
2014-02-10 14:45:54 -05:00
gc_collect_end ( ) ;
printf ( " After GC: \n " ) ;
2014-02-26 17:55:59 -05:00
gc_dump_alloc_table ( ) ;
2013-10-21 18:45:08 -04:00
}
2014-02-10 14:45:54 -05:00
# endif
2014-01-07 10:20:33 -05:00
# endif // MICROPY_ENABLE_GC