common-lisp/sbcl.sbcl
1
0
Fork 0
mirror of git://git.code.sf.net/p/sbcl/sbcl synced 2026-09-10 07:26:40 -04:00
Code Issues Actions 8 Packages Projects Releases Wiki Activity

Improve GC auto-trigger for foreign -> lisp call

Browse source 
This fixes the problem raised in
https://groups.google.com/g/sbcl-devel/c/1gFmWawGSUE/m/jMYbd9IuBQAJ
and probably fixes lp#1906191. I'll assume it does.
This commit is contained in:
Douglas Katzman 2022-11-02 15:14:26 -04:00
parent 8ed662fbfe
commit f4ff026a0e
3 changed files with 31 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/runtime/gencgc-internal.h
View file

@ -59,8 +59,10 @@ int gencgc_handle_wp_violation(void*, void*);
#endif
typedef unsigned short page_words_t;
// The flags control the behavior of sync_close_regions()
#define LOCK_PAGE_TABLE 1
#define LOCK_CODE_ALLOCATOR 2
#define CONSUME_REMAINDER 4
/* New objects are allocated to PAGE_TYPE_MIXED or PAGE_TYPE_CONS */
/* If you change these constants, then possibly also change the following

36
src/runtime/gencgc.c
View file

@ -5304,29 +5304,49 @@ NO_SANITIZE_MEMORY lispobj listify_rest_arg(lispobj* context, sword_t context_by
typedef struct { struct alloc_region* r; int type; } close_region_arg;
static void sync_close_regions(int block_signals, int locking,
static void sync_close_regions(int block_signals, int options,
close_region_arg* a, int count)
{
sigset_t savedmask;
__attribute__((unused)) int result;
if (block_signals) block_blockable_signals(&savedmask);
if (locking & LOCK_CODE_ALLOCATOR) {
if (options & LOCK_CODE_ALLOCATOR) {
result = mutex_acquire(&code_allocator_lock);
gc_dcheck(result);
}
if (locking & LOCK_PAGE_TABLE) {
if (options & LOCK_PAGE_TABLE) {
result = mutex_acquire(&free_pages_lock);
gc_dcheck(result);
}
int i;
for (i=0; i<count; ++i)
if (find_page_index(a[i].r->start_addr) >= 0)
ensure_region_closed(a[i].r, a[i].type);
if (locking & LOCK_PAGE_TABLE) {
for (i=0; i<count; ++i) {
page_index_t p = find_page_index(a[i].r->start_addr);
if (p < 0) continue;
/* Potentially use up all remaining bytes in the TLAB before closing.
* Pages below the alloc_start for the page type cannot possibly be used,
* but we didn't properly account for that space, which has a bad effect
* on the GC auto-trigger. Wasting but counting it works better.
* But (FIXME) - why does it _NOT_ _WORK_ to include the test of 'p<alloc_start' here?
* With that test in, I still see heap exhaustions, but without the test
* - so using up the remainder of the TLAB always - we do NOT get exhaustions.
* It can't be a race, because we're holding the mutex *
if ((options & CONSUME_REMAINDER) /* && p < get_alloc_start_page(a[i].type) */ ) {
char* freeptr = a[i].r->free_pointer;
char* new_end =
(a[i].type == PAGE_TYPE_CONS) ?
PTR_ALIGN_DOWN(freeptr, GENCGC_PAGE_BYTES) + CONS_PAGE_USABLE_BYTES
: PTR_ALIGN_UP(freeptr, GENCGC_PAGE_BYTES);
// fillers may not be needed. This anticipates non-zero-filed pages though.
deposit_filler(freeptr, new_end);
a[i].r->free_pointer = new_end;
}
ensure_region_closed(a[i].r, a[i].type);
}
if (options & LOCK_PAGE_TABLE) {
result = mutex_release(&free_pages_lock);
gc_dcheck(result);
}
if (locking & LOCK_CODE_ALLOCATOR) {
if (options & LOCK_CODE_ALLOCATOR) {
result = mutex_release(&code_allocator_lock);
gc_dcheck(result);
}

2
src/runtime/thread.c
View file

@ -448,7 +448,7 @@ unregister_thread(struct thread *th,
int lock_ret;
block_blockable_signals(0);
gc_close_thread_regions(th, LOCK_PAGE_TABLE);
gc_close_thread_regions(th, LOCK_PAGE_TABLE|CONSUME_REMAINDER);
#ifdef LISP_FEATURE_SB_SAFEPOINT
pop_gcing_safety(&scribble->safety);
#else