Places: prevent mastergc gen0.big_pages from leaking onto a places gen0.big_pages list

collects/tests/racket/benchmarks/places/place-channel.rktl

@@ -0,0 +1,45 @@
+#lang racket
+;; stress tests for place-channels
+
+
+(define (splat txt fn)
+  (call-with-output-file fn #:exists 'replace
+      (lambda (out)
+        (fprintf out "~a" txt))))
+
+(splat
+#<<END
+(module pct1 scheme
+  (provide place-main)
+
+  (define (place-main ch)
+  (define message-size (* 4024 1024))
+  (define count 500)
+    (define fourk-b-message (make-bytes message-size 66))
+    (for ([i (in-range count)])
+     (place-channel-recv ch)
+     (place-channel-send ch fourk-b-message)
+))
+)
+END
+"pct1.ss")
+
+(let ([pl (place "pct1.ss" 'place-main)])
+  (define message-size (* 4024 1024))
+  (define four-k-message (make-bytes message-size 65))
+  (define count 500)
+  (define-values (r t1 t2 t3)
+    (time-apply (lambda ()
+    (for ([i (in-range count)])
+      (place-channel-send pl four-k-message)
+      (place-channel-recv pl)
+    )) null))
+
+
+  (define B/sE (/ (* 2 count message-size) (/ t2 1000)))
+  (printf "~a ~a ~a ~a~n" r t1 t2 t3)
+  (printf "B/s ~a~n" (exact->inexact B/sE))
+  (printf "MB/s ~a~n" (exact->inexact (/ B/sE (* 1024 1024))))
+  (place-wait pl)
+)
+

src/racket/gc2/gclist.h

@@ -1,7 +1,7 @@
 #ifndef GCLIST_H
 #define GCLIST_H
 
-/* design take form the linux double linked list implementation in include/linux/list.h  */
+/* design taken from the linux double linked list implementation in include/linux/list.h  */
 
 typedef struct GCList {
   struct GCList *next;

src/racket/gc2/newgc.c

@@ -656,6 +656,12 @@ static size_t round_to_apage_size(size_t sizeb)
   return sizeb;
 }
 
+#if 0
+static inline size_t size_in_apages(mpage *page) {
+  return (page->size_class > 1) ? (round_to_apage_size(page->size) / APAGE_SIZE) : 1;
+}
+#endif
+
 static mpage *malloc_mpage()
 {
   mpage *page;
@@ -2467,6 +2473,31 @@ long GC_get_memory_use(void *o)
 /* Garbage collection proper ... and all the mess therein                    */
 /*****************************************************************************/
 
+static void promote_marked_gen0_big_page(NewGC *gc, mpage *page) {
+  page->generation = 1;
+
+  /* remove page */
+  if(page->prev) page->prev->next = page->next; else
+    gc->gen0.big_pages = page->next;
+  if(page->next) page->next->prev = page->prev;
+  
+  GCVERBOSEPAGE(gc, "MOVING BIG PAGE TO GEN1", page);
+
+  backtrace_new_page(gc, page);
+
+  /* add to gen1 */
+  page->next = gc->gen1_pages[PAGE_BIG]; 
+  page->prev = NULL;
+  if(page->next) page->next->prev = page;
+  gc->gen1_pages[PAGE_BIG] = page;
+
+  /* if we're doing memory accounting, then we need to make sure the
+     btc_mark is right */
+#ifdef NEWGC_BTC_ACCOUNT
+  BTC_set_btc_mark(gc, BIG_PAGE_TO_OBJHEAD(page));
+#endif
+}
+
 /* We use two mark routines to handle propagation. Why two? The first is the
    one that we export out, and it does a metric crapload of work. The second
    we use internally, and it doesn't do nearly as much. */
@@ -2474,6 +2505,9 @@ long GC_get_memory_use(void *o)
 /* This is the first mark routine. It's a bit complicated. */
 void GC_mark2(const void *const_p, struct NewGC *gc)
 {
+#ifdef MZ_USE_PLACES
+  int is_a_master_page = 0;
+#endif
   mpage *page;
   void *p = (void*)const_p;
 
@@ -2490,6 +2524,11 @@ void GC_mark2(const void *const_p, struct NewGC *gc)
       GCDEBUG((DEBUGOUTF,"Not marking %p (no page)\n",p));
       return;
     }
+#ifdef MZ_USE_PLACES
+    else {
+      is_a_master_page = 1;      
+    }
+#endif
   }
 
   /* toss this over to the BTC mark routine if we're doing accounting */
@@ -2513,29 +2552,13 @@ void GC_mark2(const void *const_p, struct NewGC *gc)
       page->size_class = 3;
 
       /* if this is in the nursery, we want to move it out of the nursery */
-      if(!page->generation) {
+#ifdef MZ_USE_PLACES
-        page->generation = 1;
+      if(!page->generation && !is_a_master_page) 
-
+#else
-        /* remove page */
+      if(!page->generation) 
-        if(page->prev) page->prev->next = page->next; else
-          gc->gen0.big_pages = page->next;
-        if(page->next) page->next->prev = page->prev;
-        
-        GCVERBOSEPAGE(gc, "MOVING BIG PAGE TO GEN1", page);
-
-        backtrace_new_page(gc, page);
-
-        /* add to gen1 */
-        page->next = gc->gen1_pages[PAGE_BIG]; 
-        page->prev = NULL;
-        if(page->next) page->next->prev = page;
-        gc->gen1_pages[PAGE_BIG] = page;
-
-        /* if we're doing memory accounting, then we need to make sure the
-           btc_mark is right */
-#ifdef NEWGC_BTC_ACCOUNT
-        BTC_set_btc_mark(gc, BIG_PAGE_TO_OBJHEAD(page));
 #endif
+      {
+        promote_marked_gen0_big_page(gc, page);
       }
 
       page->marked_on = 1;
@@ -2764,6 +2787,19 @@ static void propagate_marks(NewGC *gc)
     propagate_marks_worker(gc, mark_table, p);
   }
 }
+#ifdef MZ_USE_PLACES
+static void promote_marked_gen0_big_pages(NewGC *gc) {
+  mpage *page;
+  mpage *next;
+
+  for (page = gc->gen0.big_pages; page ; page = next) {
+    next = page->next;
+    if (page->marked_on) {
+      promote_marked_gen0_big_page(gc, page);
+    }
+  }
+}
+#endif
 
 void *GC_resolve(void *p)
 {
@@ -3930,6 +3966,10 @@ static void garbage_collect(NewGC *gc, int force_full, int switching_master)
   if (premaster_or_place_gc(gc))
 #endif
     GC_mark_variable_stack(GC_variable_stack, 0, get_stack_base(gc), NULL);
+#ifdef MZ_USE_PLACES
+  if (postmaster_and_master_gc(gc))
+    promote_marked_gen0_big_pages(gc);
+#endif
 
   TIME_STEP("stacked");
 

src/racket/gc2/sighand.c

@@ -49,6 +49,7 @@ void fault_handler(int sn, struct siginfo *si, void *ctx)
   if (si->si_code != SEGV_ACCERR) { /*SEGV_MAPERR*/
     if (c == SEGV_MAPERR) {
       printf("SIGSEGV MAPERR si_code %i fault on addr %p\n", c, p);
+      printf("This often means %p isn't getting marked, and was prematurely freed\n", p);
     }
     if (c == 0) {
       /* I have no idea why this happens on linux */

src/racket/mk-gdbinit.rkt

@@ -216,6 +216,10 @@ define psoq
         set $scharp = $scharp + 4
       end
     end
+    if ( $OT == <<scheme_byte_string_type>>)
+      printf "scheme_byte_string_type "
+      p (char *)((Scheme_Simple_Object *)$O)->u.byte_str_val.string_val
+    end
     if ( $OT == <<scheme_unix_path_type>>)
       printf "scheme_unix_path_type "
       p (char *)((Scheme_Simple_Object *)$O)->u.byte_str_val.string_val