Use safepoints for INTERRUPT-THREAD
[sbcl.git] / src / runtime / gencgc.c
old mode 100644 (file)
new mode 100755 (executable)
index 856b553..175be61
@@ -168,7 +168,7 @@ boolean gc_active_p = 0;
 static boolean conservative_stack = 1;
 
 /* An array of page structures is allocated on gc initialization.
- * This helps quickly map between an address its page structure.
+ * This helps to quickly map between an address and its page structure.
  * page_table_pages is set from the size of the dynamic space. */
 page_index_t page_table_pages;
 struct page *page_table;
@@ -917,7 +917,8 @@ size_t max_new_areas;
 static void
 add_new_area(page_index_t first_page, size_t offset, size_t size)
 {
-    size_t new_area_start, c, i;
+    size_t new_area_start, c;
+    ssize_t i;
 
     /* Ignore if full. */
     if (new_areas_index >= NUM_NEW_AREAS)
@@ -1610,13 +1611,13 @@ static lispobj trans_boxed(lispobj object);
  * Currently only absolute fixups to the constant vector, or to the
  * code area are checked. */
 void
-sniff_code_object(struct code *code, unsigned long displacement)
+sniff_code_object(struct code *code, os_vm_size_t displacement)
 {
 #ifdef LISP_FEATURE_X86
     long nheader_words, ncode_words, nwords;
-    void *p;
-    void *constants_start_addr = NULL, *constants_end_addr;
-    void *code_start_addr, *code_end_addr;
+    os_vm_address_t constants_start_addr = NULL, constants_end_addr, p;
+    os_vm_address_t code_start_addr, code_end_addr;
+    os_vm_address_t code_addr = (os_vm_address_t)code;
     int fixup_found = 0;
 
     if (!check_code_fixups)
@@ -1628,10 +1629,10 @@ sniff_code_object(struct code *code, unsigned long displacement)
     nheader_words = HeaderValue(*(lispobj *)code);
     nwords = ncode_words + nheader_words;
 
-    constants_start_addr = (void *)code + 5*N_WORD_BYTES;
-    constants_end_addr = (void *)code + nheader_words*N_WORD_BYTES;
-    code_start_addr = (void *)code + nheader_words*N_WORD_BYTES;
-    code_end_addr = (void *)code + nwords*N_WORD_BYTES;
+    constants_start_addr = code_addr + 5*N_WORD_BYTES;
+    constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
+    code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
+    code_end_addr = code_addr + nwords*N_WORD_BYTES;
 
     /* Work through the unboxed code. */
     for (p = code_start_addr; p < code_end_addr; p++) {
@@ -1648,8 +1649,8 @@ sniff_code_object(struct code *code, unsigned long displacement)
         /* Check for code references. */
         /* Check for a 32 bit word that looks like an absolute
            reference to within the code adea of the code object. */
-        if ((data >= (code_start_addr-displacement))
-            && (data < (code_end_addr-displacement))) {
+        if ((data >= (void*)(code_start_addr-displacement))
+            && (data < (void*)(code_end_addr-displacement))) {
             /* function header */
             if ((d4 == 0x5e)
                 && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
@@ -1691,8 +1692,8 @@ sniff_code_object(struct code *code, unsigned long displacement)
         /* Check for a 32 bit word that looks like an absolute
            reference to within the constant vector. Constant references
            will be aligned. */
-        if ((data >= (constants_start_addr-displacement))
-            && (data < (constants_end_addr-displacement))
+        if ((data >= (void*)(constants_start_addr-displacement))
+            && (data < (void*)(constants_end_addr-displacement))
             && (((unsigned)data & 0x3) == 0)) {
             /*  Mov eax,m32 */
             if (d1 == 0xa1) {
@@ -1790,11 +1791,12 @@ gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
 /* x86-64 uses pc-relative addressing instead of this kludge */
 #ifndef LISP_FEATURE_X86_64
     long nheader_words, ncode_words, nwords;
-    void *constants_start_addr, *constants_end_addr;
-    void *code_start_addr, *code_end_addr;
+    os_vm_address_t constants_start_addr, constants_end_addr;
+    os_vm_address_t code_start_addr, code_end_addr;
+    os_vm_address_t code_addr = (os_vm_address_t)new_code;
+    os_vm_address_t old_addr = (os_vm_address_t)old_code;
+    os_vm_size_t displacement = code_addr - old_addr;
     lispobj fixups = NIL;
-    unsigned long displacement =
-        (unsigned long)new_code - (unsigned long)old_code;
     struct vector *fixups_vector;
 
     ncode_words = fixnum_value(new_code->code_size);
@@ -1803,10 +1805,10 @@ gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
     /* FSHOW((stderr,
              "/compiled code object at %x: header words = %d, code words = %d\n",
              new_code, nheader_words, ncode_words)); */
-    constants_start_addr = (void *)new_code + 5*N_WORD_BYTES;
-    constants_end_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
-    code_start_addr = (void *)new_code + nheader_words*N_WORD_BYTES;
-    code_end_addr = (void *)new_code + nwords*N_WORD_BYTES;
+    constants_start_addr = code_addr + 5*N_WORD_BYTES;
+    constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
+    code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
+    code_end_addr = code_addr + nwords*N_WORD_BYTES;
     /*
     FSHOW((stderr,
            "/const start = %x, end = %x\n",
@@ -1854,24 +1856,22 @@ gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
         long length = fixnum_value(fixups_vector->length);
         long i;
         for (i = 0; i < length; i++) {
-            unsigned long offset = fixups_vector->data[i];
+            long offset = fixups_vector->data[i];
             /* Now check the current value of offset. */
-            unsigned long old_value =
-                *(unsigned long *)((unsigned long)code_start_addr + offset);
+            os_vm_address_t old_value = *(os_vm_address_t *)(code_start_addr + offset);
 
             /* If it's within the old_code object then it must be an
              * absolute fixup (relative ones are not saved) */
-            if ((old_value >= (unsigned long)old_code)
-                && (old_value < ((unsigned long)old_code
-                                 + nwords*N_WORD_BYTES)))
+            if ((old_value >= old_addr)
+                && (old_value < (old_addr + nwords*N_WORD_BYTES)))
                 /* So add the dispacement. */
-                *(unsigned long *)((unsigned long)code_start_addr + offset) =
+                *(os_vm_address_t *)(code_start_addr + offset) =
                     old_value + displacement;
             else
                 /* It is outside the old code object so it must be a
                  * relative fixup (absolute fixups are not saved). So
                  * subtract the displacement. */
-                *(unsigned long *)((unsigned long)code_start_addr + offset) =
+                *(os_vm_address_t *)(code_start_addr + offset) =
                     old_value - displacement;
         }
     } else {
@@ -2848,8 +2848,8 @@ print_ptr(lispobj *addr)
     page_index_t pi1 = find_page_index((void*)addr);
 
     if (pi1 != -1)
-        fprintf(stderr,"  %x: page %d  alloc %d  gen %d  bytes_used %d  offset %lu  dont_move %d\n",
-                (unsigned long) addr,
+        fprintf(stderr,"  %p: page %d  alloc %d  gen %d  bytes_used %d  offset %lu  dont_move %d\n",
+                addr,
                 pi1,
                 page_table[pi1].allocated,
                 page_table[pi1].gen,
@@ -3423,7 +3423,41 @@ garbage_collect_generation(generation_index_t generation, int raise)
         for_each_thread(th) {
             void **ptr;
             void **esp=(void **)-1;
-#ifdef LISP_FEATURE_SB_THREAD
+            if (th->state == STATE_DEAD)
+                continue;
+# if defined(LISP_FEATURE_SB_SAFEPOINT)
+            /* Conservative collect_garbage is always invoked with a
+             * foreign C call or an interrupt handler on top of every
+             * existing thread, so the stored SP in each thread
+             * structure is valid, no matter which thread we are looking
+             * at.  For threads that were running Lisp code, the pitstop
+             * and edge functions maintain this value within the
+             * interrupt or exception handler. */
+            esp = os_get_csp(th);
+            assert_on_stack(th, esp);
+
+            /* In addition to pointers on the stack, also preserve the
+             * return PC, the only value from the context that we need
+             * in addition to the SP.  The return PC gets saved by the
+             * foreign call wrapper, and removed from the control stack
+             * into a register. */
+            preserve_pointer(th->pc_around_foreign_call);
+
+            /* And on platforms with interrupts: scavenge ctx registers. */
+
+            /* Disabled on Windows, because it does not have an explicit
+             * stack of `interrupt_contexts'.  The reported CSP has been
+             * chosen so that the current context on the stack is
+             * covered by the stack scan.  See also set_csp_from_context(). */
+#  ifndef LISP_FEATURE_WIN32
+            if (th != arch_os_get_current_thread()) {
+                long k = fixnum_value(
+                    SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
+                while (k > 0)
+                    preserve_context_registers(th->interrupt_contexts[--k]);
+            }
+#  endif
+# elif defined(LISP_FEATURE_SB_THREAD)
             long i,free;
             if(th==arch_os_get_current_thread()) {
                 /* Somebody is going to burn in hell for this, but casting
@@ -3442,9 +3476,12 @@ garbage_collect_generation(generation_index_t generation, int raise)
                     }
                 }
             }
-#else
+# else
             esp = (void **)((void *)&raise);
-#endif
+# endif
+            if (!esp || esp == (void*) -1)
+                lose("garbage_collect: no SP known for thread %x (OS %x)",
+                     th, th->os_thread);
             for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp;  ptr--) {
                 preserve_pointer(*ptr);
             }
@@ -3848,7 +3885,7 @@ collect_garbage(generation_index_t last_gen)
 
     /* Update auto_gc_trigger. Make sure we trigger the next GC before
      * running out of heap! */
-    if (bytes_consed_between_gcs >= dynamic_space_size - bytes_allocated)
+    if (bytes_consed_between_gcs <= (dynamic_space_size - bytes_allocated))
         auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
     else
         auto_gc_trigger = bytes_allocated + (dynamic_space_size - bytes_allocated)/2;
@@ -4038,7 +4075,8 @@ gc_init(void)
         generations[i].num_gc = 0;
         generations[i].cum_sum_bytes_allocated = 0;
         /* the tune-able parameters */
-        generations[i].bytes_consed_between_gc = bytes_consed_between_gcs;
+        generations[i].bytes_consed_between_gc
+            = bytes_consed_between_gcs/(os_vm_size_t)HIGHEST_NORMAL_GENERATION;
         generations[i].number_of_gcs_before_promotion = 1;
         generations[i].minimum_age_before_gc = 0.75;
     }
@@ -4126,6 +4164,7 @@ general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *reg
 #endif
     void *new_obj;
     void *new_free_pointer;
+    os_vm_size_t trigger_bytes = 0;
 
     gc_assert(nbytes>0);
 
@@ -4147,10 +4186,19 @@ general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *reg
         return(new_obj);        /* yup */
     }
 
+    /* We don't want to count nbytes against auto_gc_trigger unless we
+     * have to: it speeds up the tenuring of objects and slows down
+     * allocation. However, unless we do so when allocating _very_
+     * large objects we are in danger of exhausting the heap without
+     * running sufficient GCs.
+     */
+    if (nbytes >= bytes_consed_between_gcs)
+        trigger_bytes = nbytes;
+
     /* we have to go the long way around, it seems. Check whether we
      * should GC in the near future
      */
-    if (auto_gc_trigger && bytes_allocated+nbytes > auto_gc_trigger) {
+    if (auto_gc_trigger && (bytes_allocated+trigger_bytes > auto_gc_trigger)) {
         /* Don't flood the system with interrupts if the need to gc is
          * already noted. This can happen for example when SUB-GC
          * allocates or after a gc triggered in a WITHOUT-GCING. */
@@ -4159,6 +4207,9 @@ general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *reg
              * section */
             SetSymbolValue(GC_PENDING,T,thread);
             if (SymbolValue(GC_INHIBIT,thread) == NIL) {
+#ifdef LISP_FEATURE_SB_SAFEPOINT
+                thread_register_gc_trigger();
+#else
                 set_pseudo_atomic_interrupted(thread);
 #ifdef LISP_FEATURE_PPC
                 /* PPC calls alloc() from a trap or from pa_alloc(),
@@ -4172,12 +4223,14 @@ general_alloc_internal(long nbytes, int page_type_flag, struct alloc_region *reg
 #else
                 maybe_save_gc_mask_and_block_deferrables(NULL);
 #endif
+#endif
             }
         }
     }
     new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);
 
 #ifndef LISP_FEATURE_WIN32
+    /* for sb-prof, and not supported on Windows yet */
     alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
     if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
         if ((signed long) alloc_signal <= 0) {