0.8.6.5
[sbcl.git] / src / runtime / gencgc.c
index aec65cd..3d14487 100644 (file)
 #include "arch.h"
 #include "gc.h"
 #include "gc-internal.h"
+#include "thread.h"
 #include "genesis/vector.h"
 #include "genesis/weak-pointer.h"
 #include "genesis/simple-fun.h"
-#include "genesis/static-symbols.h"
-#include "genesis/symbol.h"
+
+#ifdef LISP_FEATURE_SB_THREAD
+#include <sys/ptrace.h>
+#include <linux/user.h>                /* threading is presently linux-only */
+#endif
+
 /* assembly language stub that executes trap_PendingInterrupt */
 void do_pending_interrupt(void);
 
+/* forward declarations */
+int gc_find_freeish_pages(int *restart_page_ptr, int nbytes, int unboxed, struct alloc_region *alloc_region);
+void  gc_set_region_empty(struct alloc_region *region);
+void gc_alloc_update_all_page_tables(void);
+static void  gencgc_pickup_dynamic(void);
+boolean interrupt_maybe_gc_int(int, siginfo_t *, void *);
+
 \f
 /*
  * GC parameters
@@ -75,7 +87,9 @@ boolean gencgc_unmap_zero = 1;
 #endif
 
 /* the minimum size (in bytes) for a large object*/
-unsigned large_object_size = 4 * 4096;
+/* FIXME: Should this really be PAGE_BYTES? */
+unsigned large_object_size = 4 * PAGE_BYTES;
+
 \f
 /*
  * debugging
@@ -124,7 +138,8 @@ boolean gencgc_zero_check_during_free_heap = 0;
 
 /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
 unsigned long bytes_allocated = 0;
-static unsigned long auto_gc_trigger = 0;
+extern unsigned long bytes_consed_between_gcs; /* gc-common.c */
+unsigned long auto_gc_trigger = 0;
 
 /* the source and destination generations. These are set before a GC starts
  * scavenging. */
@@ -132,12 +147,6 @@ int from_space;
 int new_space;
 
 
-/* FIXME: It would be nice to use this symbolic constant instead of
- * bare 4096 almost everywhere. We could also use an assertion that
- * it's equal to getpagesize(). */
-
-#define PAGE_BYTES 4096
-
 /* An array of page structures is statically allocated.
  * This helps quickly map between an address its page structure.
  * NUM_PAGES is set from the size of the dynamic space. */
@@ -152,7 +161,7 @@ static void *heap_base = NULL;
 inline void *
 page_address(int page_num)
 {
-    return (heap_base + (page_num * 4096));
+    return (heap_base + (page_num * PAGE_BYTES));
 }
 
 /* Find the page index within the page_table for the given
@@ -163,7 +172,7 @@ find_page_index(void *addr)
     int index = addr-heap_base;
 
     if (index >= 0) {
-       index = ((unsigned int)index)/4096;
+       index = ((unsigned int)index)/PAGE_BYTES;
        if (index < NUM_PAGES)
            return (index);
     }
@@ -247,6 +256,16 @@ unsigned int  gencgc_oldest_gen_to_gc = NUM_GENERATIONS-1;
  * integrated with the Lisp code. */
 static int  last_free_page;
 \f
+/* This lock is to prevent multiple threads from simultaneously
+ * allocating new regions which overlap each other.  Note that the
+ * majority of GC is single-threaded, but alloc() may be called from
+ * >1 thread at a time and must be thread-safe.  This lock must be
+ * seized before all accesses to generations[] or to parts of
+ * page_table[] that other threads may want to see */
+
+static lispobj free_pages_lock=0;
+
+\f
 /*
  * miscellaneous heap functions
  */
@@ -321,6 +340,8 @@ gen_av_mem_age(int gen)
        / ((double)generations[gen].bytes_allocated);
 }
 
+void fpu_save(int *);          /* defined in x86-assem.S */
+void fpu_restore(int *);       /* defined in x86-assem.S */
 /* The verbose argument controls how much to print: 0 for normal
  * level of detail; 1 for debugging. */
 static void
@@ -379,7 +400,7 @@ print_generation_stats(int verbose) /* FIXME: should take FILE argument */
                i,
                boxed_cnt, unboxed_cnt, large_boxed_cnt, large_unboxed_cnt,
                generations[i].bytes_allocated,
-               (count_generation_pages(i)*4096
+               (count_generation_pages(i)*PAGE_BYTES
                 - generations[i].bytes_allocated),
                generations[i].gc_trigger,
                count_write_protect_generation_pages(i),
@@ -490,7 +511,7 @@ gc_alloc_new_region(int nbytes, int unboxed, struct alloc_region *alloc_region)
     gc_assert((alloc_region->first_page == 0)
              && (alloc_region->last_page == -1)
              && (alloc_region->free_pointer == alloc_region->end_addr));
-
+    get_spinlock(&free_pages_lock,(int) alloc_region);
     if (unboxed) {
        first_page =
            generations[gc_alloc_generation].alloc_unboxed_start_page;
@@ -499,8 +520,8 @@ gc_alloc_new_region(int nbytes, int unboxed, struct alloc_region *alloc_region)
            generations[gc_alloc_generation].alloc_start_page;
     }
     last_page=gc_find_freeish_pages(&first_page,nbytes,unboxed,alloc_region);
-    bytes_found=(4096 - page_table[first_page].bytes_used)
-           + 4096*(last_page-first_page);
+    bytes_found=(PAGE_BYTES - page_table[first_page].bytes_used)
+           + PAGE_BYTES*(last_page-first_page);
 
     /* Set up the alloc_region. */
     alloc_region->first_page = first_page;
@@ -510,20 +531,6 @@ gc_alloc_new_region(int nbytes, int unboxed, struct alloc_region *alloc_region)
     alloc_region->free_pointer = alloc_region->start_addr;
     alloc_region->end_addr = alloc_region->start_addr + bytes_found;
 
-    if (gencgc_zero_check) {
-       int *p;
-       for (p = (int *)alloc_region->start_addr;
-           p < (int *)alloc_region->end_addr; p++) {
-           if (*p != 0) {
-               /* KLUDGE: It would be nice to use %lx and explicit casts
-                * (long) in code like this, so that it is less likely to
-                * break randomly when running on a machine with different
-                * word sizes. -- WHN 19991129 */
-               lose("The new region at %x is not zero.", p);
-           }
-       }
-    }
-
     /* Set up the pages. */
 
     /* The first page may have already been in use. */
@@ -559,15 +566,32 @@ gc_alloc_new_region(int nbytes, int unboxed, struct alloc_region *alloc_region)
            alloc_region->start_addr - page_address(i);
        page_table[i].allocated |= OPEN_REGION_PAGE ;
     }
-
     /* Bump up last_free_page. */
     if (last_page+1 > last_free_page) {
        last_free_page = last_page+1;
        SetSymbolValue(ALLOCATION_POINTER,
-                      (lispobj)(((char *)heap_base) + last_free_page*4096));
+                      (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES),
+                      0);
+    }
+    release_spinlock(&free_pages_lock);
+    
+    /* we can do this after releasing free_pages_lock */
+    if (gencgc_zero_check) {
+       int *p;
+       for (p = (int *)alloc_region->start_addr;
+            p < (int *)alloc_region->end_addr; p++) {
+           if (*p != 0) {
+               /* KLUDGE: It would be nice to use %lx and explicit casts
+                * (long) in code like this, so that it is less likely to
+                * break randomly when running on a machine with different
+                * word sizes. -- WHN 19991129 */
+               lose("The new region at %x is not zero.", p);
+           }
     }
 }
 
+}
+
 /* If the record_new_objects flag is 2 then all new regions created
  * are recorded.
  *
@@ -619,13 +643,13 @@ add_new_area(int first_page, int offset, int size)
        gc_abort();
     }
 
-    new_area_start = 4096*first_page + offset;
+    new_area_start = PAGE_BYTES*first_page + offset;
 
     /* Search backwards for a prior area that this follows from. If
        found this will save adding a new area. */
     for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
        unsigned area_end =
-           4096*((*new_areas)[i].page)
+           PAGE_BYTES*((*new_areas)[i].page)
            + (*new_areas)[i].offset
            + (*new_areas)[i].size;
        /*FSHOW((stderr,
@@ -659,7 +683,7 @@ add_new_area(int first_page, int offset, int size)
        max_new_areas = new_areas_index;
 }
 
-/* Update the tables for the alloc_region. The region maybe added to
+/* Update the tables for the alloc_region. The region may be added to
  * the new_areas.
  *
  * When done the alloc_region is set up so that the next quick alloc
@@ -691,8 +715,9 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region)
 
     next_page = first_page+1;
 
-    /* Skip if no bytes were allocated. */
+    get_spinlock(&free_pages_lock,(int) alloc_region);
     if (alloc_region->free_pointer != alloc_region->start_addr) {
+       /* some bytes were allocated in the region */
        orig_first_page_bytes_used = page_table[first_page].bytes_used;
 
        gc_assert(alloc_region->start_addr == (page_address(first_page) + page_table[first_page].bytes_used));
@@ -719,8 +744,8 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region)
        /* Calculate the number of bytes used in this page. This is not
         * always the number of new bytes, unless it was free. */
        more = 0;
-       if ((bytes_used = (alloc_region->free_pointer - page_address(first_page)))>4096) {
-           bytes_used = 4096;
+       if ((bytes_used = (alloc_region->free_pointer - page_address(first_page)))>PAGE_BYTES) {
+           bytes_used = PAGE_BYTES;
            more = 1;
        }
        page_table[first_page].bytes_used = bytes_used;
@@ -746,8 +771,8 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region)
            /* Calculate the number of bytes used in this page. */
            more = 0;
            if ((bytes_used = (alloc_region->free_pointer
-                              - page_address(next_page)))>4096) {
-               bytes_used = 4096;
+                              - page_address(next_page)))>PAGE_BYTES) {
+               bytes_used = PAGE_BYTES;
                more = 1;
            }
            page_table[next_page].bytes_used = bytes_used;
@@ -794,7 +819,8 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region)
        page_table[next_page].allocated = FREE_PAGE;
        next_page++;
     }
-
+    release_spinlock(&free_pages_lock);
+    /* alloc_region is per-thread, we're ok to do this unlocked */
     gc_set_region_empty(alloc_region);
 }
 
@@ -836,6 +862,8 @@ gc_alloc_large(int nbytes, int unboxed, struct alloc_region *alloc_region)
        index ahead of the current region and bumped up here to save a
        lot of re-scanning. */
 
+    get_spinlock(&free_pages_lock,(int) alloc_region);
+
     if (unboxed) {
        first_page =
            generations[gc_alloc_generation].alloc_large_unboxed_start_page;
@@ -882,8 +910,8 @@ gc_alloc_large(int nbytes, int unboxed, struct alloc_region *alloc_region)
     /* Calc. the number of bytes used in this page. This is not
      * always the number of new bytes, unless it was free. */
     more = 0;
-    if ((bytes_used = nbytes+orig_first_page_bytes_used) > 4096) {
-       bytes_used = 4096;
+    if ((bytes_used = nbytes+orig_first_page_bytes_used) > PAGE_BYTES) {
+       bytes_used = PAGE_BYTES;
        more = 1;
     }
     page_table[first_page].bytes_used = bytes_used;
@@ -905,12 +933,12 @@ gc_alloc_large(int nbytes, int unboxed, struct alloc_region *alloc_region)
        page_table[next_page].large_object = large;
 
        page_table[next_page].first_object_offset =
-           orig_first_page_bytes_used - 4096*(next_page-first_page);
+           orig_first_page_bytes_used - PAGE_BYTES*(next_page-first_page);
 
        /* Calculate the number of bytes used in this page. */
        more = 0;
-       if ((bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt) > 4096) {
-           bytes_used = 4096;
+       if ((bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt) > PAGE_BYTES) {
+           bytes_used = PAGE_BYTES;
            more = 1;
        }
        page_table[next_page].bytes_used = bytes_used;
@@ -932,8 +960,9 @@ gc_alloc_large(int nbytes, int unboxed, struct alloc_region *alloc_region)
     if (last_page+1 > last_free_page) {
        last_free_page = last_page+1;
        SetSymbolValue(ALLOCATION_POINTER,
-                      (lispobj)(((char *)heap_base) + last_free_page*4096));
+                      (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES),0);
     }
+    release_spinlock(&free_pages_lock);
 
     return((void *)(page_address(first_page)+orig_first_page_bytes_used));
 }
@@ -951,6 +980,7 @@ gc_find_freeish_pages(int *restart_page_ptr, int nbytes, int unboxed, struct all
     int num_pages;
     int large = !alloc_region && (nbytes >= large_object_size);
 
+    gc_assert(free_pages_lock);
     /* Search for a contiguous free space of at least nbytes. If it's a
        large object then align it on a page boundary by searching for a
        free page. */
@@ -979,7 +1009,7 @@ gc_find_freeish_pages(int *restart_page_ptr, int nbytes, int unboxed, struct all
                   (page_table[first_page].large_object == 0) &&
                   (gc_alloc_generation == 0) &&
                   (page_table[first_page].gen == gc_alloc_generation) &&
-                  (page_table[first_page].bytes_used < (4096-32)) &&
+                  (page_table[first_page].bytes_used < (PAGE_BYTES-32)) &&
                   (page_table[first_page].write_protected == 0) &&
                   (page_table[first_page].dont_move == 0))
                    break;
@@ -997,7 +1027,7 @@ gc_find_freeish_pages(int *restart_page_ptr, int nbytes, int unboxed, struct all
        gc_assert(page_table[first_page].write_protected == 0);
 
        last_page = first_page;
-       bytes_found = 4096 - page_table[first_page].bytes_used;
+       bytes_found = PAGE_BYTES - page_table[first_page].bytes_used;
        num_pages = 1;
        while (((bytes_found < nbytes) 
                || (alloc_region && (num_pages < 2)))
@@ -1005,12 +1035,12 @@ gc_find_freeish_pages(int *restart_page_ptr, int nbytes, int unboxed, struct all
               && (page_table[last_page+1].allocated == FREE_PAGE)) {
            last_page++;
            num_pages++;
-           bytes_found += 4096;
+           bytes_found += PAGE_BYTES;
            gc_assert(page_table[last_page].write_protected == 0);
        }
 
-       region_size = (4096 - page_table[first_page].bytes_used)
-           + 4096*(last_page-first_page);
+       region_size = (PAGE_BYTES - page_table[first_page].bytes_used)
+           + PAGE_BYTES*(last_page-first_page);
 
        gc_assert(bytes_found == region_size);
        restart_page = last_page + 1;
@@ -1224,23 +1254,23 @@ copy_large_object(lispobj object, int nwords)
 
        next_page = first_page;
        remaining_bytes = nwords*4;
-       while (remaining_bytes > 4096) {
+       while (remaining_bytes > PAGE_BYTES) {
            gc_assert(page_table[next_page].gen == from_space);
            gc_assert(page_table[next_page].allocated == BOXED_PAGE);
            gc_assert(page_table[next_page].large_object);
            gc_assert(page_table[next_page].first_object_offset==
-                     -4096*(next_page-first_page));
-           gc_assert(page_table[next_page].bytes_used == 4096);
+                     -PAGE_BYTES*(next_page-first_page));
+           gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
 
            page_table[next_page].gen = new_space;
 
            /* Remove any write-protection. We should be able to rely
             * on the write-protect flag to avoid redundant calls. */
            if (page_table[next_page].write_protected) {
-               os_protect(page_address(next_page), 4096, OS_VM_PROT_ALL);
+               os_protect(page_address(next_page), PAGE_BYTES, OS_VM_PROT_ALL);
                page_table[next_page].write_protected = 0;
            }
-           remaining_bytes -= 4096;
+           remaining_bytes -= PAGE_BYTES;
            next_page++;
        }
 
@@ -1261,12 +1291,12 @@ copy_large_object(lispobj object, int nwords)
 
        /* Free any remaining pages; needs care. */
        next_page++;
-       while ((old_bytes_used == 4096) &&
+       while ((old_bytes_used == PAGE_BYTES) &&
               (page_table[next_page].gen == from_space) &&
               (page_table[next_page].allocated == BOXED_PAGE) &&
               page_table[next_page].large_object &&
               (page_table[next_page].first_object_offset ==
-               -(next_page - first_page)*4096)) {
+               -(next_page - first_page)*PAGE_BYTES)) {
            /* Checks out OK, free the page. Don't need to bother zeroing
             * pages as this should have been done before shrinking the
             * object. These pages shouldn't be write-protected as they
@@ -1389,18 +1419,18 @@ copy_large_unboxed_object(lispobj object, int nwords)
 
        next_page = first_page;
        remaining_bytes = nwords*4;
-       while (remaining_bytes > 4096) {
+       while (remaining_bytes > PAGE_BYTES) {
            gc_assert(page_table[next_page].gen == from_space);
            gc_assert((page_table[next_page].allocated == UNBOXED_PAGE)
                      || (page_table[next_page].allocated == BOXED_PAGE));
            gc_assert(page_table[next_page].large_object);
            gc_assert(page_table[next_page].first_object_offset==
-                     -4096*(next_page-first_page));
-           gc_assert(page_table[next_page].bytes_used == 4096);
+                     -PAGE_BYTES*(next_page-first_page));
+           gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
 
            page_table[next_page].gen = new_space;
            page_table[next_page].allocated = UNBOXED_PAGE;
-           remaining_bytes -= 4096;
+           remaining_bytes -= PAGE_BYTES;
            next_page++;
        }
 
@@ -1421,13 +1451,13 @@ copy_large_unboxed_object(lispobj object, int nwords)
 
        /* Free any remaining pages; needs care. */
        next_page++;
-       while ((old_bytes_used == 4096) &&
+       while ((old_bytes_used == PAGE_BYTES) &&
               (page_table[next_page].gen == from_space) &&
               ((page_table[next_page].allocated == UNBOXED_PAGE)
                || (page_table[next_page].allocated == BOXED_PAGE)) &&
               page_table[next_page].large_object &&
               (page_table[next_page].first_object_offset ==
-               -(next_page - first_page)*4096)) {
+               -(next_page - first_page)*PAGE_BYTES)) {
            /* Checks out OK, free the page. Don't need to both zeroing
             * pages as this should have been done before shrinking the
             * object. These pages shouldn't be write-protected, even if
@@ -2084,21 +2114,21 @@ search_space(lispobj *start, size_t words, lispobj *pointer)
     return (NULL);
 }
 
-static lispobj*
+lispobj*
 search_read_only_space(lispobj *pointer)
 {
     lispobj* start = (lispobj*)READ_ONLY_SPACE_START;
-    lispobj* end = (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER);
+    lispobj* end = (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
     if ((pointer < start) || (pointer >= end))
        return NULL;
     return (search_space(start, (pointer+2)-start, pointer));
 }
 
-static lispobj *
+lispobj *
 search_static_space(lispobj *pointer)
 {
     lispobj* start = (lispobj*)STATIC_SPACE_START;
-    lispobj* end = (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER);
+    lispobj* end = (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
     if ((pointer < start) || (pointer >= end))
        return NULL;
     return (search_space(start, (pointer+2)-start, pointer));
@@ -2122,7 +2152,8 @@ search_dynamic_space(lispobj *pointer)
 
 /* Is there any possibility that pointer is a valid Lisp object
  * reference, and/or something else (e.g. subroutine call return
- * address) which should prevent us from moving the referred-to thing? */
+ * address) which should prevent us from moving the referred-to thing?
+ * This is called from preserve_pointers() */
 static int
 possibly_valid_dynamic_space_pointer(lispobj *pointer)
 {
@@ -2149,21 +2180,7 @@ possibly_valid_dynamic_space_pointer(lispobj *pointer)
 
     /* Check that the object pointed to is consistent with the pointer
      * low tag.
-     *
-     * FIXME: It's not safe to rely on the result from this check
-     * before an object is initialized. Thus, if we were interrupted
-     * just as an object had been allocated but not initialized, the
-     * GC relying on this result could bogusly reclaim the memory.
-     * However, we can't really afford to do without this check. So
-     * we should make it safe somehow. 
-     *   (1) Perhaps just review the code to make sure
-     *       that WITHOUT-GCING or WITHOUT-INTERRUPTS or some such
-     *       thing is wrapped around critical sections where allocated
-     *       memory type bits haven't been set.
-     *   (2) Perhaps find some other hack to protect against this, e.g.
-     *       recording the result of the last call to allocate-lisp-memory,
-     *       and returning true from this function when *pointer is
-     *       a reference to that result. */
+     */
     switch (lowtag_of((lispobj)pointer)) {
     case FUN_POINTER_LOWTAG:
        /* Start_addr should be the enclosing code object, or a closure
@@ -2290,7 +2307,8 @@ possibly_valid_dynamic_space_pointer(lispobj *pointer)
        case COMPLEX_LONG_FLOAT_WIDETAG:
 #endif
        case SIMPLE_ARRAY_WIDETAG:
-       case COMPLEX_STRING_WIDETAG:
+       case COMPLEX_BASE_STRING_WIDETAG:
+       case COMPLEX_VECTOR_NIL_WIDETAG:
        case COMPLEX_BIT_VECTOR_WIDETAG:
        case COMPLEX_VECTOR_WIDETAG:
        case COMPLEX_ARRAY_WIDETAG:
@@ -2304,13 +2322,17 @@ possibly_valid_dynamic_space_pointer(lispobj *pointer)
 #ifdef LONG_FLOAT_WIDETAG
        case LONG_FLOAT_WIDETAG:
 #endif
-       case SIMPLE_STRING_WIDETAG:
+       case SIMPLE_BASE_STRING_WIDETAG:
        case SIMPLE_BIT_VECTOR_WIDETAG:
        case SIMPLE_ARRAY_NIL_WIDETAG:
        case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
        case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
+       case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
        case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
+       case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
        case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
+       case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
+       case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
        case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
        case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
@@ -2388,13 +2410,17 @@ maybe_adjust_large_object(lispobj *where)
        boxed = BOXED_PAGE;
        break;
     case BIGNUM_WIDETAG:
-    case SIMPLE_STRING_WIDETAG:
+    case SIMPLE_BASE_STRING_WIDETAG:
     case SIMPLE_BIT_VECTOR_WIDETAG:
     case SIMPLE_ARRAY_NIL_WIDETAG:
     case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
     case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
+    case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
     case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
+    case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
     case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
+    case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
+    case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
     case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
     case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
@@ -2444,21 +2470,21 @@ maybe_adjust_large_object(lispobj *where)
 
     next_page = first_page;
     remaining_bytes = nwords*4;
-    while (remaining_bytes > 4096) {
+    while (remaining_bytes > PAGE_BYTES) {
        gc_assert(page_table[next_page].gen == from_space);
        gc_assert((page_table[next_page].allocated == BOXED_PAGE)
                  || (page_table[next_page].allocated == UNBOXED_PAGE));
        gc_assert(page_table[next_page].large_object);
        gc_assert(page_table[next_page].first_object_offset ==
-                 -4096*(next_page-first_page));
-       gc_assert(page_table[next_page].bytes_used == 4096);
+                 -PAGE_BYTES*(next_page-first_page));
+       gc_assert(page_table[next_page].bytes_used == PAGE_BYTES);
 
        page_table[next_page].allocated = boxed;
 
        /* Shouldn't be write-protected at this stage. Essential that the
         * pages aren't. */
        gc_assert(!page_table[next_page].write_protected);
-       remaining_bytes -= 4096;
+       remaining_bytes -= PAGE_BYTES;
        next_page++;
     }
 
@@ -2480,13 +2506,13 @@ maybe_adjust_large_object(lispobj *where)
 
     /* Free any remaining pages; needs care. */
     next_page++;
-    while ((old_bytes_used == 4096) &&
+    while ((old_bytes_used == PAGE_BYTES) &&
           (page_table[next_page].gen == from_space) &&
           ((page_table[next_page].allocated == UNBOXED_PAGE)
            || (page_table[next_page].allocated == BOXED_PAGE)) &&
           page_table[next_page].large_object &&
           (page_table[next_page].first_object_offset ==
-           -(next_page - first_page)*4096)) {
+           -(next_page - first_page)*PAGE_BYTES)) {
        /* It checks out OK, free the page. We don't need to both zeroing
         * pages as this should have been done before shrinking the
         * object. These pages shouldn't be write protected as they
@@ -2550,17 +2576,15 @@ preserve_pointer(void *addr)
 
     /* quick check 2: Check the offset within the page.
      *
-     * FIXME: The mask should have a symbolic name, and ideally should
-     * be derived from page size instead of hardwired to 0xfff.
-     * (Also fix other uses of 0xfff, elsewhere.) */
-    if (((unsigned)addr & 0xfff) > page_table[addr_page_index].bytes_used)
+     */
+    if (((unsigned)addr & (PAGE_BYTES - 1)) > page_table[addr_page_index].bytes_used)
        return;
 
     /* Filter out anything which can't be a pointer to a Lisp object
      * (or, as a special case which also requires dont_move, a return
      * address referring to something in a CodeObject). This is
      * expensive but important, since it vastly reduces the
-     * probability that random garbage will be bogusly interpreter as
+     * probability that random garbage will be bogusly interpreted as
      * a pointer which prevents a page from moving. */
     if (!(possibly_valid_dynamic_space_pointer(addr)))
        return;
@@ -2579,7 +2603,7 @@ preserve_pointer(void *addr)
     while (page_table[first_page].first_object_offset != 0) {
        --first_page;
        /* Do some checks. */
-       gc_assert(page_table[first_page].bytes_used == 4096);
+       gc_assert(page_table[first_page].bytes_used == PAGE_BYTES);
        gc_assert(page_table[first_page].gen == from_space);
        gc_assert(page_table[first_page].allocated == region_allocation);
     }
@@ -2595,7 +2619,7 @@ preserve_pointer(void *addr)
        if ((page_table[addr_page_index].allocated == FREE_PAGE)
            || (page_table[addr_page_index].bytes_used == 0)
            /* Check the offset within the page. */
-           || (((unsigned)addr & 0xfff)
+           || (((unsigned)addr & (PAGE_BYTES - 1))
                > page_table[addr_page_index].bytes_used)) {
            FSHOW((stderr,
                   "weird? ignore ptr 0x%x to freed area of large object\n",
@@ -2629,8 +2653,8 @@ preserve_pointer(void *addr)
        gc_assert(!page_table[i].write_protected);
 
        /* Check whether this is the last page in this contiguous block.. */
-       if ((page_table[i].bytes_used < 4096)
-           /* ..or it is 4096 and is the last in the block */
+       if ((page_table[i].bytes_used < PAGE_BYTES)
+           /* ..or it is PAGE_BYTES and is the last in the block */
            || (page_table[i+1].allocated == FREE_PAGE)
            || (page_table[i+1].bytes_used == 0) /* next page free */
            || (page_table[i+1].gen != from_space) /* diff. gen */
@@ -2668,8 +2692,9 @@ update_page_write_prot(int page)
     gc_assert(page_table[page].allocated != FREE_PAGE);
     gc_assert(page_table[page].bytes_used != 0);
 
-    /* Skip if it's already write-protected or an unboxed page. */
+    /* Skip if it's already write-protected, pinned, or unboxed */
     if (page_table[page].write_protected
+       || page_table[page].dont_move
        || (page_table[page].allocated & UNBOXED_PAGE))
        return (0);
 
@@ -2703,7 +2728,7 @@ update_page_write_prot(int page)
        /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/
 
        os_protect((void *)page_addr,
-                  4096,
+                  PAGE_BYTES,
                   OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);
 
        /* Note the page as protected in the page tables. */
@@ -2775,8 +2800,8 @@ scavenge_generation(int generation)
            for (last_page = i; ; last_page++)
                /* Check whether this is the last page in this contiguous
                 * block. */
-               if ((page_table[last_page].bytes_used < 4096)
-                   /* Or it is 4096 and is the last in the block */
+               if ((page_table[last_page].bytes_used < PAGE_BYTES)
+                   /* Or it is PAGE_BYTES and is the last in the block */
                    || (!(page_table[last_page+1].allocated & BOXED_PAGE))
                    || (page_table[last_page+1].bytes_used == 0)
                    || (page_table[last_page+1].gen != generation)
@@ -2797,7 +2822,7 @@ scavenge_generation(int generation)
 #endif
                    {
                        scavenge(page_address(i), (page_table[last_page].bytes_used
-                                                  + (last_page-i)*4096)/4);
+                                                  + (last_page-i)*PAGE_BYTES)/4);
 
                        /* Now scan the pages and write protect those
                         * that don't have pointers to younger
@@ -2897,8 +2922,8 @@ scavenge_newspace_generation_one_scan(int generation)
            for (last_page = i; ;last_page++) {
                /* Check whether this is the last page in this
                 * contiguous block */
-               if ((page_table[last_page].bytes_used < 4096)
-                   /* Or it is 4096 and is the last in the block */
+               if ((page_table[last_page].bytes_used < PAGE_BYTES)
+                   /* Or it is PAGE_BYTES and is the last in the block */
                    || (!(page_table[last_page+1].allocated & BOXED_PAGE))
                    || (page_table[last_page+1].bytes_used == 0)
                    || (page_table[last_page+1].gen != generation)
@@ -2927,7 +2952,7 @@ scavenge_newspace_generation_one_scan(int generation)
                                - page_table[i].first_object_offset)/4;
                    else
                        size = (page_table[last_page].bytes_used
-                               + (last_page-i)*4096
+                               + (last_page-i)*PAGE_BYTES
                                - page_table[i].first_object_offset)/4;
                    
                    {
@@ -3094,7 +3119,7 @@ unprotect_oldspace(void)
            /* Remove any write-protection. We should be able to rely
             * on the write-protect flag to avoid redundant calls. */
            if (page_table[i].write_protected) {
-               os_protect(page_start, 4096, OS_VM_PROT_ALL);
+               os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
                page_table[i].write_protected = 0;
            }
        }
@@ -3142,7 +3167,7 @@ free_oldspace(void)
                void  *page_start = (void *)page_address(last_page);
        
                if (page_table[last_page].write_protected) {
-                   os_protect(page_start, 4096, OS_VM_PROT_ALL);
+                   os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
                    page_table[last_page].write_protected = 0;
                }
            }
@@ -3163,8 +3188,8 @@ free_oldspace(void)
 
            page_start = (void *)page_address(first_page);
 
-           os_invalidate(page_start, 4096*(last_page-first_page));
-           addr = os_validate(page_start, 4096*(last_page-first_page));
+           os_invalidate(page_start, PAGE_BYTES*(last_page-first_page));
+           addr = os_validate(page_start, PAGE_BYTES*(last_page-first_page));
            if (addr == NULL || addr != page_start) {
                /* Is this an error condition? I couldn't really tell from
                 * the old CMU CL code, which fprintf'ed a message with
@@ -3182,7 +3207,7 @@ free_oldspace(void)
            int *page_start;
 
            page_start = (int *)page_address(first_page);
-           i586_bzero(page_start, 4096*(last_page-first_page));
+           i586_bzero(page_start, PAGE_BYTES*(last_page-first_page));
        }
 
        first_page = last_page;
@@ -3231,7 +3256,7 @@ verify_space(lispobj *start, size_t words)
     int is_in_dynamic_space = (find_page_index((void*)start) != -1);
     int is_in_readonly_space =
        (READ_ONLY_SPACE_START <= (unsigned)start &&
-        (unsigned)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER));
+        (unsigned)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
 
     while (words > 0) {
        size_t count = 1;
@@ -3241,10 +3266,10 @@ verify_space(lispobj *start, size_t words)
            int page_index = find_page_index((void*)thing);
            int to_readonly_space =
                (READ_ONLY_SPACE_START <= thing &&
-                thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER));
+                thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
            int to_static_space =
                (STATIC_SPACE_START <= thing &&
-                thing < SymbolValue(STATIC_SPACE_FREE_POINTER));
+                thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
 
            /* Does it point to the dynamic space? */
            if (page_index != -1) {
@@ -3295,7 +3320,8 @@ verify_space(lispobj *start, size_t words)
                case RATIO_WIDETAG:
                case COMPLEX_WIDETAG:
                case SIMPLE_ARRAY_WIDETAG:
-               case COMPLEX_STRING_WIDETAG:
+               case COMPLEX_BASE_STRING_WIDETAG:
+               case COMPLEX_VECTOR_NIL_WIDETAG:
                case COMPLEX_BIT_VECTOR_WIDETAG:
                case COMPLEX_VECTOR_WIDETAG:
                case COMPLEX_ARRAY_WIDETAG:
@@ -3380,13 +3406,17 @@ verify_space(lispobj *start, size_t words)
 #ifdef COMPLEX_LONG_FLOAT_WIDETAG
                case COMPLEX_LONG_FLOAT_WIDETAG:
 #endif
-               case SIMPLE_STRING_WIDETAG:
+               case SIMPLE_BASE_STRING_WIDETAG:
                case SIMPLE_BIT_VECTOR_WIDETAG:
                case SIMPLE_ARRAY_NIL_WIDETAG:
                case SIMPLE_ARRAY_UNSIGNED_BYTE_2_WIDETAG:
                case SIMPLE_ARRAY_UNSIGNED_BYTE_4_WIDETAG:
+               case SIMPLE_ARRAY_UNSIGNED_BYTE_7_WIDETAG:
                case SIMPLE_ARRAY_UNSIGNED_BYTE_8_WIDETAG:
+               case SIMPLE_ARRAY_UNSIGNED_BYTE_15_WIDETAG:
                case SIMPLE_ARRAY_UNSIGNED_BYTE_16_WIDETAG:
+               case SIMPLE_ARRAY_UNSIGNED_BYTE_29_WIDETAG:
+               case SIMPLE_ARRAY_UNSIGNED_BYTE_31_WIDETAG:
                case SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG:
 #ifdef SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG
                case SIMPLE_ARRAY_SIGNED_BYTE_8_WIDETAG:
@@ -3439,18 +3469,20 @@ verify_gc(void)
      * to grep for all foo_size and rename the appropriate ones to
      * foo_count. */
     int read_only_space_size =
-       (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER)
+       (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
        - (lispobj*)READ_ONLY_SPACE_START;
     int static_space_size =
-       (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER)
+       (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
        - (lispobj*)STATIC_SPACE_START;
+    struct thread *th;
+    for_each_thread(th) {
     int binding_stack_size =
-       (lispobj*)SymbolValue(BINDING_STACK_POINTER)
-       - (lispobj*)BINDING_STACK_START;
-
+           (lispobj*)SymbolValue(BINDING_STACK_POINTER,th)
+           - (lispobj*)th->binding_stack_start;
+       verify_space(th->binding_stack_start, binding_stack_size);
+    }
     verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
     verify_space((lispobj*)STATIC_SPACE_START   , static_space_size);
-    verify_space((lispobj*)BINDING_STACK_START  , binding_stack_size);
 }
 
 static void
@@ -3476,8 +3508,8 @@ verify_generation(int  generation)
            for (last_page = i; ;last_page++)
                /* Check whether this is the last page in this contiguous
                 * block. */
-               if ((page_table[last_page].bytes_used < 4096)
-                   /* Or it is 4096 and is the last in the block */
+               if ((page_table[last_page].bytes_used < PAGE_BYTES)
+                   /* Or it is PAGE_BYTES and is the last in the block */
                    || (page_table[last_page+1].allocated != region_allocation)
                    || (page_table[last_page+1].bytes_used == 0)
                    || (page_table[last_page+1].gen != generation)
@@ -3485,7 +3517,7 @@ verify_generation(int  generation)
                    break;
 
            verify_space(page_address(i), (page_table[last_page].bytes_used
-                                          + (last_page-i)*4096)/4);
+                                          + (last_page-i)*PAGE_BYTES)/4);
            i = last_page;
        }
     }
@@ -3509,7 +3541,7 @@ verify_zero_fill(void)
                }
            }
        } else {
-           int free_bytes = 4096 - page_table[page].bytes_used;
+           int free_bytes = PAGE_BYTES - page_table[page].bytes_used;
            if (free_bytes > 0) {
                int *start_addr = (int *)((unsigned)page_address(page)
                                          + page_table[page].bytes_used);
@@ -3558,13 +3590,14 @@ write_protect_generation_pages(int generation)
     for (i = 0; i < last_free_page; i++)
        if ((page_table[i].allocated == BOXED_PAGE)
            && (page_table[i].bytes_used != 0)
+           && !page_table[i].dont_move
            && (page_table[i].gen == generation))  {
            void *page_start;
 
            page_start = (void *)page_address(i);
 
            os_protect(page_start,
-                      4096,
+                      PAGE_BYTES,
                       OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);
 
            /* Note the page as protected in the page tables. */
@@ -3588,7 +3621,7 @@ garbage_collect_generation(int generation, int raise)
     unsigned long bytes_freed;
     unsigned long i;
     unsigned long static_space_size;
-
+    struct thread *th;
     gc_assert(generation <= (NUM_GENERATIONS-1));
 
     /* The oldest generation can't be raised. */
@@ -3621,7 +3654,8 @@ garbage_collect_generation(int generation, int raise)
     /* Before any pointers are preserved, the dont_move flags on the
      * pages need to be cleared. */
     for (i = 0; i < last_free_page; i++)
-       page_table[i].dont_move = 0;
+       if(page_table[i].gen==from_space)
+           page_table[i].dont_move = 0;
 
     /* Un-write-protect the old-space pages. This is essential for the
      * promoted pages as they may contain pointers into the old-space
@@ -3630,12 +3664,46 @@ garbage_collect_generation(int generation, int raise)
      * be un-protected anyway before unmapping later. */
     unprotect_oldspace();
 
-    /* Scavenge the stack's conservative roots. */
-    {
+    /* Scavenge the stacks' conservative roots. */
+
+    /* there are potentially two stacks for each thread: the main
+     * stack, which may contain Lisp pointers, and the alternate stack.
+     * We don't ever run Lisp code on the altstack, but it may 
+     * host a sigcontext with lisp objects in it */
+
+    /* what we need to do: (1) find the stack pointer for the main
+     * stack; scavenge it (2) find the interrupt context on the
+     * alternate stack that might contain lisp values, and scavenge
+     * that */
+
+    /* we assume that none of the preceding applies to the thread that
+     * initiates GC.  If you ever call GC from inside an altstack
+     * handler, you will lose. */
+    for_each_thread(th) {
        void **ptr;
-       for (ptr = (void **)CONTROL_STACK_END - 1;
-            ptr > (void **)&raise;
-            ptr--) {
+       void **esp=(void **)-1;
+       int i,free;
+#ifdef LISP_FEATURE_SB_THREAD
+       if(th==arch_os_get_current_thread()) {
+           esp = (void **) &raise;
+       } else {
+           void **esp1;
+           free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
+           for(i=free-1;i>=0;i--) {
+               os_context_t *c=th->interrupt_contexts[i];
+               esp1 = (void **) *os_context_register_addr(c,reg_ESP);
+               if(esp1>=th->control_stack_start&& esp1<th->control_stack_end){
+                   if(esp1<esp) esp=esp1;
+                   for(ptr = (void **)(c+1); ptr>=(void **)c; ptr--) {
+                       preserve_pointer(*ptr);
+                   }
+               }
+           }
+       }
+#else
+       esp = (void **) &raise;
+#endif
+       for (ptr = (void **)th->control_stack_end; ptr > esp;  ptr--) {
            preserve_pointer(*ptr);
        }
     }
@@ -3646,9 +3714,7 @@ garbage_collect_generation(int generation, int raise)
        fprintf(stderr,
                "/non-movable pages due to conservative pointers = %d (%d bytes)\n",
                num_dont_move_pages,
-               /* FIXME: 4096 should be symbolic constant here and
-                * prob'ly elsewhere too. */
-               num_dont_move_pages * 4096);
+               num_dont_move_pages * PAGE_BYTES);
     }
 #endif
 
@@ -3656,18 +3722,31 @@ garbage_collect_generation(int generation, int raise)
 
     /* Scavenge the Lisp functions of the interrupt handlers, taking
      * care to avoid SIG_DFL and SIG_IGN. */
+    for_each_thread(th) {
+       struct interrupt_data *data=th->interrupt_data;
     for (i = 0; i < NSIG; i++) {
-       union interrupt_handler handler = interrupt_handlers[i];
+           union interrupt_handler handler = data->interrupt_handlers[i];
        if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
            !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
-           scavenge((lispobj *)(interrupt_handlers + i), 1);
+               scavenge((lispobj *)(data->interrupt_handlers + i), 1);
+           }
+       }
+    }
+    /* Scavenge the binding stacks. */
+ {
+     struct thread *th;
+     for_each_thread(th) {
+        long len= (lispobj *)SymbolValue(BINDING_STACK_POINTER,th) -
+            th->binding_stack_start;
+        scavenge((lispobj *) th->binding_stack_start,len);
+#ifdef LISP_FEATURE_SB_THREAD
+        /* do the tls as well */
+        len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) -
+            (sizeof (struct thread))/(sizeof (lispobj));
+         scavenge((lispobj *) (th+1),len);
+#endif
        }
     }
-
-    /* Scavenge the binding stack. */
-    scavenge((lispobj *) BINDING_STACK_START,
-            (lispobj *)SymbolValue(BINDING_STACK_POINTER) -
-            (lispobj *)BINDING_STACK_START);
 
     /* The original CMU CL code had scavenge-read-only-space code
      * controlled by the Lisp-level variable
@@ -3690,7 +3769,7 @@ garbage_collect_generation(int generation, int raise)
 
     /* Scavenge static space. */
     static_space_size =
-       (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER) -
+       (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
        (lispobj *)STATIC_SPACE_START;
     if (gencgc_verbose > 1) {
        FSHOW((stderr,
@@ -3801,7 +3880,7 @@ update_x86_dynamic_space_free_pointer(void)
     last_free_page = last_page+1;
 
     SetSymbolValue(ALLOCATION_POINTER,
-                  (lispobj)(((char *)heap_base) + last_free_page*4096));
+                  (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES),0);
     return 0; /* dummy value: return something ... */
 }
 
@@ -3923,7 +4002,10 @@ collect_garbage(unsigned last_gen)
     gc_alloc_generation = 0;
 
     update_x86_dynamic_space_free_pointer();
-
+    auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
+    if(gencgc_verbose)
+       fprintf(stderr,"Next gc when %ld bytes have been consed\n",
+               auto_gc_trigger);
     SHOW("returning from collect_garbage");
 }
 
@@ -3957,11 +4039,11 @@ gc_free_heap(void)
            page_start = (void *)page_address(page);
 
            /* First, remove any write-protection. */
-           os_protect(page_start, 4096, OS_VM_PROT_ALL);
+           os_protect(page_start, PAGE_BYTES, OS_VM_PROT_ALL);
            page_table[page].write_protected = 0;
 
-           os_invalidate(page_start,4096);
-           addr = os_validate(page_start,4096);
+           os_invalidate(page_start,PAGE_BYTES);
+           addr = os_validate(page_start,PAGE_BYTES);
            if (addr == NULL || addr != page_start) {
                lose("gc_free_heap: page moved, 0x%08x ==> 0x%08x",
                     page_start,
@@ -4005,7 +4087,7 @@ gc_free_heap(void)
     gc_set_region_empty(&unboxed_region);
 
     last_free_page = 0;
-    SetSymbolValue(ALLOCATION_POINTER, (lispobj)((char *)heap_base));
+    SetSymbolValue(ALLOCATION_POINTER, (lispobj)((char *)heap_base),0);
 
     if (verify_after_free_heap) {
        /* Check whether purify has left any bad pointers. */
@@ -4076,22 +4158,22 @@ gencgc_pickup_dynamic(void)
 {
     int page = 0;
     int addr = DYNAMIC_SPACE_START;
-    int alloc_ptr = SymbolValue(ALLOCATION_POINTER);
+    int alloc_ptr = SymbolValue(ALLOCATION_POINTER,0);
 
     /* Initialize the first region. */
     do {
        page_table[page].allocated = BOXED_PAGE;
        page_table[page].gen = 0;
-       page_table[page].bytes_used = 4096;
+       page_table[page].bytes_used = PAGE_BYTES;
        page_table[page].large_object = 0;
        page_table[page].first_object_offset =
            (void *)DYNAMIC_SPACE_START - page_address(page);
-       addr += 4096;
+       addr += PAGE_BYTES;
        page++;
     } while (addr < alloc_ptr);
 
-    generations[0].bytes_allocated = 4096*page;
-    bytes_allocated = 4096*page;
+    generations[0].bytes_allocated = PAGE_BYTES*page;
+    bytes_allocated = PAGE_BYTES*page;
 
 }
 
@@ -4104,7 +4186,6 @@ gc_initialize_pointers(void)
 
 \f
 
-extern boolean maybe_gc_pending ;
 /* alloc(..) is the external interface for memory allocation. It
  * allocates to generation 0. It is not called from within the garbage
  * collector as it is only external uses that need the check for heap
@@ -4120,19 +4201,33 @@ extern boolean maybe_gc_pending ;
 char *
 alloc(int nbytes)
 {
-    struct alloc_region *region=  &boxed_region; 
+    struct thread *th=arch_os_get_current_thread();
+    struct alloc_region *region= 
+       th ? &(th->alloc_region) : &boxed_region; 
     void *new_obj;
     void *new_free_pointer;
 
     /* Check for alignment allocation problems. */
     gc_assert((((unsigned)region->free_pointer & 0x7) == 0)
              && ((nbytes & 0x7) == 0));
-    /* At this point we should either be in pseudo-atomic, or early
-     * enough in cold initn that interrupts are not yet enabled anyway.
-     * It would be nice to assert same.
-     */
-    gc_assert(SymbolValue(PSEUDO_ATOMIC_ATOMIC));
-
+    if(all_threads)
+       /* there are a few places in the C code that allocate data in the
+        * heap before Lisp starts.  This is before interrupts are enabled,
+        * so we don't need to check for pseudo-atomic */
+#ifdef LISP_FEATURE_SB_THREAD
+       if(!SymbolValue(PSEUDO_ATOMIC_ATOMIC,th)) {
+           register u32 fs;
+           fprintf(stderr, "fatal error in thread 0x%x, pid=%d\n",
+                   th,getpid());
+           __asm__("movl %fs,%0" : "=r" (fs)  : );
+           fprintf(stderr, "fs is %x, th->tls_cookie=%x (should be identical)\n",
+                   debug_get_fs(),th->tls_cookie);
+           lose("If you see this message before 2003.12.01, mail details to sbcl-devel\n");
+       }
+#else
+    gc_assert(SymbolValue(PSEUDO_ATOMIC_ATOMIC,th));
+#endif
+    
     /* maybe we can do this quickly ... */
     new_free_pointer = region->free_pointer + nbytes;
     if (new_free_pointer <= region->end_addr) {
@@ -4145,33 +4240,20 @@ alloc(int nbytes)
      * we should GC in the near future
      */
     if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
-       auto_gc_trigger *= 2;
        /* set things up so that GC happens when we finish the PA
-        * section.  */
-       maybe_gc_pending=1;
-       SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED, make_fixnum(1));
+        * section.  We only do this if there wasn't a pending handler
+        * already, in case it was a gc.  If it wasn't a GC, the next
+        * allocation will get us back to this point anyway, so no harm done
+        */
+       struct interrupt_data *data=th->interrupt_data;
+       if(!data->pending_handler) 
+           maybe_defer_handler(interrupt_maybe_gc_int,data,0,0,0);
     }
     new_obj = gc_alloc_with_region(nbytes,0,region,0);
     return (new_obj);
 }
 
 \f
-/*
- * noise to manipulate the gc trigger stuff
- */
-
-void
-set_auto_gc_trigger(os_vm_size_t dynamic_usage)
-{
-    auto_gc_trigger += dynamic_usage;
-}
-
-void
-clear_auto_gc_trigger(void)
-{
-    auto_gc_trigger = 0;
-}
-\f
 /* Find the code object for the given pc, or return NULL on failure.
  *
  * FIXME: PC shouldn't be lispobj*, should it? Maybe void*? */
@@ -4232,23 +4314,25 @@ gencgc_handle_wp_violation(void* fault_addr)
        return 0;
 
     } else {
-
-       /* The only acceptable reason for an signal like this from the
-        * heap is that the generational GC write-protected the page. */
-       if (page_table[page_index].write_protected != 1) {
-           lose("access failure in heap page not marked as write-protected");
+       if (page_table[page_index].write_protected) {
+           /* Unprotect the page. */
+           os_protect(page_address(page_index), PAGE_BYTES, OS_VM_PROT_ALL);
+           page_table[page_index].write_protected_cleared = 1;
+           page_table[page_index].write_protected = 0;
+       } else {  
+           /* The only acceptable reason for this signal on a heap
+            * access is that GENCGC write-protected the page.
+            * However, if two CPUs hit a wp page near-simultaneously,
+            * we had better not have the second one lose here if it
+            * does this test after the first one has already set wp=0
+            */
+           if(page_table[page_index].write_protected_cleared != 1) 
+               lose("fault in heap page not marked as write-protected");
        }
-       
-       /* Unprotect the page. */
-       os_protect(page_address(page_index), 4096, OS_VM_PROT_ALL);
-       page_table[page_index].write_protected = 0;
-       page_table[page_index].write_protected_cleared = 1;
-
        /* Don't worry, we can handle it. */
        return 1;
     }
 }
-
 /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
  * it's not just a case of the program hitting the write barrier, and
  * are about to let Lisp deal with it. It's basically just a
@@ -4257,9 +4341,12 @@ void
 unhandled_sigmemoryfault()
 {}
 
-gc_alloc_update_all_page_tables(void)
+void gc_alloc_update_all_page_tables(void)
 {
     /* Flush the alloc regions updating the tables. */
+    struct thread *th;
+    for_each_thread(th) 
+        gc_alloc_update_page_tables(0, &th->alloc_region);
     gc_alloc_update_page_tables(1, &unboxed_region);
     gc_alloc_update_page_tables(0, &boxed_region);
 }