src/runtime/gencgc-internal.h

   1 /*
   2  * Generational Conservative Garbage Collector for SBCL x86
   3  *
   4  * inline functions that gc-common.c needs sight of
   5  */
   6
   7
   8 /*
   9  * This software is part of the SBCL system. See the README file for
  10  * more information.
  11  *
  12  * This software is derived from the CMU CL system, which was
  13  * written at Carnegie Mellon University and released into the
  14  * public domain. The software is in the public domain and is
  15  * provided with absolutely no warranty. See the COPYING and CREDITS
  16  * files for more information.
  17  */
  18
  19 #ifndef _GENCGC_INTERNAL_H_
  20 #define _GENCGC_INTERNAL_H_
  21
  22 #include <limits.h>
  23 #include "gc.h"
  24 #include "gencgc-alloc-region.h"
  25 #include "genesis/code.h"
  26 #include "hopscotch.h"
  27
  28 void gc_free_heap(void);
  29 extern char *page_address(page_index_t);
  30 int gencgc_handle_wp_violation(void *);
  31 \f
  32
  33 #if N_WORD_BITS == 64
  34   // It's more economical to store scan_start_offset using 4 bytes than 8.
  35   // Doing so makes struct page fit in 8 bytes if bytes_used takes 2 bytes.
  36   //   scan_start_offset = 4
  37   //   bytes_used        = 2
  38   //   flags             = 1
  39   //   gen               = 1
  40   // If bytes_used takes 4 bytes, then the above is 10 bytes which is padded to
  41   // 12, which is still an improvement over the 16 that it would have been.
  42 # define CONDENSED_PAGE_TABLE
  43 #endif
  44
  45 #if GENCGC_CARD_BYTES > USHRT_MAX
  46 # if GENCGC_CARD_BYTES > UINT_MAX
  47 #   error "GENCGC_CARD_BYTES unexpectedly large."
  48 # else
  49     typedef unsigned int page_bytes_t;
  50 # endif
  51 #else
  52   typedef unsigned short page_bytes_t;
  53 #endif
  54
  55 /* Note that this structure is also used from Lisp-side in
  56  * src/code/room.lisp, and the Lisp-side structure layout is currently
  57  * not groveled from C code but hardcoded. Any changes to the
  58  * structure layout need to be also made there.
  59  *
  60  * FIXME: We should probably just define this structure in Lisp, and
  61  * output the C version in genesis. -- JES, 2006-12-30.
  62  */
  63 struct page {
  64     /* This is the offset from the first byte of some object in memory
  65      * prior to and no closer than the start of the page to the start
  66      * of the page.  Lower values here are better, 0 is ideal.  This
  67      * is useful for determining where to start when scanning forward
  68      * through a heap page (either for conservative root validation or
  69      * for scavenging).
  70      */
  71 #ifdef CONDENSED_PAGE_TABLE
  72     // The low bit of the offset indicates the scale factor:
  73     // 0 = double-lispwords, 1 = gc cards. Large objects are card-aligned,
  74     // and this representation allows for a 32TB contiguous block using 32K
  75     // card size. Larger allocations will have pages that can't directly
  76     // store the full offset. That has to be dealt with by the accessor.
  77     unsigned int scan_start_offset_;
  78 #else
  79     os_vm_size_t scan_start_offset_;
  80 #endif
  81
  82     /* the number of bytes of this page that are used. This may be less
  83      * than the actual bytes used for pages within the current
  84      * allocation regions. It should be 0 for all unallocated pages (not
  85      * hard to achieve).
  86      * When read, the low bit has to be masked off.
  87      */
  88     page_bytes_t bytes_used_;
  89
  90     unsigned char
  91         /*  000 free
  92          *  ?01 boxed data
  93          *  ?10 unboxed data
  94          *  ?11 code
  95          *  1?? open region
  96          *
  97          * Constants for this field are defined in gc-internal.h, the
  98          * xxx_PAGE_FLAG definitions.
  99          *
 100          * If the page is free the following slots are invalid, except
 101          * for the bytes_used which must be zero. */
 102         allocated :3,
 103         /* This is set when the page is write-protected. This should
 104          * always reflect the actual write_protect status of a page.
 105          * (If the page is written into, we catch the exception, make
 106          * the page writable, and clear this flag.) */
 107         write_protected :1,
 108         /* This flag is set when the above write_protected flag is
 109          * cleared by the SIGBUS handler (or SIGSEGV handler, for some
 110          * OSes). This is useful for re-scavenging pages that are
 111          * written during a GC. */
 112         write_protected_cleared :1,
 113         /* If this page should not be moved during a GC then this flag
 114          * is set. It's only valid during a GC for allocated pages. */
 115         dont_move :1,
 116         // FIXME: this should be identical to (dont_move & !large_object),
 117         // so we don't need to store it as a bit unto itself.
 118         /* If this page is not a large object page and contains
 119          * any objects which are pinned */
 120         has_pins :1,
 121         /* If the page is part of a large object then this flag is
 122          * set. No other objects should be allocated to these pages.
 123          * This is only valid when the page is allocated. */
 124         large_object :1;
 125
 126     /* the generation that this page belongs to. This should be valid
 127      * for all pages that may have objects allocated, even current
 128      * allocation region pages - this allows the space of an object to
 129      * be easily determined. */
 130     generation_index_t gen;
 131 };
 132 extern struct page *page_table;
 133
 134 #ifndef CONDENSED_PAGE_TABLE
 135
 136 // 32-bit doesn't need magic to reduce the size of scan_start_offset.
 137 #define set_page_scan_start_offset(index,val) \
 138   page_table[index].scan_start_offset_ = val
 139 #define page_scan_start_offset(index) page_table[index].scan_start_offset_
 140
 141 #else
 142
 143 /// A "condensed" offset reduces page table size, which improves scan locality.
 144 /// As stored, the offset is scaled down either by card size or double-lispwords.
 145 /// If the offset is the maximum, then we must check if the page pointed to by
 146 /// that offset is actually the start of a region, and retry if not.
 147 /// For debugging the iterative algorithm it helps to use a max value
 148 /// that is less than UINT_MAX to get a pass/fail more quickly.
 149
 150 //#define SCAN_START_OFS_MAX 0x3fff
 151 #define SCAN_START_OFS_MAX UINT_MAX
 152
 153 #define page_scan_start_offset(index) \
 154   (page_table[index].scan_start_offset_ != SCAN_START_OFS_MAX \
 155     ? (os_vm_size_t)(page_table[index].scan_start_offset_ & ~1) \
 156        << ((page_table[index].scan_start_offset_ & 1)?(GENCGC_CARD_SHIFT-1):WORD_SHIFT) \
 157     : scan_start_offset_iterated(index))
 158
 159 static os_vm_size_t __attribute__((unused))
 160 scan_start_offset_iterated(page_index_t index)
 161 {
 162     // The low bit of the MAX is the 'scale' bit. The max pages we can look
 163     // backwards is therefore the max shifted right by 1 bit.
 164     page_index_t tot_offset_in_pages = 0;
 165     unsigned int offset;
 166     do {
 167         page_index_t lookback_page = index - tot_offset_in_pages;
 168         offset = page_table[lookback_page].scan_start_offset_;
 169         tot_offset_in_pages += offset >> 1;
 170     } while (offset == SCAN_START_OFS_MAX);
 171     return (os_vm_size_t)tot_offset_in_pages << GENCGC_CARD_SHIFT;
 172 }
 173
 174 /// This is a macro, but it could/should be an inline function.
 175 /// Problem is that we need gc_assert() which is in gc-internal,
 176 /// and it's easy enough for GC to flip around some stuff, but then
 177 /// you have a different problem that more things get messed up,
 178 /// such as {foo}-os.c. Basically we have inclusion order issues
 179 /// that nobody ever bothers to deal with, in addition to the fact
 180 /// that a something-internal header is *directly* included by others.
 181 /// (Indirect inclusion should be allowed, direct should not be)
 182 #define set_page_scan_start_offset(index, ofs) \
 183   { os_vm_size_t ofs_ = ofs; \
 184     unsigned int lsb_ = ofs_!=0 && !(ofs_ & (GENCGC_CARD_BYTES-1)); \
 185     os_vm_size_t scaled_ = \
 186      (ofs_ >> (lsb_ ? GENCGC_CARD_SHIFT-1 : WORD_SHIFT)) | lsb_; \
 187     if (scaled_ > SCAN_START_OFS_MAX) \
 188      { gc_assert(lsb_ == 1); scaled_ = SCAN_START_OFS_MAX; } \
 189     page_table[index].scan_start_offset_ = scaled_; }
 190
 191 #endif
 192
 193 /// There is some additional cleverness that could potentially be had -
 194 /// the "need_to_zero" bit (a/k/a "page dirty") is obviously 1 if the page
 195 /// contains objects. Only for an empty page must we distinguish between pages
 196 /// not needing be zero-filled before next use and those which must be.
 197 /// Thus, masking off the dirty bit could be avoided by not storing it for
 198 /// any in-use page. But since that's not what we do - we set the bit to 1
 199 /// as soon as a page is used - we do have to mask off the bit.
 200 #define page_bytes_used(index) (page_table[index].bytes_used_ & ~1)
 201 #define page_need_to_zero(index) (page_table[index].bytes_used_ & 1)
 202 #define set_page_bytes_used(index,val) \
 203   page_table[index].bytes_used_ = (val) | page_need_to_zero(index)
 204 #define set_page_need_to_zero(index,val) \
 205   page_table[index].bytes_used_ = page_bytes_used(index) | val
 206
 207 /* values for the page.allocated field */
 208
 209 \f
 210 extern page_index_t page_table_pages;
 211
 212 \f
 213 /* forward declarations */
 214 #ifdef LISP_FEATURE_X86
 215 void sniff_code_object(struct code *code, os_vm_size_t displacement);
 216 void gencgc_apply_code_fixups(struct code *old_code, struct code *new_code);
 217 #endif
 218
 219 sword_t update_dynamic_space_free_pointer(void);
 220 void gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region);
 221 void gc_alloc_update_all_page_tables(int);
 222 void gc_set_region_empty(struct alloc_region *region);
 223
 224 /*
 225  * predicates
 226  */
 227
 228 /* Find the page index within the page_table for the given
 229  * address. Return -1 on failure. */
 230 static inline page_index_t
 231 find_page_index(void *addr)
 232 {
 233     if (addr >= (void*)DYNAMIC_SPACE_START) {
 234         page_index_t index = ((uintptr_t)addr -
 235                               (uintptr_t)DYNAMIC_SPACE_START) / GENCGC_CARD_BYTES;
 236         if (index < page_table_pages)
 237             return (index);
 238     }
 239     return (-1);
 240 }
 241
 242 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 243 static inline boolean pinned_p(lispobj obj, page_index_t page)
 244 {
 245     extern struct hopscotch_table pinned_objects;
 246     return page_table[page].has_pins
 247         && hopscotch_containsp(&pinned_objects, obj);
 248 }
 249 #else
 250 #  define pinned_p(obj, page) (0)
 251 #endif
 252
 253 // Return true only if 'obj' must be *physically* transported to survive gc.
 254 // Return false if obj is in the immobile space regardless of its generation.
 255 // Pretend pinned objects are not in oldspace so that they don't get moved.
 256 static boolean __attribute__((unused))
 257 from_space_p(lispobj obj)
 258 {
 259     page_index_t page_index = find_page_index((void*)obj);
 260     return page_index >= 0
 261         && page_table[page_index].gen == from_space
 262         && !pinned_p(obj, page_index);
 263 }
 264
 265 extern page_index_t last_free_page;
 266 extern boolean gencgc_partial_pickup;
 267
 268 #endif
 269
 270 extern uword_t
 271 walk_generation(uword_t (*proc)(lispobj*,lispobj*,uword_t),
 272                 generation_index_t generation, uword_t extra);