innodb_opt_lru_count.patch

   1 # name       : innodb_opt_lru_count.patch
   2 # introduced : 11 or before
   3 # maintainer : Yasufumi
   4 #
   5 #!!! notice !!!
   6 # Any small change to this file in the main branch
   7 # should be done or reviewed by the maintainer!
   8 --- a/storage/innobase/buf/buf0buddy.c
   9 +++ b/storage/innobase/buf/buf0buddy.c
  10 @@ -123,7 +123,7 @@
  11
  12         ut_d(BUF_BUDDY_LIST_VALIDATE(buf_pool, i));
  13
  14 -       bpage = UT_LIST_GET_FIRST(buf_pool->zip_free[i]);
  15 +       bpage = UT_LIST_GET_LAST(buf_pool->zip_free[i]);
  16
  17         if (bpage) {
  18                 ut_a(buf_page_get_state(bpage) == BUF_BLOCK_ZIP_FREE);
  19 --- a/storage/innobase/buf/buf0buf.c
  20 +++ b/storage/innobase/buf/buf0buf.c
  21 @@ -890,9 +890,9 @@
  22         block->page.in_zip_hash = FALSE;
  23         block->page.in_flush_list = FALSE;
  24         block->page.in_free_list = FALSE;
  25 -       block->page.in_LRU_list = FALSE;
  26         block->in_unzip_LRU_list = FALSE;
  27  #endif /* UNIV_DEBUG */
  28 +       block->page.in_LRU_list = FALSE;
  29  #if defined UNIV_AHI_DEBUG || defined UNIV_DEBUG
  30         block->n_pointers = 0;
  31  #endif /* UNIV_AHI_DEBUG || UNIV_DEBUG */
  32 @@ -1403,7 +1403,7 @@
  33
  34         memcpy(dpage, bpage, sizeof *dpage);
  35
  36 -       ut_d(bpage->in_LRU_list = FALSE);
  37 +       bpage->in_LRU_list = FALSE;
  38         ut_d(bpage->in_page_hash = FALSE);
  39
  40         /* relocate buf_pool->LRU */
  41 @@ -3225,8 +3225,8 @@
  42                 bpage->in_zip_hash = FALSE;
  43                 bpage->in_flush_list = FALSE;
  44                 bpage->in_free_list = FALSE;
  45 -               bpage->in_LRU_list = FALSE;
  46  #endif /* UNIV_DEBUG */
  47 +               bpage->in_LRU_list = FALSE;
  48
  49                 ut_d(bpage->in_page_hash = TRUE);
  50
  51 @@ -3391,7 +3391,7 @@
  52         ibuf_merge_or_delete_for_page(NULL, space, offset, zip_size, TRUE);
  53
  54         /* Flush pages from the end of the LRU list if necessary */
  55 -       buf_flush_free_margin(buf_pool);
  56 +       buf_flush_free_margin(buf_pool, FALSE);
  57
  58         frame = block->frame;
  59
  60 --- a/storage/innobase/buf/buf0flu.c
  61 +++ b/storage/innobase/buf/buf0flu.c
  62 @@ -431,19 +431,21 @@
  63                                 buf_page_in_file(bpage) and in the LRU list */
  64  {
  65  #ifdef UNIV_DEBUG
  66 -       buf_pool_t*     buf_pool = buf_pool_from_bpage(bpage);
  67 -       ut_ad(buf_pool_mutex_own(buf_pool));
  68 +       //buf_pool_t*   buf_pool = buf_pool_from_bpage(bpage);
  69 +       //ut_ad(buf_pool_mutex_own(buf_pool));
  70  #endif
  71 -       ut_ad(mutex_own(buf_page_get_mutex(bpage)));
  72 -       ut_ad(bpage->in_LRU_list);
  73 +       //ut_ad(mutex_own(buf_page_get_mutex(bpage)));
  74 +       //ut_ad(bpage->in_LRU_list);
  75
  76 -       if (UNIV_LIKELY(buf_page_in_file(bpage))) {
  77 +       if (UNIV_LIKELY(bpage->in_LRU_list && buf_page_in_file(bpage))) {
  78
  79                 return(bpage->oldest_modification == 0
  80                        && buf_page_get_io_fix(bpage) == BUF_IO_NONE
  81                        && bpage->buf_fix_count == 0);
  82         }
  83
  84 +       /* permited not to own LRU_mutex..  */
  85 +/*
  86         ut_print_timestamp(stderr);
  87         fprintf(stderr,
  88                 "  InnoDB: Error: buffer block state %lu"
  89 @@ -451,6 +453,7 @@
  90                 (ulong) buf_page_get_state(bpage));
  91         ut_print_buf(stderr, bpage, sizeof(buf_page_t));
  92         putc('\n', stderr);
  93 +*/
  94
  95         return(FALSE);
  96  }
  97 @@ -2049,8 +2052,14 @@
  98         buf_page_t*     bpage;
  99         ulint           n_replaceable;
 100         ulint           distance        = 0;
 101 +       ibool           have_LRU_mutex = FALSE;
 102
 103 -       buf_pool_mutex_enter(buf_pool);
 104 +       if(UT_LIST_GET_LEN(buf_pool->unzip_LRU))
 105 +               have_LRU_mutex = TRUE;
 106 +retry:
 107 +       //buf_pool_mutex_enter(buf_pool);
 108 +       if (have_LRU_mutex)
 109 +               buf_pool_mutex_enter(buf_pool);
 110
 111         n_replaceable = UT_LIST_GET_LEN(buf_pool->free);
 112
 113 @@ -2061,7 +2070,13 @@
 114                    + BUF_FLUSH_EXTRA_MARGIN(buf_pool))
 115                && (distance < BUF_LRU_FREE_SEARCH_LEN(buf_pool))) {
 116
 117 -               mutex_t* block_mutex = buf_page_get_mutex(bpage);
 118 +               mutex_t* block_mutex;
 119 +               if (!bpage->in_LRU_list) {
 120 +                       /* reatart. but it is very optimistic */
 121 +                       bpage = UT_LIST_GET_LAST(buf_pool->LRU);
 122 +                       continue;
 123 +               }
 124 +               block_mutex = buf_page_get_mutex(bpage);
 125
 126                 mutex_enter(block_mutex);
 127
 128 @@ -2076,11 +2091,18 @@
 129                 bpage = UT_LIST_GET_PREV(LRU, bpage);
 130         }
 131
 132 -       buf_pool_mutex_exit(buf_pool);
 133 +       //buf_pool_mutex_exit(buf_pool);
 134 +       if (have_LRU_mutex)
 135 +               buf_pool_mutex_exit(buf_pool);
 136
 137         if (n_replaceable >= BUF_FLUSH_FREE_BLOCK_MARGIN(buf_pool)) {
 138
 139                 return(0);
 140 +       } else if (!have_LRU_mutex) {
 141 +               /* confirm it again with LRU_mutex for exactness */
 142 +               have_LRU_mutex = TRUE;
 143 +               distance = 0;
 144 +               goto retry;
 145         }
 146
 147         return(BUF_FLUSH_FREE_BLOCK_MARGIN(buf_pool)
 148 @@ -2098,7 +2120,8 @@
 149  void
 150  buf_flush_free_margin(
 151  /*==================*/
 152 -       buf_pool_t*     buf_pool)               /*!< in: Buffer pool instance */
 153 +       buf_pool_t*     buf_pool,               /*!< in: Buffer pool instance */
 154 +       ibool           wait)
 155  {
 156         ulint   n_to_flush;
 157
 158 @@ -2109,7 +2132,7 @@
 159
 160                 n_flushed = buf_flush_LRU(buf_pool, n_to_flush);
 161
 162 -               if (n_flushed == ULINT_UNDEFINED) {
 163 +               if (wait && n_flushed == ULINT_UNDEFINED) {
 164                         /* There was an LRU type flush batch already running;
 165                         let us wait for it to end */
 166
 167 @@ -2122,8 +2145,9 @@
 168  Flushes pages from the end of all the LRU lists. */
 169  UNIV_INTERN
 170  void
 171 -buf_flush_free_margins(void)
 172 +buf_flush_free_margins(
 173  /*========================*/
 174 +       ibool   wait)
 175  {
 176         ulint   i;
 177
 178 @@ -2132,7 +2156,7 @@
 179
 180                 buf_pool = buf_pool_from_array(i);
 181
 182 -               buf_flush_free_margin(buf_pool);
 183 +               buf_flush_free_margin(buf_pool, wait);
 184         }
 185  }
 186
 187 --- a/storage/innobase/buf/buf0lru.c
 188 +++ b/storage/innobase/buf/buf0lru.c
 189 @@ -934,7 +934,7 @@
 190
 191         /* No free block was found: try to flush the LRU list */
 192
 193 -       buf_flush_free_margin(buf_pool);
 194 +       buf_flush_free_margin(buf_pool, TRUE);
 195         ++srv_buf_pool_wait_free;
 196
 197         os_aio_simulated_wake_handler_threads();
 198 @@ -1131,7 +1131,7 @@
 199
 200         /* Remove the block from the LRU list */
 201         UT_LIST_REMOVE(LRU, buf_pool->LRU, bpage);
 202 -       ut_d(bpage->in_LRU_list = FALSE);
 203 +       bpage->in_LRU_list = FALSE;
 204
 205         buf_unzip_LRU_remove_block_if_needed(bpage);
 206
 207 @@ -1210,7 +1210,7 @@
 208
 209         ut_ad(!bpage->in_LRU_list);
 210         UT_LIST_ADD_LAST(LRU, buf_pool->LRU, bpage);
 211 -       ut_d(bpage->in_LRU_list = TRUE);
 212 +       bpage->in_LRU_list = TRUE;
 213
 214         if (UT_LIST_GET_LEN(buf_pool->LRU) > BUF_LRU_OLD_MIN_LEN) {
 215
 216 @@ -1280,7 +1280,7 @@
 217                 buf_pool->LRU_old_len++;
 218         }
 219
 220 -       ut_d(bpage->in_LRU_list = TRUE);
 221 +       bpage->in_LRU_list = TRUE;
 222
 223         if (UT_LIST_GET_LEN(buf_pool->LRU) > BUF_LRU_OLD_MIN_LEN) {
 224
 225 @@ -1524,7 +1524,7 @@
 226                                 buf_page_set_old(b, buf_page_is_old(b));
 227  #endif /* UNIV_LRU_DEBUG */
 228                         } else {
 229 -                               ut_d(b->in_LRU_list = FALSE);
 230 +                               b->in_LRU_list = FALSE;
 231                                 buf_LRU_add_block_low(b, buf_page_is_old(b));
 232                         }
 233
 234 --- a/storage/innobase/buf/buf0rea.c
 235 +++ b/storage/innobase/buf/buf0rea.c
 236 @@ -367,7 +367,7 @@
 237         }
 238
 239         /* Flush pages from the end of the LRU list if necessary */
 240 -       buf_flush_free_margin(buf_pool);
 241 +       buf_flush_free_margin(buf_pool, TRUE);
 242
 243         /* Increment number of I/O operations used for LRU policy. */
 244         buf_LRU_stat_inc_io();
 245 @@ -641,7 +641,7 @@
 246         os_aio_simulated_wake_handler_threads();
 247
 248         /* Flush pages from the end of the LRU list if necessary */
 249 -       buf_flush_free_margin(buf_pool);
 250 +       buf_flush_free_margin(buf_pool, TRUE);
 251
 252  #ifdef UNIV_DEBUG
 253         if (buf_debug_prints && (count > 0)) {
 254 @@ -729,7 +729,7 @@
 255         os_aio_simulated_wake_handler_threads();
 256
 257         /* Flush pages from the end of all the LRU lists if necessary */
 258 -       buf_flush_free_margins();
 259 +       buf_flush_free_margins(FALSE);
 260
 261  #ifdef UNIV_DEBUG
 262         if (buf_debug_prints) {
 263 @@ -823,7 +823,7 @@
 264         os_aio_simulated_wake_handler_threads();
 265
 266         /* Flush pages from the end of all the LRU lists if necessary */
 267 -       buf_flush_free_margins();
 268 +       buf_flush_free_margins(FALSE);
 269
 270  #ifdef UNIV_DEBUG
 271         if (buf_debug_prints) {
 272 --- a/storage/innobase/include/buf0buf.h
 273 +++ b/storage/innobase/include/buf0buf.h
 274 @@ -1424,11 +1424,11 @@
 275
 276         UT_LIST_NODE_T(buf_page_t) LRU;
 277                                         /*!< node of the LRU list */
 278 -#ifdef UNIV_DEBUG
 279 +//#ifdef UNIV_DEBUG
 280         ibool           in_LRU_list;    /*!< TRUE if the page is in
 281                                         the LRU list; used in
 282                                         debugging */
 283 -#endif /* UNIV_DEBUG */
 284 +//#endif /* UNIV_DEBUG */
 285         unsigned        old:1;          /*!< TRUE if the block is in the old
 286                                         blocks in buf_pool->LRU_old */
 287         unsigned        freed_page_clock:31;/*!< the value of
 288 --- a/storage/innobase/include/buf0flu.h
 289 +++ b/storage/innobase/include/buf0flu.h
 290 @@ -65,13 +65,15 @@
 291  void
 292  buf_flush_free_margin(
 293  /*==================*/
 294 -        buf_pool_t*    buf_pool);
 295 +        buf_pool_t*    buf_pool,
 296 +       ibool           wait);
 297  /*********************************************************************//**
 298  Flushes pages from the end of all the LRU lists. */
 299  UNIV_INTERN
 300  void
 301 -buf_flush_free_margins(void);
 302 +buf_flush_free_margins(
 303  /*=========================*/
 304 +       ibool           wait);
 305  #endif /* !UNIV_HOTBACKUP */
 306  /********************************************************************//**
 307  Initializes a page for writing to the tablespace. */