ldiskfs/kernel_patches/patches/rhel7.6/ext4-simple-blockalloc.patch

   1 Index: linux-stage/fs/ext4/mballoc.c
   2 ===================================================================
   3 --- linux-stage.orig/fs/ext4/mballoc.c
   4 +++ linux-stage/fs/ext4/mballoc.c
   5 @@ -2078,6 +2078,21 @@ static int ext4_mb_good_group(struct ext
   6         return 0;
   7  }
   8
   9 +static u64 available_blocks_count(struct ext4_sb_info *sbi)
  10 +{
  11 +       ext4_fsblk_t resv_blocks;
  12 +       u64 bfree;
  13 +       struct ext4_super_block *es = sbi->s_es;
  14 +
  15 +       resv_blocks = EXT4_C2B(sbi, atomic64_read(&sbi->s_resv_clusters));
  16 +       bfree = percpu_counter_sum_positive(&sbi->s_freeclusters_counter) -
  17 +                percpu_counter_sum_positive(&sbi->s_dirtyclusters_counter);
  18 +
  19 +       bfree = EXT4_C2B(sbi, max_t(s64, bfree, 0));
  20 +       return bfree - (ext4_r_blocks_count(es) + resv_blocks);
  21 +}
  22 +
  23 +
  24  static noinline_for_stack int
  25  ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
  26  {
  27 @@ -2087,6 +2102,7 @@ ext4_mb_regular_allocator(struct ext4_al
  28         struct ext4_sb_info *sbi;
  29         struct super_block *sb;
  30         struct ext4_buddy e4b;
  31 +       ext4_fsblk_t avail_blocks;
  32
  33         sb = ac->ac_sb;
  34         sbi = EXT4_SB(sb);
  35 @@ -2136,6 +2152,21 @@ ext4_mb_regular_allocator(struct ext4_al
  36
  37         /* Let's just scan groups to find more-less suitable blocks */
  38         cr = ac->ac_2order ? 0 : 1;
  39 +
  40 +       /* Choose what loop to pass based on disk fullness */
  41 +       avail_blocks = available_blocks_count(sbi) ;
  42 +
  43 +       if (avail_blocks < sbi->s_mb_c3_blocks) {
  44 +               cr = 3;
  45 +               atomic64_inc(&sbi->s_bal_cX_skipped[2]);
  46 +       } else if(avail_blocks < sbi->s_mb_c2_blocks) {
  47 +               cr = 2;
  48 +               atomic64_inc(&sbi->s_bal_cX_skipped[1]);
  49 +       } else if(avail_blocks < sbi->s_mb_c1_blocks) {
  50 +               cr = 1;
  51 +               atomic64_inc(&sbi->s_bal_cX_skipped[0]);
  52 +       }
  53 +
  54         /*
  55          * cr == 0 try to get exact allocation,
  56          * cr == 3  try to get anything
  57 @@ -2193,6 +2224,9 @@ repeat:
  58                         if (ac->ac_status != AC_STATUS_CONTINUE)
  59                                 break;
  60                 }
  61 +               /* Processed all groups and haven't found blocks */
  62 +               if (i == ngroups)
  63 +                       atomic64_inc(&sbi->s_bal_cX_failed[cr]);
  64         }
  65
  66         if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
  67 @@ -2316,6 +2350,93 @@ static const struct seq_operations ext4_
  68         .show   = ext4_mb_seq_groups_show,
  69  };
  70
  71 +static int mb_seq_alloc_show(struct seq_file *seq, void *v)
  72 +{
  73 +       struct super_block *sb = seq->private;
  74 +       struct ext4_sb_info *sbi = EXT4_SB(sb);
  75 +
  76 +       seq_printf(seq, "mballoc:\n");
  77 +       seq_printf(seq, "\tblocks: %u\n", atomic_read(&sbi->s_bal_allocated));
  78 +       seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs));
  79 +       seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success));
  80 +
  81 +       seq_printf(seq, "\textents_scanned: %u\n",
  82 +                  atomic_read(&sbi->s_bal_ex_scanned));
  83 +       seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals));
  84 +       seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders));
  85 +       seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks));
  86 +       seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks));
  87 +
  88 +       seq_printf(seq, "\tuseless_c1_loops: %lu\n",
  89 +                  atomic64_read(&sbi->s_bal_cX_failed[0]));
  90 +       seq_printf(seq, "\tuseless_c2_loops: %lu\n",
  91 +                  atomic64_read(&sbi->s_bal_cX_failed[1]));
  92 +       seq_printf(seq, "\tuseless_c3_loops: %lu\n",
  93 +                  atomic64_read(&sbi->s_bal_cX_failed[2]));
  94 +       seq_printf(seq, "\tskipped_c1_loops: %lu\n",
  95 +                  atomic64_read(&sbi->s_bal_cX_skipped[0]));
  96 +       seq_printf(seq, "\tskipped_c2_loops: %lu\n",
  97 +                  atomic64_read(&sbi->s_bal_cX_skipped[1]));
  98 +       seq_printf(seq, "\tskipped_c3_loops: %lu\n",
  99 +                  atomic64_read(&sbi->s_bal_cX_skipped[2]));
 100 +       seq_printf(seq, "\tbuddies_generated: %lu\n",
 101 +                  sbi->s_mb_buddies_generated);
 102 +       seq_printf(seq, "\tbuddies_time_used: %llu\n", sbi->s_mb_generation_time);
 103 +       seq_printf(seq, "\tpreallocated: %u\n",
 104 +                  atomic_read(&sbi->s_mb_preallocated));
 105 +       seq_printf(seq, "\tdiscarded: %u\n",
 106 +                  atomic_read(&sbi->s_mb_discarded));
 107 +       return 0;
 108 +}
 109 +
 110 +static ssize_t mb_seq_alloc_write(struct file *file,
 111 +                             const char __user *buf,
 112 +                             size_t cnt, loff_t *pos)
 113 +{
 114 +       struct ext4_sb_info *sbi = EXT4_SB(PDE_DATA(file_inode(file)));
 115 +
 116 +       atomic_set(&sbi->s_bal_allocated, 0),
 117 +       atomic_set(&sbi->s_bal_reqs, 0),
 118 +       atomic_set(&sbi->s_bal_success, 0);
 119 +
 120 +       atomic_set(&sbi->s_bal_ex_scanned, 0),
 121 +       atomic_set(&sbi->s_bal_goals, 0),
 122 +       atomic_set(&sbi->s_bal_2orders, 0),
 123 +       atomic_set(&sbi->s_bal_breaks, 0),
 124 +       atomic_set(&sbi->s_mb_lost_chunks, 0);
 125 +
 126 +       atomic64_set(&sbi->s_bal_cX_failed[0], 0),
 127 +       atomic64_set(&sbi->s_bal_cX_failed[1], 0),
 128 +       atomic64_set(&sbi->s_bal_cX_failed[2], 0);
 129 +
 130 +       atomic64_set(&sbi->s_bal_cX_skipped[0], 0),
 131 +       atomic64_set(&sbi->s_bal_cX_skipped[1], 0),
 132 +       atomic64_set(&sbi->s_bal_cX_skipped[2], 0);
 133 +
 134 +
 135 +       sbi->s_mb_buddies_generated = 0;
 136 +       sbi->s_mb_generation_time = 0;
 137 +
 138 +       atomic_set(&sbi->s_mb_preallocated, 0),
 139 +       atomic_set(&sbi->s_mb_discarded, 0);
 140 +
 141 +       return cnt;
 142 +}
 143 +
 144 +static int mb_seq_alloc_open(struct inode *inode, struct file *file)
 145 +{
 146 +       return single_open(file, mb_seq_alloc_show, PDE_DATA(inode));
 147 +}
 148 +
 149 +static const struct file_operations ext4_mb_seq_alloc_fops = {
 150 +       .owner          = THIS_MODULE,
 151 +       .open           = mb_seq_alloc_open,
 152 +       .read           = seq_read,
 153 +       .llseek         = seq_lseek,
 154 +       .release        = single_release,
 155 +       .write          = mb_seq_alloc_write,
 156 +};
 157 +
 158  #define EXT4_MB_PREALLOC_TABLE          "prealloc_table"
 159
 160  static int ext4_mb_check_and_update_prealloc(struct ext4_sb_info *sbi,
 161 @@ -2730,6 +2851,7 @@ static int ext4_groupinfo_create_slab(si
 162         return 0;
 163  }
 164
 165 +#define THRESHOLD_BLOCKS(ts) (ext4_blocks_count(sbi->s_es) / 100 * ts)
 166  int ext4_mb_init(struct super_block *sb)
 167  {
 168         struct ext4_sb_info *sbi = EXT4_SB(sb);
 169 @@ -2781,6 +2903,9 @@ int ext4_mb_init(struct super_block *sb)
 170         sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
 171         sbi->s_mb_stats = MB_DEFAULT_STATS;
 172         sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
 173 +       sbi->s_mb_c1_blocks = THRESHOLD_BLOCKS(MB_DEFAULT_C1_THRESHOLD);
 174 +       sbi->s_mb_c2_blocks = THRESHOLD_BLOCKS(MB_DEFAULT_C2_THRESHOLD);
 175 +       sbi->s_mb_c3_blocks = THRESHOLD_BLOCKS(MB_DEFAULT_C3_THRESHOLD);
 176         /*
 177          * The default group preallocation is 512, which for 4k block
 178          * sizes translates to 2 megabytes.  However for bigalloc file
 179 @@ -2853,6 +2978,8 @@ int ext4_mb_init(struct super_block *sb)
 180                 proc_create_data(EXT4_MB_PREALLOC_TABLE, S_IFREG | S_IRUGO |
 181                                  S_IWUSR, sbi->s_proc,
 182                                  &ext4_mb_prealloc_seq_fops, sb);
 183 +               proc_create_data("mb_alloc_stats", S_IFREG | S_IRUGO | S_IWUSR,
 184 +                                sbi->s_proc, &ext4_mb_seq_alloc_fops, sb);
 185                 proc_create_data("mb_last_group", S_IFREG | S_IRUGO |
 186                                  S_IWUSR, sbi->s_proc,
 187                                  &ext4_mb_seq_last_group_fops, sb);
 188 @@ -2906,6 +3033,7 @@ int ext4_mb_release(struct super_block *
 189                 remove_proc_entry("mb_last_group", sbi->s_proc);
 190                 remove_proc_entry("mb_last_start", sbi->s_proc);
 191                 remove_proc_entry(EXT4_MB_PREALLOC_TABLE, sbi->s_proc);
 192 +               remove_proc_entry("mb_alloc_stats", sbi->s_proc);
 193         }
 194
 195         if (sbi->s_group_info) {
 196 @@ -2936,6 +3064,16 @@ int ext4_mb_release(struct super_block *
 197                                 atomic_read(&sbi->s_bal_reqs),
 198                                 atomic_read(&sbi->s_bal_success));
 199                 ext4_msg(sb, KERN_INFO,
 200 +                       "mballoc: (%lu, %lu, %lu) useless c(0,1,2) loops",
 201 +                               atomic64_read(&sbi->s_bal_cX_failed[0]),
 202 +                               atomic64_read(&sbi->s_bal_cX_failed[1]),
 203 +                               atomic64_read(&sbi->s_bal_cX_failed[2]));
 204 +               ext4_msg(sb, KERN_INFO,
 205 +                       "mballoc: (%lu, %lu, %lu) skipped c(0,1,2) loops",
 206 +                               atomic64_read(&sbi->s_bal_cX_skipped[0]),
 207 +                               atomic64_read(&sbi->s_bal_cX_skipped[1]),
 208 +                               atomic64_read(&sbi->s_bal_cX_skipped[2]));
 209 +               ext4_msg(sb, KERN_INFO,
 210                       "mballoc: %u extents scanned, %u goal hits, "
 211                                 "%u 2^N hits, %u breaks, %u lost",
 212                                 atomic_read(&sbi->s_bal_ex_scanned),
 213 Index: linux-stage/fs/ext4/ext4.h
 214 ===================================================================
 215 --- linux-stage.orig/fs/ext4/ext4.h
 216 +++ linux-stage/fs/ext4/ext4.h
 217 @@ -1409,6 +1409,9 @@ struct ext4_sb_info {
 218         unsigned int s_mb_min_to_scan;
 219         unsigned int s_mb_stats;
 220         unsigned int s_mb_order2_reqs;
 221 +       ext4_fsblk_t s_mb_c1_blocks;
 222 +       ext4_fsblk_t s_mb_c2_blocks;
 223 +       ext4_fsblk_t s_mb_c3_blocks;
 224         unsigned long *s_mb_prealloc_table;
 225         unsigned int s_mb_group_prealloc;
 226         unsigned int s_max_dir_size_kb;
 227 @@ -1425,6 +1428,9 @@ struct ext4_sb_info {
 228         atomic_t s_bal_goals;   /* goal hits */
 229         atomic_t s_bal_breaks;  /* too long searches */
 230         atomic_t s_bal_2orders; /* 2^order hits */
 231 +       /* cX loop didn't find blocks */
 232 +       atomic64_t s_bal_cX_failed[3];
 233 +       atomic64_t s_bal_cX_skipped[3];
 234         spinlock_t s_bal_lock;
 235         unsigned long s_mb_buddies_generated;
 236         unsigned long long s_mb_generation_time;
 237 Index: linux-stage/fs/ext4/super.c
 238 ===================================================================
 239 --- linux-stage.orig/fs/ext4/super.c
 240 +++ linux-stage/fs/ext4/super.c
 241 @@ -2734,6 +2734,73 @@ static ssize_t sbi_deprecated_show(struc
 242         return snprintf(buf, PAGE_SIZE, "%d\n", a->u.deprecated_val);
 243  }
 244
 245 +static int save_threshold(struct ext4_sb_info *sbi, const char *buf,
 246 +                        ext4_fsblk_t *blocks) {
 247 +       unsigned long long val;
 248 +
 249 +       if (!parse_strtoull(buf, 100, &val)) {
 250 +               *blocks = val * ext4_blocks_count(sbi->s_es) / 100;
 251 +               return 0;
 252 +       }
 253 +
 254 +       return -EINVAL;
 255 +}
 256 +
 257 +#define THRESHOLD_PERCENT(ts) (ts * 100 / ext4_blocks_count(sbi->s_es))
 258 +static ssize_t mb_c1_threshold_store(struct ext4_attr *a,
 259 +                                   struct ext4_sb_info *sbi,
 260 +                                   const char *buf, size_t count)
 261 +{
 262 +       int ret;
 263 +
 264 +       ret = save_threshold(sbi, buf, &sbi->s_mb_c1_blocks);
 265 +
 266 +       return ret ? ret : count;
 267 +}
 268 +
 269 +static ssize_t mb_c1_threshold_show(struct ext4_attr *a,
 270 +                                  struct ext4_sb_info *sbi, char *buf)
 271 +{
 272 +       return snprintf(buf, PAGE_SIZE, "%llu\n",
 273 +                       THRESHOLD_PERCENT(sbi->s_mb_c1_blocks));
 274 +}
 275 +
 276 +static ssize_t mb_c2_threshold_store(struct ext4_attr *a,
 277 +                                   struct ext4_sb_info *sbi,
 278 +                                   const char *buf, size_t count)
 279 +{
 280 +       int ret;
 281 +
 282 +       ret = save_threshold(sbi, buf, &sbi->s_mb_c2_blocks);
 283 +       return ret ? ret : count;
 284 +}
 285 +
 286 +static ssize_t mb_c2_threshold_show(struct ext4_attr *a,
 287 +                                  struct ext4_sb_info *sbi, char *buf)
 288 +{
 289 +               return snprintf(buf, PAGE_SIZE, "%llu\n",
 290 +                               THRESHOLD_PERCENT(sbi->s_mb_c2_blocks));
 291 +}
 292 +
 293 +static ssize_t mb_c3_threshold_store(struct ext4_attr *a,
 294 +                                   struct ext4_sb_info *sbi,
 295 +                                   const char *buf, size_t count)
 296 +{
 297 +       int ret;
 298 +
 299 +       ret = save_threshold(sbi, buf, &sbi->s_mb_c3_blocks);
 300 +
 301 +       return ret ? ret : count;
 302 +}
 303 +
 304 +static ssize_t mb_c3_threshold_show(struct ext4_attr *a,
 305 +                                  struct ext4_sb_info *sbi, char *buf)
 306 +{
 307 +               return snprintf(buf, PAGE_SIZE, "%llu\n",
 308 +                               THRESHOLD_PERCENT(sbi->s_mb_c3_blocks));
 309 +}
 310 +
 311 +
 312  #define EXT4_ATTR_OFFSET(_name,_mode,_show,_store,_elname) \
 313  static struct ext4_attr ext4_attr_##_name = {                  \
 314         .attr = {.name = __stringify(_name), .mode = _mode },   \
 315 @@ -2790,6 +2857,9 @@ EXT4_RW_ATTR_SBI_UI(mb_stats, s_mb_stats
 316  EXT4_RW_ATTR_SBI_UI(mb_max_to_scan, s_mb_max_to_scan);
 317  EXT4_RW_ATTR_SBI_UI(mb_min_to_scan, s_mb_min_to_scan);
 318  EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs);
 319 +EXT4_RW_ATTR(mb_c1_threshold);
 320 +EXT4_RW_ATTR(mb_c2_threshold);
 321 +EXT4_RW_ATTR(mb_c3_threshold);
 322  EXT4_RW_ATTR_SBI_UI(mb_small_req, s_mb_small_req);
 323  EXT4_RW_ATTR_SBI_UI(mb_large_req, s_mb_large_req);
 324  EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
 325 @@ -2820,6 +2890,9 @@ static struct attribute *ext4_attrs[] =
 326         ATTR_LIST(mb_max_to_scan),
 327         ATTR_LIST(mb_min_to_scan),
 328         ATTR_LIST(mb_order2_req),
 329 +       ATTR_LIST(mb_c1_threshold),
 330 +       ATTR_LIST(mb_c2_threshold),
 331 +       ATTR_LIST(mb_c3_threshold),
 332         ATTR_LIST(mb_small_req),
 333         ATTR_LIST(mb_large_req),
 334         ATTR_LIST(mb_group_prealloc),
 335 Index: linux-stage/fs/ext4/mballoc.h
 336 ===================================================================
 337 --- linux-stage.orig/fs/ext4/mballoc.h
 338 +++ linux-stage/fs/ext4/mballoc.h
 339 @@ -84,6 +84,9 @@ extern ushort ext4_mballoc_debug;
 340   * for which requests use 2^N search using buddies
 341   */
 342  #define MB_DEFAULT_ORDER2_REQS         8
 343 +#define MB_DEFAULT_C1_THRESHOLD                25
 344 +#define MB_DEFAULT_C2_THRESHOLD                15
 345 +#define MB_DEFAULT_C3_THRESHOLD                5
 346
 347  /*
 348   * default group prealloc size 512 blocks