lustre/kernel_patches/patches/sd_iostats-2.6-suse.patch

   1 Index: linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik/drivers/scsi/Kconfig
   2 ===================================================================
   3 --- linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik.orig/drivers/scsi/Kconfig   2004-11-11 07:28:52.000000000 -0800
   4 +++ linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik/drivers/scsi/Kconfig        2005-01-06 22:58:42.338770968 -0800
   5 @@ -55,6 +55,14 @@
   6           In this case, do not compile the driver for your SCSI host adapter
   7           (below) as a module either.
   8
   9 +config SD_IOSTATS
  10 +   bool "Enable SCSI disk I/O stats"
  11 +   depends on BLK_DEV_SD
  12 +   default y
  13 +   ---help---
  14 +     This enables SCSI disk I/O stats collection.  You must also enable
  15 +     /proc file system support if you want this feature.
  16 +
  17  config CHR_DEV_ST
  18         tristate "SCSI tape support"
  19         depends on SCSI
  20 Index: linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik/drivers/scsi/sd.c
  21 ===================================================================
  22 --- linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik.orig/drivers/scsi/sd.c      2004-11-11 07:28:28.000000000 -0800
  23 +++ linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik/drivers/scsi/sd.c   2005-01-07 01:29:47.033727872 -0800
  24 @@ -59,12 +59,44 @@
  25
  26  #include "scsi_logging.h"
  27
  28 +#if (defined(CONFIG_SD_IOSTATS) && defined(CONFIG_PROC_FS))
  29 +# include <linux/proc_fs.h>
  30 +# include <linux/seq_file.h>
  31 +
  32 +typedef struct {
  33 +        unsigned long long iostat_size;
  34 +        unsigned long long iostat_count;
  35 +} iostat_counter_t;
  36 +
  37 +#define IOSTAT_NCOUNTERS 16
  38 +typedef struct {
  39 +        iostat_counter_t        iostat_read_histogram[IOSTAT_NCOUNTERS];
  40 +        iostat_counter_t        iostat_write_histogram[IOSTAT_NCOUNTERS];
  41 +        struct timeval          iostat_timeval;
  42 +} iostat_stats_t;
  43 +
  44 +iostat_stats_t       **sd_iostats;
  45 +spinlock_t             sd_iostats_lock;
  46 +struct proc_dir_entry *sd_iostats_procdir;
  47 +char                   sd_iostats_procdir_name[] = "sd_iostats";
  48 +
  49 +extern void sd_iostats_init(void);
  50 +extern void sd_iostats_init_disk(struct gendisk *);
  51 +extern void sd_iostats_fini(void);
  52 +extern void sd_iostats_bump(int disk, unsigned int nsect, int iswrite);
  53 +#else
  54 +static inline void sd_iostats_init(void) {}
  55 +static inline void sd_iostats_init_disk(struct gendisk *disk) {}
  56 +static inline void sd_iostats_fini(void) {}
  57 +static inline void sd_iostats_bump(int disk, unsigned int nsect, int iswrite) {}
  58 +#endif
  59
  60  /*
  61   * Remaining dev_t-handling stuff
  62   */
  63  #define SD_MAJORS      16
  64  #define SD_DISKS       32768   /* anything between 256 and 262144 */
  65 +#define SD_STATS 256
  66
  67  /*
  68   * Time out in seconds for disks and Magneto-opticals (which are slower).
  69 @@ -264,6 +296,9 @@
  70         SCSI_LOG_HLQUEUE(2, printk("%s : block=%llu\n",
  71                                    disk->disk_name, (unsigned long long)block));
  72
  73 +   sd_iostats_bump(scsi_disk(disk)->index, this_count,
  74 +                   rq_data_dir(SCpnt->request) == WRITE);
  75 +
  76         /*
  77          * If we have a 1K hardware sectorsize, prevent access to single
  78          * 512 byte sectors.  In theory we could handle this - in fact
  79 @@ -460,6 +495,7 @@
  80                         scsi_set_medium_removal(sdev, SCSI_REMOVAL_PREVENT);
  81         }
  82
  83 +   sd_iostats_init_disk(disk);
  84         return 0;
  85
  86  error_out:
  87 @@ -1548,6 +1584,327 @@
  88         sd_sync_cache(sdp);
  89  }
  90
  91 +#if (defined(CONFIG_SD_IOSTATS) && defined(CONFIG_PROC_FS))
  92 +static int
  93 +sd_iostats_seq_show(struct seq_file *seq, void *v)
  94 +{
  95 +        struct timeval     now;
  96 +        struct gendisk *disk = seq->private;
  97 +        iostat_stats_t    *stats;
  98 +        unsigned long long read_len;
  99 +        unsigned long long read_len_tot;
 100 +        unsigned long      read_num;
 101 +        unsigned long      read_num_tot;
 102 +        unsigned long long write_len;
 103 +        unsigned long long write_len_tot;
 104 +        unsigned long      write_num;
 105 +        unsigned long      write_num_tot;
 106 +        int                i;
 107 +        int                maxi;
 108 +
 109 +        if (sd_iostats == NULL) {
 110 +                printk(KERN_ERR "sd_iostats_seq_show: NULL stats array\n");
 111 +                BUG();
 112 +        }
 113 +
 114 +        stats = sd_iostats[scsi_disk(disk)->index];
 115 +        if (stats == NULL) {
 116 +                printk(KERN_ERR "sd_iostats_seq_show: NULL stats entry\n");
 117 +                BUG();
 118 +        }
 119 +
 120 +        do_gettimeofday(&now);
 121 +        now.tv_sec -= stats->iostat_timeval.tv_sec;
 122 +        now.tv_usec -= stats->iostat_timeval.tv_usec;
 123 +        if (now.tv_usec < 0) {
 124 +                now.tv_usec += 1000000;
 125 +                now.tv_sec--;
 126 +        }
 127 +
 128 +        /* this sampling races with updates */
 129 +        seq_printf(seq, "index:        %lu   snapshot_time:         %lu.%06lu\n",
 130 +                   scsi_disk(disk)->index, now.tv_sec, now.tv_usec);
 131 +
 132 +        for (i = IOSTAT_NCOUNTERS - 1; i > 0; i--)
 133 +                if (stats->iostat_read_histogram[i].iostat_count != 0 ||
 134 +                    stats->iostat_write_histogram[i].iostat_count != 0)
 135 +                        break;
 136 +        maxi = i;
 137 +
 138 +        seq_printf(seq, "%8s %8s %12s %8s %12s\n", "size",
 139 +                   "reads", "total", "writes", "total");
 140 +
 141 +        read_len_tot = write_len_tot = 0;
 142 +        read_num_tot = write_num_tot = 0;
 143 +        for (i = 0; i <= maxi; i++) {
 144 +                read_len = stats->iostat_read_histogram[i].iostat_size;
 145 +                read_len_tot += read_len;
 146 +                read_num = stats->iostat_read_histogram[i].iostat_count;
 147 +                read_num_tot += read_num;
 148 +
 149 +                write_len = stats->iostat_write_histogram[i].iostat_size;
 150 +                write_len_tot += write_len;
 151 +                write_num = stats->iostat_write_histogram[i].iostat_count;
 152 +                write_num_tot += write_num;
 153 +
 154 +                seq_printf (seq, "%8d %8lu %12llu %8lu %12llu\n",
 155 +                            512<<i, read_num, read_len, write_num, write_len);
 156 +        }
 157 +
 158 +        seq_printf(seq, "%8s %8lu %12llu %8lu %12llu\n", "total",
 159 +                   read_num_tot, read_len_tot,
 160 +                   write_num_tot, write_len_tot);
 161 +        return 0;
 162 +}
 163 +
 164 +static void *
 165 +sd_iostats_seq_start(struct seq_file *p, loff_t *pos)
 166 +{
 167 +        return (*pos == 0) ? (void *)1 : NULL;
 168 +}
 169 +
 170 +static void *
 171 +sd_iostats_seq_next(struct seq_file *p, void *v, loff_t *pos)
 172 +{
 173 +        ++*pos;
 174 +        return NULL;
 175 +}
 176 +
 177 +static void
 178 +sd_iostats_seq_stop(struct seq_file *p, void *v)
 179 +{
 180 +}
 181 +
 182 +static struct seq_operations sd_iostats_seqops = {
 183 +        .start = sd_iostats_seq_start,
 184 +        .stop  = sd_iostats_seq_stop,
 185 +        .next  = sd_iostats_seq_next,
 186 +        .show  = sd_iostats_seq_show,
 187 +};
 188 +
 189 +static int
 190 +sd_iostats_seq_open (struct inode *inode, struct file *file)
 191 +{
 192 +        int                    rc;
 193 +
 194 +        rc = seq_open(file, &sd_iostats_seqops);
 195 +        if (rc != 0)
 196 +                return rc;
 197 +
 198 +        ((struct seq_file *)file->private_data)->private = PDE(inode)->data;
 199 +        return 0;
 200 +}
 201 +
 202 +static ssize_t
 203 +sd_iostats_seq_write(struct file *file, const char *buffer,
 204 +                     size_t len, loff_t *off)
 205 +{
 206 +        struct seq_file   *seq = file->private_data;
 207 +        struct gendisk *disk = seq->private;
 208 +        iostat_stats_t    *stats = sd_iostats[scsi_disk(disk)->index];
 209 +        unsigned long      flags;
 210 +
 211 +
 212 +        spin_lock_irqsave (&sd_iostats_lock, flags);
 213 +        memset (stats, 0, sizeof(*stats));
 214 +        do_gettimeofday(&stats->iostat_timeval);
 215 +        spin_unlock_irqrestore (&sd_iostats_lock, flags);
 216 +
 217 +        return len;
 218 +}
 219 +
 220 +static struct file_operations sd_iostats_proc_fops = {
 221 +        .owner   = THIS_MODULE,
 222 +        .open    = sd_iostats_seq_open,
 223 +        .read    = seq_read,
 224 +        .write   = sd_iostats_seq_write,
 225 +        .llseek  = seq_lseek,
 226 +        .release = seq_release,
 227 +};
 228 +
 229 +extern struct proc_dir_entry *proc_scsi;
 230 +
 231 +void
 232 +sd_iostats_init(void)
 233 +{
 234 +        int    i;
 235 +
 236 +        spin_lock_init(&sd_iostats_lock);
 237 +
 238 +        sd_iostats = kmalloc(SD_STATS * sizeof(iostat_stats_t *), GFP_KERNEL);
 239 +        if (sd_iostats == NULL) {
 240 +                printk(KERN_WARNING "Can't keep sd iostats: "
 241 +                       "ENOMEM allocating stats array size %ld\n",
 242 +                       SD_STATS * sizeof(iostat_stats_t *));
 243 +                return;
 244 +        }
 245 +
 246 +        for (i = 0; i < SD_STATS; i++)
 247 +                sd_iostats[i] = NULL;
 248 +
 249 +        if (proc_scsi == NULL) {
 250 +                printk(KERN_WARNING "No access to sd iostats: "
 251 +                       "proc_scsi is NULL\n");
 252 +                return;
 253 +        }
 254 +
 255 +        sd_iostats_procdir = create_proc_entry(sd_iostats_procdir_name,
 256 +                                               S_IFDIR | S_IRUGO | S_IXUGO,
 257 +                                               proc_scsi);
 258 +        if (sd_iostats_procdir == NULL) {
 259 +                printk(KERN_WARNING "No access to sd iostats: "
 260 +                       "can't create /proc/scsi/%s\n", sd_iostats_procdir_name);
 261 +                return;
 262 +        }
 263 +}
 264 +
 265 +void
 266 +sd_iostats_init_disk(struct gendisk *disk)
 267 +{
 268 +        struct proc_dir_entry *pde;
 269 +        unsigned long          flags;
 270 +        iostat_stats_t        *stats;
 271 +
 272 +        if (sd_iostats == NULL ||
 273 +            sd_iostats_procdir == NULL)
 274 +                return;
 275 +
 276 +        if (scsi_disk(disk)->index > SD_STATS) {
 277 +                printk(KERN_ERR "sd_iostats_init_disk: "
 278 +                       "unexpected disk index %d(%d)\n",
 279 +                       scsi_disk(disk)->index, SD_STATS);
 280 +                                   return;
 281 +        }
 282 +
 283 +        if (sd_iostats[scsi_disk(disk)->index] != NULL)
 284 +                return;
 285 +
 286 +        stats = kmalloc(sizeof(*stats), GFP_KERNEL);
 287 +        if (stats == NULL) {
 288 +                printk(KERN_WARNING "Can't keep %s iostats: "
 289 +                       "ENOMEM allocating stats size %ld\n",
 290 +                       disk->disk_name, sizeof(*stats));
 291 +                return;
 292 +        }
 293 +
 294 +        memset (stats, 0, sizeof(*stats));
 295 +        do_gettimeofday(&stats->iostat_timeval);
 296 +
 297 +        spin_lock_irqsave(&sd_iostats_lock, flags);
 298 +
 299 +        if (sd_iostats[scsi_disk(disk)->index] != NULL) {
 300 +                spin_unlock_irqrestore(&sd_iostats_lock, flags);
 301 +                kfree (stats);
 302 +                return;
 303 +        }
 304 +
 305 +        sd_iostats[scsi_disk(disk)->index] = stats;
 306 +
 307 +        spin_unlock_irqrestore(&sd_iostats_lock, flags);
 308 +
 309 +        pde = create_proc_entry(disk->disk_name, S_IRUGO | S_IWUSR,
 310 +                                sd_iostats_procdir);
 311 +        if (pde == NULL) {
 312 +                printk(KERN_WARNING "Can't create /proc/scsi/%s/%s\n",
 313 +                       sd_iostats_procdir_name, disk->disk_name);
 314 +        } else {
 315 +                pde->proc_fops = &sd_iostats_proc_fops;
 316 +                pde->data = disk;
 317 +        }
 318 +}
 319 +
 320 +static void sd_devname(unsigned int disknum, char *buffer)
 321 +{
 322 +        if (disknum < 26)
 323 +                sprintf(buffer, "sd%c", 'a' + disknum);
 324 +        else {
 325 +                unsigned int min1;
 326 +                unsigned int min2;
 327 +                /*
 328 +                 * For larger numbers of disks, we need to go to a new
 329 +                 * naming scheme.
 330 +                 */
 331 +                min1 = disknum / 26;
 332 +                min2 = disknum % 26;
 333 +                sprintf(buffer, "sd%c%c", 'a' + min1 - 1, 'a' + min2);
 334 +        }
 335 +}
 336 +
 337 +void
 338 +sd_iostats_fini(void)
 339 +{
 340 +        char name[6];
 341 +        int  i;
 342 +
 343 +        if (sd_iostats_procdir != NULL) {
 344 +                for (i = 0; i < SD_STATS; i++) {
 345 +                        sd_devname(i, name);
 346 +                        remove_proc_entry(name, sd_iostats_procdir);
 347 +                }
 348 +
 349 +                if (proc_scsi == NULL) {
 350 +                        printk(KERN_ERR "sd_iostats_fini: proc_scsi NULL\n");
 351 +                        BUG();
 352 +                }
 353 +                remove_proc_entry(sd_iostats_procdir_name,
 354 +                                  proc_scsi);
 355 +
 356 +                sd_iostats_procdir = NULL;
 357 +        }
 358 +
 359 +        if (sd_iostats != NULL) {
 360 +                for (i = 0; i < SD_STATS; i++) {
 361 +                        if (sd_iostats[i] != NULL)
 362 +                                kfree (sd_iostats[i]);
 363 +                }
 364 +
 365 +                kfree(sd_iostats);
 366 +                sd_iostats = NULL;
 367 +        }
 368 +}
 369 +
 370 +void
 371 +sd_iostats_bump(int disk, unsigned int nsect, int iswrite)
 372 +{
 373 +        iostat_stats_t    *stats;
 374 +        iostat_counter_t  *counter;
 375 +        int                bucket;
 376 +        int                tmp;
 377 +        unsigned long      irqflags;
 378 +
 379 +        if (sd_iostats == NULL)
 380 +                return;
 381 +
 382 +        if (disk < 0 || disk >= SD_STATS) {
 383 +                printk(KERN_ERR "sd_iostats_bump: unexpected disk index %d([0-%d])\n",
 384 +                       disk, SD_STATS);
 385 +                BUG();
 386 +        }
 387 +
 388 +        for (bucket = 0, tmp = nsect; tmp > 1; bucket++)
 389 +                tmp /= 2;
 390 +
 391 +        if (bucket >= IOSTAT_NCOUNTERS) {
 392 +                printk (KERN_ERR "sd_iostats_bump: nsect %d too big\n", nsect);
 393 +                BUG();
 394 +        }
 395 +
 396 +        spin_lock_irqsave(&sd_iostats_lock, irqflags);
 397 +
 398 +        stats = sd_iostats[disk];
 399 +        if (stats != NULL) {
 400 +                counter = iswrite ?
 401 +                          &stats->iostat_write_histogram[bucket] :
 402 +                          &stats->iostat_read_histogram[bucket];
 403 +
 404 +                counter->iostat_size += nsect;
 405 +                counter->iostat_count++;
 406 +        }
 407 +
 408 +        spin_unlock_irqrestore(&sd_iostats_lock, irqflags);
 409 +}
 410 +#endif
 411 +
 412  /**
 413   *     init_sd - entry point for this driver (both when built in or when
 414   *     a module).
 415 @@ -1557,6 +1914,7 @@
 416  static int __init init_sd(void)
 417  {
 418         int majors = 0, i;
 419 +   int rc = 0;
 420
 421         SCSI_LOG_HLQUEUE(3, printk("init_sd: sd driver entry point\n"));
 422
 423 @@ -1567,7 +1925,10 @@
 424         if (!majors)
 425                 return -ENODEV;
 426
 427 -       return scsi_register_driver(&sd_template.gendrv);
 428 +   rc = scsi_register_driver(&sd_template.gendrv);
 429 +   if (rc == 0)
 430 +      sd_iostats_init();
 431 +   return rc;
 432  }
 433
 434  /**
 435 @@ -1581,6 +1942,7 @@
 436
 437         SCSI_LOG_HLQUEUE(3, printk("exit_sd: exiting sd driver\n"));
 438
 439 +   sd_iostats_fini();
 440         scsi_unregister_driver(&sd_template.gendrv);
 441         for (i = 0; i < SD_MAJORS; i++)
 442                 unregister_blkdev(sd_major(i), "sd");
 443 Index: linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik/drivers/scsi/scsi_proc.c
 444 ===================================================================
 445 --- linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik.orig/drivers/scsi/scsi_proc.c       2004-04-03 19:36:17.000000000 -0800
 446 +++ linux-2.6.5-SLES9_SP1_BRANCH_2004111114454891_lustre.1.4.0-phik/drivers/scsi/scsi_proc.c    2005-01-07 00:15:53.905665776 -0800
 447 @@ -37,7 +37,8 @@
 448  /* 4K page size, but our output routines, use some slack for overruns */
 449  #define PROC_BLOCK_SIZE (3*1024)
 450
 451 -static struct proc_dir_entry *proc_scsi;
 452 +struct proc_dir_entry *proc_scsi;
 453 +EXPORT_SYMBOL(proc_scsi);
 454
 455  /* Protect sht->present and sht->proc_dir */
 456  static DECLARE_MUTEX(global_host_template_sem);