lustre/obdfs/flushd.c

   1 /*
   2  * OBDFS Super operations - also used for Lustre file system
   3  *
   4  *
   5  *  Copyright (C) 1991, 1992  Linus Torvalds
   6  * Copryright (C) 1999 Stelias Computing Inc. <braam@stelias.com>
   7  * Copryright (C) 1999 Seagate Technology Inc.
   8  *
   9  */
  10 #define __NO_VERSION__
  11 #include <linux/module.h>
  12 #include <linux/sched.h>
  13 #include <linux/fs.h>
  14 #include <linux/malloc.h>
  15 #include <linux/locks.h>
  16 #include <linux/errno.h>
  17 #include <linux/swap.h>
  18 #include <linux/smp_lock.h>
  19 #include <linux/vmalloc.h>
  20 #include <linux/blkdev.h>
  21 #include <linux/sysrq.h>
  22 #include <linux/file.h>
  23 #include <linux/init.h>
  24 #include <linux/quotaops.h>
  25 #include <linux/iobuf.h>
  26 #include <linux/highmem.h>
  27
  28 #include <asm/uaccess.h>
  29 #include <asm/io.h>
  30 #include <asm/bitops.h>
  31 #include <asm/mmu_context.h>
  32
  33 #include <linux/obd_support.h>
  34 #include <linux/obd_class.h>
  35 #include <linux/obdfs.h>
  36
  37
  38 struct {
  39         int nfract;  /* Percentage of buffer cache dirty to
  40                         activate bdflush */
  41         int ndirty;  /* Maximum number of dirty blocks to write out per
  42                         wake-cycle */
  43         int nrefill; /* Number of clean buffers to try to obtain
  44                                 each time we call refill */
  45         int nref_dirt; /* Dirty buffer threshold for activating bdflush
  46                           when trying to refill buffers. */
  47         int interval; /* jiffies delay between kupdate flushes */
  48         int age_buffer;  /* Time for normal buffer to age before we flush it */
  49         int age_super;  /* Time for superblock to age before we flush it */
  50 /* } pupd_prm = {40, 500, 64, 256, 5*HZ, 30*HZ, 5*HZ };  */
  51 } pupd_prm = {40, 500, 64, 256, 10*HZ, 30*HZ, 5*HZ };
  52
  53 /* Called with the superblock list lock */
  54 static int obdfs_enqueue_pages(struct inode *inode, struct obdo **obdo,
  55                                int nr_slots, struct page **pages, char **bufs,
  56                                obd_size *counts, obd_off *offsets,
  57                                obd_flag *flag, int check_time)
  58 {
  59         struct list_head *page_list = obdfs_iplist(inode);
  60         struct list_head *tmp;
  61         int num = 0;
  62
  63         ENTRY;
  64         OIDEBUG(inode);
  65
  66         *obdo = obdo_fromid(IID(inode), inode->i_ino, OBD_MD_FLNOTOBD);
  67         if ( IS_ERR(*obdo) ) {
  68                 EXIT;
  69                 return PTR_ERR(*obdo);
  70         }
  71
  72         obdfs_from_inode(*obdo, inode); /* FIXME revisit fromid & from_inode */
  73         *flag = OBD_BRW_CREATE;
  74
  75         tmp = page_list;
  76         while ( ((tmp = tmp->next) != page_list) && (num < nr_slots) ) {
  77                 struct obdfs_pgrq *req;
  78                 struct page *page;
  79
  80                 req = list_entry(tmp, struct obdfs_pgrq, rq_plist);
  81                 page = req->rq_page;
  82
  83
  84                 if (check_time &&
  85                     (jiffies - req->rq_jiffies) < pupd_prm.age_buffer)
  86                         continue;
  87
  88                 /* Remove request from list before write to avoid conflict.
  89                  * Note that obdfs_pgrq_del() also deletes the request.
  90                  */
  91                 obdfs_pgrq_del(req);
  92                 if ( !page ) {
  93                         CDEBUG(D_INODE, "no page \n");
  94                         continue;
  95                 }
  96
  97                 bufs[num] = (char *)page_address(page);
  98                 pages[num] = page;
  99                 counts[num] = PAGE_SIZE;
 100                 offsets[num] = ((obd_off)page->index) << PAGE_SHIFT;
 101                 CDEBUG(D_INODE, "ENQ inode %ld, page %p addr %p to vector\n",
 102                        inode->i_ino, page, (char *)page_address(page));
 103                 num++;
 104         }
 105
 106         if (!list_empty(page_list))
 107                 CDEBUG(D_INODE, "inode %ld list not empty\n", inode->i_ino);
 108         CDEBUG(D_INODE, "added %d page(s) to vector\n", num);
 109
 110         EXIT;
 111         return num;
 112 } /* obdfs_enqueue_pages */
 113
 114 /* dequeue requests for a dying inode */
 115 void obdfs_dequeue_reqs(struct inode *inode)
 116 {
 117
 118         struct list_head *tmp;
 119
 120         obd_down(&obdfs_i2sbi(inode)->osi_list_mutex);
 121         tmp = obdfs_islist(inode);
 122         if ( list_empty(tmp) ) {
 123                 obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 124                 EXIT;
 125                 return;
 126         }
 127
 128         /* take it out of the super list */
 129         list_del(tmp);
 130         INIT_LIST_HEAD(obdfs_islist(inode));
 131
 132         tmp = obdfs_iplist(inode);
 133         while ( (tmp = tmp->next) != obdfs_iplist(inode) ) {
 134                 struct obdfs_pgrq *req;
 135                 struct page *page;
 136
 137                 req = list_entry(tmp, struct obdfs_pgrq, rq_plist);
 138                 page = req->rq_page;
 139                 /* take it out of the list and free */
 140                 obdfs_pgrq_del(req);
 141                 /* now put the page away */
 142                 put_page(page);
 143         }
 144         obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 145 } /* obdfs_dequeue_reqs */
 146
 147 /* Remove writeback requests for the superblock */
 148 int obdfs_flush_reqs(struct list_head *inode_list, int check_time)
 149 {
 150         struct list_head *tmp;
 151         int               total_io = 0;
 152         obd_count         num_io = 0;
 153         obd_count         num_obdos = 0;
 154         struct inode     *inodes[MAX_IOVEC];    /* write data back to these */
 155         struct page      *pages[MAX_IOVEC];     /* call put_page on these */
 156         struct obdo      *obdos[MAX_IOVEC];
 157         char             *bufs[MAX_IOVEC];
 158         obd_size          counts[MAX_IOVEC];
 159         obd_off           offsets[MAX_IOVEC];
 160         obd_flag          flags[MAX_IOVEC];
 161         obd_count         bufs_per_obdo[MAX_IOVEC];
 162         int               err = 0;
 163         struct obdfs_sb_info *sbi;
 164
 165
 166         ENTRY;
 167
 168         if (!inode_list) {
 169                 CDEBUG(D_INODE, "no list\n");
 170                 EXIT;
 171                 return 0;
 172         }
 173
 174         sbi = list_entry(inode_list, struct obdfs_sb_info, osi_inodes);
 175
 176         obd_down(&sbi->osi_list_mutex);
 177         if ( list_empty(inode_list)) {
 178                 CDEBUG(D_INODE, "list empty\n");
 179                 obd_up(&sbi->osi_list_mutex);
 180                 EXIT;
 181                 return 0;
 182         }
 183
 184         /* add each inode's dirty pages to a write vector, and write it */
 185  again:
 186         tmp = inode_list;
 187         while ( (tmp = tmp->next) != inode_list &&
 188                 total_io < pupd_prm.ndirty) {
 189                 struct obdfs_inode_info *ii;
 190                 struct inode *inode;
 191                 int res;
 192
 193                 ii = list_entry(tmp, struct obdfs_inode_info, oi_inodes);
 194                 inode = list_entry(ii, struct inode, u);
 195                 inodes[num_obdos] = inode;
 196                 CDEBUG(D_INODE, "checking inode %ld pages\n", inode->i_ino);
 197
 198                 res = 1;
 199
 200                 /* Loop on this inode until we can't get more pages from it
 201                  * (either no more pages, or the pages aren't old enough).
 202                  * Make sure we reference "inode" and not "inodes[num_obdos]",
 203                  * as num_obdos will change after the loop is run.
 204                  */
 205                 while (!list_empty(obdfs_iplist(inode)) && res &&
 206                        total_io < pupd_prm.ndirty ) {
 207                         res = obdfs_enqueue_pages(inode, &obdos[num_obdos],
 208                                                   MAX_IOVEC - num_io,
 209                                                   &pages[num_io], &bufs[num_io],
 210                                                   &counts[num_io],
 211                                                   &offsets[num_io],
 212                                                   &flags[num_obdos],
 213                                                   check_time);
 214                         CDEBUG(D_INODE, "FLUSHED inode %ld, pages flushed: %d\n",
 215                                inode->i_ino, res);
 216                         if ( res < 0 ) {
 217                                 obd_up(&sbi->osi_list_mutex);
 218                                 err = res;
 219                                 goto ERR;
 220                         }
 221
 222                         num_io += res;
 223                         total_io += res;
 224                         bufs_per_obdo[num_obdos] = res;
 225                         num_obdos++;
 226
 227                         if ( num_io == MAX_IOVEC ) {
 228                                 obd_up(&sbi->osi_list_mutex);
 229                                 err = obdfs_do_vec_wr(inodes, num_io, num_obdos,
 230                                                       obdos, bufs_per_obdo,
 231                                                       pages, bufs, counts,
 232                                                       offsets, flags);
 233                                 if ( err ) {
 234                                         EXIT;
 235                                         goto ERR;
 236                                 }
 237                                 inodes[0] = inode;
 238                                 num_io = 0;
 239                                 num_obdos = 0;
 240                                 obd_down(&sbi->osi_list_mutex);
 241                                 goto again;
 242                         }
 243                 }
 244         }
 245
 246         obd_up(&sbi->osi_list_mutex);
 247
 248         /* flush any remaining I/Os */
 249         if ( num_io ) {
 250                 err = obdfs_do_vec_wr(inodes, num_io, num_obdos, obdos,
 251                                       bufs_per_obdo, pages, bufs, counts,
 252                                       offsets, flags);
 253         }
 254
 255         /* Remove inode from superblock dirty list when no more pages.
 256          * Make sure we don't point at the current inode with tmp
 257          * when we re-init the list on the inode, or we will loop.
 258          */
 259         obd_down(&sbi->osi_list_mutex);
 260         tmp = inode_list;
 261         while ( (tmp = tmp->next) != inode_list ) {
 262                 struct obdfs_inode_info *ii;
 263                 struct inode *inode;
 264
 265                 ii = list_entry(tmp, struct obdfs_inode_info, oi_inodes);
 266                 inode = list_entry(ii, struct inode, u);
 267                 CDEBUG(D_INODE, "checking inode %ld empty\n", inode->i_ino);
 268                 if (list_empty(obdfs_iplist(inode))) {
 269                         CDEBUG(D_INODE, "remove inode %ld from dirty list\n",
 270                                inode->i_ino);
 271                         tmp = tmp->prev;
 272                         list_del(obdfs_islist(inode));
 273                         iput(inode);
 274                         INIT_LIST_HEAD(obdfs_islist(inode));
 275                 }
 276         }
 277         obd_up(&sbi->osi_list_mutex);
 278
 279         CDEBUG(D_INODE, "flushed %d pages in total\n", total_io);
 280         EXIT;
 281 ERR:
 282         return err;
 283 } /* obdfs_remove_pages_from_cache */
 284
 285
 286 void obdfs_flush_dirty_pages(int check_time)
 287 {
 288         struct list_head *sl;
 289
 290         ENTRY;
 291         sl = &obdfs_super_list;
 292         while ( (sl = sl->next) != &obdfs_super_list ) {
 293                 struct obdfs_sb_info *sbi =
 294                         list_entry(sl, struct obdfs_sb_info, osi_list);
 295
 296                 /* walk write requests here, use the sb, check the time */
 297                 obdfs_flush_reqs(&sbi->osi_inodes, check_time);
 298         }
 299         EXIT;
 300 }
 301
 302
 303 static struct task_struct *pupdated;
 304
 305 static int pupdate(void *unused)
 306 {
 307         struct task_struct * tsk = current;
 308         int interval;
 309
 310         pupdated = current;
 311
 312         exit_files(current);
 313         exit_mm(current);
 314
 315         tsk->session = 1;
 316         tsk->pgrp = 1;
 317         sprintf(tsk->comm, "pupdated");
 318         pupdated = current;
 319
 320         MOD_INC_USE_COUNT;      /* XXX until send_sig works */
 321         printk("pupdated activated...\n");
 322
 323         /* sigstop and sigcont will stop and wakeup pupdate */
 324         spin_lock_irq(&tsk->sigmask_lock);
 325         sigfillset(&tsk->blocked);
 326         siginitsetinv(&tsk->blocked, sigmask(SIGTERM));
 327         recalc_sigpending(tsk);
 328         spin_unlock_irq(&tsk->sigmask_lock);
 329
 330         for (;;) {
 331                 /* update interval */
 332                 interval = pupd_prm.interval;
 333                 if (interval)
 334                 {
 335                         tsk->state = TASK_INTERRUPTIBLE;
 336                         schedule_timeout(interval);
 337                 }
 338                 else
 339                 {
 340                 stop_pupdate:
 341                         tsk->state = TASK_STOPPED;
 342                         MOD_DEC_USE_COUNT; /* XXX until send_sig works */
 343                         printk("pupdated stopped...\n");
 344                         return 0;
 345                 }
 346                 /* check for sigstop */
 347                 if (signal_pending(tsk))
 348                 {
 349                         int stopped = 0;
 350                         spin_lock_irq(&tsk->sigmask_lock);
 351                         if (sigismember(&tsk->signal, SIGTERM))
 352                         {
 353                                 sigdelset(&tsk->signal, SIGTERM);
 354                                 stopped = 1;
 355                         }
 356                         recalc_sigpending(tsk);
 357                         spin_unlock_irq(&tsk->sigmask_lock);
 358                         if (stopped)
 359                                 goto stop_pupdate;
 360                 }
 361                 /* asynchronous setattr etc for the future ...
 362                 flush_inodes();
 363                  */
 364                 /* we don't currently check the time on the pages
 365                 obdfs_flush_dirty_pages(1);
 366                  */
 367                 obdfs_flush_dirty_pages(0);
 368         }
 369 }
 370
 371
 372 int obdfs_flushd_init(void)
 373 {
 374         /*
 375         kernel_thread(bdflush, NULL, CLONE_FS | CLONE_FILES | CLONE_SIGHAND);
 376          */
 377         kernel_thread(pupdate, NULL, 0);
 378         CDEBUG(D_PSDEV, __FUNCTION__ ": flushd inited\n");
 379         return 0;
 380 }
 381
 382 int obdfs_flushd_cleanup(void)
 383 {
 384         ENTRY;
 385         /* deliver a signal to pupdated to shut it down
 386            XXX need to kill it from user space for now XXX
 387         if (pupdated) {
 388                 send_sig_info(SIGTERM, 1, pupdated);
 389         }
 390          */
 391
 392         EXIT;
 393         /* not reached */
 394         return 0;
 395
 396 }