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
  39 struct {
  40         int nfract;  /* Percentage of buffer cache dirty to
  41                         activate bdflush */
  42         int ndirty;  /* Maximum number of dirty blocks to write out per
  43                         wake-cycle */
  44         int nrefill; /* Number of clean buffers to try to obtain
  45                                 each time we call refill */
  46         int nref_dirt; /* Dirty buffer threshold for activating bdflush
  47                           when trying to refill buffers. */
  48         int interval; /* jiffies delay between kupdate flushes */
  49         int age_buffer;  /* Time for normal buffer to age before we flush it */
  50         int age_super;  /* Time for superblock to age before we flush it */
  51 /* } pupd_prm = {40, 500, 64, 256, 5*HZ, 30*HZ, 5*HZ };  */
  52 } pupd_prm = {40, 500, 64, 256, 10*HZ, 30*HZ, 5*HZ };
  53
  54
  55 /* Called with the superblock list lock */
  56 static int obdfs_enqueue_pages(struct inode *inode, struct obdo **obdo,
  57                                int nr_slots, struct page **pages, char **bufs,
  58                                obd_size *counts, obd_off *offsets,
  59                                obd_flag *flag, int check_time)
  60 {
  61         struct list_head *page_list = obdfs_iplist(inode);
  62         struct list_head *tmp;
  63         int num = 0;
  64
  65         ENTRY;
  66         OIDEBUG(inode);
  67
  68         *obdo = obdo_fromid(IID(inode), inode->i_ino, OBD_MD_FLNOTOBD);
  69         if ( IS_ERR(*obdo) ) {
  70                 EXIT;
  71                 return PTR_ERR(*obdo);
  72         }
  73
  74         obdfs_from_inode(*obdo, inode); /* FIXME revisit fromid & from_inode */
  75         *flag = OBD_BRW_CREATE;
  76
  77         tmp = page_list;
  78         while ( ((tmp = tmp->next) != page_list) && (num < nr_slots) ) {
  79                 struct obdfs_pgrq *req;
  80                 struct page *page;
  81
  82                 req = list_entry(tmp, struct obdfs_pgrq, rq_plist);
  83                 page = req->rq_page;
  84
  85
  86                 if (check_time &&
  87                     (jiffies - req->rq_jiffies) < pupd_prm.age_buffer)
  88                         continue;
  89
  90                 /* Remove request from list before write to avoid conflict.
  91                  * Note that obdfs_pgrq_del() also deletes the request.
  92                  */
  93                 obdfs_pgrq_del(req);
  94                 if ( !page ) {
  95                         CDEBUG(D_INODE, "no page \n");
  96                         continue;
  97                 }
  98
  99                 bufs[num] = (char *)page_address(page);
 100                 pages[num] = page;
 101                 counts[num] = PAGE_SIZE;
 102                 offsets[num] = ((obd_off)page->index) << PAGE_SHIFT;
 103                 CDEBUG(D_INODE, "ENQ inode %ld, page %p addr %p to vector\n",
 104                        inode->i_ino, page, (char *)page_address(page));
 105                 num++;
 106         }
 107
 108         if (!list_empty(page_list))
 109                 CDEBUG(D_INODE, "inode %ld list not empty\n", inode->i_ino);
 110         CDEBUG(D_INODE, "added %d page(s) to vector\n", num);
 111
 112         EXIT;
 113         return num;
 114 }
 115
 116 /* dequeue requests for a dying inode */
 117 void obdfs_dequeue_reqs(struct inode *inode)
 118 {
 119
 120         struct list_head *tmp;
 121
 122         obd_down(&obdfs_i2sbi(inode)->osi_list_mutex);
 123         tmp = obdfs_islist(inode);
 124         if ( list_empty(tmp) ) {
 125                 obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 126                 EXIT;
 127                 return;
 128         }
 129
 130         /* take it out of the super list */
 131         list_del(tmp);
 132         INIT_LIST_HEAD(obdfs_islist(inode));
 133
 134         tmp = obdfs_iplist(inode);
 135         while ( (tmp = tmp->next) != obdfs_iplist(inode) ) {
 136                 struct obdfs_pgrq *req;
 137                 struct page *page;
 138
 139                 req = list_entry(tmp, struct obdfs_pgrq, rq_plist);
 140                 page = req->rq_page;
 141                 /* take it out of the list and free */
 142                 obdfs_pgrq_del(req);
 143                 /* now put the page away */
 144                 put_page(page);
 145         }
 146         iput(inode);
 147         obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 148 }
 149
 150 /* Remove writeback requests for the superblock */
 151 int obdfs_flush_reqs(struct list_head *inode_list, int check_time)
 152 {
 153         struct list_head *tmp;
 154         int               total_io = 0;
 155         obd_count         num_io = 0;
 156         obd_count         num_obdos = 0;
 157         struct inode     *inodes[MAX_IOVEC];    /* write data back to these */
 158         struct page      *pages[MAX_IOVEC];     /* call put_page on these */
 159         struct obdo      *obdos[MAX_IOVEC];
 160         char             *bufs[MAX_IOVEC];
 161         obd_size          counts[MAX_IOVEC];
 162         obd_off           offsets[MAX_IOVEC];
 163         obd_flag          flags[MAX_IOVEC];
 164         obd_count         bufs_per_obdo[MAX_IOVEC];
 165         int               err = 0;
 166         struct obdfs_sb_info *sbi;
 167
 168
 169         ENTRY;
 170
 171         if (!inode_list) {
 172                 CDEBUG(D_INODE, "no list\n");
 173                 EXIT;
 174                 return 0;
 175         }
 176
 177         sbi = list_entry(inode_list, struct obdfs_sb_info, osi_inodes);
 178
 179         obd_down(&sbi->osi_list_mutex);
 180         if ( list_empty(inode_list)) {
 181                 CDEBUG(D_INODE, "list empty\n");
 182                 obd_up(&sbi->osi_list_mutex);
 183                 EXIT;
 184                 return 0;
 185         }
 186
 187         /* add each inode's dirty pages to a write vector, and write it */
 188  again:
 189         tmp = inode_list;
 190         while ( (tmp = tmp->next) != inode_list &&
 191                 total_io < pupd_prm.ndirty) {
 192                 struct obdfs_inode_info *ii;
 193                 struct inode *inode;
 194                 int res;
 195
 196                 ii = list_entry(tmp, struct obdfs_inode_info, oi_inodes);
 197                 inode = list_entry(ii, struct inode, u);
 198                 inodes[num_obdos] = inode;
 199                 CDEBUG(D_INODE, "checking inode %ld pages\n", inode->i_ino);
 200
 201                 res = 1;
 202
 203                 /* Loop on this inode until we can't get more pages from it
 204                  * (either no more pages, or the pages aren't old enough).
 205                  * Make sure we reference "inode" and not "inodes[num_obdos]",
 206                  * as num_obdos will change after the loop is run.
 207                  */
 208                 while (!list_empty(obdfs_iplist(inode)) && res &&
 209                        total_io < pupd_prm.ndirty ) {
 210                         res = obdfs_enqueue_pages(inode, &obdos[num_obdos],
 211                                                   MAX_IOVEC - num_io,
 212                                                   &pages[num_io], &bufs[num_io],
 213                                                   &counts[num_io],
 214                                                   &offsets[num_io],
 215                                                   &flags[num_obdos], 1);
 216                         CDEBUG(D_INODE, "FLUSHED inode %ld, pages flushed: %d\n",
 217                                inode->i_ino, res);
 218                         if ( res < 0 ) {
 219                                 obd_up(&sbi->osi_list_mutex);
 220                                 err = res;
 221                                 goto ERR;
 222                         }
 223
 224                         num_io += res;
 225                         total_io += res;
 226                         bufs_per_obdo[num_obdos] = res;
 227                         num_obdos++;
 228
 229                         if ( num_io == MAX_IOVEC ) {
 230                                 obd_up(&sbi->osi_list_mutex);
 231                                 err = obdfs_do_vec_wr(inodes, num_io, num_obdos,
 232                                                       obdos, bufs_per_obdo,
 233                                                       pages, bufs, counts,
 234                                                       offsets, flags);
 235                                 if ( err ) {
 236                                         EXIT;
 237                                         goto ERR;
 238                                 }
 239                                 inodes[0] = inode;
 240                                 num_io = 0;
 241                                 num_obdos = 0;
 242                                 obd_down(&sbi->osi_list_mutex);
 243                                 goto again;
 244                         }
 245                 }
 246         }
 247
 248         obd_up(&sbi->osi_list_mutex);
 249
 250         /* flush any remaining I/Os */
 251         if ( num_io ) {
 252                 err = obdfs_do_vec_wr(inodes, num_io, num_obdos, obdos,
 253                                       bufs_per_obdo, pages, bufs, counts,
 254                                       offsets, flags);
 255         }
 256
 257         /* Remove inode from superblock dirty list when no more pages.
 258          * Make sure we don't point at the current inode with tmp
 259          * when we re-init the list on the inode, or we will loop.
 260          */
 261         obd_down(&sbi->osi_list_mutex);
 262         tmp = inode_list;
 263         while ( (tmp = tmp->next) != inode_list ) {
 264                 struct obdfs_inode_info *ii;
 265                 struct inode *inode;
 266
 267                 ii = list_entry(tmp, struct obdfs_inode_info, oi_inodes);
 268                 inode = list_entry(ii, struct inode, u);
 269                 CDEBUG(D_INODE, "checking inode %ld empty\n", inode->i_ino);
 270                 if (list_empty(obdfs_iplist(inode))) {
 271                         CDEBUG(D_INODE, "remove inode %ld from dirty list\n",
 272                                inode->i_ino);
 273                         tmp = tmp->prev;
 274                         list_del(obdfs_islist(inode));
 275                         iput(inode);
 276                         INIT_LIST_HEAD(obdfs_islist(inode));
 277                 }
 278         }
 279         obd_up(&sbi->osi_list_mutex);
 280
 281         CDEBUG(D_INODE, "flushed %d pages in total\n", total_io);
 282         EXIT;
 283 ERR:
 284         return err;
 285 } /* obdfs_remove_pages_from_cache */
 286
 287
 288 void obdfs_flush_dirty_pages(int check_time)
 289 {
 290         struct list_head *sl;
 291
 292         sl = &obdfs_super_list;
 293         while ( (sl = sl->next) != &obdfs_super_list ) {
 294                 struct obdfs_sb_info *sbi =
 295                         list_entry(sl, struct obdfs_sb_info, osi_list);
 296
 297                 /* walk write requests here, use the sb, check the time */
 298                 obdfs_flush_reqs(&sbi->osi_inodes, 0);
 299         }
 300
 301 }
 302
 303
 304 static struct task_struct *pupdated;
 305
 306
 307 static int pupdate(void *unused)
 308 {
 309         struct task_struct * tsk = current;
 310         int interval;
 311
 312         pupdated = current;
 313
 314         exit_files(current);
 315         exit_mm(current);
 316
 317         tsk->session = 1;
 318         tsk->pgrp = 1;
 319         sprintf(tsk->comm, "pupdated");
 320         pupdated = current;
 321
 322         printk("pupdated activated...\n");
 323
 324         /* sigstop and sigcont will stop and wakeup pupdate */
 325         spin_lock_irq(&tsk->sigmask_lock);
 326         sigfillset(&tsk->blocked);
 327         siginitsetinv(&tsk->blocked, sigmask(SIGTERM));
 328         recalc_sigpending(tsk);
 329         spin_unlock_irq(&tsk->sigmask_lock);
 330
 331         for (;;) {
 332                 /* update interval */
 333                 interval = pupd_prm.interval;
 334                 if (interval)
 335                 {
 336                         tsk->state = TASK_INTERRUPTIBLE;
 337                         schedule_timeout(interval);
 338                 }
 339                 else
 340                 {
 341                 stop_pupdate:
 342                         obdfs_flush_dirty_pages(0);
 343                         tsk->state = TASK_STOPPED;
 344                         /* MOD_DEC_USE_COUNT; */
 345                         printk("pupdated stopped...\n");
 346                         return 0;
 347                 }
 348                 /* check for sigstop */
 349                 if (signal_pending(tsk))
 350                 {
 351                         int stopped = 0;
 352                         spin_lock_irq(&tsk->sigmask_lock);
 353                         if (sigismember(&tsk->signal, SIGTERM))
 354                         {
 355                                 sigdelset(&tsk->signal, SIGTERM);
 356                                 stopped = 1;
 357                         }
 358                         recalc_sigpending(tsk);
 359                         spin_unlock_irq(&tsk->sigmask_lock);
 360                         if (stopped)
 361                                 goto stop_pupdate;
 362                 }
 363                 /* asynchronous setattr etc for the future ... */
 364                 /* flush_inodes(); */
 365                 obdfs_flush_dirty_pages(1);
 366         }
 367 }
 368
 369
 370 int flushd_init(void)
 371 {
 372         /*
 373         kernel_thread(bdflush, NULL, CLONE_FS | CLONE_FILES | CLONE_SIGHAND);
 374          */
 375         /* MOD_INC_USE_COUNT; */
 376         kernel_thread(pupdate, NULL, 0);
 377         printk("flushd inited\n");
 378         return 0;
 379 }
 380
 381 int flushd_cleanup(void)
 382 {
 383         /* this should deliver a signal to */
 384
 385
 386         /* XXX Andreas, we will do this later, for now, you must kill
 387            pupdated with a SIGTERM from userland, before unloading obdfs.o
 388         */
 389         if (pupdated) {
 390                 /* then let it run at least once, before continuing */
 391
 392                 /* XXX need to do something like this here:
 393                 send_sig(SIGTERM, current, 0);
 394                  */
 395                 1;
 396                 /*obdfs_flush_dirty_pages(0); */
 397         }
 398
 399         /* not reached */
 400         return 0;
 401
 402 }