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 pupdate 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, 2*HZ, 30*HZ, 5*HZ };
  51
  52 /* Called with the superblock list lock */
  53 static int obdfs_enqueue_pages(struct inode *inode, struct obdo **obdo,
  54                                int nr_slots, struct page **pages, char **bufs,
  55                                obd_size *counts, obd_off *offsets,
  56                                obd_flag *flag, int check_time)
  57 {
  58         struct list_head *page_list = obdfs_iplist(inode);
  59         struct list_head *tmp;
  60         int num = 0;
  61
  62         ENTRY;
  63         OIDEBUG(inode);
  64
  65         *obdo = obdo_fromid(IID(inode), inode->i_ino, OBD_MD_FLNOTOBD);
  66         if ( IS_ERR(*obdo) ) {
  67                 EXIT;
  68                 return PTR_ERR(*obdo);
  69         }
  70
  71         obdfs_from_inode(*obdo, inode); /* FIXME revisit fromid & from_inode */
  72         *flag = OBD_BRW_CREATE;
  73
  74         tmp = page_list;
  75         while ( ((tmp = tmp->next) != page_list) && (num < nr_slots) ) {
  76                 struct obdfs_pgrq *req;
  77                 struct page *page;
  78
  79                 req = list_entry(tmp, struct obdfs_pgrq, rq_plist);
  80                 page = req->rq_page;
  81
  82
  83                 if (check_time &&
  84                     (jiffies - req->rq_jiffies) < pupd_prm.age_buffer)
  85                         continue;
  86
  87                 /* Remove request from list before write to avoid conflict.
  88                  * Note that obdfs_pgrq_del() also deletes the request.
  89                  */
  90                 obdfs_pgrq_del(req);
  91                 if ( !page ) {
  92                         CDEBUG(D_INODE, "no page \n");
  93                         continue;
  94                 }
  95
  96                 bufs[num] = (char *)page_address(page);
  97                 pages[num] = page;
  98                 counts[num] = PAGE_SIZE;
  99                 offsets[num] = ((obd_off)page->index) << PAGE_SHIFT;
 100                 CDEBUG(D_INODE, "ENQ inode %ld, page %p addr %p to vector\n",
 101                        inode->i_ino, page, (char *)page_address(page));
 102                 num++;
 103         }
 104
 105         if (!list_empty(page_list))
 106                 CDEBUG(D_INODE, "inode %ld list not empty\n", inode->i_ino);
 107         CDEBUG(D_INODE, "added %d page(s) to vector\n", num);
 108
 109         EXIT;
 110         return num;
 111 } /* obdfs_enqueue_pages */
 112
 113 /* Remove writeback requests for the superblock */
 114 int obdfs_flush_reqs(struct list_head *inode_list, int check_time)
 115 {
 116         struct list_head *tmp;
 117         int               total_io = 0;
 118         obd_count         num_io = 0;
 119         obd_count         num_obdos = 0;
 120         struct inode     *inodes[MAX_IOVEC];    /* write data back to these */
 121         struct page      *pages[MAX_IOVEC];     /* call put_page on these */
 122         struct obdo      *obdos[MAX_IOVEC];
 123         char             *bufs[MAX_IOVEC];
 124         obd_size          counts[MAX_IOVEC];
 125         obd_off           offsets[MAX_IOVEC];
 126         obd_flag          flags[MAX_IOVEC];
 127         obd_count         bufs_per_obdo[MAX_IOVEC];
 128         int               err = 0;
 129         struct obdfs_sb_info *sbi;
 130
 131         ENTRY;
 132         if (!inode_list) {
 133                 CDEBUG(D_INODE, "no list\n");
 134                 EXIT;
 135                 return 0;
 136         }
 137
 138         sbi = list_entry(inode_list, struct obdfs_sb_info, osi_inodes);
 139
 140         obd_down(&sbi->osi_list_mutex);
 141         if ( list_empty(inode_list)) {
 142                 CDEBUG(D_INODE, "list empty\n");
 143                 obd_up(&sbi->osi_list_mutex);
 144                 EXIT;
 145                 return 0;
 146         }
 147
 148         /* add each inode's dirty pages to a write vector, and write it */
 149  again:
 150         tmp = inode_list;
 151         while ( (tmp = tmp->next) != inode_list &&
 152                 total_io < pupd_prm.ndirty) {
 153                 struct obdfs_inode_info *ii;
 154                 struct inode *inode;
 155                 int res;
 156
 157                 ii = list_entry(tmp, struct obdfs_inode_info, oi_inodes);
 158                 inode = list_entry(ii, struct inode, u);
 159                 inodes[num_obdos] = inode;
 160                 CDEBUG(D_INODE, "checking inode %ld pages\n", inode->i_ino);
 161
 162                 res = 1;
 163
 164                 /* Loop on this inode until we can't get more pages from it
 165                  * (either no more pages, or the pages aren't old enough).
 166                  * Make sure we reference "inode" and not "inodes[num_obdos]",
 167                  * as num_obdos will change after the loop is run.
 168                  */
 169                 while (!list_empty(obdfs_iplist(inode)) && res &&
 170                        total_io < pupd_prm.ndirty ) {
 171                         res = obdfs_enqueue_pages(inode, &obdos[num_obdos],
 172                                                   MAX_IOVEC - num_io,
 173                                                   &pages[num_io], &bufs[num_io],
 174                                                   &counts[num_io],
 175                                                   &offsets[num_io],
 176                                                   &flags[num_obdos],
 177                                                   check_time);
 178                         CDEBUG(D_INODE, "FLUSHED inode %ld, pages flushed: %d\n",
 179                                inode->i_ino, res);
 180                         if ( res < 0 ) {
 181                                 obd_up(&sbi->osi_list_mutex);
 182                                 err = res;
 183                                 goto ERR;
 184                         }
 185
 186                         num_io += res;
 187                         total_io += res;
 188                         bufs_per_obdo[num_obdos] = res;
 189                         num_obdos++;
 190
 191                         if ( num_io == MAX_IOVEC ) {
 192                                 obd_up(&sbi->osi_list_mutex);
 193                                 err = obdfs_do_vec_wr(inodes, num_io, num_obdos,
 194                                                       obdos, bufs_per_obdo,
 195                                                       pages, bufs, counts,
 196                                                       offsets, flags);
 197                                 if ( err ) {
 198                                         EXIT;
 199                                         goto ERR;
 200                                 }
 201                                 inodes[0] = inode;
 202                                 num_io = 0;
 203                                 num_obdos = 0;
 204                                 obd_down(&sbi->osi_list_mutex);
 205                                 goto again;
 206                         }
 207                 }
 208         }
 209
 210         obd_up(&sbi->osi_list_mutex);
 211
 212         /* flush any remaining I/Os */
 213         if ( num_io ) {
 214                 err = obdfs_do_vec_wr(inodes, num_io, num_obdos, obdos,
 215                                       bufs_per_obdo, pages, bufs, counts,
 216                                       offsets, flags);
 217         }
 218
 219         /* Remove inode from superblock dirty list when no more pages.
 220          * Make sure we don't point at the current inode with tmp
 221          * when we re-init the list on the inode, or we will loop.
 222          */
 223         obd_down(&sbi->osi_list_mutex);
 224         tmp = inode_list;
 225         while ( (tmp = tmp->next) != inode_list ) {
 226                 struct obdfs_inode_info *ii;
 227                 struct inode *inode;
 228
 229                 ii = list_entry(tmp, struct obdfs_inode_info, oi_inodes);
 230                 inode = list_entry(ii, struct inode, u);
 231                 CDEBUG(D_INODE, "checking inode %ld empty\n", inode->i_ino);
 232                 if (list_empty(obdfs_iplist(inode))) {
 233                         CDEBUG(D_INODE, "remove inode %ld from dirty list\n",
 234                                inode->i_ino);
 235                         tmp = tmp->prev;
 236                         list_del(obdfs_islist(inode));
 237                         /* decrement inode reference for page cache */
 238                         inode->i_count--;
 239                         INIT_LIST_HEAD(obdfs_islist(inode));
 240                 }
 241         }
 242         obd_up(&sbi->osi_list_mutex);
 243
 244         CDEBUG(D_INODE, "flushed %d pages in total\n", total_io);
 245         EXIT;
 246 ERR:
 247         return err;
 248 } /* obdfs_flush_reqs */
 249
 250
 251 void obdfs_flush_dirty_pages(int check_time)
 252 {
 253         struct list_head *sl;
 254
 255         ENTRY;
 256         sl = &obdfs_super_list;
 257         while ( (sl = sl->next) != &obdfs_super_list ) {
 258                 struct obdfs_sb_info *sbi =
 259                         list_entry(sl, struct obdfs_sb_info, osi_list);
 260
 261                 /* walk write requests here, use the sb, check the time */
 262                 obdfs_flush_reqs(&sbi->osi_inodes, check_time);
 263         }
 264         EXIT;
 265 } /* obdfs_flush_dirty_pages */
 266
 267
 268 static struct task_struct *pupdated;
 269
 270 static int pupdate(void *unused)
 271 {
 272         struct task_struct * tsk = current;
 273         int interval;
 274
 275         pupdated = current;
 276
 277         exit_files(current);
 278         exit_mm(current);
 279
 280         tsk->session = 1;
 281         tsk->pgrp = 1;
 282         sprintf(tsk->comm, "pupdated");
 283         pupdated = current;
 284
 285         MOD_INC_USE_COUNT;      /* XXX until send_sig works */
 286         printk("pupdated activated...\n");
 287
 288         /* sigstop and sigcont will stop and wakeup pupdate */
 289         spin_lock_irq(&tsk->sigmask_lock);
 290         sigfillset(&tsk->blocked);
 291         siginitsetinv(&tsk->blocked, sigmask(SIGTERM));
 292         recalc_sigpending(tsk);
 293         spin_unlock_irq(&tsk->sigmask_lock);
 294
 295         for (;;) {
 296                 /* update interval */
 297                 interval = pupd_prm.interval;
 298                 if (interval)
 299                 {
 300                         tsk->state = TASK_INTERRUPTIBLE;
 301                         schedule_timeout(interval);
 302                 }
 303                 else
 304                 {
 305                 stop_pupdate:
 306                         tsk->state = TASK_STOPPED;
 307                         MOD_DEC_USE_COUNT; /* XXX until send_sig works */
 308                         printk("pupdated stopped...\n");
 309                         return 0;
 310                 }
 311                 /* check for sigstop */
 312                 if (signal_pending(tsk))
 313                 {
 314                         int stopped = 0;
 315                         spin_lock_irq(&tsk->sigmask_lock);
 316                         if (sigismember(&tsk->signal, SIGTERM))
 317                         {
 318                                 sigdelset(&tsk->signal, SIGTERM);
 319                                 stopped = 1;
 320                         }
 321                         recalc_sigpending(tsk);
 322                         spin_unlock_irq(&tsk->sigmask_lock);
 323                         if (stopped)
 324                                 goto stop_pupdate;
 325                 }
 326                 /* asynchronous setattr etc for the future ...
 327                 flush_inodes();
 328                  */
 329                 obdfs_flush_dirty_pages(1);
 330         }
 331 }
 332
 333
 334 int obdfs_flushd_init(void)
 335 {
 336         /*
 337         kernel_thread(bdflush, NULL, CLONE_FS | CLONE_FILES | CLONE_SIGHAND);
 338          */
 339         kernel_thread(pupdate, NULL, 0);
 340         CDEBUG(D_PSDEV, __FUNCTION__ ": flushd inited\n");
 341         return 0;
 342 }
 343
 344 int obdfs_flushd_cleanup(void)
 345 {
 346         ENTRY;
 347         /* deliver a signal to pupdated to shut it down
 348            XXX need to kill it from user space for now XXX
 349         if (pupdated) {
 350                 send_sig_info(SIGTERM, 1, pupdated);
 351         }
 352          */
 353
 354         EXIT;
 355         /* not reached */
 356         return 0;
 357
 358 }