Subversion Repositories or1k

/************************************************************************

 * raid1.c : Multiple Devices driver for Linux

 *           Copyright (C) 1996 Ingo Molnar, Miguel de Icaza, Gadi Oxman

 *

 * RAID-1 management functions.

 *

 * This program is free software; you can redistribute it and/or modify

 * it under the terms of the GNU General Public License as published by

 * the Free Software Foundation; either version 2, or (at your option)

 * any later version.

 *

 * You should have received a copy of the GNU General Public License

 * (for example /usr/src/linux/COPYING); if not, write to the Free

 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.

 */

#include <linux/module.h>

#include <linux/locks.h>

#include <linux/malloc.h>

#include <linux/md.h>

#include <linux/raid1.h>

#include <asm/bitops.h>

#include <asm/atomic.h>

#define MAJOR_NR MD_MAJOR

#define MD_DRIVER

#define MD_PERSONALITY

/*

 * The following can be used to debug the driver

 */

/*#define RAID1_DEBUG*/

#ifdef RAID1_DEBUG

#define PRINTK(x)   do { printk x; } while (0);

#else

#define PRINTK(x)   do { ; } while (0);

#endif

static struct md_personality raid1_personality;

static struct md_thread *raid1_thread = NULL;

struct buffer_head *raid1_retry_list = NULL;

static int __raid1_map (struct md_dev *mddev, kdev_t *rdev,

                        unsigned long *rsector, unsigned long size)

{

        struct raid1_data *raid_conf = (struct raid1_data *) mddev->private;

        int i, n = raid_conf->raid_disks;

        /*

         * Later we do read balancing on the read side

         * now we use the first available disk.

         */

        PRINTK(("raid1_map().\n"));

        for (i=0; i<n; i++) {

                if (raid_conf->mirrors[i].operational) {

                        *rdev = raid_conf->mirrors[i].dev;

                        return (0);

                }

        }

        printk (KERN_ERR "raid1_map(): huh, no more operational devices?\n");

        return (-1);

}

static int raid1_map (struct md_dev *mddev, kdev_t *rdev,

                      unsigned long *rsector, unsigned long size)

{

        return 0;

}

void raid1_reschedule_retry (struct buffer_head *bh)

{

        struct raid1_bh * r1_bh = (struct raid1_bh *)(bh->private_bh);

        PRINTK(("raid1_reschedule_retry().\n"));

        r1_bh->next_retry = raid1_retry_list;

        raid1_retry_list = bh;

        md_wakeup_thread(raid1_thread);

}

/*

 * raid1_end_buffer_io() is called when we have finished servicing a mirrored

 * operation and are ready to return a success/failture code to the buffer

 * cache layer.

 */

static inline void raid1_end_buffer_io (struct buffer_head *bh, int uptodate)

{

        /*

         * kfree() can sleep and we try to keep this bh operation atomic.

         */

        struct raid1_bh * tmp = (struct raid1_bh *) bh->private_bh;

        clear_bit (BH_MD, &bh->b_state);

        bh->private_bh = NULL;

        bh->personality = NULL;

        mark_buffer_uptodate(bh, uptodate);

        unlock_buffer(bh);

        kfree(tmp);

}

void raid1_end_request (struct buffer_head *bh, int uptodate)

{

        struct raid1_bh * r1_bh = (struct raid1_bh *)(bh->private_bh);

        unsigned long flags;

        save_flags(flags);

        cli();

        PRINTK(("raid1_end_request().\n"));

        /*

         * this branch is our 'one mirror IO has finished' event handler:

         */

        if (!uptodate)

                md_error (bh->b_dev, bh->b_rdev);

        else {

                /*

                 * Set BH_Uptodate in our master buffer_head, so that

                 * we will return a good error code for to the higher

                 * levels even if IO on some other mirrored buffer fails.

                 *

                 * The 'master' represents the complex operation to

                 * user-side. So if something waits for IO, then it will

                 * wait for the 'master' buffer_head.

                 */

                set_bit (BH_Uptodate, &r1_bh->state);

        }

        /*

         * We split up the read and write side, imho they are

         * conceptually different.

         */

        if ( (r1_bh->cmd == READ) || (r1_bh->cmd == READA) ) {

                PRINTK(("raid1_end_request(), read branch.\n"));

                /*

                 * we have only one buffer_head on the read side

                 */

                if (uptodate) {

                        PRINTK(("raid1_end_request(), read branch, uptodate.\n"));

                        raid1_end_buffer_io (bh, uptodate);

                        restore_flags(flags);

                        return;

                }

                /*

                 * oops, read error:

                 */

                printk(KERN_ERR "raid1: %s: rescheduling block %lu\n",

                                 kdevname(bh->b_dev), bh->b_blocknr);

                raid1_reschedule_retry (bh);

                restore_flags(flags);

                return;

        }

        /*

         * WRITE or WRITEA.

         */

        PRINTK(("raid1_end_request(), write branch.\n"));

        /*

         * lets see if all mirrored write operations have finished

         * already [we have irqs off, so we can decrease]:

         */

        if (!--r1_bh->remaining) {

                struct md_dev *mddev = r1_bh->mddev;

                struct raid1_data *raid_conf = (struct raid1_data *) mddev->private;

                int i, n = raid_conf->raid_disks;

                PRINTK(("raid1_end_request(), remaining == 0.\n"));

                /*

                 * kfree() can sleep? really? if yes then we are

                 * doomed here ...

                 */

                for ( i=0; i<n; i++) {

                        if (r1_bh->mirror_bh[i]) kfree(r1_bh->mirror_bh[i]);

                }

                /*

                 * the 'master' bh is the one that is used in page IO,

                 * perhaps someone is waiting on it. Lets erase all

                 * signs of mirroring, and lets finish the bh operation:

                 *

                 * In particular, the "uptodate" value which we return

                 * to the higher level represents the entire mirror set.

                 *

                 * yes, and this is why i want to use the 'master' bh as

                 * a 'representative'. Thats why i think it's not clean to

                 * use the master bh for real IO. We mix concepts, which

                 * isnt too good.

                 *

                 * a buffer_head is basically a user-side file buffer.

                 * Normally it has direct relationship with the physical

                 * device, but as in this case, we have an abstract mapping

                 * between the file buffer and the physical layout. So i've

                 * reverted all changes that do this mixing.

                 *

                 * we 'waste' about 76 bytes for the one more buffer_head,

                 * but note that we will do the mirror bh allocation at once

                 * in the future, so this isnt really a valid point, i think.

                 *

                 * Also i dont like the current way of mixing the user-side buffer

                 * concept with the 'real' physical layout like raid0.c does

                 * now: it increases the size of buffer_head even for nonstriped

                 * devices, etc.

                 *

                 * IMHO, in the future, we should have a lightweight buffer_head

                 * structure, which holds almost no physical device information.

                 * Abstract relationship between buffers:

                 * =====================================

                 *

                 *           [user]

                 *              |

                 *              |

                 *    ['master' buffer_head] + [private_buffer_head]

                 *                                      |

                 *                                      |

                 *                                      |

                 *                        [additional 'sub'-buffer_heads]

                 *                           |          |           |

                 *                         [dev1]     [dev2]      [dev3]

                 *

                 * In this scheme it's not clean to use the 'master' as one of

                 * the 'sub' buffer_heads. If you think about it, currently we can

                 * do this only because raid0 introduced it's own private_buffer_head

                 * structure in buffer_head: rdev,rsector. And raid0 has a 1:1

                 * relationship to the physical device. But this is really just a

                 * special case. Once we have our megafast bh pools running, we could

                 * clean up raid0.c too :))

                 *

                 * Not that it isnt clean, it is lethal if in the future we insert our

                 * sub buffer_heads into the global block cache. The master request

                 * should be an IO operation label for the complex operation, nothing

                 * more.

                 *

                 * So we have almost no performance arguments, and alot of cleanness

                 * arguments.

                 *

                 * Comments? Gonna change it back to your way again if you can convince

                 * me :)) --mingo

                 *

                 */

                raid1_end_buffer_io ( r1_bh->master_bh,

                                test_bit (BH_Uptodate, &r1_bh->state));

        }

        else PRINTK(("raid1_end_request(), remaining == %u.\n", r1_bh->remaining));

        restore_flags(flags);

}

/* This routine checks if the undelying device is an md device and in that

 * case it maps the blocks before putting the request on the queue

 */

static inline void

map_and_make_request (int rw, struct buffer_head *bh)

{

        if (MAJOR (bh->b_rdev) == MD_MAJOR){

                md_map (MINOR (bh->b_rdev), &bh->b_rdev, &bh->b_rsector, bh->b_size >> 9);

        }

        make_request (MAJOR (bh->b_rdev), rw, bh);

}

static int

raid1_make_request (struct md_dev *mddev, int rw, struct buffer_head * bh)

{

        struct raid1_data *raid_conf = (struct raid1_data *) mddev->private;

        struct buffer_head *mirror_bh[MD_SB_DISKS];

        struct raid1_bh * r1_bh;

        int n = raid_conf->raid_disks, i, sum_bhs = 0, switch_disks = 0, sectors;

        struct mirror_info *mirror;

        PRINTK(("raid1_make_request().\n"));

/*

 * We put allocations at the beginning, to avoid sleeping while doing

 * atomic operations of buffer heads. This might or might not make much

 * difference, but lets rather be careful.

 *

 * but this has two side effects (probably non harmless):

 *

 *      1.      The buffer will not be locked while we sleep.

 *      2.      The rest of the kernel will see BH_Req without

 *              BH_Lock.

 */

        while (!( /* FIXME: now we are rather fault tolerant than nice */

        r1_bh = kmalloc (sizeof (struct raid1_bh), GFP_KERNEL)

        ) )

                printk ("raid1_make_request(#1): out of memory\n");

        memset (r1_bh, 0, sizeof (struct raid1_bh));

/*

 * make_request() can abort the operation when READA or WRITEA are being

 * used and no empty request is available.

 *

 * Currently, just replace the command with READ/WRITE.

 */

        if (rw == READA) rw = READ;

        if (rw == WRITEA) rw = WRITE;

        if (rw == WRITE || rw == WRITEA)

                mark_buffer_clean(bh);          /* Too early ? */

/*

 * i think the read and write branch should be separated completely, since we want

 * to do read balancing on the read side for example. Comments? :) --mingo

 */

        r1_bh->master_bh=bh;

        r1_bh->mddev=mddev;

        r1_bh->cmd = rw;

        set_bit (BH_MD, &bh->b_state);

        bh->personality  = &raid1_personality;

        bh->private_bh   = (void*)(r1_bh);

        if (rw==READ || rw==READA) {

                int last_used = raid_conf->last_used;

                PRINTK(("raid1_make_request(), read branch.\n"));

                mirror = raid_conf->mirrors + last_used;

                bh->b_rdev = mirror->dev;

                sectors = bh->b_size >> 9;

                if (bh->b_blocknr * sectors == raid_conf->next_sect) {

                        raid_conf->sect_count += sectors;

                        if (raid_conf->sect_count >= mirror->sect_limit)

                                switch_disks = 1;

                } else

                        switch_disks = 1;

                raid_conf->next_sect = (bh->b_blocknr + 1) * sectors;

                if (switch_disks) {

                        PRINTK(("read-balancing: switching %d -> %d (%d sectors)\n", last_used, mirror->next, raid_conf->sect_count));

                        raid_conf->sect_count = 0;

                        raid_conf->last_used = mirror->next;

                }

                PRINTK (("raid1 read queue: %d %d\n", MAJOR (bh->b_rdev), MINOR (bh->b_rdev)));

                clear_bit (BH_Lock, &bh->b_state);

                map_and_make_request (rw, bh);

                return 0;

        }

        /*

         * WRITE or WRITEA.

         */

/*

 * btw, we have no more master disk. 'slave' is gone too :) [i hate that word :))]

 *

 * We are now using the master bh for a real IO. It seems important that:

 *

 * 1.   lock_buffer() will be called when we start to handle the request,

 *      before we do anything (done by ll_rw_blk.c).

 *

 * 2.   It seems that Linus took great care to set mark_buffer_clean()

 *      atomically with cli() in effect just when the buffer was placed

 *      into the queue. To be compatible with this behavior, it would be

 *      best to lock the buffer *first*, but mark it clean *last*, and to

 *      do this by passing through the exact logic in ll_rw_blk.c.

 *

 * Note: i've reverted this #3 thing, see the big comment in this file.

 *

 * 3.   We are now called from within make_request(), so the real bh

 *      will be automatically handled last when we return, so we only need

 *      to add the rest of the buffers (but remember to include the

 *      master bh in the remaining count).

 */

        PRINTK(("raid1_make_request(n=%d), write branch.\n",n));

        for (i = 0; i < n; i++) {

                if (!raid_conf->mirrors [i].operational) {

                        /*

                         * the r1_bh->mirror_bh[i] pointer remains NULL

                         */

                        mirror_bh[i] = NULL;

                        continue;

                }

        /*

         * We should use a private pool (size depending on NR_REQUEST),

         * to avoid writes filling up the memory with bhs

         *

         * Such pools are much faster than kmalloc anyways (so we waste almost

         * nothing by not using the master bh when writing and win alot of cleanness)

         *

         * but for now we are cool enough. --mingo

         *

         * It's safe to sleep here, buffer heads cannot be used in a shared

         * manner in the write branch. Look how we lock the buffer at the beginning

         * of this function to grok the difference ;)

         */

                while (!( /* FIXME: now we are rather fault tolerant than nice */

                mirror_bh[i] = kmalloc (sizeof (struct buffer_head), GFP_KERNEL)

                ) )

                        printk ("raid1_make_request(#2): out of memory\n");

                memset (mirror_bh[i], 0, sizeof (struct buffer_head));

        /*

         * prepare mirrored bh (fields ordered for max mem throughput):

         */

                mirror_bh [i]->b_blocknr    = bh->b_blocknr;

                mirror_bh [i]->b_dev        = bh->b_dev;

                mirror_bh [i]->b_rdev       = raid_conf->mirrors [i].dev;

                mirror_bh [i]->b_rsector    = bh->b_rsector;

                mirror_bh [i]->b_state      =   (1<<BH_MD)      | (1<<BH_Req) |

                                                (1<<BH_Touched) | (1<<BH_Dirty);

                mirror_bh [i]->b_count      = 1;

                mirror_bh [i]->b_size       = bh->b_size;

                mirror_bh [i]->b_data       = bh->b_data;

                mirror_bh [i]->b_list       = BUF_LOCKED;

                mirror_bh [i]->personality  = &raid1_personality;

                mirror_bh [i]->private_bh   = (void*)(r1_bh);

                r1_bh->mirror_bh[i] = mirror_bh[i];

                sum_bhs++;

        }

        r1_bh->remaining = sum_bhs;

        PRINTK(("raid1_make_request(), write branch, sum_bhs=%d.\n",sum_bhs));

        /*

         * We have to be a bit careful about the semaphore above, thats why we

         * start the requests separately. Since kmalloc() could fail, sleep and

         * make_request() can sleep too, this is the safer solution. Imagine,

         * end_request decreasing the semaphore before we could have set it up ...

         * We could play tricks with the semaphore (presetting it and correcting

         * at the end if sum_bhs is not 'n' but we have to do end_request by hand

         * if all requests finish until we had a chance to set up the semaphore

         * correctly ... lots of races).

         */

        for (i = 0; i < n; i++)

                if (mirror_bh [i] != NULL)

                        map_and_make_request (rw, mirror_bh [i]);

        return (0);

}

static int raid1_status (char *page, int minor, struct md_dev *mddev)

{

        struct raid1_data *raid_conf = (struct raid1_data *) mddev->private;

        int sz = 0, i;

        sz += sprintf (page+sz, " [%d/%d] [", raid_conf->raid_disks, raid_conf->working_disks);

        for (i = 0; i < raid_conf->raid_disks; i++)

                sz += sprintf (page+sz, "%s", raid_conf->mirrors [i].operational ? "U" : "_");

        sz += sprintf (page+sz, "]");

        return sz;

}

static void raid1_fix_links (struct raid1_data *raid_conf, int failed_index)

{

        int disks = raid_conf->raid_disks;

        int j;

        for (j = 0; j < disks; j++)

                if (raid_conf->mirrors [j].next == failed_index)

                        raid_conf->mirrors [j].next = raid_conf->mirrors [failed_index].next;

}

static int raid1_error (struct md_dev *mddev, kdev_t dev)

{

        struct raid1_data *raid_conf = (struct raid1_data *) mddev->private;

        struct mirror_info *mirror;

        md_superblock_t *sb = mddev->sb;

        int disks = raid_conf->raid_disks;

        int i;

        PRINTK(("raid1_error called\n"));

        if (raid_conf->working_disks == 1) {

                /*

                 * Uh oh, we can do nothing if this is our last disk, but

                 * first check if this is a queued request for a device

                 * which has just failed.

                 */

                for (i = 0, mirror = raid_conf->mirrors; i < disks; i++, mirror++)

                        if (mirror->dev == dev && !mirror->operational)

                                return 0;

                printk (KERN_ALERT "RAID1: only one disk left and IO error.\n");

                return 0;

        }

        /* Mark disk as unusable */

        for (i = 0, mirror = raid_conf->mirrors; i < disks; i++, mirror++) {

                if (mirror->dev == dev && mirror->operational){

                        mirror->operational = 0;

                        raid1_fix_links (raid_conf, i);

                        sb->disks[mirror->number].state |= (1 << MD_FAULTY_DEVICE);

                        sb->disks[mirror->number].state &= ~(1 << MD_SYNC_DEVICE);

                        sb->disks[mirror->number].state &= ~(1 << MD_ACTIVE_DEVICE);

                        sb->active_disks--;

                        sb->working_disks--;

                        sb->failed_disks++;

                        mddev->sb_dirty = 1;

                        md_wakeup_thread(raid1_thread);

                        raid_conf->working_disks--;

                        printk (KERN_ALERT

                                "RAID1: Disk failure on %s, disabling device."

                                "Operation continuing on %d devices\n",

                                kdevname (dev), raid_conf->working_disks);

                }

        }

        return 0;

}

/*

 * This is a kernel thread which:

 *

 *      1.      Retries failed read operations on working mirrors.

 *      2.      Updates the raid superblock when problems are encountered.

 */

void raid1d (void *data)

{

        struct buffer_head *bh;

        kdev_t dev;

        unsigned long flags;

        struct raid1_bh * r1_bh;

        struct md_dev *mddev;

        PRINTK(("raid1d() active\n"));

        save_flags(flags);

        cli();

        while (raid1_retry_list) {

                bh = raid1_retry_list;

                r1_bh = (struct raid1_bh *)(bh->private_bh);

                raid1_retry_list = r1_bh->next_retry;

                restore_flags(flags);

                mddev = md_dev + MINOR(bh->b_dev);

                if (mddev->sb_dirty) {

                        mddev->sb_dirty = 0;

                        md_update_sb(MINOR(bh->b_dev));

                }

                dev = bh->b_rdev;

                __raid1_map (md_dev + MINOR(bh->b_dev), &bh->b_rdev, &bh->b_rsector, bh->b_size >> 9);

                if (bh->b_rdev == dev) {

                        printk (KERN_ALERT

                                        "raid1: %s: unrecoverable I/O read error for block %lu\n",

                                                kdevname(bh->b_dev), bh->b_blocknr);

                        raid1_end_buffer_io (bh, 0);

                } else {

                        printk (KERN_ERR "raid1: %s: redirecting sector %lu to another mirror\n",

                                          kdevname(bh->b_dev), bh->b_blocknr);

                        clear_bit (BH_Lock, &bh->b_state);

                        map_and_make_request (r1_bh->cmd, bh);

                }

                cli();

        }

        restore_flags(flags);

}

/*

 * This will catch the scenario in which one of the mirrors was

 * mounted as a normal device rather than as a part of a raid set.

 */

static int check_consistenty (struct md_dev *mddev)

{

        struct raid1_data *raid_conf = mddev->private;

        kdev_t dev;

        struct buffer_head *bh = NULL;

        int i, rc = 0;

        char *buffer = NULL;

        for (i = 0; i < raid_conf->raid_disks; i++) {

                if (!raid_conf->mirrors[i].operational)

                        continue;

                dev = raid_conf->mirrors[i].dev;

                set_blocksize(dev, 4096);

                if ((bh = bread(dev, 0, 4096)) == NULL)

                        break;

                if (!buffer) {

                        buffer = (char *) __get_free_page(GFP_KERNEL);

                        if (!buffer)

                                break;

                        memcpy(buffer, bh->b_data, 4096);

                } else if (memcmp(buffer, bh->b_data, 4096)) {

                        rc = 1;

                        break;

                }

                bforget(bh);

                fsync_dev(dev);

                invalidate_buffers(dev);

                bh = NULL;

        }

        if (buffer)

                free_page((unsigned long) buffer);

        if (bh) {

                dev = bh->b_dev;

                bforget(bh);

                fsync_dev(dev);

                invalidate_buffers(dev);

        }

        return rc;

}

static int raid1_run (int minor, struct md_dev *mddev)

{

        struct raid1_data *raid_conf;

        int i, j, raid_disk;

        md_superblock_t *sb = mddev->sb;

        md_descriptor_t *descriptor;

        struct real_dev *realdev;

        MOD_INC_USE_COUNT;

        if (sb->level != 1) {

                printk("raid1: %s: raid level not set to mirroring (%d)\n", kdevname(MKDEV(MD_MAJOR, minor)), sb->level);

                MOD_DEC_USE_COUNT;

                return -EIO;

        }

        /****

         * copy the now verified devices into our private RAID1 bookkeeping area:

         *

         * [whatever we allocate in raid1_run(), should be freed in raid1_stop()]

         */

        while (!( /* FIXME: now we are rather fault tolerant than nice */

        mddev->private = kmalloc (sizeof (struct raid1_data), GFP_KERNEL)

        ) )

                printk ("raid1_run(): out of memory\n");

        raid_conf = mddev->private;

        memset(raid_conf, 0, sizeof(*raid_conf));

        PRINTK(("raid1_run(%d) called.\n", minor));

        for (i = 0; i < mddev->nb_dev; i++) {

                realdev = &mddev->devices[i];

                if (!realdev->sb) {

                        printk(KERN_ERR "raid1: disabled mirror %s (couldn't access raid superblock)\n", kdevname(realdev->dev));

                        continue;

                }

                /*

                 * This is important -- we are using the descriptor on

                 * the disk only to get a pointer to the descriptor on

                 * the main superblock, which might be more recent.

                 */

                descriptor = &sb->disks[realdev->sb->descriptor.number];

                if (descriptor->state & (1 << MD_FAULTY_DEVICE)) {

                        printk(KERN_ERR "raid1: disabled mirror %s (errors detected)\n", kdevname(realdev->dev));

                        continue;

                }

                if (descriptor->state & (1 << MD_ACTIVE_DEVICE)) {

                        if (!(descriptor->state & (1 << MD_SYNC_DEVICE))) {

                                printk(KERN_ERR "raid1: disabled mirror %s (not in sync)\n", kdevname(realdev->dev));

                                continue;

                        }

                        raid_disk = descriptor->raid_disk;

                        if (descriptor->number > sb->nr_disks || raid_disk > sb->raid_disks) {

                                printk(KERN_ERR "raid1: disabled mirror %s (inconsistent descriptor)\n", kdevname(realdev->dev));

                                continue;

                        }

                        if (raid_conf->mirrors[raid_disk].operational) {

                                printk(KERN_ERR "raid1: disabled mirror %s (mirror %d already operational)\n", kdevname(realdev->dev), raid_disk);

                                continue;

                        }

                        printk(KERN_INFO "raid1: device %s operational as mirror %d\n", kdevname(realdev->dev), raid_disk);

                        raid_conf->mirrors[raid_disk].number = descriptor->number;

                        raid_conf->mirrors[raid_disk].raid_disk = raid_disk;

                        raid_conf->mirrors[raid_disk].dev = mddev->devices [i].dev;

                        raid_conf->mirrors[raid_disk].operational = 1;

                        raid_conf->mirrors[raid_disk].sect_limit = 128;

                        raid_conf->working_disks++;

                }

        }

        if (!raid_conf->working_disks) {

                printk(KERN_ERR "raid1: no operational mirrors for %s\n", kdevname(MKDEV(MD_MAJOR, minor)));