BBR Target. --- diff/drivers/md/Config.in 2004-12-16 13:25:08.317323488 -0600 +++ source/drivers/md/Config.in 2004-12-16 13:25:32.576635512 -0600 @@ -16,5 +16,8 @@ dep_tristate ' Logical volume manager (LVM) support' CONFIG_BLK_DEV_LVM $CONFIG_MD dep_tristate ' Device-mapper support' CONFIG_BLK_DEV_DM $CONFIG_MD dep_tristate ' Mirror (RAID-1) support' CONFIG_BLK_DEV_DM_MIRROR $CONFIG_BLK_DEV_DM +if [ "$CONFIG_EXPERIMENTAL" = "y" ]; then + dep_tristate ' Bad Block Relocation Device Target (EXPERIMENTAL)' CONFIG_BLK_DEV_DM_BBR $CONFIG_BLK_DEV_DM +fi endmenu --- diff/drivers/md/Makefile 2004-12-16 13:25:08.326322120 -0600 +++ source/drivers/md/Makefile 2004-12-16 13:25:32.577635360 -0600 @@ -30,6 +30,7 @@ obj-$(CONFIG_BLK_DEV_DM) += dm-mod.o obj-$(CONFIG_BLK_DEV_DM_MIRROR) += dm-mirror.o +obj-$(CONFIG_BLK_DEV_DM_BBR) += dm-bbr.o include $(TOPDIR)/Rules.make --- diff/drivers/md/dm-bbr.c 1969-12-31 18:00:00.000000000 -0600 +++ source/drivers/md/dm-bbr.c 2004-12-16 13:25:37.149940264 -0600 @@ -0,0 +1,1013 @@ +/* + * (C) Copyright IBM Corp. 2002, 2004 + * + * 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 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See + * the GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA + * + * linux/drivers/md/dm-bbr.c + * + * Bad-block-relocation (BBR) target for device-mapper. + * + * The BBR target is designed to remap I/O write failures to another safe + * location on disk. Note that most disk drives have BBR built into them, + * this means that our software BBR will be only activated when all hardware + * BBR replacement sectors have been used. + */ + +#include +#include +#include +#include +#include +#include + +#include "dm.h" +#include "dm-bh-list.h" +#include "dm-bh-record.h" +#include "dm-bbr.h" +#include "dm-io.h" +#include "dm-daemon.h" + +static struct dm_daemon bbr_daemon; +static LIST_HEAD(bbr_daemon_list); +static DECLARE_MUTEX(bbr_daemon_list_lock); +static kmem_cache_t *bbr_remap_cache; +static kmem_cache_t *bbr_io_cache; +static mempool_t *bbr_io_pool; + +/** + * bbr_binary_tree_destroy + * + * Destroy the binary tree. + **/ +static void bbr_binary_tree_destroy(struct bbr_runtime_remap *root) +{ + struct bbr_runtime_remap **link = NULL; + struct bbr_runtime_remap *node = root; + + while (node) { + if (node->left) { + link = &(node->left); + node = node->left; + continue; + } + if (node->right) { + link = &(node->right); + node = node->right; + continue; + } + + kmem_cache_free(bbr_remap_cache, node); + if (node == root) { + /* If root is deleted, we're done. */ + break; + } + + /* Back to root. */ + node = root; + *link = NULL; + } +} + +static void bbr_free_remap(struct bbr_private *bbr_id) +{ + spin_lock_irq(&bbr_id->remap_root_lock); + bbr_binary_tree_destroy(bbr_id->remap_root); + bbr_id->remap_root = NULL; + spin_unlock_irq(&bbr_id->remap_root_lock); +} + +static struct bbr_private *bbr_alloc_private(void) +{ + struct bbr_private *bbr_id; + + bbr_id = kmalloc(sizeof(*bbr_id), GFP_KERNEL); + if (bbr_id) { + memset(bbr_id, 0, sizeof(*bbr_id)); + INIT_LIST_HEAD(&bbr_id->list); + bbr_id->remap_root_lock = SPIN_LOCK_UNLOCKED; + bbr_id->remap_ios_lock = SPIN_LOCK_UNLOCKED; + bbr_id->in_use_replacement_blks = (atomic_t)ATOMIC_INIT(0); + } + + return bbr_id; +} + +static void bbr_free_private(struct bbr_private *bbr_id) +{ + if (bbr_id->bbr_table) { + kfree(bbr_id->bbr_table); + } + bbr_free_remap(bbr_id); + kfree(bbr_id); +} + +static u32 crc_table[256]; +static u32 crc_table_built = 0; + +static void build_crc_table(void) +{ + u32 i, j, crc; + + for (i = 0; i <= 255; i++) { + crc = i; + for (j = 8; j > 0; j--) { + if (crc & 1) + crc = (crc >> 1) ^ CRC_POLYNOMIAL; + else + crc >>= 1; + } + crc_table[i] = crc; + } + crc_table_built = 1; +} + +static u32 calculate_crc(u32 crc, void *buffer, u32 buffersize) +{ + unsigned char *current_byte; + u32 temp1, temp2, i; + + current_byte = (unsigned char *) buffer; + /* Make sure the crc table is available */ + if (!crc_table_built) + build_crc_table(); + /* Process each byte in the buffer. */ + for (i = 0; i < buffersize; i++) { + temp1 = (crc >> 8) & 0x00FFFFFF; + temp2 = crc_table[(crc ^ (u32) * current_byte) & + (u32) 0xff]; + current_byte++; + crc = temp1 ^ temp2; + } + return crc; +} + +/** + * le_bbr_table_sector_to_cpu + * + * Convert bbr meta data from on-disk (LE) format + * to the native cpu endian format. + **/ +static void le_bbr_table_sector_to_cpu(struct bbr_table *p) +{ + int i; + p->signature = le32_to_cpup(&p->signature); + p->crc = le32_to_cpup(&p->crc); + p->sequence_number = le32_to_cpup(&p->sequence_number); + p->in_use_cnt = le32_to_cpup(&p->in_use_cnt); + for (i = 0; i < BBR_ENTRIES_PER_SECT; i++) { + p->entries[i].bad_sect = + le64_to_cpup(&p->entries[i].bad_sect); + p->entries[i].replacement_sect = + le64_to_cpup(&p->entries[i].replacement_sect); + } +} + +/** + * cpu_bbr_table_sector_to_le + * + * Convert bbr meta data from cpu endian format to on-disk (LE) format + **/ +static void cpu_bbr_table_sector_to_le(struct bbr_table *p, + struct bbr_table *le) +{ + int i; + le->signature = cpu_to_le32p(&p->signature); + le->crc = cpu_to_le32p(&p->crc); + le->sequence_number = cpu_to_le32p(&p->sequence_number); + le->in_use_cnt = cpu_to_le32p(&p->in_use_cnt); + for (i = 0; i < BBR_ENTRIES_PER_SECT; i++) { + le->entries[i].bad_sect = + cpu_to_le64p(&p->entries[i].bad_sect); + le->entries[i].replacement_sect = + cpu_to_le64p(&p->entries[i].replacement_sect); + } +} + +/** + * validate_bbr_table_sector + * + * Check the specified BBR table sector for a valid signature and CRC. If it's + * valid, endian-convert the table sector. + **/ +static int validate_bbr_table_sector(struct bbr_table *p) +{ + int rc = 0; + int org_crc, final_crc; + + if (le32_to_cpup(&p->signature) != BBR_TABLE_SIGNATURE) { + DMERR("BBR table signature doesn't match!"); + DMERR("Found 0x%x. Expecting 0x%x", + le32_to_cpup(&p->signature), BBR_TABLE_SIGNATURE); + rc = -EINVAL; + goto out; + } + + if (!p->crc) { + DMERR("BBR table sector has no CRC!"); + rc = -EINVAL; + goto out; + } + + org_crc = le32_to_cpup(&p->crc); + p->crc = 0; + final_crc = calculate_crc(INITIAL_CRC, (void *)p, sizeof(*p)); + if (final_crc != org_crc) { + DMERR("CRC failed!"); + DMERR("Found 0x%x. Expecting 0x%x", + org_crc, final_crc); + rc = -EINVAL; + goto out; + } + + p->crc = cpu_to_le32p(&org_crc); + le_bbr_table_sector_to_cpu(p); + +out: + return rc; +} + +/** + * bbr_binary_tree_insert + * + * Insert a node into the binary tree. + **/ +static void bbr_binary_tree_insert(struct bbr_runtime_remap **root, + struct bbr_runtime_remap *newnode) +{ + struct bbr_runtime_remap **node = root; + while (node && *node) { + if (newnode->remap.bad_sect > (*node)->remap.bad_sect) { + node = &((*node)->right); + } else { + node = &((*node)->left); + } + } + + newnode->left = newnode->right = NULL; + *node = newnode; +} + +/** + * bbr_binary_search + * + * Search for a node that contains bad_sect == lsn. + **/ +static struct bbr_runtime_remap *bbr_binary_search( + struct bbr_runtime_remap *root, + u64 lsn) +{ + struct bbr_runtime_remap *node = root; + while (node) { + if (node->remap.bad_sect == lsn) { + break; + } + if (lsn > node->remap.bad_sect) { + node = node->right; + } else { + node = node->left; + } + } + return node; +} + +/** + * bbr_insert_remap_entry + * + * Create a new remap entry and add it to the binary tree for this node. + **/ +static int bbr_insert_remap_entry(struct bbr_private *bbr_id, + struct bbr_table_entry *new_bbr_entry) +{ + struct bbr_runtime_remap *newnode; + + newnode = kmem_cache_alloc(bbr_remap_cache, GFP_NOIO); + if (!newnode) { + DMERR("Could not allocate from remap cache!"); + return -ENOMEM; + } + newnode->remap.bad_sect = new_bbr_entry->bad_sect; + newnode->remap.replacement_sect = new_bbr_entry->replacement_sect; + spin_lock_irq(&bbr_id->remap_root_lock); + bbr_binary_tree_insert(&bbr_id->remap_root, newnode); + spin_unlock_irq(&bbr_id->remap_root_lock); + return 0; +} + +/** + * bbr_table_to_remap_list + * + * The on-disk bbr table is sorted by the replacement sector LBA. In order to + * improve run time performance, the in memory remap list must be sorted by + * the bad sector LBA. This function is called at discovery time to initialize + * the remap list. This function assumes that at least one copy of meta data + * is valid. + **/ +static u32 bbr_table_to_remap_list(struct bbr_private *bbr_id) +{ + u32 in_use_blks = 0; + int i, j; + struct bbr_table *p; + + for (i = 0, p = bbr_id->bbr_table; + i < bbr_id->nr_sects_bbr_table; + i++, p++) { + if (!p->in_use_cnt) { + break; + } + in_use_blks += p->in_use_cnt; + for (j = 0; j < p->in_use_cnt; j++) { + bbr_insert_remap_entry(bbr_id, &p->entries[j]); + } + } + if (in_use_blks) { + DMWARN("There are %u BBR entries for device %s", + in_use_blks, dm_kdevname(bbr_id->dev->dev)); + } + + return in_use_blks; +} + +/** + * bbr_search_remap_entry + * + * Search remap entry for the specified sector. If found, return a pointer to + * the table entry. Otherwise, return NULL. + **/ +static struct bbr_table_entry *bbr_search_remap_entry( + struct bbr_private *bbr_id, + u64 lsn) +{ + struct bbr_runtime_remap *p; + + spin_lock_irq(&bbr_id->remap_root_lock); + p = bbr_binary_search(bbr_id->remap_root, lsn); + spin_unlock_irq(&bbr_id->remap_root_lock); + if (p) { + return (&p->remap); + } else { + return NULL; + } +} + +/** + * bbr_remap + * + * If *lsn is in the remap table, return TRUE and modify *lsn, + * else, return FALSE. + **/ +static inline int bbr_remap(struct bbr_private *bbr_id, + u64 *lsn) +{ + struct bbr_table_entry *e; + + if (atomic_read(&bbr_id->in_use_replacement_blks)) { + e = bbr_search_remap_entry(bbr_id, *lsn); + if (e) { + *lsn = e->replacement_sect; + return 1; + } + } + return 0; +} + +/** + * bbr_remap_probe + * + * If any of the sectors in the range [lsn, lsn+nr_sects] are in the remap + * table return TRUE, Else, return FALSE. + **/ +static inline int bbr_remap_probe(struct bbr_private *bbr_id, + u64 lsn, u64 nr_sects) +{ + u64 tmp, cnt; + + if (atomic_read(&bbr_id->in_use_replacement_blks)) { + for (cnt = 0, tmp = lsn; + cnt < nr_sects; + cnt += bbr_id->blksize_in_sects, tmp = lsn + cnt) { + if (bbr_remap(bbr_id,&tmp)) { + return 1; + } + } + } + return 0; +} + +/** + * bbr_setup + * + * Read the remap tables from disk and set up the initial remap tree. + **/ +static int bbr_setup(struct bbr_private *bbr_id) +{ + struct bbr_table *table = bbr_id->bbr_table; + struct page *page; + struct io_region job; + unsigned int error, offset; + int i, rc = 0; + + job.dev = bbr_id->dev->dev; + job.count = 1; + + /* Read and verify each BBR table sector individually. */ + for (i = 0; i < bbr_id->nr_sects_bbr_table; i++, table++) { + job.sector = bbr_id->lba_table1 + i; + page = virt_to_page(table); + offset = (unsigned long)table & ~PAGE_MASK; + rc = dm_io_sync(1, &job, READ, page, offset, &error); + if (rc && bbr_id->lba_table2) { + job.sector = bbr_id->lba_table2 + i; + rc = dm_io_sync(1, &job, READ, page, offset, &error); + } + if (rc) { + goto out; + } + + rc = validate_bbr_table_sector(table); + if (rc) { + goto out; + } + } + atomic_set(&bbr_id->in_use_replacement_blks, + bbr_table_to_remap_list(bbr_id)); + +out: + if (rc) { + DMERR("dm-bbr: error during device setup: %d", rc); + } + return rc; +} + +/** + * bbr_io_remap_error + * @bbr_id: Private data for the BBR node. + * @rw: READ or WRITE. + * @starting_lsn: Starting sector of request to remap. + * @count: Number of sectors in the request. + * @buffer: Data buffer for the request. + * + * For the requested range, try to write each sector individually. For each + * sector that fails, find the next available remap location and write the + * data to that new location. Then update the table and write both copies + * of the table to disk. Finally, update the in-memory mapping and do any + * other necessary bookkeeping. + **/ +static int bbr_io_remap_error(struct bbr_private *bbr_id, + int rw, + u64 starting_lsn, + u64 count, + char *buffer) +{ + struct bbr_table *bbr_table; + struct io_region job; + struct page *page; + unsigned long table_sector_index; + unsigned long table_sector_offset; + unsigned long index; + unsigned int offset_in_page, error; + u64 lsn, new_lsn; + int rc; + + if (rw == READ) { + /* Nothing can be done about read errors. */ + return -EIO; + } + + job.dev = bbr_id->dev->dev; + job.count = 1; + + /* For each sector in the request. */ + for (lsn = 0; lsn < count; lsn++, buffer += SECTOR_SIZE) { + job.sector = starting_lsn + lsn; + page = virt_to_page(buffer); + offset_in_page = (unsigned long)buffer & ~PAGE_MASK; + rc = dm_io_sync(1, &job, rw, page, offset_in_page, &error); + while (rc) { + /* Find the next available relocation sector. */ + new_lsn = atomic_read(&bbr_id->in_use_replacement_blks); + if (new_lsn >= bbr_id->nr_replacement_blks) { + /* No more replacement sectors available. */ + return -EIO; + } + new_lsn += bbr_id->start_replacement_sect; + + /* Write the data to its new location. */ + DMWARN("dm-bbr: device %s: Trying to remap bad sector "PFU64" to sector "PFU64, + dm_kdevname(bbr_id->dev->dev), + starting_lsn + lsn, new_lsn); + job.sector = new_lsn; + rc = dm_io_sync(1, &job, rw, page, offset_in_page, &error); + if (rc) { + /* This replacement sector is bad. + * Try the next one. + */ + DMERR("dm-bbr: device %s: replacement sector "PFU64" is bad. Skipping.", + dm_kdevname(bbr_id->dev->dev), new_lsn); + atomic_inc(&bbr_id->in_use_replacement_blks); + continue; + } + + /* Add this new entry to the on-disk table. */ + table_sector_index = new_lsn - + bbr_id->start_replacement_sect; + table_sector_offset = table_sector_index / + BBR_ENTRIES_PER_SECT; + index = table_sector_index % BBR_ENTRIES_PER_SECT; + + bbr_table = &bbr_id->bbr_table[table_sector_offset]; + bbr_table->entries[index].bad_sect = starting_lsn + lsn; + bbr_table->entries[index].replacement_sect = new_lsn; + bbr_table->in_use_cnt++; + bbr_table->sequence_number++; + bbr_table->crc = 0; + bbr_table->crc = calculate_crc(INITIAL_CRC, + bbr_table, + sizeof(struct bbr_table)); + + /* Write the table to disk. */ + cpu_bbr_table_sector_to_le(bbr_table, bbr_table); + page = virt_to_page(bbr_table); + offset_in_page = (unsigned long)bbr_table & ~PAGE_MASK; + if (bbr_id->lba_table1) { + job.sector = bbr_id->lba_table1 + table_sector_offset; + rc = dm_io_sync(1, &job, WRITE, page, offset_in_page, &error); + } + if (bbr_id->lba_table2) { + job.sector = bbr_id->lba_table2 + table_sector_offset; + rc |= dm_io_sync(1, &job, WRITE, page, offset_in_page, &error); + } + le_bbr_table_sector_to_cpu(bbr_table); + + if (rc) { + /* Error writing one of the tables to disk. */ + DMERR("dm-bbr: device %s: error updating BBR tables on disk.", + dm_kdevname(bbr_id->dev->dev)); + return rc; + } + + /* Insert a new entry in the remapping binary-tree. */ + rc = bbr_insert_remap_entry(bbr_id, + &bbr_table->entries[index]); + if (rc) { + DMERR("dm-bbr: device %s: error adding new entry to remap tree.", + dm_kdevname(bbr_id->dev->dev)); + return rc; + } + + atomic_inc(&bbr_id->in_use_replacement_blks); + } + } + + return 0; +} + +/** + * bbr_io_process_request + * + * For each sector in this request, check if the sector has already + * been remapped. If so, process all previous sectors in the request, + * followed by the remapped sector. Then reset the starting lsn and + * count, and keep going with the rest of the request as if it were + * a whole new request. If any of the sync_io's return an error, + * call the remapper to relocate the bad sector(s). + **/ +static int bbr_io_process_request(struct bbr_private *bbr_id, + struct buffer_head *bh, int rw) +{ + struct io_region job; + u64 starting_lsn = bh->b_rsector; + u64 count, lsn, remapped_lsn; + char *buffer; + struct page *page; + unsigned int offset_in_page; + unsigned int error; + int rc = 0; + + count = bh->b_size >> SECTOR_SHIFT; + buffer = bh->b_data; + page = virt_to_page(buffer); + offset_in_page = (unsigned long)buffer & ~PAGE_MASK; + + job.dev = bbr_id->dev->dev; + + /* For each sector in this request, check if this sector has + * already been remapped. If so, process all previous sectors + * in this request, followed by the remapped sector. Then reset + * the starting lsn and count and keep going with the rest of + * the request as if it were a whole new request. + */ + for (lsn = 0; lsn < count; lsn++) { + remapped_lsn = starting_lsn + lsn; + rc = bbr_remap(bbr_id, &remapped_lsn); + if (!rc) { + /* This sector is fine. */ + continue; + } + + /* Process all sectors in the request up to this one. */ + if (lsn > 0) { + job.sector = starting_lsn; + job.count = lsn; + rc = dm_io_sync(1, &job, rw, page, + offset_in_page, &error); + if (rc) { + /* If this I/O failed, then one of the + * sectors in this request needs to be + * relocated. + */ + rc = bbr_io_remap_error(bbr_id, rw, + starting_lsn, + lsn, buffer); + if (rc) { + return rc; + } + } + buffer += (lsn << SECTOR_SHIFT); + page = virt_to_page(buffer); + offset_in_page = (unsigned long)buffer & ~PAGE_MASK; + } + + /* Process the remapped sector. */ + job.sector = remapped_lsn; + job.count = 1; + rc = dm_io_sync(1, &job, rw, page, offset_in_page, &error); + if (rc) { + /* BUGBUG - Need more processing if this caused + * an error. If this I/O failed, then the + * existing remap is now bad, and we need to + * find a new remap. Can't use + * bbr_io_remap_error(), because the existing + * map entry needs to be changed, not added + * again, and the original table entry also + * needs to be changed. + */ + return rc; + } + + buffer += SECTOR_SIZE; + starting_lsn += (lsn + 1); + count -= (lsn + 1); + lsn = -1; + page = virt_to_page(buffer); + offset_in_page = (unsigned long)buffer & ~PAGE_MASK; + } + + /* Check for any remaining sectors after the last split. This + * could potentially be the whole request, but that should be a + * rare case because requests should only be processed by the + * thread if we know an error occurred or they contained one or + * more remapped sectors. + */ + if (count) { + job.sector = starting_lsn; + job.count = count; + rc = dm_io_sync(1, &job, rw, page, offset_in_page, &error); + if (rc) { + /* If this I/O failed, then one of the sectors + * in this request needs to be relocated. + */ + rc = bbr_io_remap_error(bbr_id, rw, starting_lsn, + count, buffer); + if (rc) { + return rc; + } + } + } + + return 0; +} + +static void bbr_io_process_requests(struct bbr_private *bbr_id, + struct buffer_head *bh, int rw) +{ + struct buffer_head *next; + int rc; + + while (bh) { + next = bh->b_reqnext; + bh->b_reqnext = NULL; + + rc = bbr_io_process_request(bbr_id, bh, rw); + + if (bh->b_end_io) + bh->b_end_io(bh, rc ? 0 : 1); + + bh = next; + } +} + +/** + * bbr_remap_handler + * + * This is the handler for the bbr daemon. + * + * I/O requests should only be sent to this handler if we know that: + * a) the request contains at least one remapped sector. + * or + * b) the request caused an error on the normal I/O path. + * + * This function uses synchronous I/O, so sending a request to this + * thread that doesn't need special processing will cause severe + * performance degredation. + **/ +static void bbr_remap_handler(struct bbr_private *bbr_id) +{ + struct buffer_head *rbh, *wbh; + unsigned long flags; + + spin_lock_irqsave(&bbr_id->remap_ios_lock, flags); + rbh = bh_list_get(&bbr_id->remap_ios_r); + wbh = bh_list_get(&bbr_id->remap_ios_w); + spin_unlock_irqrestore(&bbr_id->remap_ios_lock, flags); + + bbr_io_process_requests(bbr_id, rbh, READ); + bbr_io_process_requests(bbr_id, wbh, WRITE); +} + +static void do_work(void) +{ + struct bbr_private *bbr_id; + + down(&bbr_daemon_list_lock); + list_for_each_entry(bbr_id, &bbr_daemon_list, list) { + bbr_remap_handler(bbr_id); + } + up(&bbr_daemon_list_lock); +} + +/** + * bbr_endio + * + * This is the callback for normal write requests. Check for an error + * during the I/O, and send to the thread for processing if necessary. + **/ +static int bbr_endio(struct dm_target *ti, struct buffer_head *bh, int rw, + int error, union map_info *map_context) +{ + struct bbr_private *bbr_id = ti->private; + struct dm_bh_details *bbr_io = map_context->ptr; + + if (error && rw == WRITE && bbr_io) { + unsigned long flags; + + dm_bh_restore(bbr_io, bh); + map_context->ptr = NULL; + + DMERR("dm-bbr: device %s: Write failure on sector %lu. " + "Scheduling for retry.", + dm_kdevname(bbr_id->dev->dev), + (unsigned long)bh->b_rsector); + + spin_lock_irqsave(&bbr_id->remap_ios_lock, flags); + bh_list_add(&bbr_id->remap_ios_w, bh); + spin_unlock_irqrestore(&bbr_id->remap_ios_lock, flags); + + dm_daemon_wake(&bbr_daemon); + + error = 1; + } + + if (bbr_io) + mempool_free(bbr_io, bbr_io_pool); + + return error; +} + +/** + * Construct a bbr mapping + **/ +static int bbr_ctr(struct dm_target *ti, unsigned int argc, char **argv) +{ + struct bbr_private *bbr_id; + unsigned long block_size; + char *end; + int rc = -EINVAL; + + if (argc != 8) { + ti->error = "dm-bbr requires exactly 8 arguments: " + "device offset table1_lsn table2_lsn table_size start_replacement nr_replacement_blks block_size"; + goto out1; + } + + bbr_id = bbr_alloc_private(); + if (!bbr_id) { + ti->error = "dm-bbr: Error allocating bbr private data."; + goto out1; + } + + bbr_id->offset = simple_strtoull(argv[1], &end, 10); + bbr_id->lba_table1 = simple_strtoull(argv[2], &end, 10); + bbr_id->lba_table2 = simple_strtoull(argv[3], &end, 10); + bbr_id->nr_sects_bbr_table = simple_strtoull(argv[4], &end, 10); + bbr_id->start_replacement_sect = simple_strtoull(argv[5], &end, 10); + bbr_id->nr_replacement_blks = simple_strtoull(argv[6], &end, 10); + block_size = simple_strtoul(argv[7], &end, 10); + bbr_id->blksize_in_sects = (block_size >> SECTOR_SHIFT); + + bbr_id->bbr_table = kmalloc(bbr_id->nr_sects_bbr_table << SECTOR_SHIFT, + GFP_KERNEL); + if (!bbr_id->bbr_table) { + ti->error = "dm-bbr: Error allocating bbr table."; + goto out2; + } + + if (dm_get_device(ti, argv[0], 0, ti->len, + dm_table_get_mode(ti->table), &bbr_id->dev)) { + ti->error = "dm-bbr: Device lookup failed"; + goto out2; + } + + rc = bbr_setup(bbr_id); + if (rc) { + ti->error = "dm-bbr: Device setup failed"; + goto out3; + } + + down(&bbr_daemon_list_lock); + list_add_tail(&bbr_id->list, &bbr_daemon_list); + up(&bbr_daemon_list_lock); + + ti->private = bbr_id; + return 0; + +out3: + dm_put_device(ti, bbr_id->dev); +out2: + bbr_free_private(bbr_id); +out1: + return rc; +} + +static void bbr_dtr(struct dm_target *ti) +{ + struct bbr_private *bbr_id = ti->private; + + down(&bbr_daemon_list_lock); + list_del(&bbr_id->list); + up(&bbr_daemon_list_lock); + + dm_put_device(ti, bbr_id->dev); + bbr_free_private(bbr_id); +} + +static int bbr_map(struct dm_target *ti, struct buffer_head *bh, int rw, + union map_info *map_context) +{ + struct bbr_private *bbr_id = ti->private; + struct dm_bh_details *bbr_io; + unsigned long flags; + int rc = 1; + + bh->b_rsector += bbr_id->offset; + + if (atomic_read(&bbr_id->in_use_replacement_blks) == 0 || + !bbr_remap_probe(bbr_id, bh->b_rsector, bh->b_size >> SECTOR_SHIFT)) { + /* No existing remaps or this request doesn't + * contain any remapped sectors. + */ + bh->b_rdev = bbr_id->dev->dev; + + bbr_io = mempool_alloc(bbr_io_pool, GFP_NOIO); + dm_bh_record(bbr_io, bh); + map_context->ptr = bbr_io; + } else { + /* This request has at least one remapped sector. + * Give it to the daemon for processing. + */ + map_context->ptr = NULL; + spin_lock_irqsave(&bbr_id->remap_ios_lock, flags); + if (rw == READ) + bh_list_add(&bbr_id->remap_ios_r, bh); + else if (rw == WRITE) + bh_list_add(&bbr_id->remap_ios_w, bh); + spin_unlock_irqrestore(&bbr_id->remap_ios_lock, flags); + + dm_daemon_wake(&bbr_daemon); + rc = 0; + } + + return rc; +} + +static int bbr_status(struct dm_target *ti, status_type_t type, + char *result, unsigned int maxlen) +{ + struct bbr_private *bbr_id = ti->private; + + switch (type) { + case STATUSTYPE_INFO: + result[0] = '\0'; + break; + + case STATUSTYPE_TABLE: + snprintf(result, maxlen, "%s "PFU64" "PFU64" "PFU64" "PFU64" "PFU64" "PFU64" %u", + dm_kdevname(bbr_id->dev->dev), + bbr_id->offset, bbr_id->lba_table1, bbr_id->lba_table2, + bbr_id->nr_sects_bbr_table, + bbr_id->start_replacement_sect, + bbr_id->nr_replacement_blks, + bbr_id->blksize_in_sects << SECTOR_SHIFT); + break; + } + return 0; +} + +static struct target_type bbr_target = { + .name = "bbr", + .version= {1, 0, 1}, + .module = THIS_MODULE, + .ctr = bbr_ctr, + .dtr = bbr_dtr, + .map = bbr_map, + .end_io = bbr_endio, + .status = bbr_status, +}; + +int __init dm_bbr_init(void) +{ + int rc; + + rc = dm_register_target(&bbr_target); + if (rc) { + DMERR("dm-bbr: error registering target."); + goto err1; + } + + bbr_remap_cache = kmem_cache_create("bbr-remap", + sizeof(struct bbr_runtime_remap), + 0, SLAB_HWCACHE_ALIGN, NULL, NULL); + if (!bbr_remap_cache) { + DMERR("dm-bbr: error creating remap cache."); + rc = ENOMEM; + goto err2; + } + + bbr_io_cache = kmem_cache_create("bbr-io", sizeof(struct dm_bh_details), + 0, SLAB_HWCACHE_ALIGN, NULL, NULL); + if (!bbr_io_cache) { + DMERR("dm-bbr: error creating io cache."); + rc = ENOMEM; + goto err3; + } + + bbr_io_pool = mempool_create(256, mempool_alloc_slab, + mempool_free_slab, bbr_io_cache); + if (!bbr_io_pool) { + DMERR("dm-bbr: error creating io mempool."); + rc = ENOMEM; + goto err4; + } + + rc = dm_daemon_start(&bbr_daemon, "dm-bbr", do_work); + if (rc) { + DMERR("dm-bbr: error creating daemon."); + goto err5; + } + + rc = dm_io_get(1); + if (rc) { + DMERR("dm-bbr: error initializing I/O service."); + goto err6; + } + + return 0; + +err6: + dm_daemon_stop(&bbr_daemon); +err5: + mempool_destroy(bbr_io_pool); +err4: + kmem_cache_destroy(bbr_io_cache); +err3: + kmem_cache_destroy(bbr_remap_cache); +err2: + dm_unregister_target(&bbr_target); +err1: + return rc; +} + +void __exit dm_bbr_exit(void) +{ + dm_io_put(1); + dm_daemon_stop(&bbr_daemon); + mempool_destroy(bbr_io_pool); + kmem_cache_destroy(bbr_io_cache); + kmem_cache_destroy(bbr_remap_cache); + dm_unregister_target(&bbr_target); +} + +module_init(dm_bbr_init); +module_exit(dm_bbr_exit); +MODULE_LICENSE("GPL"); --- diff/drivers/md/dm-bbr.h 1969-12-31 18:00:00.000000000 -0600 +++ source/drivers/md/dm-bbr.h 2004-12-16 13:25:32.582634600 -0600 @@ -0,0 +1,128 @@ +/* + * (C) Copyright IBM Corp. 2002, 2004 + * + * 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 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See + * the GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA + * + * linux/drivers/md/dm-bbr.h + * + * Bad-block-relocation (BBR) target for device-mapper. + * + * The BBR target is designed to remap I/O write failures to another safe + * location on disk. Note that most disk drives have BBR built into them, + * this means that our software BBR will be only activated when all hardware + * BBR replacement sectors have been used. + */ + +#define BBR_TABLE_SIGNATURE 0x42627254 /* BbrT */ +#define BBR_ENTRIES_PER_SECT 31 +#define INITIAL_CRC 0xFFFFFFFF +#define CRC_POLYNOMIAL 0xEDB88320L + +/** + * Macros to cleanly print 64-bit numbers on both 32-bit and 64-bit machines. + * Use these in place of %Ld, %Lu, and %Lx. + **/ +#if BITS_PER_LONG > 32 +#define PFU64 "%lu" +#else +#define PFU64 "%Lu" +#endif + +/** + * struct bbr_table_entry + * @bad_sect: LBA of bad location. + * @replacement_sect: LBA of new location. + * + * Structure to describe one BBR remap. + **/ +struct bbr_table_entry { + u64 bad_sect; + u64 replacement_sect; +}; + +/** + * struct bbr_table + * @signature: Signature on each BBR table sector. + * @crc: CRC for this table sector. + * @sequence_number: Used to resolve conflicts when primary and secondary + * tables do not match. + * @in_use_cnt: Number of in-use table entries. + * @entries: Actual table of remaps. + * + * Structure to describe each sector of the metadata table. Each sector in this + * table can describe 31 remapped sectors. + **/ +struct bbr_table { + u32 signature; + u32 crc; + u32 sequence_number; + u32 in_use_cnt; + struct bbr_table_entry entries[BBR_ENTRIES_PER_SECT]; +}; + +/** + * struct bbr_runtime_remap + * + * Node in the binary tree used to keep track of remaps. + **/ +struct bbr_runtime_remap { + struct bbr_table_entry remap; + struct bbr_runtime_remap *left; + struct bbr_runtime_remap *right; +}; + +/** + * struct bbr_private + * @list: List of all BBR devices. + * @dev: Info about underlying device. + * @bbr_table: Copy of metadata table. + * @remap_root: Binary tree containing all remaps. + * @remap_root_lock: Lock for the binary tree. + * @remap_ios_r: List of read I/Os for the daemon to handle. + * @remap_ios_w: List of read I/Os for the daemon to handle. + * @remap_ios_lock: Lock for the remap_ios list. + * @offset: LBA of data area. + * @lba_table1: LBA of primary BBR table. + * @lba_table2: LBA of secondary BBR table. + * @nr_sects_bbr_table: Size of each BBR table. + * @nr_replacement_blks: Number of replacement blocks. + * @start_replacement_sect: LBA of start of replacement blocks. + * @blksize_in_sects: Size of each block. + * @in_use_replacement_blks: Current number of remapped blocks. + * + * Private data for each BBR target. + **/ +struct bbr_private { + struct list_head list; + + struct dm_dev *dev; + struct bbr_table *bbr_table; + struct bbr_runtime_remap *remap_root; + spinlock_t remap_root_lock; + + struct bh_list remap_ios_r; + struct bh_list remap_ios_w; + spinlock_t remap_ios_lock; + + u64 offset; + u64 lba_table1; + u64 lba_table2; + u64 nr_sects_bbr_table; + u64 start_replacement_sect; + u64 nr_replacement_blks; + u32 blksize_in_sects; + atomic_t in_use_replacement_blks; +}; + --- diff/drivers/md/dm-bh-list.h 1969-12-31 18:00:00.000000000 -0600 +++ source/drivers/md/dm-bh-list.h 2004-12-16 13:25:32.583634448 -0600 @@ -0,0 +1,68 @@ +/* + * Copyright (C) 2004 Red Hat UK Ltd. + * + * This file is released under the GPL. + */ + +#ifndef DM_BH_LIST_H +#define DM_BH_LIST_H + +#include + +struct bh_list { + struct buffer_head *head; + struct buffer_head *tail; +}; + +static inline void bh_list_init(struct bh_list *bl) +{ + bl->head = bl->tail = NULL; +} + +static inline void bh_list_add(struct bh_list *bl, struct buffer_head *bh) +{ + bh->b_reqnext = NULL; + + if (bl->tail) + bl->tail->b_reqnext = bh; + else + bl->head = bh; + + bl->tail = bh; +} + +static inline void bh_list_merge(struct bh_list *bl, struct bh_list *bl2) +{ + if (bl->tail) + bl->tail->b_reqnext = bl2->head; + else + bl->head = bl2->head; + + bl->tail = bl2->tail; +} + +static inline struct buffer_head *bh_list_pop(struct bh_list *bl) +{ + struct buffer_head *bh = bl->head; + + if (bh) { + bl->head = bl->head->b_reqnext; + if (!bl->head) + bl->tail = NULL; + + bh->b_reqnext = NULL; + } + + return bh; +} + +static inline struct buffer_head *bh_list_get(struct bh_list *bl) +{ + struct buffer_head *bh = bl->head; + + bl->head = bl->tail = NULL; + + return bh; +} + +#endif --- diff/drivers/md/dm-bh-record.h 1969-12-31 18:00:00.000000000 -0600 +++ source/drivers/md/dm-bh-record.h 2004-12-16 13:25:32.584634296 -0600 @@ -0,0 +1,38 @@ +/* + * Copyright (C) 2004 Red Hat UK Ltd. + * + * This file is released under the GPL. + */ + +#ifndef DM_BH_RECORD_H +#define DM_BH_RECORD_H + +#include + +/* + * There are lots of mutable fields in the buffer-head struct that get + * changed by the lower levels of the block layer. Some targets, + * such as multipath, may wish to resubmit a buffer-head on error. The + * functions in this file help the target record and restore the + * original buffer-head state. + */ +struct dm_bh_details { + unsigned long b_rsector; + kdev_t b_rdev; +}; + +static inline void dm_bh_record(struct dm_bh_details *bd, + struct buffer_head *bh) +{ + bd->b_rsector = bh->b_rsector; + bd->b_rdev = bh->b_rdev; +} + +static inline void dm_bh_restore(struct dm_bh_details *bd, + struct buffer_head *bh) +{ + bh->b_rsector = bd->b_rsector; + bh->b_rdev = bd->b_rdev; +} + +#endif