badblocks: Add core badblock management code
authorVishal Verma <vishal.l.verma@intel.com>
Fri, 25 Dec 2015 02:20:32 +0000 (19:20 -0700)
committerDan Williams <dan.j.williams@intel.com>
Sat, 9 Jan 2016 14:35:12 +0000 (06:35 -0800)
Take the core badblocks implementation from md, and make it generally
available. This follows the same style as kernel implementations of
linked lists, rb-trees etc, where you can have a structure that can be
embedded anywhere, and accessor functions to manipulate the data.

The only changes in this copy of the code are ones to generalize
function/variable names from md-specific ones. Also add init and free
functions.

Signed-off-by: Vishal Verma <vishal.l.verma@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
block/Makefile
block/badblocks.c [new file with mode: 0644]
include/linux/badblocks.h [new file with mode: 0644]

index 00ecc97629db783d735798c961579ee1a531c367..db5f622c9d67e05a4e2a15778c922b367e9cd82b 100644 (file)
@@ -8,7 +8,7 @@ obj-$(CONFIG_BLOCK) := bio.o elevator.o blk-core.o blk-tag.o blk-sysfs.o \
                        blk-iopoll.o blk-lib.o blk-mq.o blk-mq-tag.o \
                        blk-mq-sysfs.o blk-mq-cpu.o blk-mq-cpumap.o ioctl.o \
                        genhd.o scsi_ioctl.o partition-generic.o ioprio.o \
-                       partitions/
+                       badblocks.o partitions/
 
 obj-$(CONFIG_BOUNCE)   += bounce.o
 obj-$(CONFIG_BLK_DEV_BSG)      += bsg.o
diff --git a/block/badblocks.c b/block/badblocks.c
new file mode 100644 (file)
index 0000000..96aeb91
--- /dev/null
@@ -0,0 +1,561 @@
+/*
+ * Bad block management
+ *
+ * - Heavily based on MD badblocks code from Neil Brown
+ *
+ * Copyright (c) 2015, Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope 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.
+ */
+
+#include <linux/badblocks.h>
+#include <linux/seqlock.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/stddef.h>
+#include <linux/types.h>
+#include <linux/slab.h>
+
+/**
+ * badblocks_check() - check a given range for bad sectors
+ * @bb:                the badblocks structure that holds all badblock information
+ * @s:         sector (start) at which to check for badblocks
+ * @sectors:   number of sectors to check for badblocks
+ * @first_bad: pointer to store location of the first badblock
+ * @bad_sectors: pointer to store number of badblocks after @first_bad
+ *
+ * We can record which blocks on each device are 'bad' and so just
+ * fail those blocks, or that stripe, rather than the whole device.
+ * Entries in the bad-block table are 64bits wide.  This comprises:
+ * Length of bad-range, in sectors: 0-511 for lengths 1-512
+ * Start of bad-range, sector offset, 54 bits (allows 8 exbibytes)
+ *  A 'shift' can be set so that larger blocks are tracked and
+ *  consequently larger devices can be covered.
+ * 'Acknowledged' flag - 1 bit. - the most significant bit.
+ *
+ * Locking of the bad-block table uses a seqlock so badblocks_check
+ * might need to retry if it is very unlucky.
+ * We will sometimes want to check for bad blocks in a bi_end_io function,
+ * so we use the write_seqlock_irq variant.
+ *
+ * When looking for a bad block we specify a range and want to
+ * know if any block in the range is bad.  So we binary-search
+ * to the last range that starts at-or-before the given endpoint,
+ * (or "before the sector after the target range")
+ * then see if it ends after the given start.
+ *
+ * Return:
+ *  0: there are no known bad blocks in the range
+ *  1: there are known bad block which are all acknowledged
+ * -1: there are bad blocks which have not yet been acknowledged in metadata.
+ * plus the start/length of the first bad section we overlap.
+ */
+int badblocks_check(struct badblocks *bb, sector_t s, int sectors,
+                       sector_t *first_bad, int *bad_sectors)
+{
+       int hi;
+       int lo;
+       u64 *p = bb->page;
+       int rv;
+       sector_t target = s + sectors;
+       unsigned seq;
+
+       if (bb->shift > 0) {
+               /* round the start down, and the end up */
+               s >>= bb->shift;
+               target += (1<<bb->shift) - 1;
+               target >>= bb->shift;
+               sectors = target - s;
+       }
+       /* 'target' is now the first block after the bad range */
+
+retry:
+       seq = read_seqbegin(&bb->lock);
+       lo = 0;
+       rv = 0;
+       hi = bb->count;
+
+       /* Binary search between lo and hi for 'target'
+        * i.e. for the last range that starts before 'target'
+        */
+       /* INVARIANT: ranges before 'lo' and at-or-after 'hi'
+        * are known not to be the last range before target.
+        * VARIANT: hi-lo is the number of possible
+        * ranges, and decreases until it reaches 1
+        */
+       while (hi - lo > 1) {
+               int mid = (lo + hi) / 2;
+               sector_t a = BB_OFFSET(p[mid]);
+
+               if (a < target)
+                       /* This could still be the one, earlier ranges
+                        * could not.
+                        */
+                       lo = mid;
+               else
+                       /* This and later ranges are definitely out. */
+                       hi = mid;
+       }
+       /* 'lo' might be the last that started before target, but 'hi' isn't */
+       if (hi > lo) {
+               /* need to check all range that end after 's' to see if
+                * any are unacknowledged.
+                */
+               while (lo >= 0 &&
+                      BB_OFFSET(p[lo]) + BB_LEN(p[lo]) > s) {
+                       if (BB_OFFSET(p[lo]) < target) {
+                               /* starts before the end, and finishes after
+                                * the start, so they must overlap
+                                */
+                               if (rv != -1 && BB_ACK(p[lo]))
+                                       rv = 1;
+                               else
+                                       rv = -1;
+                               *first_bad = BB_OFFSET(p[lo]);
+                               *bad_sectors = BB_LEN(p[lo]);
+                       }
+                       lo--;
+               }
+       }
+
+       if (read_seqretry(&bb->lock, seq))
+               goto retry;
+
+       return rv;
+}
+EXPORT_SYMBOL_GPL(badblocks_check);
+
+/**
+ * badblocks_set() - Add a range of bad blocks to the table.
+ * @bb:                the badblocks structure that holds all badblock information
+ * @s:         first sector to mark as bad
+ * @sectors:   number of sectors to mark as bad
+ * @acknowledged: weather to mark the bad sectors as acknowledged
+ *
+ * This might extend the table, or might contract it if two adjacent ranges
+ * can be merged. We binary-search to find the 'insertion' point, then
+ * decide how best to handle it.
+ *
+ * Return:
+ *  0: success
+ *  1: failed to set badblocks (out of space)
+ */
+int badblocks_set(struct badblocks *bb, sector_t s, int sectors,
+                       int acknowledged)
+{
+       u64 *p;
+       int lo, hi;
+       int rv = 0;
+       unsigned long flags;
+
+       if (bb->shift < 0)
+               /* badblocks are disabled */
+               return 0;
+
+       if (bb->shift) {
+               /* round the start down, and the end up */
+               sector_t next = s + sectors;
+
+               s >>= bb->shift;
+               next += (1<<bb->shift) - 1;
+               next >>= bb->shift;
+               sectors = next - s;
+       }
+
+       write_seqlock_irqsave(&bb->lock, flags);
+
+       p = bb->page;
+       lo = 0;
+       hi = bb->count;
+       /* Find the last range that starts at-or-before 's' */
+       while (hi - lo > 1) {
+               int mid = (lo + hi) / 2;
+               sector_t a = BB_OFFSET(p[mid]);
+
+               if (a <= s)
+                       lo = mid;
+               else
+                       hi = mid;
+       }
+       if (hi > lo && BB_OFFSET(p[lo]) > s)
+               hi = lo;
+
+       if (hi > lo) {
+               /* we found a range that might merge with the start
+                * of our new range
+                */
+               sector_t a = BB_OFFSET(p[lo]);
+               sector_t e = a + BB_LEN(p[lo]);
+               int ack = BB_ACK(p[lo]);
+
+               if (e >= s) {
+                       /* Yes, we can merge with a previous range */
+                       if (s == a && s + sectors >= e)
+                               /* new range covers old */
+                               ack = acknowledged;
+                       else
+                               ack = ack && acknowledged;
+
+                       if (e < s + sectors)
+                               e = s + sectors;
+                       if (e - a <= BB_MAX_LEN) {
+                               p[lo] = BB_MAKE(a, e-a, ack);
+                               s = e;
+                       } else {
+                               /* does not all fit in one range,
+                                * make p[lo] maximal
+                                */
+                               if (BB_LEN(p[lo]) != BB_MAX_LEN)
+                                       p[lo] = BB_MAKE(a, BB_MAX_LEN, ack);
+                               s = a + BB_MAX_LEN;
+                       }
+                       sectors = e - s;
+               }
+       }
+       if (sectors && hi < bb->count) {
+               /* 'hi' points to the first range that starts after 's'.
+                * Maybe we can merge with the start of that range
+                */
+               sector_t a = BB_OFFSET(p[hi]);
+               sector_t e = a + BB_LEN(p[hi]);
+               int ack = BB_ACK(p[hi]);
+
+               if (a <= s + sectors) {
+                       /* merging is possible */
+                       if (e <= s + sectors) {
+                               /* full overlap */
+                               e = s + sectors;
+                               ack = acknowledged;
+                       } else
+                               ack = ack && acknowledged;
+
+                       a = s;
+                       if (e - a <= BB_MAX_LEN) {
+                               p[hi] = BB_MAKE(a, e-a, ack);
+                               s = e;
+                       } else {
+                               p[hi] = BB_MAKE(a, BB_MAX_LEN, ack);
+                               s = a + BB_MAX_LEN;
+                       }
+                       sectors = e - s;
+                       lo = hi;
+                       hi++;
+               }
+       }
+       if (sectors == 0 && hi < bb->count) {
+               /* we might be able to combine lo and hi */
+               /* Note: 's' is at the end of 'lo' */
+               sector_t a = BB_OFFSET(p[hi]);
+               int lolen = BB_LEN(p[lo]);
+               int hilen = BB_LEN(p[hi]);
+               int newlen = lolen + hilen - (s - a);
+
+               if (s >= a && newlen < BB_MAX_LEN) {
+                       /* yes, we can combine them */
+                       int ack = BB_ACK(p[lo]) && BB_ACK(p[hi]);
+
+                       p[lo] = BB_MAKE(BB_OFFSET(p[lo]), newlen, ack);
+                       memmove(p + hi, p + hi + 1,
+                               (bb->count - hi - 1) * 8);
+                       bb->count--;
+               }
+       }
+       while (sectors) {
+               /* didn't merge (it all).
+                * Need to add a range just before 'hi'
+                */
+               if (bb->count >= MAX_BADBLOCKS) {
+                       /* No room for more */
+                       rv = 1;
+                       break;
+               } else {
+                       int this_sectors = sectors;
+
+                       memmove(p + hi + 1, p + hi,
+                               (bb->count - hi) * 8);
+                       bb->count++;
+
+                       if (this_sectors > BB_MAX_LEN)
+                               this_sectors = BB_MAX_LEN;
+                       p[hi] = BB_MAKE(s, this_sectors, acknowledged);
+                       sectors -= this_sectors;
+                       s += this_sectors;
+               }
+       }
+
+       bb->changed = 1;
+       if (!acknowledged)
+               bb->unacked_exist = 1;
+       write_sequnlock_irqrestore(&bb->lock, flags);
+
+       return rv;
+}
+EXPORT_SYMBOL_GPL(badblocks_set);
+
+/**
+ * badblocks_clear() - Remove a range of bad blocks to the table.
+ * @bb:                the badblocks structure that holds all badblock information
+ * @s:         first sector to mark as bad
+ * @sectors:   number of sectors to mark as bad
+ *
+ * This may involve extending the table if we spilt a region,
+ * but it must not fail.  So if the table becomes full, we just
+ * drop the remove request.
+ *
+ * Return:
+ *  0: success
+ *  1: failed to clear badblocks
+ */
+int badblocks_clear(struct badblocks *bb, sector_t s, int sectors)
+{
+       u64 *p;
+       int lo, hi;
+       sector_t target = s + sectors;
+       int rv = 0;
+
+       if (bb->shift > 0) {
+               /* When clearing we round the start up and the end down.
+                * This should not matter as the shift should align with
+                * the block size and no rounding should ever be needed.
+                * However it is better the think a block is bad when it
+                * isn't than to think a block is not bad when it is.
+                */
+               s += (1<<bb->shift) - 1;
+               s >>= bb->shift;
+               target >>= bb->shift;
+               sectors = target - s;
+       }
+
+       write_seqlock_irq(&bb->lock);
+
+       p = bb->page;
+       lo = 0;
+       hi = bb->count;
+       /* Find the last range that starts before 'target' */
+       while (hi - lo > 1) {
+               int mid = (lo + hi) / 2;
+               sector_t a = BB_OFFSET(p[mid]);
+
+               if (a < target)
+                       lo = mid;
+               else
+                       hi = mid;
+       }
+       if (hi > lo) {
+               /* p[lo] is the last range that could overlap the
+                * current range.  Earlier ranges could also overlap,
+                * but only this one can overlap the end of the range.
+                */
+               if (BB_OFFSET(p[lo]) + BB_LEN(p[lo]) > target) {
+                       /* Partial overlap, leave the tail of this range */
+                       int ack = BB_ACK(p[lo]);
+                       sector_t a = BB_OFFSET(p[lo]);
+                       sector_t end = a + BB_LEN(p[lo]);
+
+                       if (a < s) {
+                               /* we need to split this range */
+                               if (bb->count >= MAX_BADBLOCKS) {
+                                       rv = -ENOSPC;
+                                       goto out;
+                               }
+                               memmove(p+lo+1, p+lo, (bb->count - lo) * 8);
+                               bb->count++;
+                               p[lo] = BB_MAKE(a, s-a, ack);
+                               lo++;
+                       }
+                       p[lo] = BB_MAKE(target, end - target, ack);
+                       /* there is no longer an overlap */
+                       hi = lo;
+                       lo--;
+               }
+               while (lo >= 0 &&
+                      BB_OFFSET(p[lo]) + BB_LEN(p[lo]) > s) {
+                       /* This range does overlap */
+                       if (BB_OFFSET(p[lo]) < s) {
+                               /* Keep the early parts of this range. */
+                               int ack = BB_ACK(p[lo]);
+                               sector_t start = BB_OFFSET(p[lo]);
+
+                               p[lo] = BB_MAKE(start, s - start, ack);
+                               /* now low doesn't overlap, so.. */
+                               break;
+                       }
+                       lo--;
+               }
+               /* 'lo' is strictly before, 'hi' is strictly after,
+                * anything between needs to be discarded
+                */
+               if (hi - lo > 1) {
+                       memmove(p+lo+1, p+hi, (bb->count - hi) * 8);
+                       bb->count -= (hi - lo - 1);
+               }
+       }
+
+       bb->changed = 1;
+out:
+       write_sequnlock_irq(&bb->lock);
+       return rv;
+}
+EXPORT_SYMBOL_GPL(badblocks_clear);
+
+/**
+ * ack_all_badblocks() - Acknowledge all bad blocks in a list.
+ * @bb:                the badblocks structure that holds all badblock information
+ *
+ * This only succeeds if ->changed is clear.  It is used by
+ * in-kernel metadata updates
+ */
+void ack_all_badblocks(struct badblocks *bb)
+{
+       if (bb->page == NULL || bb->changed)
+               /* no point even trying */
+               return;
+       write_seqlock_irq(&bb->lock);
+
+       if (bb->changed == 0 && bb->unacked_exist) {
+               u64 *p = bb->page;
+               int i;
+
+               for (i = 0; i < bb->count ; i++) {
+                       if (!BB_ACK(p[i])) {
+                               sector_t start = BB_OFFSET(p[i]);
+                               int len = BB_LEN(p[i]);
+
+                               p[i] = BB_MAKE(start, len, 1);
+                       }
+               }
+               bb->unacked_exist = 0;
+       }
+       write_sequnlock_irq(&bb->lock);
+}
+EXPORT_SYMBOL_GPL(ack_all_badblocks);
+
+/**
+ * badblocks_show() - sysfs access to bad-blocks list
+ * @bb:                the badblocks structure that holds all badblock information
+ * @page:      buffer received from sysfs
+ * @unack:     weather to show unacknowledged badblocks
+ *
+ * Return:
+ *  Length of returned data
+ */
+ssize_t badblocks_show(struct badblocks *bb, char *page, int unack)
+{
+       size_t len;
+       int i;
+       u64 *p = bb->page;
+       unsigned seq;
+
+       if (bb->shift < 0)
+               return 0;
+
+retry:
+       seq = read_seqbegin(&bb->lock);
+
+       len = 0;
+       i = 0;
+
+       while (len < PAGE_SIZE && i < bb->count) {
+               sector_t s = BB_OFFSET(p[i]);
+               unsigned int length = BB_LEN(p[i]);
+               int ack = BB_ACK(p[i]);
+
+               i++;
+
+               if (unack && ack)
+                       continue;
+
+               len += snprintf(page+len, PAGE_SIZE-len, "%llu %u\n",
+                               (unsigned long long)s << bb->shift,
+                               length << bb->shift);
+       }
+       if (unack && len == 0)
+               bb->unacked_exist = 0;
+
+       if (read_seqretry(&bb->lock, seq))
+               goto retry;
+
+       return len;
+}
+EXPORT_SYMBOL_GPL(badblocks_show);
+
+/**
+ * badblocks_store() - sysfs access to bad-blocks list
+ * @bb:                the badblocks structure that holds all badblock information
+ * @page:      buffer received from sysfs
+ * @len:       length of data received from sysfs
+ * @unack:     weather to show unacknowledged badblocks
+ *
+ * Return:
+ *  Length of the buffer processed or -ve error.
+ */
+ssize_t badblocks_store(struct badblocks *bb, const char *page, size_t len,
+                       int unack)
+{
+       unsigned long long sector;
+       int length;
+       char newline;
+
+       switch (sscanf(page, "%llu %d%c", &sector, &length, &newline)) {
+       case 3:
+               if (newline != '\n')
+                       return -EINVAL;
+       case 2:
+               if (length <= 0)
+                       return -EINVAL;
+               break;
+       default:
+               return -EINVAL;
+       }
+
+       if (badblocks_set(bb, sector, length, !unack))
+               return -ENOSPC;
+       else
+               return len;
+}
+EXPORT_SYMBOL_GPL(badblocks_store);
+
+/**
+ * badblocks_init() - initialize the badblocks structure
+ * @bb:                the badblocks structure that holds all badblock information
+ * @enable:    weather to enable badblocks accounting
+ *
+ * Return:
+ *  0: success
+ *  -ve errno: on error
+ */
+int badblocks_init(struct badblocks *bb, int enable)
+{
+       bb->count = 0;
+       if (enable)
+               bb->shift = 0;
+       else
+               bb->shift = -1;
+       bb->page = kzalloc(PAGE_SIZE, GFP_KERNEL);
+       if (bb->page == (u64 *)0) {
+               bb->shift = -1;
+               return -ENOMEM;
+       }
+       seqlock_init(&bb->lock);
+
+       return 0;
+}
+EXPORT_SYMBOL_GPL(badblocks_init);
+
+/**
+ * badblocks_free() - free the badblocks structure
+ * @bb:                the badblocks structure that holds all badblock information
+ */
+void badblocks_free(struct badblocks *bb)
+{
+       kfree(bb->page);
+       bb->page = NULL;
+}
+EXPORT_SYMBOL_GPL(badblocks_free);
diff --git a/include/linux/badblocks.h b/include/linux/badblocks.h
new file mode 100644 (file)
index 0000000..9293446
--- /dev/null
@@ -0,0 +1,53 @@
+#ifndef _LINUX_BADBLOCKS_H
+#define _LINUX_BADBLOCKS_H
+
+#include <linux/seqlock.h>
+#include <linux/kernel.h>
+#include <linux/stddef.h>
+#include <linux/types.h>
+
+#define BB_LEN_MASK    (0x00000000000001FFULL)
+#define BB_OFFSET_MASK (0x7FFFFFFFFFFFFE00ULL)
+#define BB_ACK_MASK    (0x8000000000000000ULL)
+#define BB_MAX_LEN     512
+#define BB_OFFSET(x)   (((x) & BB_OFFSET_MASK) >> 9)
+#define BB_LEN(x)      (((x) & BB_LEN_MASK) + 1)
+#define BB_ACK(x)      (!!((x) & BB_ACK_MASK))
+#define BB_MAKE(a, l, ack) (((a)<<9) | ((l)-1) | ((u64)(!!(ack)) << 63))
+
+/* Bad block numbers are stored sorted in a single page.
+ * 64bits is used for each block or extent.
+ * 54 bits are sector number, 9 bits are extent size,
+ * 1 bit is an 'acknowledged' flag.
+ */
+#define MAX_BADBLOCKS  (PAGE_SIZE/8)
+
+struct badblocks {
+       int count;              /* count of bad blocks */
+       int unacked_exist;      /* there probably are unacknowledged
+                                * bad blocks.  This is only cleared
+                                * when a read discovers none
+                                */
+       int shift;              /* shift from sectors to block size
+                                * a -ve shift means badblocks are
+                                * disabled.*/
+       u64 *page;              /* badblock list */
+       int changed;
+       seqlock_t lock;
+       sector_t sector;
+       sector_t size;          /* in sectors */
+};
+
+int badblocks_check(struct badblocks *bb, sector_t s, int sectors,
+                  sector_t *first_bad, int *bad_sectors);
+int badblocks_set(struct badblocks *bb, sector_t s, int sectors,
+                       int acknowledged);
+int badblocks_clear(struct badblocks *bb, sector_t s, int sectors);
+void ack_all_badblocks(struct badblocks *bb);
+ssize_t badblocks_show(struct badblocks *bb, char *page, int unack);
+ssize_t badblocks_store(struct badblocks *bb, const char *page, size_t len,
+                       int unack);
+int badblocks_init(struct badblocks *bb, int enable);
+void badblocks_free(struct badblocks *bb);
+
+#endif