return 0;
}
+/**
+ * leb_write_lock - lock logical eraseblock for writing.
+ * @ubi: UBI device description object
+ * @vol_id: volume ID
+ * @lnum: logical eraseblock number
+ *
+ * This function locks a logical eraseblock for writing if there is no
+ * contention and does nothing if there is contention. Returns %0 in case of
+ * success, %1 in case of contention, and and a negative error code in case of
+ * failure.
+ */
+static int leb_write_trylock(struct ubi_device *ubi, int vol_id, int lnum)
+{
+ int free;
+ struct ubi_ltree_entry *le;
+
+ le = ltree_add_entry(ubi, vol_id, lnum);
+ if (IS_ERR(le))
+ return PTR_ERR(le);
+ if (down_write_trylock(&le->mutex))
+ return 0;
+
+ /* Contention, cancel */
+ spin_lock(&ubi->ltree_lock);
+ le->users -= 1;
+ ubi_assert(le->users >= 0);
+ if (le->users == 0) {
+ rb_erase(&le->rb, &ubi->ltree);
+ free = 1;
+ } else
+ free = 0;
+ spin_unlock(&ubi->ltree_lock);
+ if (free)
+ kmem_cache_free(ubi_ltree_slab, le);
+
+ return 1;
+}
+
/**
* leb_write_unlock - unlock logical eraseblock.
* @ubi: UBI device description object
*
* This function copies logical eraseblock from physical eraseblock @from to
* physical eraseblock @to. The @vid_hdr buffer may be changed by this
- * function. Returns zero in case of success, %UBI_IO_BITFLIPS if the operation
- * was canceled because bit-flips were detected at the target PEB, and a
- * negative error code in case of failure.
+ * function. Returns:
+ * o %0 in case of success;
+ * o %1 if the operation was canceled and should be tried later (e.g.,
+ * because a bit-flip was detected at the target PEB);
+ * o %2 if the volume is being deleted and this LEB should not be moved.
*/
int ubi_eba_copy_leb(struct ubi_device *ubi, int from, int to,
struct ubi_vid_hdr *vid_hdr)
{
- int err, vol_id, lnum, data_size, aldata_size, pnum, idx;
+ int err, vol_id, lnum, data_size, aldata_size, idx;
struct ubi_volume *vol;
uint32_t crc;
data_size = aldata_size =
ubi->leb_size - be32_to_cpu(vid_hdr->data_pad);
- /*
- * We do not want anybody to write to this logical eraseblock while we
- * are moving it, so we lock it.
- */
- err = leb_write_lock(ubi, vol_id, lnum);
- if (err)
- return err;
-
- mutex_lock(&ubi->buf_mutex);
-
- /*
- * But the logical eraseblock might have been put by this time.
- * Cancel if it is true.
- */
idx = vol_id2idx(ubi, vol_id);
-
+ spin_lock(&ubi->volumes_lock);
/*
- * We may race with volume deletion/re-size, so we have to hold
- * @ubi->volumes_lock.
- *
- * Note, it is not a problem if we race with volume deletion or re-size
- * here. If the volume is deleted or re-sized while we are moving an
- * eraseblock which belongs to this volume, we'll end up with finding
- * out that this LEB was unmapped at the end (see WL), and drop this
- * PEB.
+ * Note, we may race with volume deletion, which means that the volume
+ * this logical eraseblock belongs to might be being deleted. Since the
+ * volume deletion unmaps all the volume's logical eraseblocks, it will
+ * be locked in 'ubi_wl_put_peb()' and wait for the WL worker to finish.
*/
- spin_lock(&ubi->volumes_lock);
vol = ubi->volumes[idx];
if (!vol) {
- dbg_eba("volume %d was removed meanwhile", vol_id);
+ /* No need to do further work, cancel */
+ dbg_eba("volume %d is being removed, cancel", vol_id);
spin_unlock(&ubi->volumes_lock);
- goto out_unlock;
+ return 2;
}
+ spin_unlock(&ubi->volumes_lock);
- pnum = vol->eba_tbl[lnum];
- if (pnum != from) {
- dbg_eba("LEB %d:%d is no longer mapped to PEB %d, mapped to "
- "PEB %d, cancel", vol_id, lnum, from, pnum);
- spin_unlock(&ubi->volumes_lock);
- goto out_unlock;
+ /*
+ * We do not want anybody to write to this logical eraseblock while we
+ * are moving it, so lock it.
+ *
+ * Note, we are using non-waiting locking here, because we cannot sleep
+ * on the LEB, since it may cause deadlocks. Indeed, imagine a task is
+ * unmapping the LEB which is mapped to the PEB we are going to move
+ * (@from). This task locks the LEB and goes sleep in the
+ * 'ubi_wl_put_peb()' function on the @ubi->move_mutex. In turn, we are
+ * holding @ubi->move_mutex and go sleep on the LEB lock. So, if the
+ * LEB is already locked, we just do not move it and return %1.
+ */
+ err = leb_write_trylock(ubi, vol_id, lnum);
+ if (err) {
+ dbg_eba("contention on LEB %d:%d, cancel", vol_id, lnum);
+ return err;
}
- spin_unlock(&ubi->volumes_lock);
- /* OK, now the LEB is locked and we can safely start moving it */
+ /*
+ * The LEB might have been put meanwhile, and the task which put it is
+ * probably waiting on @ubi->move_mutex. No need to continue the work,
+ * cancel it.
+ */
+ if (vol->eba_tbl[lnum] != from) {
+ dbg_eba("LEB %d:%d is no longer mapped to PEB %d, mapped to "
+ "PEB %d, cancel", vol_id, lnum, from,
+ vol->eba_tbl[lnum]);
+ err = 1;
+ goto out_unlock_leb;
+ }
+ /*
+ * OK, now the LEB is locked and we can safely start moving iy. Since
+ * this function utilizes thie @ubi->peb1_buf buffer which is shared
+ * with some other functions, so lock the buffer by taking the
+ * @ubi->buf_mutex.
+ */
+ mutex_lock(&ubi->buf_mutex);
dbg_eba("read %d bytes of data", aldata_size);
err = ubi_io_read_data(ubi, ubi->peb_buf1, from, 0, aldata_size);
if (err && err != UBI_IO_BITFLIPS) {
ubi_warn("error %d while reading data from PEB %d",
err, from);
- goto out_unlock;
+ goto out_unlock_buf;
}
/*
err = ubi_io_write_vid_hdr(ubi, to, vid_hdr);
if (err)
- goto out_unlock;
+ goto out_unlock_buf;
cond_resched();
if (err) {
if (err != UBI_IO_BITFLIPS)
ubi_warn("cannot read VID header back from PEB %d", to);
- goto out_unlock;
+ else
+ err = 1;
+ goto out_unlock_buf;
}
if (data_size > 0) {
err = ubi_io_write_data(ubi, ubi->peb_buf1, to, 0, aldata_size);
if (err)
- goto out_unlock;
+ goto out_unlock_buf;
cond_resched();
if (err != UBI_IO_BITFLIPS)
ubi_warn("cannot read data back from PEB %d",
to);
- goto out_unlock;
+ else
+ err = 1;
+ goto out_unlock_buf;
}
cond_resched();
if (memcmp(ubi->peb_buf1, ubi->peb_buf2, aldata_size)) {
ubi_warn("read data back from PEB %d - it is different",
to);
- goto out_unlock;
+ goto out_unlock_buf;
}
}
ubi_assert(vol->eba_tbl[lnum] == from);
vol->eba_tbl[lnum] = to;
-out_unlock:
+out_unlock_buf:
mutex_unlock(&ubi->buf_mutex);
+out_unlock_leb:
leb_write_unlock(ubi, vol_id, lnum);
return err;
}
int err;
struct ubi_work *wrk;
+ cond_resched();
+
spin_lock(&ubi->wl_lock);
if (list_empty(&ubi->works)) {
* prot_tree_del - remove a physical eraseblock from the protection trees
* @ubi: UBI device description object
* @pnum: the physical eraseblock to remove
+ *
+ * This function returns PEB @pnum from the protection trees and returns zero
+ * in case of success and %-ENODEV if the PEB was not found in the protection
+ * trees.
*/
-static void prot_tree_del(struct ubi_device *ubi, int pnum)
+static int prot_tree_del(struct ubi_device *ubi, int pnum)
{
struct rb_node *p;
struct ubi_wl_prot_entry *pe = NULL;
pe = rb_entry(p, struct ubi_wl_prot_entry, rb_pnum);
if (pnum == pe->e->pnum)
- break;
+ goto found;
if (pnum < pe->e->pnum)
p = p->rb_left;
p = p->rb_right;
}
+ return -ENODEV;
+
+found:
ubi_assert(pe->e->pnum == pnum);
rb_erase(&pe->rb_aec, &ubi->prot.aec);
rb_erase(&pe->rb_pnum, &ubi->prot.pnum);
kfree(pe);
+ return 0;
}
/**
static int wear_leveling_worker(struct ubi_device *ubi, struct ubi_work *wrk,
int cancel)
{
- int err, put = 0;
+ int err, put = 0, scrubbing = 0, protect = 0;
+ struct ubi_wl_prot_entry *pe;
struct ubi_wl_entry *e1, *e2;
struct ubi_vid_hdr *vid_hdr;
if (!vid_hdr)
return -ENOMEM;
+ mutex_lock(&ubi->move_mutex);
spin_lock(&ubi->wl_lock);
+ ubi_assert(!ubi->move_from && !ubi->move_to);
+ ubi_assert(!ubi->move_to_put);
- /*
- * Only one WL worker at a time is supported at this implementation, so
- * make sure a PEB is not being moved already.
- */
- if (ubi->move_to || !ubi->free.rb_node ||
+ if (!ubi->free.rb_node ||
(!ubi->used.rb_node && !ubi->scrub.rb_node)) {
/*
- * Only one WL worker at a time is supported at this
- * implementation, so if a LEB is already being moved, cancel.
- *
- * No free physical eraseblocks? Well, we cancel wear-leveling
- * then. It will be triggered again when a free physical
- * eraseblock appears.
+ * No free physical eraseblocks? Well, they must be waiting in
+ * the queue to be erased. Cancel movement - it will be
+ * triggered again when a free physical eraseblock appears.
*
* No used physical eraseblocks? They must be temporarily
* protected from being moved. They will be moved to the
*/
dbg_wl("cancel WL, a list is empty: free %d, used %d",
!ubi->free.rb_node, !ubi->used.rb_node);
- ubi->wl_scheduled = 0;
- spin_unlock(&ubi->wl_lock);
- ubi_free_vid_hdr(ubi, vid_hdr);
- return 0;
+ goto out_cancel;
}
if (!ubi->scrub.rb_node) {
if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD)) {
dbg_wl("no WL needed: min used EC %d, max free EC %d",
e1->ec, e2->ec);
- ubi->wl_scheduled = 0;
- spin_unlock(&ubi->wl_lock);
- ubi_free_vid_hdr(ubi, vid_hdr);
- return 0;
+ goto out_cancel;
}
paranoid_check_in_wl_tree(e1, &ubi->used);
rb_erase(&e1->rb, &ubi->used);
dbg_wl("move PEB %d EC %d to PEB %d EC %d",
e1->pnum, e1->ec, e2->pnum, e2->ec);
} else {
+ /* Perform scrubbing */
+ scrubbing = 1;
e1 = rb_entry(rb_first(&ubi->scrub), struct ubi_wl_entry, rb);
e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
paranoid_check_in_wl_tree(e1, &ubi->scrub);
paranoid_check_in_wl_tree(e2, &ubi->free);
rb_erase(&e2->rb, &ubi->free);
- ubi_assert(!ubi->move_from && !ubi->move_to);
- ubi_assert(!ubi->move_to_put && !ubi->move_from_put);
ubi->move_from = e1;
ubi->move_to = e2;
spin_unlock(&ubi->wl_lock);
* We so far do not know which logical eraseblock our physical
* eraseblock (@e1) belongs to. We have to read the volume identifier
* header first.
+ *
+ * Note, we are protected from this PEB being unmapped and erased. The
+ * 'ubi_wl_put_peb()' would wait for moving to be finished if the PEB
+ * which is being moved was unmapped.
*/
err = ubi_io_read_vid_hdr(ubi, e1->pnum, vid_hdr, 0);
* likely have the VID header in place.
*/
dbg_wl("PEB %d has no VID header", e1->pnum);
- err = 0;
- } else {
- ubi_err("error %d while reading VID header from PEB %d",
- err, e1->pnum);
- if (err > 0)
- err = -EIO;
+ goto out_not_moved;
}
- goto error;
+
+ ubi_err("error %d while reading VID header from PEB %d",
+ err, e1->pnum);
+ if (err > 0)
+ err = -EIO;
+ goto out_error;
}
err = ubi_eba_copy_leb(ubi, e1->pnum, e2->pnum, vid_hdr);
if (err) {
- if (err == UBI_IO_BITFLIPS)
- err = 0;
- goto error;
+
+ if (err < 0)
+ goto out_error;
+ if (err == 1)
+ goto out_not_moved;
+
+ /*
+ * For some reason the LEB was not moved - it might be because
+ * the volume is being deleted. We should prevent this PEB from
+ * being selected for wear-levelling movement for some "time",
+ * so put it to the protection tree.
+ */
+
+ dbg_wl("cancelled moving PEB %d", e1->pnum);
+ pe = kmalloc(sizeof(struct ubi_wl_prot_entry), GFP_NOFS);
+ if (!pe) {
+ err = -ENOMEM;
+ goto out_error;
+ }
+
+ protect = 1;
}
ubi_free_vid_hdr(ubi, vid_hdr);
spin_lock(&ubi->wl_lock);
+ if (protect)
+ prot_tree_add(ubi, e1, pe, protect);
if (!ubi->move_to_put)
wl_tree_add(e2, &ubi->used);
else
put = 1;
ubi->move_from = ubi->move_to = NULL;
- ubi->move_from_put = ubi->move_to_put = 0;
- ubi->wl_scheduled = 0;
+ ubi->move_to_put = ubi->wl_scheduled = 0;
spin_unlock(&ubi->wl_lock);
if (put) {
*/
dbg_wl("PEB %d was put meanwhile, erase", e2->pnum);
err = schedule_erase(ubi, e2, 0);
- if (err) {
- kmem_cache_free(ubi_wl_entry_slab, e2);
- ubi_ro_mode(ubi);
- }
+ if (err)
+ goto out_error;
}
- err = schedule_erase(ubi, e1, 0);
- if (err) {
- kmem_cache_free(ubi_wl_entry_slab, e1);
- ubi_ro_mode(ubi);
+ if (!protect) {
+ err = schedule_erase(ubi, e1, 0);
+ if (err)
+ goto out_error;
}
+
dbg_wl("done");
- return err;
+ mutex_unlock(&ubi->move_mutex);
+ return 0;
/*
- * Some error occurred. @e1 was not changed, so return it back. @e2
- * might be changed, schedule it for erasure.
+ * For some reasons the LEB was not moved, might be an error, might be
+ * something else. @e1 was not changed, so return it back. @e2 might
+ * be changed, schedule it for erasure.
*/
-error:
- if (err)
- dbg_wl("error %d occurred, cancel operation", err);
- ubi_assert(err <= 0);
-
+out_not_moved:
ubi_free_vid_hdr(ubi, vid_hdr);
spin_lock(&ubi->wl_lock);
- ubi->wl_scheduled = 0;
- if (ubi->move_from_put)
- put = 1;
+ if (scrubbing)
+ wl_tree_add(e1, &ubi->scrub);
else
wl_tree_add(e1, &ubi->used);
ubi->move_from = ubi->move_to = NULL;
- ubi->move_from_put = ubi->move_to_put = 0;
+ ubi->move_to_put = ubi->wl_scheduled = 0;
spin_unlock(&ubi->wl_lock);
- if (put) {
- /*
- * Well, the target PEB was put meanwhile, schedule it for
- * erasure.
- */
- dbg_wl("PEB %d was put meanwhile, erase", e1->pnum);
- err = schedule_erase(ubi, e1, 0);
- if (err) {
- kmem_cache_free(ubi_wl_entry_slab, e1);
- ubi_ro_mode(ubi);
- }
- }
-
err = schedule_erase(ubi, e2, 0);
- if (err) {
- kmem_cache_free(ubi_wl_entry_slab, e2);
- ubi_ro_mode(ubi);
- }
+ if (err)
+ goto out_error;
+
+ mutex_unlock(&ubi->move_mutex);
+ return 0;
+
+out_error:
+ ubi_err("error %d while moving PEB %d to PEB %d",
+ err, e1->pnum, e2->pnum);
- yield();
+ ubi_free_vid_hdr(ubi, vid_hdr);
+ spin_lock(&ubi->wl_lock);
+ ubi->move_from = ubi->move_to = NULL;
+ ubi->move_to_put = ubi->wl_scheduled = 0;
+ spin_unlock(&ubi->wl_lock);
+
+ kmem_cache_free(ubi_wl_entry_slab, e1);
+ kmem_cache_free(ubi_wl_entry_slab, e2);
+ ubi_ro_mode(ubi);
+
+ mutex_unlock(&ubi->move_mutex);
return err;
+
+out_cancel:
+ ubi->wl_scheduled = 0;
+ spin_unlock(&ubi->wl_lock);
+ mutex_unlock(&ubi->move_mutex);
+ ubi_free_vid_hdr(ubi, vid_hdr);
+ return 0;
}
/**
}
/**
- * ubi_wl_put_peb - return a physical eraseblock to the wear-leveling
- * unit.
+ * ubi_wl_put_peb - return a physical eraseblock to the wear-leveling unit.
* @ubi: UBI device description object
* @pnum: physical eraseblock to return
* @torture: if this physical eraseblock has to be tortured
* This function is called to return physical eraseblock @pnum to the pool of
* free physical eraseblocks. The @torture flag has to be set if an I/O error
* occurred to this @pnum and it has to be tested. This function returns zero
- * in case of success and a negative error code in case of failure.
+ * in case of success, and a negative error code in case of failure.
*/
int ubi_wl_put_peb(struct ubi_device *ubi, int pnum, int torture)
{
ubi_assert(pnum >= 0);
ubi_assert(pnum < ubi->peb_count);
+retry:
spin_lock(&ubi->wl_lock);
-
e = ubi->lookuptbl[pnum];
if (e == ubi->move_from) {
/*
* be moved. It will be scheduled for erasure in the
* wear-leveling worker.
*/
- dbg_wl("PEB %d is being moved", pnum);
- ubi_assert(!ubi->move_from_put);
- ubi->move_from_put = 1;
+ dbg_wl("PEB %d is being moved, wait", pnum);
spin_unlock(&ubi->wl_lock);
- return 0;
+
+ /* Wait for the WL worker by taking the @ubi->move_mutex */
+ mutex_lock(&ubi->move_mutex);
+ mutex_unlock(&ubi->move_mutex);
+ goto retry;
} else if (e == ubi->move_to) {
/*
* User is putting the physical eraseblock which was selected
} else if (in_wl_tree(e, &ubi->scrub)) {
paranoid_check_in_wl_tree(e, &ubi->scrub);
rb_erase(&e->rb, &ubi->scrub);
- } else
- prot_tree_del(ubi, e->pnum);
+ } else {
+ err = prot_tree_del(ubi, e->pnum);
+ if (err) {
+ ubi_err("PEB %d not found", pnum);
+ ubi_ro_mode(ubi);
+ spin_unlock(&ubi->wl_lock);
+ return err;
+ }
+ }
}
spin_unlock(&ubi->wl_lock);
if (in_wl_tree(e, &ubi->used)) {
paranoid_check_in_wl_tree(e, &ubi->used);
rb_erase(&e->rb, &ubi->used);
- } else
- prot_tree_del(ubi, pnum);
+ } else {
+ int err;
+
+ err = prot_tree_del(ubi, e->pnum);
+ if (err) {
+ ubi_err("PEB %d not found", pnum);
+ ubi_ro_mode(ubi);
+ spin_unlock(&ubi->wl_lock);
+ return err;
+ }
+ }
wl_tree_add(e, &ubi->scrub);
spin_unlock(&ubi->wl_lock);
ubi->used = ubi->free = ubi->scrub = RB_ROOT;
ubi->prot.pnum = ubi->prot.aec = RB_ROOT;
spin_lock_init(&ubi->wl_lock);
+ mutex_init(&ubi->move_mutex);
ubi->max_ec = si->max_ec;
INIT_LIST_HEAD(&ubi->works);
projects / linux-2.6 / commitdiff