* UBI wear-leveling sub-system.
*
* This sub-system is responsible for wear-leveling. It works in terms of
- * physical* eraseblocks and erase counters and knows nothing about logical
+ * physical eraseblocks and erase counters and knows nothing about logical
* eraseblocks, volumes, etc. From this sub-system's perspective all physical
* eraseblocks are of two types - used and free. Used physical eraseblocks are
* those that were "get" by the 'ubi_wl_get_peb()' function, and free physical
*
* As it was said, for the UBI sub-system all physical eraseblocks are either
* "free" or "used". Free eraseblock are kept in the @wl->free RB-tree, while
- * used eraseblocks are kept in a set of different RB-trees: @wl->used,
- * @wl->prot.pnum, @wl->prot.aec, and @wl->scrub.
+ * used eraseblocks are kept in @wl->used or @wl->scrub RB-trees, or
+ * (temporarily) in the @wl->pq queue.
+ *
+ * When the WL sub-system returns a physical eraseblock, the physical
+ * eraseblock is protected from being moved for some "time". For this reason,
+ * the physical eraseblock is not directly moved from the @wl->free tree to the
+ * @wl->used tree. There is a protection queue in between where this
+ * physical eraseblock is temporarily stored (@wl->pq).
+ *
+ * All this protection stuff is needed because:
+ * o we don't want to move physical eraseblocks just after we have given them
+ * to the user; instead, we first want to let users fill them up with data;
+ *
+ * o there is a chance that the user will put the physical eraseblock very
+ * soon, so it makes sense not to move it for some time, but wait; this is
+ * especially important in case of "short term" physical eraseblocks.
+ *
+ * Physical eraseblocks stay protected only for limited time. But the "time" is
+ * measured in erase cycles in this case. This is implemented with help of the
+ * protection queue. Eraseblocks are put to the tail of this queue when they
+ * are returned by the 'ubi_wl_get_peb()', and eraseblocks are removed from the
+ * head of the queue on each erase operation (for any eraseblock). So the
+ * length of the queue defines how may (global) erase cycles PEBs are protected.
+ *
+ * To put it differently, each physical eraseblock has 2 main states: free and
+ * used. The former state corresponds to the @wl->free tree. The latter state
+ * is split up on several sub-states:
+ * o the WL movement is allowed (@wl->used tree);
+ * o the WL movement is temporarily prohibited (@wl->pq queue);
+ * o scrubbing is needed (@wl->scrub tree).
+ *
+ * Depending on the sub-state, wear-leveling entries of the used physical
+ * eraseblocks may be kept in one of those structures.
*
* Note, in this implementation, we keep a small in-RAM object for each physical
* eraseblock. This is surely not a scalable solution. But it appears to be good
* target PEB, we pick a PEB with the highest EC if our PEB is "old" and we
* pick target PEB with an average EC if our PEB is not very "old". This is a
* room for future re-works of the WL sub-system.
- *
- * Note: the stuff with protection trees looks too complex and is difficult to
- * understand. Should be fixed.
*/
#include <linux/slab.h>
#define WL_RESERVED_PEBS 1
/*
- * How many erase cycles are short term, unknown, and long term physical
- * eraseblocks protected.
- */
-#define ST_PROTECTION 16
-#define U_PROTECTION 10
-#define LT_PROTECTION 4
-
-/*
* Maximum difference between two erase counters. If this threshold is
* exceeded, the WL sub-system starts moving data from used physical
* eraseblocks with low erase counter to free physical eraseblocks with high
#define WL_MAX_FAILURES 32
/**
- * struct ubi_wl_prot_entry - PEB protection entry.
- * @rb_pnum: link in the @wl->prot.pnum RB-tree
- * @rb_aec: link in the @wl->prot.aec RB-tree
- * @abs_ec: the absolute erase counter value when the protection ends
- * @e: the wear-leveling entry of the physical eraseblock under protection
- *
- * When the WL sub-system returns a physical eraseblock, the physical
- * eraseblock is protected from being moved for some "time". For this reason,
- * the physical eraseblock is not directly moved from the @wl->free tree to the
- * @wl->used tree. There is one more tree in between where this physical
- * eraseblock is temporarily stored (@wl->prot).
- *
- * All this protection stuff is needed because:
- * o we don't want to move physical eraseblocks just after we have given them
- * to the user; instead, we first want to let users fill them up with data;
- *
- * o there is a chance that the user will put the physical eraseblock very
- * soon, so it makes sense not to move it for some time, but wait; this is
- * especially important in case of "short term" physical eraseblocks.
- *
- * Physical eraseblocks stay protected only for limited time. But the "time" is
- * measured in erase cycles in this case. This is implemented with help of the
- * absolute erase counter (@wl->abs_ec). When it reaches certain value, the
- * physical eraseblocks are moved from the protection trees (@wl->prot.*) to
- * the @wl->used tree.
- *
- * Protected physical eraseblocks are searched by physical eraseblock number
- * (when they are put) and by the absolute erase counter (to check if it is
- * time to move them to the @wl->used tree). So there are actually 2 RB-trees
- * storing the protected physical eraseblocks: @wl->prot.pnum and
- * @wl->prot.aec. They are referred to as the "protection" trees. The
- * first one is indexed by the physical eraseblock number. The second one is
- * indexed by the absolute erase counter. Both trees store
- * &struct ubi_wl_prot_entry objects.
- *
- * Each physical eraseblock has 2 main states: free and used. The former state
- * corresponds to the @wl->free tree. The latter state is split up on several
- * sub-states:
- * o the WL movement is allowed (@wl->used tree);
- * o the WL movement is temporarily prohibited (@wl->prot.pnum and
- * @wl->prot.aec trees);
- * o scrubbing is needed (@wl->scrub tree).
- *
- * Depending on the sub-state, wear-leveling entries of the used physical
- * eraseblocks may be kept in one of those trees.
- */
-struct ubi_wl_prot_entry {
- struct rb_node rb_pnum;
- struct rb_node rb_aec;
- unsigned long long abs_ec;
- struct ubi_wl_entry *e;
-};
-
-/**
* struct ubi_work - UBI work description data structure.
* @list: a link in the list of pending works
* @func: worker function
- * @priv: private data of the worker function
* @e: physical eraseblock to erase
* @torture: if the physical eraseblock has to be tortured
*
static int paranoid_check_ec(struct ubi_device *ubi, int pnum, int ec);
static int paranoid_check_in_wl_tree(struct ubi_wl_entry *e,
struct rb_root *root);
+static int paranoid_check_in_pq(struct ubi_device *ubi, struct ubi_wl_entry *e);
#else
#define paranoid_check_ec(ubi, pnum, ec) 0
#define paranoid_check_in_wl_tree(e, root)
+#define paranoid_check_in_pq(ubi, e) 0
#endif
/**
}
/**
- * prot_tree_add - add physical eraseblock to protection trees.
+ * prot_queue_add - add physical eraseblock to the protection queue.
* @ubi: UBI device description object
* @e: the physical eraseblock to add
- * @pe: protection entry object to use
- * @ec: for how many erase operations this PEB should be protected
*
- * @wl->lock has to be locked.
+ * This function adds @e to the tail of the protection queue @ubi->pq, where
+ * @e will stay for %UBI_PROT_QUEUE_LEN erase operations and will be
+ * temporarily protected from the wear-leveling worker. Note, @wl->lock has to
+ * be locked.
*/
-static void prot_tree_add(struct ubi_device *ubi, struct ubi_wl_entry *e,
- struct ubi_wl_prot_entry *pe, int ec)
+static void prot_queue_add(struct ubi_device *ubi, struct ubi_wl_entry *e)
{
- struct rb_node **p, *parent = NULL;
- struct ubi_wl_prot_entry *pe1;
+ int pq_tail = ubi->pq_head - 1;
- pe->e = e;
- pe->abs_ec = ubi->abs_ec + ec;
-
- p = &ubi->prot.pnum.rb_node;
- while (*p) {
- parent = *p;
- pe1 = rb_entry(parent, struct ubi_wl_prot_entry, rb_pnum);
-
- if (e->pnum < pe1->e->pnum)
- p = &(*p)->rb_left;
- else
- p = &(*p)->rb_right;
- }
- rb_link_node(&pe->rb_pnum, parent, p);
- rb_insert_color(&pe->rb_pnum, &ubi->prot.pnum);
-
- p = &ubi->prot.aec.rb_node;
- parent = NULL;
- while (*p) {
- parent = *p;
- pe1 = rb_entry(parent, struct ubi_wl_prot_entry, rb_aec);
-
- if (pe->abs_ec < pe1->abs_ec)
- p = &(*p)->rb_left;
- else
- p = &(*p)->rb_right;
- }
- rb_link_node(&pe->rb_aec, parent, p);
- rb_insert_color(&pe->rb_aec, &ubi->prot.aec);
+ if (pq_tail < 0)
+ pq_tail = UBI_PROT_QUEUE_LEN - 1;
+ ubi_assert(pq_tail >= 0 && pq_tail < UBI_PROT_QUEUE_LEN);
+ list_add_tail(&e->u.list, &ubi->pq[pq_tail]);
+ dbg_wl("added PEB %d EC %d to the protection queue", e->pnum, e->ec);
}
/**
*/
int ubi_wl_get_peb(struct ubi_device *ubi, int dtype)
{
- int err, protect, medium_ec;
+ int err, medium_ec;
struct ubi_wl_entry *e, *first, *last;
- struct ubi_wl_prot_entry *pe;
ubi_assert(dtype == UBI_LONGTERM || dtype == UBI_SHORTTERM ||
dtype == UBI_UNKNOWN);
- pe = kmalloc(sizeof(struct ubi_wl_prot_entry), GFP_NOFS);
- if (!pe)
- return -ENOMEM;
-
retry:
spin_lock(&ubi->wl_lock);
if (!ubi->free.rb_node) {
ubi_assert(list_empty(&ubi->works));
ubi_err("no free eraseblocks");
spin_unlock(&ubi->wl_lock);
- kfree(pe);
return -ENOSPC;
}
spin_unlock(&ubi->wl_lock);
err = produce_free_peb(ubi);
- if (err < 0) {
- kfree(pe);
+ if (err < 0)
return err;
- }
goto retry;
}
* %WL_FREE_MAX_DIFF.
*/
e = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
- protect = LT_PROTECTION;
break;
case UBI_UNKNOWN:
/*
medium_ec = (first->ec + WL_FREE_MAX_DIFF)/2;
e = find_wl_entry(&ubi->free, medium_ec);
}
- protect = U_PROTECTION;
break;
case UBI_SHORTTERM:
/*
* lowest erase counter as we expect it will be erased soon.
*/
e = rb_entry(rb_first(&ubi->free), struct ubi_wl_entry, u.rb);
- protect = ST_PROTECTION;
break;
default:
- protect = 0;
- e = NULL;
BUG();
}
+ paranoid_check_in_wl_tree(e, &ubi->free);
+
/*
- * Move the physical eraseblock to the protection trees where it will
+ * Move the physical eraseblock to the protection queue where it will
* be protected from being moved for some time.
*/
- paranoid_check_in_wl_tree(e, &ubi->free);
rb_erase(&e->u.rb, &ubi->free);
- prot_tree_add(ubi, e, pe, protect);
-
- dbg_wl("PEB %d EC %d, protection %d", e->pnum, e->ec, protect);
+ dbg_wl("PEB %d EC %d", e->pnum, e->ec);
+ prot_queue_add(ubi, e);
spin_unlock(&ubi->wl_lock);
-
return e->pnum;
}
/**
- * prot_tree_del - remove a physical eraseblock from the protection trees
+ * prot_queue_del - remove a physical eraseblock from the protection queue.
* @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.
+ * This function deletes PEB @pnum from the protection queue and returns zero
+ * in case of success and %-ENODEV if the PEB was not found.
*/
-static int prot_tree_del(struct ubi_device *ubi, int pnum)
+static int prot_queue_del(struct ubi_device *ubi, int pnum)
{
- struct rb_node *p;
- struct ubi_wl_prot_entry *pe = NULL;
-
- p = ubi->prot.pnum.rb_node;
- while (p) {
-
- pe = rb_entry(p, struct ubi_wl_prot_entry, rb_pnum);
-
- if (pnum == pe->e->pnum)
- goto found;
+ struct ubi_wl_entry *e;
- if (pnum < pe->e->pnum)
- p = p->rb_left;
- else
- p = p->rb_right;
- }
+ e = ubi->lookuptbl[pnum];
+ if (!e)
+ return -ENODEV;
- return -ENODEV;
+ if (paranoid_check_in_pq(ubi, e))
+ 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);
+ list_del(&e->u.list);
+ dbg_wl("deleted PEB %d from the protection queue", e->pnum);
return 0;
}
}
/**
- * check_protection_over - check if it is time to stop protecting some PEBs.
+ * serve_prot_queue - check if it is time to stop protecting PEBs.
* @ubi: UBI device description object
*
- * This function is called after each erase operation, when the absolute erase
- * counter is incremented, to check if some physical eraseblock have not to be
- * protected any longer. These physical eraseblocks are moved from the
- * protection trees to the used tree.
+ * This function is called after each erase operation and removes PEBs from the
+ * tail of the protection queue. These PEBs have been protected for long enough
+ * and should be moved to the used tree.
*/
-static void check_protection_over(struct ubi_device *ubi)
+static void serve_prot_queue(struct ubi_device *ubi)
{
- struct ubi_wl_prot_entry *pe;
+ struct ubi_wl_entry *e, *tmp;
+ int count;
/*
* There may be several protected physical eraseblock to remove,
* process them all.
*/
- while (1) {
- spin_lock(&ubi->wl_lock);
- if (!ubi->prot.aec.rb_node) {
- spin_unlock(&ubi->wl_lock);
- break;
- }
-
- pe = rb_entry(rb_first(&ubi->prot.aec),
- struct ubi_wl_prot_entry, rb_aec);
+repeat:
+ count = 0;
+ spin_lock(&ubi->wl_lock);
+ list_for_each_entry_safe(e, tmp, &ubi->pq[ubi->pq_head], u.list) {
+ dbg_wl("PEB %d EC %d protection over, move to used tree",
+ e->pnum, e->ec);
- if (pe->abs_ec > ubi->abs_ec) {
+ list_del(&e->u.list);
+ wl_tree_add(e, &ubi->used);
+ if (count++ > 32) {
+ /*
+ * Let's be nice and avoid holding the spinlock for
+ * too long.
+ */
spin_unlock(&ubi->wl_lock);
- break;
+ cond_resched();
+ goto repeat;
}
-
- dbg_wl("PEB %d protection over, abs_ec %llu, PEB abs_ec %llu",
- pe->e->pnum, ubi->abs_ec, pe->abs_ec);
- rb_erase(&pe->rb_aec, &ubi->prot.aec);
- rb_erase(&pe->rb_pnum, &ubi->prot.pnum);
- wl_tree_add(pe->e, &ubi->used);
- spin_unlock(&ubi->wl_lock);
-
- kfree(pe);
- cond_resched();
}
+
+ ubi->pq_head += 1;
+ if (ubi->pq_head == UBI_PROT_QUEUE_LEN)
+ ubi->pq_head = 0;
+ ubi_assert(ubi->pq_head >= 0 && ubi->pq_head < UBI_PROT_QUEUE_LEN);
+ spin_unlock(&ubi->wl_lock);
}
/**
* @ubi: UBI device description object
* @wrk: the work to schedule
*
- * This function enqueues a work defined by @wrk to the tail of the pending
- * works list.
+ * This function adds a work defined by @wrk to the tail of the pending works
+ * list.
*/
static void schedule_ubi_work(struct ubi_device *ubi, struct ubi_work *wrk)
{
int cancel)
{
int err, scrubbing = 0, torture = 0;
- struct ubi_wl_prot_entry *uninitialized_var(pe);
struct ubi_wl_entry *e1, *e2;
struct ubi_vid_hdr *vid_hdr;
* The LEB has not been moved because the volume is being
* deleted or the PEB has been put meanwhile. We should prevent
* this PEB from being selected for wear-leveling movement
- * again, so put it to the protection tree.
+ * again, so put it to the protection queue.
*/
dbg_wl("canceled moving PEB %d", e1->pnum);
ubi_assert(err == 1);
- pe = kmalloc(sizeof(struct ubi_wl_prot_entry), GFP_NOFS);
- if (!pe) {
- err = -ENOMEM;
- goto out_error;
- }
-
ubi_free_vid_hdr(ubi, vid_hdr);
vid_hdr = NULL;
spin_lock(&ubi->wl_lock);
- prot_tree_add(ubi, e1, pe, U_PROTECTION);
+ prot_queue_add(ubi, e1);
ubi_assert(!ubi->move_to_put);
ubi->move_from = ubi->move_to = NULL;
ubi->wl_scheduled = 0;
kfree(wl_wrk);
spin_lock(&ubi->wl_lock);
- ubi->abs_ec += 1;
wl_tree_add(e, &ubi->free);
spin_unlock(&ubi->wl_lock);
* One more erase operation has happened, take care about
* protected physical eraseblocks.
*/
- check_protection_over(ubi);
+ serve_prot_queue(ubi);
/* And take care about wear-leveling */
err = ensure_wear_leveling(ubi);
paranoid_check_in_wl_tree(e, &ubi->scrub);
rb_erase(&e->u.rb, &ubi->scrub);
} else {
- err = prot_tree_del(ubi, e->pnum);
+ err = prot_queue_del(ubi, e->pnum);
if (err) {
ubi_err("PEB %d not found", pnum);
ubi_ro_mode(ubi);
} else {
int err;
- err = prot_tree_del(ubi, e->pnum);
+ err = prot_queue_del(ubi, e->pnum);
if (err) {
ubi_err("PEB %d not found", pnum);
ubi_ro_mode(ubi);
int err;
/*
- * Erase while the pending works queue is not empty, but not more then
+ * Erase while the pending works queue is not empty, but not more than
* the number of currently pending works.
*/
dbg_wl("flush (%d pending works)", ubi->works_count);
*/
int ubi_wl_init_scan(struct ubi_device *ubi, struct ubi_scan_info *si)
{
- int err;
+ int err, i;
struct rb_node *rb1, *rb2;
struct ubi_scan_volume *sv;
struct ubi_scan_leb *seb, *tmp;
struct ubi_wl_entry *e;
-
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);
init_rwsem(&ubi->work_sem);
if (!ubi->lookuptbl)
return err;
+ for (i = 0; i < UBI_PROT_QUEUE_LEN; i++)
+ INIT_LIST_HEAD(&ubi->pq[i]);
+ ubi->pq_head = 0;
+
list_for_each_entry_safe(seb, tmp, &si->erase, u.list) {
cond_resched();
}
/**
- * protection_trees_destroy - destroy the protection RB-trees.
+ * protection_queue_destroy - destroy the protection queue.
* @ubi: UBI device description object
*/
-static void protection_trees_destroy(struct ubi_device *ubi)
+static void protection_queue_destroy(struct ubi_device *ubi)
{
- struct rb_node *rb;
- struct ubi_wl_prot_entry *pe;
+ int i;
+ struct ubi_wl_entry *e, *tmp;
- rb = ubi->prot.aec.rb_node;
- while (rb) {
- if (rb->rb_left)
- rb = rb->rb_left;
- else if (rb->rb_right)
- rb = rb->rb_right;
- else {
- pe = rb_entry(rb, struct ubi_wl_prot_entry, rb_aec);
-
- rb = rb_parent(rb);
- if (rb) {
- if (rb->rb_left == &pe->rb_aec)
- rb->rb_left = NULL;
- else
- rb->rb_right = NULL;
- }
-
- kmem_cache_free(ubi_wl_entry_slab, pe->e);
- kfree(pe);
+ for (i = 0; i < UBI_PROT_QUEUE_LEN; ++i) {
+ list_for_each_entry_safe(e, tmp, &ubi->pq[i], u.list) {
+ list_del(&e->u.list);
+ kmem_cache_free(ubi_wl_entry_slab, e);
}
}
}
{
dbg_wl("close the WL sub-system");
cancel_pending(ubi);
- protection_trees_destroy(ubi);
+ protection_queue_destroy(ubi);
tree_destroy(&ubi->used);
tree_destroy(&ubi->free);
tree_destroy(&ubi->scrub);
return 1;
}
+/**
+ * paranoid_check_in_pq - check if wear-leveling entry is in the protection
+ * queue.
+ * @ubi: UBI device description object
+ * @e: the wear-leveling entry to check
+ *
+ * This function returns zero if @e is in @ubi->pq and %1 if it is not.
+ */
+static int paranoid_check_in_pq(struct ubi_device *ubi, struct ubi_wl_entry *e)
+{
+ struct ubi_wl_entry *p;
+ int i;
+
+ for (i = 0; i < UBI_PROT_QUEUE_LEN; ++i)
+ list_for_each_entry(p, &ubi->pq[i], u.list)
+ if (p == e)
+ return 0;
+
+ ubi_err("paranoid check failed for PEB %d, EC %d, Protect queue",
+ e->pnum, e->ec);
+ ubi_dbg_dump_stack();
+ return 1;
+}
#endif /* CONFIG_MTD_UBI_DEBUG_PARANOID */
git://ftp.safe.ca / safe / jmp / linux-2.6 / commitdiff