wl.c

基于linux-2.6.28的mtd驱动

/*
 * Copyright (c) International Business Machines Corp., 2006
 *
 * 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
 *
 * Authors: Artem Bityutskiy (Битюцкий Артём), Thomas Gleixner
 */

/*
 * 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
 * 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
 * eraseblocks are those that were put by the 'ubi_wl_put_peb()' function.
 *
 * Physical eraseblocks returned by 'ubi_wl_get_peb()' have only erase counter
 * header. The rest of the physical eraseblock contains only %0xFF bytes.
 *
 * When physical eraseblocks are returned to the WL sub-system by means of the
 * 'ubi_wl_put_peb()' function, they are scheduled for erasure. The erasure is
 * done asynchronously in context of the per-UBI device background thread,
 * which is also managed by the WL sub-system.
 *
 * The wear-leveling is ensured by means of moving the contents of used
 * physical eraseblocks with low erase counter to free physical eraseblocks
 * with high erase counter.
 *
 * The 'ubi_wl_get_peb()' function accepts data type hints which help to pick
 * an "optimal" physical eraseblock. For example, when it is known that the
 * physical eraseblock will be "put" soon because it contains short-term data,
 * the WL sub-system may pick a free physical eraseblock with low erase
 * counter, and so forth.
 *
 * If the WL sub-system fails to erase a physical eraseblock, it marks it as
 * bad.
 *
 * This sub-system is also responsible for scrubbing. If a bit-flip is detected
 * in a physical eraseblock, it has to be moved. Technically this is the same
 * as moving it for wear-leveling reasons.
 *
 * 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.
 *
 * 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
 * enough for moderately large flashes and it is simple. In future, one may
 * re-work this sub-system and make it more scalable.
 *
 * At the moment this sub-system does not utilize the sequence number, which
 * was introduced relatively recently. But it would be wise to do this because
 * the sequence number of a logical eraseblock characterizes how old is it. For
 * example, when we move a PEB with low erase counter, and we need to pick the
 * 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>
#include <linux/crc32.h>
#include <linux/freezer.h>
#include <linux/kthread.h>
#include "ubi.h"

/* Number of physical eraseblocks reserved for wear-leveling purposes */
#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
 * erase counter.
 */
#define UBI_WL_THRESHOLD CONFIG_MTD_UBI_WL_THRESHOLD

/*
 * When a physical eraseblock is moved, the WL sub-system has to pick the target
 * physical eraseblock to move to. The simplest way would be just to pick the
 * one with the highest erase counter. But in certain workloads this could lead
 * to an unlimited wear of one or few physical eraseblock. Indeed, imagine a
 * situation when the picked physical eraseblock is constantly erased after the
 * data is written to it. So, we have a constant which limits the highest erase
 * counter of the free physical eraseblock to pick. Namely, the WL sub-system
 * does not pick eraseblocks with erase counter greater then the lowest erase
 * counter plus %WL_FREE_MAX_DIFF.
 */
#define WL_FREE_MAX_DIFF (2*UBI_WL_THRESHOLD)

/*
 * Maximum number of consecutive background thread failures which is enough to
 * switch to read-only mode.
 */
#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
 *
 * The @func pointer points to the worker function. If the @cancel argument is
 * not zero, the worker has to free the resources and exit immediately. The
 * worker has to return zero in case of success and a negative error code in
 * case of failure.
 */
struct ubi_work {
	struct list_head list;
	int (*func)(struct ubi_device *ubi, struct ubi_work *wrk, int cancel);
	/* The below fields are only relevant to erasure works */
	struct ubi_wl_entry *e;
	int torture;
};

#ifdef CONFIG_MTD_UBI_DEBUG_PARANOID
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);
#else
#define paranoid_check_ec(ubi, pnum, ec) 0
#define paranoid_check_in_wl_tree(e, root)
#endif

/**
 * wl_tree_add - add a wear-leveling entry to a WL RB-tree.
 * @e: the wear-leveling entry to add
 * @root: the root of the tree
 *
 * Note, we use (erase counter, physical eraseblock number) pairs as keys in
 * the @ubi->used and @ubi->free RB-trees.
 */
static void wl_tree_add(struct ubi_wl_entry *e, struct rb_root *root)
{
	struct rb_node **p, *parent = NULL;

	p = &root->rb_node;
	while (*p) {
		struct ubi_wl_entry *e1;

		parent = *p;
		e1 = rb_entry(parent, struct ubi_wl_entry, rb);

		if (e->ec < e1->ec)
			p = &(*p)->rb_left;
		else if (e->ec > e1->ec)
			p = &(*p)->rb_right;
		else {
			ubi_assert(e->pnum != e1->pnum);
			if (e->pnum < e1->pnum)
				p = &(*p)->rb_left;
			else
				p = &(*p)->rb_right;
		}
	}

	rb_link_node(&e->rb, parent, p);
	rb_insert_color(&e->rb, root);
}

/**
 * do_work - do one pending work.
 * @ubi: UBI device description object
 *
 * This function returns zero in case of success and a negative error code in
 * case of failure.
 */
static int do_work(struct ubi_device *ubi)
{
	int err;
	struct ubi_work *wrk;

	cond_resched();

	/*
	 * @ubi->work_sem is used to synchronize with the workers. Workers take
	 * it in read mode, so many of them may be doing works at a time. But
	 * the queue flush code has to be sure the whole queue of works is
	 * done, and it takes the mutex in write mode.
	 */
	down_read(&ubi->work_sem);
	spin_lock(&ubi->wl_lock);
	if (list_empty(&ubi->works)) {
		spin_unlock(&ubi->wl_lock);
		up_read(&ubi->work_sem);
		return 0;
	}

	wrk = list_entry(ubi->works.next, struct ubi_work, list);
	list_del(&wrk->list);
	ubi->works_count -= 1;
	ubi_assert(ubi->works_count >= 0);
	spin_unlock(&ubi->wl_lock);

	/*
	 * Call the worker function. Do not touch the work structure
	 * after this call as it will have been freed or reused by that
	 * time by the worker function.
	 */
	err = wrk->func(ubi, wrk, 0);
	if (err)
		ubi_err("work failed with error code %d", err);
	up_read(&ubi->work_sem);

	return err;
}

/**
 * produce_free_peb - produce a free physical eraseblock.
 * @ubi: UBI device description object
 *
 * This function tries to make a free PEB by means of synchronous execution of
 * pending works. This may be needed if, for example the background thread is
 * disabled. Returns zero in case of success and a negative error code in case
 * of failure.
 */
static int produce_free_peb(struct ubi_device *ubi)
{
	int err;

	spin_lock(&ubi->wl_lock);
	while (!ubi->free.rb_node) {
		spin_unlock(&ubi->wl_lock);

		dbg_wl("do one work synchronously");
		err = do_work(ubi);
		if (err)
			return err;

		spin_lock(&ubi->wl_lock);
	}
	spin_unlock(&ubi->wl_lock);

	return 0;
}

/**
 * in_wl_tree - check if wear-leveling entry is present in a WL RB-tree.
 * @e: the wear-leveling entry to check
 * @root: the root of the tree
 *
 * This function returns non-zero if @e is in the @root RB-tree and zero if it
 * is not.
 */
static int in_wl_tree(struct ubi_wl_entry *e, struct rb_root *root)
{
	struct rb_node *p;

	p = root->rb_node;
	while (p) {
		struct ubi_wl_entry *e1;

		e1 = rb_entry(p, struct ubi_wl_entry, rb);

		if (e->pnum == e1->pnum) {
			ubi_assert(e == e1);
			return 1;
		}

		if (e->ec < e1->ec)
			p = p->rb_left;
		else if (e->ec > e1->ec)
			p = p->rb_right;
		else {
			ubi_assert(e->pnum != e1->pnum);
			if (e->pnum < e1->pnum)
				p = p->rb_left;
			else
				p = p->rb_right;
		}
	}

	return 0;
}

/**
 * prot_tree_add - add physical eraseblock to protection trees.
 * @ubi: UBI device description object
 * @e: the physical eraseblock to add
 * @pe: protection entry object to use
 * @abs_ec: absolute erase counter value when this physical eraseblock has
 * to be removed from the protection trees.
 *
 * @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 abs_ec)
{
	struct rb_node **p, *parent = NULL;
	struct ubi_wl_prot_entry *pe1;

	pe->e = e;
	pe->abs_ec = ubi->abs_ec + abs_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);
}

/**
 * find_wl_entry - find wear-leveling entry closest to certain erase counter.
 * @root: the RB-tree where to look for
 * @max: highest possible erase counter
 *
 * This function looks for a wear leveling entry with erase counter closest to
 * @max and less then @max.
 */
static struct ubi_wl_entry *find_wl_entry(struct rb_root *root, int max)
{
	struct rb_node *p;
	struct ubi_wl_entry *e;

	e = rb_entry(rb_first(root), struct ubi_wl_entry, rb);
	max += e->ec;

	p = root->rb_node;
	while (p) {
		struct ubi_wl_entry *e1;

		e1 = rb_entry(p, struct ubi_wl_entry, rb);
		if (e1->ec >= max)
			p = p->rb_left;
		else {
			p = p->rb_right;
			e = e1;
		}
	}

	return e;
}

/**
 * ubi_wl_get_peb - get a physical eraseblock.
 * @ubi: UBI device description object
 * @dtype: type of data which will be stored in this physical eraseblock
 *
 * This function returns a physical eraseblock in case of success and a
 * negative error code in case of failure. Might sleep.
 */
int ubi_wl_get_peb(struct ubi_device *ubi, int dtype)
{
	int err, protect, 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) {
		if (ubi->works_count == 0) {
			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);
			return err;
		}
		goto retry;
	}

	switch (dtype) {
	case UBI_LONGTERM:
		/*
		 * For long term data we pick a physical eraseblock with high
		 * erase counter. But the highest erase counter we can pick is
		 * bounded by the the lowest erase counter plus
		 * %WL_FREE_MAX_DIFF.
		 */
		e = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
		protect = LT_PROTECTION;
		break;
	case UBI_UNKNOWN:
		/*
		 * For unknown data we pick a physical eraseblock with medium
		 * erase counter. But we by no means can pick a physical
		 * eraseblock with erase counter greater or equivalent than the
		 * lowest erase counter plus %WL_FREE_MAX_DIFF.
		 */
		first = rb_entry(rb_first(&ubi->free), struct ubi_wl_entry, rb);
		last = rb_entry(rb_last(&ubi->free), struct ubi_wl_entry, rb);

		if (last->ec - first->ec < WL_FREE_MAX_DIFF)
			e = rb_entry(ubi->free.rb_node,
					struct ubi_wl_entry, rb);
		else {
			medium_ec = (first->ec + WL_FREE_MAX_DIFF)/2;
			e = find_wl_entry(&ubi->free, medium_ec);
		}
		protect = U_PROTECTION;
		break;
	case UBI_SHORTTERM:
		/*
		 * For short term data we pick a physical eraseblock with the
		 * lowest erase counter as we expect it will be erased soon.
		 */
		e = rb_entry(rb_first(&ubi->free), struct ubi_wl_entry, rb);
		protect = ST_PROTECTION;
		break;
	default:
		protect = 0;
		e = NULL;
		BUG();
	}

	/*
	 * Move the physical eraseblock to the protection trees where it will
	 * be protected from being moved for some time.
	 */
	paranoid_check_in_wl_tree(e, &ubi->free);
	rb_erase(&e->rb, &ubi->free);
	prot_tree_add(ubi, e, pe, protect);

	dbg_wl("PEB %d EC %d, protection %d", e->pnum, e->ec, protect);
	spin_unlock(&ubi->wl_lock);

	return e->pnum;
}

/**
 * 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 int prot_tree_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;