#ifdef CONFIG_CGROUP_WRITEBACK
+/* parameters for foreign inode detection, see wb_detach_inode() */
+#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
+#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
+#define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
+#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
+
+#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
+#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
+ /* each slot's duration is 2s / 16 */
+#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
+ /* if foreign slots >= 8, switch */
+#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
+ /* one round can affect upto 5 slots */
+
void __inode_attach_wb(struct inode *inode, struct page *page)
{
struct backing_dev_info *bdi = inode_to_bdi(inode);
wb_put(wb);
}
+/**
+ * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
+ * @inode: inode of interest with i_lock held
+ *
+ * Returns @inode's wb with its list_lock held. @inode->i_lock must be
+ * held on entry and is released on return. The returned wb is guaranteed
+ * to stay @inode's associated wb until its list_lock is released.
+ */
+static struct bdi_writeback *
+locked_inode_to_wb_and_lock_list(struct inode *inode)
+ __releases(&inode->i_lock)
+ __acquires(&wb->list_lock)
+{
+ while (true) {
+ struct bdi_writeback *wb = inode_to_wb(inode);
+
+ /*
+ * inode_to_wb() association is protected by both
+ * @inode->i_lock and @wb->list_lock but list_lock nests
+ * outside i_lock. Drop i_lock and verify that the
+ * association hasn't changed after acquiring list_lock.
+ */
+ wb_get(wb);
+ spin_unlock(&inode->i_lock);
+ spin_lock(&wb->list_lock);
+ wb_put(wb); /* not gonna deref it anymore */
+
+ /* i_wb may have changed inbetween, can't use inode_to_wb() */
+ if (likely(wb == inode->i_wb))
+ return wb; /* @inode already has ref */
+
+ spin_unlock(&wb->list_lock);
+ cpu_relax();
+ spin_lock(&inode->i_lock);
+ }
+}
+
+/**
+ * inode_to_wb_and_lock_list - determine an inode's wb and lock it
+ * @inode: inode of interest
+ *
+ * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
+ * on entry.
+ */
+static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
+ __acquires(&wb->list_lock)
+{
+ spin_lock(&inode->i_lock);
+ return locked_inode_to_wb_and_lock_list(inode);
+}
+
+struct inode_switch_wbs_context {
+ struct inode *inode;
+ struct bdi_writeback *new_wb;
+
+ struct rcu_head rcu_head;
+ struct work_struct work;
+};
+
+static void inode_switch_wbs_work_fn(struct work_struct *work)
+{
+ struct inode_switch_wbs_context *isw =
+ container_of(work, struct inode_switch_wbs_context, work);
+ struct inode *inode = isw->inode;
+ struct address_space *mapping = inode->i_mapping;
+ struct bdi_writeback *old_wb = inode->i_wb;
+ struct bdi_writeback *new_wb = isw->new_wb;
+ struct radix_tree_iter iter;
+ bool switched = false;
+ void **slot;
+
+ /*
+ * By the time control reaches here, RCU grace period has passed
+ * since I_WB_SWITCH assertion and all wb stat update transactions
+ * between unlocked_inode_to_wb_begin/end() are guaranteed to be
+ * synchronizing against mapping->tree_lock.
+ *
+ * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
+ * gives us exclusion against all wb related operations on @inode
+ * including IO list manipulations and stat updates.
+ */
+ if (old_wb < new_wb) {
+ spin_lock(&old_wb->list_lock);
+ spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
+ } else {
+ spin_lock(&new_wb->list_lock);
+ spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
+ }
+ spin_lock(&inode->i_lock);
+ spin_lock_irq(&mapping->tree_lock);
+
+ /*
+ * Once I_FREEING is visible under i_lock, the eviction path owns
+ * the inode and we shouldn't modify ->i_wb_list.
+ */
+ if (unlikely(inode->i_state & I_FREEING))
+ goto skip_switch;
+
+ /*
+ * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
+ * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
+ * pages actually under underwriteback.
+ */
+ radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
+ PAGECACHE_TAG_DIRTY) {
+ struct page *page = radix_tree_deref_slot_protected(slot,
+ &mapping->tree_lock);
+ if (likely(page) && PageDirty(page)) {
+ __dec_wb_stat(old_wb, WB_RECLAIMABLE);
+ __inc_wb_stat(new_wb, WB_RECLAIMABLE);
+ }
+ }
+
+ radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
+ PAGECACHE_TAG_WRITEBACK) {
+ struct page *page = radix_tree_deref_slot_protected(slot,
+ &mapping->tree_lock);
+ if (likely(page)) {
+ WARN_ON_ONCE(!PageWriteback(page));
+ __dec_wb_stat(old_wb, WB_WRITEBACK);
+ __inc_wb_stat(new_wb, WB_WRITEBACK);
+ }
+ }
+
+ wb_get(new_wb);
+
+ /*
+ * Transfer to @new_wb's IO list if necessary. The specific list
+ * @inode was on is ignored and the inode is put on ->b_dirty which
+ * is always correct including from ->b_dirty_time. The transfer
+ * preserves @inode->dirtied_when ordering.
+ */
+ if (!list_empty(&inode->i_wb_list)) {
+ struct inode *pos;
+
+ inode_wb_list_del_locked(inode, old_wb);
+ inode->i_wb = new_wb;
+ list_for_each_entry(pos, &new_wb->b_dirty, i_wb_list)
+ if (time_after_eq(inode->dirtied_when,
+ pos->dirtied_when))
+ break;
+ inode_wb_list_move_locked(inode, new_wb, pos->i_wb_list.prev);
+ } else {
+ inode->i_wb = new_wb;
+ }
+
+ /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
+ inode->i_wb_frn_winner = 0;
+ inode->i_wb_frn_avg_time = 0;
+ inode->i_wb_frn_history = 0;
+ switched = true;
+skip_switch:
+ /*
+ * Paired with load_acquire in unlocked_inode_to_wb_begin() and
+ * ensures that the new wb is visible if they see !I_WB_SWITCH.
+ */
+ smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
+
+ spin_unlock_irq(&mapping->tree_lock);
+ spin_unlock(&inode->i_lock);
+ spin_unlock(&new_wb->list_lock);
+ spin_unlock(&old_wb->list_lock);
+
+ if (switched) {
+ wb_wakeup(new_wb);
+ wb_put(old_wb);
+ }
+ wb_put(new_wb);
+
+ iput(inode);
+ kfree(isw);
+}
+
+static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
+{
+ struct inode_switch_wbs_context *isw = container_of(rcu_head,
+ struct inode_switch_wbs_context, rcu_head);
+
+ /* needs to grab bh-unsafe locks, bounce to work item */
+ INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
+ schedule_work(&isw->work);
+}
+
+/**
+ * inode_switch_wbs - change the wb association of an inode
+ * @inode: target inode
+ * @new_wb_id: ID of the new wb
+ *
+ * Switch @inode's wb association to the wb identified by @new_wb_id. The
+ * switching is performed asynchronously and may fail silently.
+ */
+static void inode_switch_wbs(struct inode *inode, int new_wb_id)
+{
+ struct backing_dev_info *bdi = inode_to_bdi(inode);
+ struct cgroup_subsys_state *memcg_css;
+ struct inode_switch_wbs_context *isw;
+
+ /* noop if seems to be already in progress */
+ if (inode->i_state & I_WB_SWITCH)
+ return;
+
+ isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
+ if (!isw)
+ return;
+
+ /* find and pin the new wb */
+ rcu_read_lock();
+ memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
+ if (memcg_css)
+ isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
+ rcu_read_unlock();
+ if (!isw->new_wb)
+ goto out_free;
+
+ /* while holding I_WB_SWITCH, no one else can update the association */
+ spin_lock(&inode->i_lock);
+ if (inode->i_state & (I_WB_SWITCH | I_FREEING) ||
+ inode_to_wb(inode) == isw->new_wb) {
+ spin_unlock(&inode->i_lock);
+ goto out_free;
+ }
+ inode->i_state |= I_WB_SWITCH;
+ spin_unlock(&inode->i_lock);
+
+ ihold(inode);
+ isw->inode = inode;
+
+ /*
+ * In addition to synchronizing among switchers, I_WB_SWITCH tells
+ * the RCU protected stat update paths to grab the mapping's
+ * tree_lock so that stat transfer can synchronize against them.
+ * Let's continue after I_WB_SWITCH is guaranteed to be visible.
+ */
+ call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
+ return;
+
+out_free:
+ if (isw->new_wb)
+ wb_put(isw->new_wb);
+ kfree(isw);
+}
+
/**
* wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
* @wbc: writeback_control of interest
void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
struct inode *inode)
{
+ if (!inode_cgwb_enabled(inode)) {
+ spin_unlock(&inode->i_lock);
+ return;
+ }
+
wbc->wb = inode_to_wb(inode);
+ wbc->inode = inode;
+
+ wbc->wb_id = wbc->wb->memcg_css->id;
+ wbc->wb_lcand_id = inode->i_wb_frn_winner;
+ wbc->wb_tcand_id = 0;
+ wbc->wb_bytes = 0;
+ wbc->wb_lcand_bytes = 0;
+ wbc->wb_tcand_bytes = 0;
+
wb_get(wbc->wb);
spin_unlock(&inode->i_lock);
+
+ /*
+ * A dying wb indicates that the memcg-blkcg mapping has changed
+ * and a new wb is already serving the memcg. Switch immediately.
+ */
+ if (unlikely(wb_dying(wbc->wb)))
+ inode_switch_wbs(inode, wbc->wb_id);
}
/**
- * wbc_detach_inode - disassociate wbc from its target inode
- * @wbc: writeback_control of interest
+ * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
+ * @wbc: writeback_control of the just finished writeback
*
* To be called after a writeback attempt of an inode finishes and undoes
* wbc_attach_and_unlock_inode(). Can be called under any context.
+ *
+ * As concurrent write sharing of an inode is expected to be very rare and
+ * memcg only tracks page ownership on first-use basis severely confining
+ * the usefulness of such sharing, cgroup writeback tracks ownership
+ * per-inode. While the support for concurrent write sharing of an inode
+ * is deemed unnecessary, an inode being written to by different cgroups at
+ * different points in time is a lot more common, and, more importantly,
+ * charging only by first-use can too readily lead to grossly incorrect
+ * behaviors (single foreign page can lead to gigabytes of writeback to be
+ * incorrectly attributed).
+ *
+ * To resolve this issue, cgroup writeback detects the majority dirtier of
+ * an inode and transfers the ownership to it. To avoid unnnecessary
+ * oscillation, the detection mechanism keeps track of history and gives
+ * out the switch verdict only if the foreign usage pattern is stable over
+ * a certain amount of time and/or writeback attempts.
+ *
+ * On each writeback attempt, @wbc tries to detect the majority writer
+ * using Boyer-Moore majority vote algorithm. In addition to the byte
+ * count from the majority voting, it also counts the bytes written for the
+ * current wb and the last round's winner wb (max of last round's current
+ * wb, the winner from two rounds ago, and the last round's majority
+ * candidate). Keeping track of the historical winner helps the algorithm
+ * to semi-reliably detect the most active writer even when it's not the
+ * absolute majority.
+ *
+ * Once the winner of the round is determined, whether the winner is
+ * foreign or not and how much IO time the round consumed is recorded in
+ * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
+ * over a certain threshold, the switch verdict is given.
*/
void wbc_detach_inode(struct writeback_control *wbc)
{
+ struct bdi_writeback *wb = wbc->wb;
+ struct inode *inode = wbc->inode;
+ unsigned long avg_time, max_bytes, max_time;
+ u16 history;
+ int max_id;
+
+ if (!wb)
+ return;
+
+ history = inode->i_wb_frn_history;
+ avg_time = inode->i_wb_frn_avg_time;
+
+ /* pick the winner of this round */
+ if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
+ wbc->wb_bytes >= wbc->wb_tcand_bytes) {
+ max_id = wbc->wb_id;
+ max_bytes = wbc->wb_bytes;
+ } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
+ max_id = wbc->wb_lcand_id;
+ max_bytes = wbc->wb_lcand_bytes;
+ } else {
+ max_id = wbc->wb_tcand_id;
+ max_bytes = wbc->wb_tcand_bytes;
+ }
+
+ /*
+ * Calculate the amount of IO time the winner consumed and fold it
+ * into the running average kept per inode. If the consumed IO
+ * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
+ * deciding whether to switch or not. This is to prevent one-off
+ * small dirtiers from skewing the verdict.
+ */
+ max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
+ wb->avg_write_bandwidth);
+ if (avg_time)
+ avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
+ (avg_time >> WB_FRN_TIME_AVG_SHIFT);
+ else
+ avg_time = max_time; /* immediate catch up on first run */
+
+ if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
+ int slots;
+
+ /*
+ * The switch verdict is reached if foreign wb's consume
+ * more than a certain proportion of IO time in a
+ * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
+ * history mask where each bit represents one sixteenth of
+ * the period. Determine the number of slots to shift into
+ * history from @max_time.
+ */
+ slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
+ (unsigned long)WB_FRN_HIST_MAX_SLOTS);
+ history <<= slots;
+ if (wbc->wb_id != max_id)
+ history |= (1U << slots) - 1;
+
+ /*
+ * Switch if the current wb isn't the consistent winner.
+ * If there are multiple closely competing dirtiers, the
+ * inode may switch across them repeatedly over time, which
+ * is okay. The main goal is avoiding keeping an inode on
+ * the wrong wb for an extended period of time.
+ */
+ if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
+ inode_switch_wbs(inode, max_id);
+ }
+
+ /*
+ * Multiple instances of this function may race to update the
+ * following fields but we don't mind occassional inaccuracies.
+ */
+ inode->i_wb_frn_winner = max_id;
+ inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
+ inode->i_wb_frn_history = history;
+
wb_put(wbc->wb);
wbc->wb = NULL;
}
+/**
+ * wbc_account_io - account IO issued during writeback
+ * @wbc: writeback_control of the writeback in progress
+ * @page: page being written out
+ * @bytes: number of bytes being written out
+ *
+ * @bytes from @page are about to written out during the writeback
+ * controlled by @wbc. Keep the book for foreign inode detection. See
+ * wbc_detach_inode().
+ */
+void wbc_account_io(struct writeback_control *wbc, struct page *page,
+ size_t bytes)
+{
+ int id;
+
+ /*
+ * pageout() path doesn't attach @wbc to the inode being written
+ * out. This is intentional as we don't want the function to block
+ * behind a slow cgroup. Ultimately, we want pageout() to kick off
+ * regular writeback instead of writing things out itself.
+ */
+ if (!wbc->wb)
+ return;
+
+ rcu_read_lock();
+ id = mem_cgroup_css_from_page(page)->id;
+ rcu_read_unlock();
+
+ if (id == wbc->wb_id) {
+ wbc->wb_bytes += bytes;
+ return;
+ }
+
+ if (id == wbc->wb_lcand_id)
+ wbc->wb_lcand_bytes += bytes;
+
+ /* Boyer-Moore majority vote algorithm */
+ if (!wbc->wb_tcand_bytes)
+ wbc->wb_tcand_id = id;
+ if (id == wbc->wb_tcand_id)
+ wbc->wb_tcand_bytes += bytes;
+ else
+ wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
+}
+
/**
* inode_congested - test whether an inode is congested
* @inode: inode to test for congestion
*/
int inode_congested(struct inode *inode, int cong_bits)
{
- if (inode) {
- struct bdi_writeback *wb = inode_to_wb(inode);
- if (wb)
- return wb_congested(wb, cong_bits);
+ /*
+ * Once set, ->i_wb never becomes NULL while the inode is alive.
+ * Start transaction iff ->i_wb is visible.
+ */
+ if (inode && inode_to_wb_is_valid(inode)) {
+ struct bdi_writeback *wb;
+ bool locked, congested;
+
+ wb = unlocked_inode_to_wb_begin(inode, &locked);
+ congested = wb_congested(wb, cong_bits);
+ unlocked_inode_to_wb_end(inode, locked);
+ return congested;
}
return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
#else /* CONFIG_CGROUP_WRITEBACK */
+static struct bdi_writeback *
+locked_inode_to_wb_and_lock_list(struct inode *inode)
+ __releases(&inode->i_lock)
+ __acquires(&wb->list_lock)
+{
+ struct bdi_writeback *wb = inode_to_wb(inode);
+
+ spin_unlock(&inode->i_lock);
+ spin_lock(&wb->list_lock);
+ return wb;
+}
+
+static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
+ __acquires(&wb->list_lock)
+{
+ struct bdi_writeback *wb = inode_to_wb(inode);
+
+ spin_lock(&wb->list_lock);
+ return wb;
+}
+
static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
{
return nr_pages;
*/
void inode_wb_list_del(struct inode *inode)
{
- struct bdi_writeback *wb = inode_to_wb(inode);
+ struct bdi_writeback *wb;
- spin_lock(&wb->list_lock);
+ wb = inode_to_wb_and_lock_list(inode);
inode_wb_list_del_locked(inode, wb);
spin_unlock(&wb->list_lock);
}
* reposition it (that would break b_dirty time-ordering).
*/
if (!was_dirty) {
- struct bdi_writeback *wb = inode_to_wb(inode);
+ struct bdi_writeback *wb;
struct list_head *dirty_list;
bool wakeup_bdi = false;
- spin_unlock(&inode->i_lock);
- spin_lock(&wb->list_lock);
+ wb = locked_inode_to_wb_and_lock_list(inode);
WARN(bdi_cap_writeback_dirty(wb->bdi) &&
!test_bit(WB_registered, &wb->state),