]> www.infradead.org Git - users/hch/xfs.git/commitdiff
xfs: convert busy extent tracking to the generic group structure
authorChristoph Hellwig <hch@lst.de>
Fri, 20 Sep 2024 13:40:57 +0000 (15:40 +0200)
committerChristoph Hellwig <hch@lst.de>
Sun, 22 Sep 2024 05:37:32 +0000 (07:37 +0200)
Split busy extent tracking from struct xfs_perag into its own private
structure, which can be pointed to by the generic group structure.

Note that this structure is now dynamically allocated instead of embedded
as the upcoming zone XFS code doesn't need it and will also have an
unusually high number of groups due to hardware constraints.  Dynamically
allocating the structure this is a big memory saver for this case.

Signed-off-by: Christoph Hellwig <hch@lst.de>
12 files changed:
fs/xfs/libxfs/xfs_ag.c
fs/xfs/libxfs/xfs_ag.h
fs/xfs/libxfs/xfs_alloc.c
fs/xfs/libxfs/xfs_alloc_btree.c
fs/xfs/libxfs/xfs_group.c
fs/xfs/libxfs/xfs_group.h
fs/xfs/libxfs/xfs_rmap_btree.c
fs/xfs/scrub/alloc_repair.c
fs/xfs/scrub/reap.c
fs/xfs/xfs_discard.c
fs/xfs/xfs_extent_busy.c
fs/xfs/xfs_extent_busy.h

index 56bc30e53738696f3dc6dca58c4b132d5d5ffcc9..e36c740bcb99f5cf5b5a77dd0b148f023156f403 100644 (file)
@@ -207,11 +207,8 @@ xfs_perag_alloc(
 #ifdef __KERNEL__
        /* Place kernel structure only init below this point. */
        spin_lock_init(&pag->pag_ici_lock);
-       spin_lock_init(&pag->pagb_lock);
        INIT_DELAYED_WORK(&pag->pag_blockgc_work, xfs_blockgc_worker);
        INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
-       init_waitqueue_head(&pag->pagb_wait);
-       pag->pagb_tree = RB_ROOT;
 #endif /* __KERNEL__ */
 
        error = xfs_buf_cache_init(&pag->pag_bcache);
index a8c27906de98a2e1a7ecc5a063dd4c31725593dd..242346872351c16339f2d34a3aa844a523f061f2 100644 (file)
@@ -80,11 +80,6 @@ struct xfs_perag {
        uint8_t         pagf_repair_rmap_level;
 #endif
 
-       spinlock_t      pagb_lock;      /* lock for pagb_tree */
-       struct rb_root  pagb_tree;      /* ordered tree of busy extents */
-       unsigned int    pagb_gen;       /* generation count for pagb_tree */
-       wait_queue_head_t pagb_wait;    /* woken when pagb_gen changes */
-
        atomic_t        pagf_fstrms;    /* # of filestreams active in this AG */
 
        spinlock_t      pag_ici_lock;   /* incore inode cache lock */
index 66e38310a13129ae656ef3ca266badc30b36a1da..820ffa6ea6bd75f4e9e734c5a1162b18bc4259db 100644 (file)
@@ -331,7 +331,8 @@ xfs_alloc_compute_aligned(
        bool            busy;
 
        /* Trim busy sections out of found extent */
-       busy = xfs_extent_busy_trim(args, &bno, &len, busy_gen);
+       busy = xfs_extent_busy_trim(&args->pag->pag_group, args->minlen,
+                       args->maxlen, &bno, &len, busy_gen);
 
        /*
         * If we have a largish extent that happens to start before min_agbno,
@@ -1251,7 +1252,7 @@ xfs_alloc_ag_vextent_small(
        if (fbno == NULLAGBLOCK)
                goto out;
 
-       xfs_extent_busy_reuse(args->pag, fbno, 1,
+       xfs_extent_busy_reuse(&args->pag->pag_group, fbno, 1,
                              (args->datatype & XFS_ALLOC_NOBUSY));
 
        if (args->datatype & XFS_ALLOC_USERDATA) {
@@ -1364,7 +1365,8 @@ xfs_alloc_ag_vextent_exact(
         */
        tbno = fbno;
        tlen = flen;
-       xfs_extent_busy_trim(args, &tbno, &tlen, &busy_gen);
+       xfs_extent_busy_trim(&args->pag->pag_group, args->minlen, args->maxlen,
+                       &tbno, &tlen, &busy_gen);
 
        /*
         * Give up if the start of the extent is busy, or the freespace isn't
@@ -1757,8 +1759,9 @@ restart:
                         * the allocation can be retried.
                         */
                        trace_xfs_alloc_near_busy(args);
-                       error = xfs_extent_busy_flush(args->tp, args->pag,
-                                       acur.busy_gen, alloc_flags);
+                       error = xfs_extent_busy_flush(args->tp,
+                                       &args->pag->pag_group, acur.busy_gen,
+                                       alloc_flags);
                        if (error)
                                goto out;
 
@@ -1873,8 +1876,9 @@ restart:
                         * the allocation can be retried.
                         */
                        trace_xfs_alloc_size_busy(args);
-                       error = xfs_extent_busy_flush(args->tp, args->pag,
-                                       busy_gen, alloc_flags);
+                       error = xfs_extent_busy_flush(args->tp,
+                                       &args->pag->pag_group, busy_gen,
+                                       alloc_flags);
                        if (error)
                                goto error0;
 
@@ -1972,8 +1976,9 @@ restart:
                         * the allocation can be retried.
                         */
                        trace_xfs_alloc_size_busy(args);
-                       error = xfs_extent_busy_flush(args->tp, args->pag,
-                                       busy_gen, alloc_flags);
+                       error = xfs_extent_busy_flush(args->tp,
+                                       &args->pag->pag_group, busy_gen,
+                                       alloc_flags);
                        if (error)
                                goto error0;
 
@@ -3618,8 +3623,8 @@ xfs_alloc_vextent_finish(
                if (error)
                        goto out_drop_perag;
 
-               ASSERT(!xfs_extent_busy_search(args->pag, args->agbno,
-                               args->len));
+               ASSERT(!xfs_extent_busy_search(&args->pag->pag_group,
+                               args->agbno, args->len));
        }
 
        xfs_ag_resv_alloc_extent(args->pag, args->resv, args);
@@ -4017,7 +4022,7 @@ __xfs_free_extent(
 
        if (skip_discard)
                busy_flags |= XFS_EXTENT_BUSY_SKIP_DISCARD;
-       xfs_extent_busy_insert(tp, pag, agbno, len, busy_flags);
+       xfs_extent_busy_insert(tp, &pag->pag_group, agbno, len, busy_flags);
        return 0;
 
 err_release:
index 88e1545ed4c9dca0c2e9fa78f03343fc632adeec..843174a59036587e26cf6ce641f6d92e0ebe7875 100644 (file)
@@ -86,7 +86,7 @@ xfs_allocbt_alloc_block(
        }
 
        atomic64_inc(&cur->bc_mp->m_allocbt_blks);
-       xfs_extent_busy_reuse(cur->bc_ag.pag, bno, 1, false);
+       xfs_extent_busy_reuse(&cur->bc_ag.pag->pag_group, bno, 1, false);
 
        new->s = cpu_to_be32(bno);
 
@@ -110,7 +110,7 @@ xfs_allocbt_free_block(
                return error;
 
        atomic64_dec(&cur->bc_mp->m_allocbt_blks);
-       xfs_extent_busy_insert(cur->bc_tp, agbp->b_pag, bno, 1,
+       xfs_extent_busy_insert(cur->bc_tp, &agbp->b_pag->pag_group, bno, 1,
                              XFS_EXTENT_BUSY_SKIP_DISCARD);
        return 0;
 }
index 50e35d07a915d7362b72e1a8b85deacc14015706..ed7e3ee1d979ca098abe2732fe3c97a834ca5750 100644 (file)
@@ -10,6 +10,7 @@
 #include "xfs_mount.h"
 #include "xfs_error.h"
 #include "xfs_trace.h"
+#include "xfs_extent_busy.h"
 #include "xfs_group.h"
 
 /*
@@ -161,6 +162,7 @@ xfs_group_free(
        XFS_IS_CORRUPT(mp, atomic_read(&xg->xg_ref) != 0);
 
        xfs_defer_drain_free(&xg->xg_intents_drain);
+       kfree(xg->xg_busy_extents);
 
        if (uninit)
                uninit(xg);
@@ -185,6 +187,9 @@ xfs_group_insert(
        xg->xg_type = type;
 
 #ifdef __KERNEL__
+       xg->xg_busy_extents = xfs_extent_busy_alloc();
+       if (!xg->xg_busy_extents)
+               return -ENOMEM;
        spin_lock_init(&xg->xg_state_lock);
        xfs_hooks_init(&xg->xg_rmap_update_hooks);
 #endif
@@ -196,6 +201,7 @@ xfs_group_insert(
        error = xa_insert(&mp->m_groups[type].xa, index, xg, GFP_KERNEL);
        if (error) {
                WARN_ON_ONCE(error == -EBUSY);
+               kfree(xg->xg_busy_extents);
                xfs_defer_drain_free(&xg->xg_intents_drain);
                return error;
        }
index 485cc5ce99f03fb33d66aa1bae62c07d24d28d18..2f685b04e1d0c9346ce147c530a54fc3e6da2144 100644 (file)
@@ -15,6 +15,11 @@ struct xfs_group {
 #ifdef __KERNEL__
        /* -- kernel only structures below this line -- */
 
+       /*
+        * Track freed but not yet committed extents.
+        */
+       struct xfs_extent_busy_tree *xg_busy_extents;
+
        /*
         * Bitsets of per-ag metadata that have been checked and/or are sick.
         * Callers should hold xg_state_lock before accessing this field.
index 6fd460fc7c9c1df26db0649cf999a541fbbf2cf5..dca5e29296400bc4be89658915023227519fe52a 100644 (file)
@@ -102,7 +102,7 @@ xfs_rmapbt_alloc_block(
                return 0;
        }
 
-       xfs_extent_busy_reuse(pag, bno, 1, false);
+       xfs_extent_busy_reuse(&pag->pag_group, bno, 1, false);
 
        new->s = cpu_to_be32(bno);
        be32_add_cpu(&agf->agf_rmap_blocks, 1);
@@ -136,7 +136,7 @@ xfs_rmapbt_free_block(
        if (error)
                return error;
 
-       xfs_extent_busy_insert(cur->bc_tp, pag, bno, 1,
+       xfs_extent_busy_insert(cur->bc_tp, &pag->pag_group, bno, 1,
                              XFS_EXTENT_BUSY_SKIP_DISCARD);
 
        xfs_ag_resv_free_extent(pag, XFS_AG_RESV_RMAPBT, NULL, 1);
index f07cd93012c6750fb86f1ce581a355f896b69288..79a8adff21ffdd54a7e890740f5469b7f61b5c29 100644 (file)
@@ -132,13 +132,16 @@ int
 xrep_setup_ag_allocbt(
        struct xfs_scrub        *sc)
 {
+       struct xfs_group        *xg = &sc->sa.pag->pag_group;
+       unsigned int            busy_gen;
+
        /*
         * Make sure the busy extent list is clear because we can't put extents
         * on there twice.
         */
-       if (xfs_extent_busy_list_empty(sc->sa.pag, &busy_gen))
+       if (xfs_extent_busy_list_empty(xg, &busy_gen))
                return 0;
-       return xfs_extent_busy_flush(sc->tp, sc->sa.pag, busy_gen, 0);
+       return xfs_extent_busy_flush(sc->tp, xg, busy_gen, 0);
 }
 
 /* Check for any obvious conflicts in the free extent. */
@@ -866,7 +869,7 @@ xrep_allocbt(
         * on there twice.  In theory we cleared this before we started, but
         * let's not risk the filesystem.
         */
-       if (!xfs_extent_busy_list_empty(sc->sa.pag, &busy_gen)) {
+       if (!xfs_extent_busy_list_empty(&sc->sa.pag->pag_group, &busy_gen)) {
                error = -EDEADLOCK;
                goto out_ra;
        }
index d65ad6aa856f4d805451d0a88e618c7c651b2e3a..316cf70e3a8fd83ba87fafb12caeeabe60d0a3d1 100644 (file)
@@ -137,7 +137,7 @@ xreap_put_freelist(
                        agfl_bp, agbno, 0);
        if (error)
                return error;
-       xfs_extent_busy_insert(sc->tp, sc->sa.pag, agbno, 1,
+       xfs_extent_busy_insert(sc->tp, &sc->sa.pag->pag_group, agbno, 1,
                        XFS_EXTENT_BUSY_SKIP_DISCARD);
 
        return 0;
index 739ec69c44281cf6c68102bafb737f7ecee8c935..9c6d3b136826cc0ae593fb4b45db294f2931e9ad 100644 (file)
@@ -117,10 +117,12 @@ xfs_discard_extents(
 
        blk_start_plug(&plug);
        list_for_each_entry(busyp, &extents->extent_list, list) {
-               trace_xfs_discard_extent(busyp->pag, busyp->bno, busyp->length);
+               struct xfs_perag        *pag = to_perag(busyp->group);
+
+               trace_xfs_discard_extent(pag, busyp->bno, busyp->length);
 
                error = __blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,
-                               xfs_agbno_to_daddr(busyp->pag, busyp->bno),
+                               xfs_agbno_to_daddr(pag, busyp->bno),
                                XFS_FSB_TO_BB(mp, busyp->length),
                                GFP_KERNEL, &bio);
                if (error && error != -EOPNOTSUPP) {
@@ -271,12 +273,12 @@ xfs_trim_gather_extents(
                 * If any blocks in the range are still busy, skip the
                 * discard and try again the next time.
                 */
-               if (xfs_extent_busy_search(pag, fbno, flen)) {
+               if (xfs_extent_busy_search(&pag->pag_group, fbno, flen)) {
                        trace_xfs_discard_busy(pag, fbno, flen);
                        goto next_extent;
                }
 
-               xfs_extent_busy_insert_discard(pag, fbno, flen,
+               xfs_extent_busy_insert_discard(&pag->pag_group, fbno, flen,
                                &extents->extent_list);
 next_extent:
                if (tcur->by_bno)
index ff10307702f01165700e2351fbae4bb21809f293..d63208a87f1d7e9eb752caf829a6d56937958784 100644 (file)
 #include "xfs_log.h"
 #include "xfs_ag.h"
 
+struct xfs_extent_busy_tree {
+       spinlock_t              eb_lock;
+       struct rb_root          eb_tree;
+       unsigned int            eb_gen;
+       wait_queue_head_t       eb_wait;
+};
+
 static void
 xfs_extent_busy_insert_list(
-       struct xfs_perag        *pag,
+       struct xfs_group        *xg,
        xfs_agblock_t           bno,
        xfs_extlen_t            len,
        unsigned int            flags,
        struct list_head        *busy_list)
 {
+       struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
        struct xfs_extent_busy  *new;
        struct xfs_extent_busy  *busyp;
        struct rb_node          **rbp;
@@ -34,17 +42,17 @@ xfs_extent_busy_insert_list(
 
        new = kzalloc(sizeof(struct xfs_extent_busy),
                        GFP_KERNEL | __GFP_NOFAIL);
-       new->pag = xfs_perag_hold(pag);
+       new->group = xfs_group_hold(xg);
        new->bno = bno;
        new->length = len;
        INIT_LIST_HEAD(&new->list);
        new->flags = flags;
 
        /* trace before insert to be able to see failed inserts */
-       trace_xfs_extent_busy(&pag->pag_group, bno, len);
+       trace_xfs_extent_busy(xg, bno, len);
 
-       spin_lock(&pag->pagb_lock);
-       rbp = &pag->pagb_tree.rb_node;
+       spin_lock(&eb->eb_lock);
+       rbp = &eb->eb_tree.rb_node;
        while (*rbp) {
                parent = *rbp;
                busyp = rb_entry(parent, struct xfs_extent_busy, rb_node);
@@ -61,32 +69,32 @@ xfs_extent_busy_insert_list(
        }
 
        rb_link_node(&new->rb_node, parent, rbp);
-       rb_insert_color(&new->rb_node, &pag->pagb_tree);
+       rb_insert_color(&new->rb_node, &eb->eb_tree);
 
        /* always process discard lists in fifo order */
        list_add_tail(&new->list, busy_list);
-       spin_unlock(&pag->pagb_lock);
+       spin_unlock(&eb->eb_lock);
 }
 
 void
 xfs_extent_busy_insert(
        struct xfs_trans        *tp,
-       struct xfs_perag        *pag,
+       struct xfs_group        *xg,
        xfs_agblock_t           bno,
        xfs_extlen_t            len,
        unsigned int            flags)
 {
-       xfs_extent_busy_insert_list(pag, bno, len, flags, &tp->t_busy);
+       xfs_extent_busy_insert_list(xg, bno, len, flags, &tp->t_busy);
 }
 
 void
 xfs_extent_busy_insert_discard(
-       struct xfs_perag        *pag,
+       struct xfs_group        *xg,
        xfs_agblock_t           bno,
        xfs_extlen_t            len,
        struct list_head        *busy_list)
 {
-       xfs_extent_busy_insert_list(pag, bno, len, XFS_EXTENT_BUSY_DISCARDED,
+       xfs_extent_busy_insert_list(xg, bno, len, XFS_EXTENT_BUSY_DISCARDED,
                        busy_list);
 }
 
@@ -101,17 +109,18 @@ xfs_extent_busy_insert_discard(
  */
 int
 xfs_extent_busy_search(
-       struct xfs_perag        *pag,
+       struct xfs_group        *xg,
        xfs_agblock_t           bno,
        xfs_extlen_t            len)
 {
+       struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
        struct rb_node          *rbp;
        struct xfs_extent_busy  *busyp;
        int                     match = 0;
 
        /* find closest start bno overlap */
-       spin_lock(&pag->pagb_lock);
-       rbp = pag->pagb_tree.rb_node;
+       spin_lock(&eb->eb_lock);
+       rbp = eb->eb_tree.rb_node;
        while (rbp) {
                busyp = rb_entry(rbp, struct xfs_extent_busy, rb_node);
                if (bno < busyp->bno) {
@@ -130,7 +139,7 @@ xfs_extent_busy_search(
                        break;
                }
        }
-       spin_unlock(&pag->pagb_lock);
+       spin_unlock(&eb->eb_lock);
        return match;
 }
 
@@ -147,13 +156,15 @@ xfs_extent_busy_search(
  */
 STATIC bool
 xfs_extent_busy_update_extent(
-       struct xfs_perag        *pag,
+       struct xfs_group        *xg,
        struct xfs_extent_busy  *busyp,
        xfs_agblock_t           fbno,
        xfs_extlen_t            flen,
-       bool                    userdata) __releases(&pag->pagb_lock)
-                                         __acquires(&pag->pagb_lock)
+       bool                    userdata)
+               __releases(&eb->eb_lock)
+               __acquires(&eb->eb_lock)
 {
+       struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
        xfs_agblock_t           fend = fbno + flen;
        xfs_agblock_t           bbno = busyp->bno;
        xfs_agblock_t           bend = bbno + busyp->length;
@@ -164,9 +175,9 @@ xfs_extent_busy_update_extent(
         * and retry.
         */
        if (busyp->flags & XFS_EXTENT_BUSY_DISCARDED) {
-               spin_unlock(&pag->pagb_lock);
+               spin_unlock(&eb->eb_lock);
                delay(1);
-               spin_lock(&pag->pagb_lock);
+               spin_lock(&eb->eb_lock);
                return false;
        }
 
@@ -239,7 +250,7 @@ xfs_extent_busy_update_extent(
                 * tree root, because erasing the node can rearrange the
                 * tree topology.
                 */
-               rb_erase(&busyp->rb_node, &pag->pagb_tree);
+               rb_erase(&busyp->rb_node, &eb->eb_tree);
                busyp->length = 0;
                return false;
        } else if (fend < bend) {
@@ -278,14 +289,14 @@ xfs_extent_busy_update_extent(
                ASSERT(0);
        }
 
-       trace_xfs_extent_busy_reuse(&pag->pag_group, fbno, flen);
+       trace_xfs_extent_busy_reuse(xg, fbno, flen);
        return true;
 
 out_force_log:
-       spin_unlock(&pag->pagb_lock);
-       xfs_log_force(pag_mount(pag), XFS_LOG_SYNC);
-       trace_xfs_extent_busy_force(&pag->pag_group, fbno, flen);
-       spin_lock(&pag->pagb_lock);
+       spin_unlock(&eb->eb_lock);
+       xfs_log_force(xg->xg_mount, XFS_LOG_SYNC);
+       trace_xfs_extent_busy_force(xg, fbno, flen);
+       spin_lock(&eb->eb_lock);
        return false;
 }
 
@@ -294,17 +305,18 @@ out_force_log:
  */
 void
 xfs_extent_busy_reuse(
-       struct xfs_perag        *pag,
+       struct xfs_group        *xg,
        xfs_agblock_t           fbno,
        xfs_extlen_t            flen,
        bool                    userdata)
 {
+       struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
        struct rb_node          *rbp;
 
        ASSERT(flen > 0);
-       spin_lock(&pag->pagb_lock);
+       spin_lock(&eb->eb_lock);
 restart:
-       rbp = pag->pagb_tree.rb_node;
+       rbp = eb->eb_tree.rb_node;
        while (rbp) {
                struct xfs_extent_busy *busyp =
                        rb_entry(rbp, struct xfs_extent_busy, rb_node);
@@ -319,11 +331,11 @@ restart:
                        continue;
                }
 
-               if (!xfs_extent_busy_update_extent(pag, busyp, fbno, flen,
+               if (!xfs_extent_busy_update_extent(xg, busyp, fbno, flen,
                                                  userdata))
                        goto restart;
        }
-       spin_unlock(&pag->pagb_lock);
+       spin_unlock(&eb->eb_lock);
 }
 
 /*
@@ -332,7 +344,7 @@ restart:
  * args->minlen no suitable extent could be found, and the higher level
  * code needs to force out the log and retry the allocation.
  *
- * Return the current busy generation for the AG if the extent is busy. This
+ * Return the current busy generation for the group if the extent is busy. This
  * value can be used to wait for at least one of the currently busy extents
  * to be cleared. Note that the busy list is not guaranteed to be empty after
  * the gen is woken. The state of a specific extent must always be confirmed
@@ -340,11 +352,14 @@ restart:
  */
 bool
 xfs_extent_busy_trim(
-       struct xfs_alloc_arg    *args,
+       struct xfs_group        *xg,
+       xfs_extlen_t            minlen,
+       xfs_extlen_t            maxlen,
        xfs_agblock_t           *bno,
        xfs_extlen_t            *len,
        unsigned                *busy_gen)
 {
+       struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
        xfs_agblock_t           fbno;
        xfs_extlen_t            flen;
        struct rb_node          *rbp;
@@ -352,11 +367,11 @@ xfs_extent_busy_trim(
 
        ASSERT(*len > 0);
 
-       spin_lock(&args->pag->pagb_lock);
+       spin_lock(&eb->eb_lock);
        fbno = *bno;
        flen = *len;
-       rbp = args->pag->pagb_tree.rb_node;
-       while (rbp && flen >= args->minlen) {
+       rbp = eb->eb_tree.rb_node;
+       while (rbp && flen >= minlen) {
                struct xfs_extent_busy *busyp =
                        rb_entry(rbp, struct xfs_extent_busy, rb_node);
                xfs_agblock_t   fend = fbno + flen;
@@ -477,13 +492,13 @@ xfs_extent_busy_trim(
                         * good chance subsequent allocations will be
                         * contiguous.
                         */
-                       if (bbno - fbno >= args->maxlen) {
+                       if (bbno - fbno >= maxlen) {
                                /* left candidate fits perfect */
                                fend = bbno;
-                       } else if (fend - bend >= args->maxlen * 4) {
+                       } else if (fend - bend >= maxlen * 4) {
                                /* right candidate has enough free space */
                                fbno = bend;
-                       } else if (bbno - fbno >= args->minlen) {
+                       } else if (bbno - fbno >= minlen) {
                                /* left candidate fits minimum requirement */
                                fend = bbno;
                        } else {
@@ -496,14 +511,13 @@ xfs_extent_busy_trim(
 out:
 
        if (fbno != *bno || flen != *len) {
-               trace_xfs_extent_busy_trim(&args->pag->pag_group, *bno, *len,
-                               fbno, flen);
+               trace_xfs_extent_busy_trim(xg, *bno, *len, fbno, flen);
                *bno = fbno;
                *len = flen;
-               *busy_gen = args->pag->pagb_gen;
+               *busy_gen = eb->eb_gen;
                ret = true;
        }
-       spin_unlock(&args->pag->pagb_lock);
+       spin_unlock(&eb->eb_lock);
        return ret;
 fail:
        /*
@@ -516,23 +530,24 @@ fail:
 
 static bool
 xfs_extent_busy_clear_one(
-       struct xfs_perag        *pag,
        struct xfs_extent_busy  *busyp,
        bool                    do_discard)
 {
+       struct xfs_extent_busy_tree *eb = busyp->group->xg_busy_extents;
+
        if (busyp->length) {
                if (do_discard &&
                    !(busyp->flags & XFS_EXTENT_BUSY_SKIP_DISCARD)) {
                        busyp->flags = XFS_EXTENT_BUSY_DISCARDED;
                        return false;
                }
-               trace_xfs_extent_busy_clear(&pag->pag_group, busyp->bno,
+               trace_xfs_extent_busy_clear(busyp->group, busyp->bno,
                                busyp->length);
-               rb_erase(&busyp->rb_node, &pag->pagb_tree);
+               rb_erase(&busyp->rb_node, &eb->eb_tree);
        }
 
        list_del_init(&busyp->list);
-       xfs_perag_put(busyp->pag);
+       xfs_group_put(busyp->group);
        kfree(busyp);
        return true;
 }
@@ -554,29 +569,30 @@ xfs_extent_busy_clear(
                return;
 
        do {
-               struct xfs_perag        *pag = xfs_perag_hold(busyp->pag);
+               struct xfs_group        *xg = xfs_group_hold(busyp->group);
+               struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
                bool                    wakeup = false;
 
-               spin_lock(&pag->pagb_lock);
+               spin_lock(&eb->eb_lock);
                do {
                        next = list_next_entry(busyp, list);
-                       if (xfs_extent_busy_clear_one(pag, busyp, do_discard))
+                       if (xfs_extent_busy_clear_one(busyp, do_discard))
                                wakeup = true;
                        busyp = next;
                } while (!list_entry_is_head(busyp, list, list) &&
-                        busyp->pag == pag);
+                        busyp->group == xg);
 
                if (wakeup) {
-                       pag->pagb_gen++;
-                       wake_up_all(&pag->pagb_wait);
+                       eb->eb_gen++;
+                       wake_up_all(&eb->eb_wait);
                }
-               spin_unlock(&pag->pagb_lock);
-               xfs_perag_put(pag);
+               spin_unlock(&eb->eb_lock);
+               xfs_group_put(xg);
        } while (!list_entry_is_head(busyp, list, list));
 }
 
 /*
- * Flush out all busy extents for this AG.
+ * Flush out all busy extents for this group.
  *
  * If the current transaction is holding busy extents, the caller may not want
  * to wait for committed busy extents to resolve. If we are being told just to
@@ -592,10 +608,11 @@ xfs_extent_busy_clear(
 int
 xfs_extent_busy_flush(
        struct xfs_trans        *tp,
-       struct xfs_perag        *pag,
+       struct xfs_group        *xg,
        unsigned                busy_gen,
        uint32_t                alloc_flags)
 {
+       struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
        DEFINE_WAIT             (wait);
        int                     error;
 
@@ -608,7 +625,7 @@ xfs_extent_busy_flush(
                if (alloc_flags & XFS_ALLOC_FLAG_TRYFLUSH)
                        return 0;
 
-               if (busy_gen != READ_ONCE(pag->pagb_gen))
+               if (busy_gen != READ_ONCE(eb->eb_gen))
                        return 0;
 
                if (alloc_flags & XFS_ALLOC_FLAG_FREEING)
@@ -617,36 +634,44 @@ xfs_extent_busy_flush(
 
        /* Wait for committed busy extents to resolve. */
        do {
-               prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
-               if  (busy_gen != READ_ONCE(pag->pagb_gen))
+               prepare_to_wait(&eb->eb_wait, &wait, TASK_KILLABLE);
+               if  (busy_gen != READ_ONCE(eb->eb_gen))
                        break;
                schedule();
        } while (1);
 
-       finish_wait(&pag->pagb_wait, &wait);
+       finish_wait(&eb->eb_wait, &wait);
        return 0;
 }
 
+static void
+xfs_extent_busy_wait_group(
+       struct xfs_group        *xg)
+{
+       DEFINE_WAIT             (wait);
+       struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
+
+       do {
+               prepare_to_wait(&eb->eb_wait, &wait, TASK_KILLABLE);
+               if  (RB_EMPTY_ROOT(&eb->eb_tree))
+                       break;
+               schedule();
+       } while (1);
+       finish_wait(&eb->eb_wait, &wait);
+}
+
 void
 xfs_extent_busy_wait_all(
        struct xfs_mount        *mp)
 {
        struct xfs_perag        *pag = NULL;
-       DEFINE_WAIT             (wait);
 
-       while ((pag = xfs_perag_next(mp, pag))) {
-               do {
-                       prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);
-                       if  (RB_EMPTY_ROOT(&pag->pagb_tree))
-                               break;
-                       schedule();
-               } while (1);
-               finish_wait(&pag->pagb_wait, &wait);
-       }
+       while ((pag = xfs_perag_next(mp, pag)))
+               xfs_extent_busy_wait_group(&pag->pag_group);
 }
 
 /*
- * Callback for list_sort to sort busy extents by the AG they reside in.
+ * Callback for list_sort to sort busy extents by the group they reside in.
  */
 int
 xfs_extent_busy_ag_cmp(
@@ -660,23 +685,38 @@ xfs_extent_busy_ag_cmp(
                container_of(l2, struct xfs_extent_busy, list);
        s32 diff;
 
-       diff = pag_agno(b1->pag) - pag_agno(b2->pag);
+       diff = b1->group->xg_index - b2->group->xg_index;
        if (!diff)
                diff = b1->bno - b2->bno;
        return diff;
 }
 
-/* Are there any busy extents in this AG? */
+/* Are there any busy extents in this group? */
 bool
 xfs_extent_busy_list_empty(
-       struct xfs_perag        *pag,
+       struct xfs_group        *xg,
        unsigned                *busy_gen)
 {
+       struct xfs_extent_busy_tree *eb = xg->xg_busy_extents;
        bool                    res;
 
-       spin_lock(&pag->pagb_lock);
-       res = RB_EMPTY_ROOT(&pag->pagb_tree);
-       *busy_gen = READ_ONCE(pag->pagb_gen);
-       spin_unlock(&pag->pagb_lock);
+       spin_lock(&eb->eb_lock);
+       res = RB_EMPTY_ROOT(&eb->eb_tree);
+       *busy_gen = READ_ONCE(eb->eb_gen);
+       spin_unlock(&eb->eb_lock);
        return res;
 }
+
+struct xfs_extent_busy_tree *
+xfs_extent_busy_alloc(void)
+{
+       struct xfs_extent_busy_tree *eb;
+
+       eb = kzalloc(sizeof(*eb), GFP_KERNEL);
+       if (!eb)
+               return NULL;
+       spin_lock_init(&eb->eb_lock);
+       init_waitqueue_head(&eb->eb_wait);
+       eb->eb_tree = RB_ROOT;
+       return eb;
+}
index c803dcd124a62857302996aaae6812ee312cf6af..f069b04e8ea184ab2128f30d407084c80b8114ff 100644 (file)
@@ -8,19 +8,18 @@
 #ifndef __XFS_EXTENT_BUSY_H__
 #define        __XFS_EXTENT_BUSY_H__
 
+struct xfs_group;
 struct xfs_mount;
-struct xfs_perag;
 struct xfs_trans;
-struct xfs_alloc_arg;
 
 /*
- * Busy block/extent entry.  Indexed by a rbtree in perag to mark blocks that
- * have been freed but whose transactions aren't committed to disk yet.
+ * Busy block/extent entry.  Indexed by a rbtree in the group to mark blocks
+ * that have been freed but whose transactions aren't committed to disk yet.
  */
 struct xfs_extent_busy {
-       struct rb_node  rb_node;        /* ag by-bno indexed search tree */
+       struct rb_node  rb_node;        /* group by-bno indexed search tree */
        struct list_head list;          /* transaction busy extent list */
-       struct xfs_perag *pag;
+       struct xfs_group *group;
        xfs_agblock_t   bno;
        xfs_extlen_t    length;
        unsigned int    flags;
@@ -44,45 +43,29 @@ struct xfs_busy_extents {
        void                    *owner;
 };
 
-void
-xfs_extent_busy_insert(struct xfs_trans *tp, struct xfs_perag *pag,
-       xfs_agblock_t bno, xfs_extlen_t len, unsigned int flags);
-
-void
-xfs_extent_busy_insert_discard(struct xfs_perag *pag, xfs_agblock_t bno,
-       xfs_extlen_t len, struct list_head *busy_list);
-
-void
-xfs_extent_busy_clear(struct list_head *list, bool do_discard);
-
-int
-xfs_extent_busy_search(struct xfs_perag *pag, xfs_agblock_t bno,
+void xfs_extent_busy_insert(struct xfs_trans *tp, struct xfs_group *xg,
+               xfs_agblock_t bno, xfs_extlen_t len, unsigned int flags);
+void xfs_extent_busy_insert_discard(struct xfs_group *xg, xfs_agblock_t bno,
+               xfs_extlen_t len, struct list_head *busy_list);
+void xfs_extent_busy_clear(struct list_head *list, bool do_discard);
+int xfs_extent_busy_search(struct xfs_group *xg, xfs_agblock_t bno,
                xfs_extlen_t len);
-
-void
-xfs_extent_busy_reuse(struct xfs_perag *pag, xfs_agblock_t fbno,
+void xfs_extent_busy_reuse(struct xfs_group *xg, xfs_agblock_t fbno,
                xfs_extlen_t flen, bool userdata);
-
-bool
-xfs_extent_busy_trim(struct xfs_alloc_arg *args, xfs_agblock_t *bno,
-               xfs_extlen_t *len, unsigned *busy_gen);
-
-int
-xfs_extent_busy_flush(struct xfs_trans *tp, struct xfs_perag *pag,
+bool xfs_extent_busy_trim(struct xfs_group *xg, xfs_extlen_t minlen,
+               xfs_extlen_t maxlen, xfs_agblock_t *bno, xfs_extlen_t *len,
+               unsigned *busy_gen);
+int xfs_extent_busy_flush(struct xfs_trans *tp, struct xfs_group *xg,
                unsigned busy_gen, uint32_t alloc_flags);
+void xfs_extent_busy_wait_all(struct xfs_mount *mp);
+bool xfs_extent_busy_list_empty(struct xfs_group *xg, unsigned int *busy_gen);
+struct xfs_extent_busy_tree *xfs_extent_busy_alloc(void);
 
-void
-xfs_extent_busy_wait_all(struct xfs_mount *mp);
-
-int
-xfs_extent_busy_ag_cmp(void *priv, const struct list_head *a,
-       const struct list_head *b);
-
+int xfs_extent_busy_ag_cmp(void *priv, const struct list_head *a,
+               const struct list_head *b);
 static inline void xfs_extent_busy_sort(struct list_head *list)
 {
        list_sort(NULL, list, xfs_extent_busy_ag_cmp);
 }
 
-bool xfs_extent_busy_list_empty(struct xfs_perag *pag, unsigned int *busy_gen);
-
 #endif /* __XFS_EXTENT_BUSY_H__ */