]> git.ipfire.org Git - thirdparty/xfsprogs-dev.git/commitdiff
xfs: move perag structure and setup to libxfs/xfs_ag.[ch]
authorDave Chinner <dchinner@redhat.com>
Fri, 15 Oct 2021 20:28:25 +0000 (16:28 -0400)
committerEric Sandeen <sandeen@sandeen.net>
Fri, 15 Oct 2021 20:28:25 +0000 (16:28 -0400)
Source kernel commit: 07b6403a6873045344b0c18cbb4a4360854f6d76

Move the xfs_perag infrastructure to the libxfs files that contain
all the per AG infrastructure. This helps set up for passing perags
around all the code instead of bare agnos with minimal extra
includes for existing files.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
include/xfs_mount.h
libfrog/radix-tree.h
libxfs/init.c
libxfs/libxfs_api_defs.h
libxfs/libxfs_priv.h
libxfs/xfs_ag.c
libxfs/xfs_ag.h
libxfs/xfs_ag_resv.h
libxfs/xfs_btree.c
libxfs/xfs_shared.h

index 9e43cd230166bf786fca250ec970469ba883fad4..cc4682e2d9d5875d8d28e680b18cfb79857533ff 100644 (file)
@@ -114,76 +114,12 @@ typedef struct xfs_mount {
         * reservation from attempting to reserve allocation btree blocks.
         */
        atomic64_t              m_allocbt_blks;
+       spinlock_t              m_perag_lock;   /* lock for m_perag_tree */
 
 } xfs_mount_t;
 
 #define M_IGEO(mp)             (&(mp)->m_ino_geo)
 
-struct xfs_ag_resv {
-       /* number of blocks originally reserved here */
-       xfs_extlen_t    ar_orig_reserved;
-       /* number of blocks reserved here */
-       xfs_extlen_t    ar_reserved;
-       /* number of blocks originally asked for */
-       xfs_extlen_t    ar_asked;
-};
-
-/*
- * Per-ag incore structure, copies of information in agf and agi,
- * to improve the performance of allocation group selection.
- */
-typedef struct xfs_perag {
-       struct xfs_mount *pag_mount;    /* owner filesystem */
-       xfs_agnumber_t  pag_agno;       /* AG this structure belongs to */
-       atomic_t        pag_ref;        /* perag reference count */
-       char            pagf_init;      /* this agf's entry is initialized */
-       char            pagi_init;      /* this agi's entry is initialized */
-       char            pagf_metadata;  /* the agf is preferred to be metadata */
-       char            pagi_inodeok;   /* The agi is ok for inodes */
-       uint8_t         pagf_levels[XFS_BTNUM_AGF];
-                                       /* # of levels in bno & cnt btree */
-       bool            pagf_agflreset; /* agfl requires reset before use */
-       uint32_t        pagf_flcount;   /* count of blocks in freelist */
-       xfs_extlen_t    pagf_freeblks;  /* total free blocks */
-       xfs_extlen_t    pagf_longest;   /* longest free space */
-       uint32_t        pagf_btreeblks; /* # of blocks held in AGF btrees */
-       xfs_agino_t     pagi_freecount; /* number of free inodes */
-       xfs_agino_t     pagi_count;     /* number of allocated inodes */
-
-       /*
-        * Inode allocation search lookup optimisation.
-        * If the pagino matches, the search for new inodes
-        * doesn't need to search the near ones again straight away
-        */
-       xfs_agino_t     pagl_pagino;
-       xfs_agino_t     pagl_leftrec;
-       xfs_agino_t     pagl_rightrec;
-       int             pagb_count;     /* pagb slots in use */
-
-       /* Blocks reserved for all kinds of metadata. */
-       struct xfs_ag_resv      pag_meta_resv;
-       /* Blocks reserved for just AGFL-based metadata. */
-       struct xfs_ag_resv      pag_rmapbt_resv;
-
-       /* reference count */
-       uint8_t         pagf_refcount_level;
-} xfs_perag_t;
-
-static inline struct xfs_ag_resv *
-xfs_perag_resv(
-       struct xfs_perag        *pag,
-       enum xfs_ag_resv_type   type)
-{
-       switch (type) {
-       case XFS_AG_RESV_METADATA:
-               return &pag->pag_meta_resv;
-       case XFS_AG_RESV_RMAPBT:
-               return &pag->pag_rmapbt_resv;
-       default:
-               return NULL;
-       }
-}
-
 #define LIBXFS_MOUNT_DEBUGGER          0x0001
 #define LIBXFS_MOUNT_32BITINODES       0x0002
 #define LIBXFS_MOUNT_32BITINOOPT       0x0004
index f08156b9d03205da7f821b9b43c756b7f29b8a6d..dad5f5b72039e3b4b71198c0a4934beba390d937 100644 (file)
@@ -60,4 +60,7 @@ radix_tree_gang_lookup_tag(struct radix_tree_root *root, void **results,
 int radix_tree_tagged(struct radix_tree_root *root, unsigned int tag);
 #endif
 
+static inline int radix_tree_preload(int gfp_mask) { return 0; }
+static inline void radix_tree_preload_end(void) { }
+
 #endif /* __LIBFROG_RADIX_TREE_H__ */
index 14911596c7701b68d143a7b6594e3c19f34c7b35..0f7e89509d5f77ccb88846228015ac352b47a007 100644 (file)
@@ -573,60 +573,6 @@ xfs_set_inode_alloc(
        return (mp->m_flags & XFS_MOUNT_32BITINODES) ? maxagi : agcount;
 }
 
-static int
-libxfs_initialize_perag(
-       xfs_mount_t     *mp,
-       xfs_agnumber_t  agcount,
-       xfs_agnumber_t  *maxagi)
-{
-       xfs_agnumber_t  index;
-       xfs_agnumber_t  first_initialised = 0;
-       xfs_perag_t     *pag;
-       int             error = -ENOMEM;
-
-       /*
-        * Walk the current per-ag tree so we don't try to initialise AGs
-        * that already exist (growfs case). Allocate and insert all the
-        * AGs we don't find ready for initialisation.
-        */
-       for (index = 0; index < agcount; index++) {
-               pag = xfs_perag_get(mp, index);
-               if (pag) {
-                       xfs_perag_put(pag);
-                       continue;
-               }
-               if (!first_initialised)
-                       first_initialised = index;
-
-               pag = kmem_zalloc(sizeof(*pag), KM_MAYFAIL);
-               if (!pag)
-                       goto out_unwind;
-               pag->pag_agno = index;
-               pag->pag_mount = mp;
-
-               if (radix_tree_insert(&mp->m_perag_tree, index, pag)) {
-                       error = -EEXIST;
-                       goto out_unwind;
-               }
-       }
-
-       index = xfs_set_inode_alloc(mp, agcount);
-
-       if (maxagi)
-               *maxagi = index;
-
-       mp->m_ag_prealloc_blocks = xfs_prealloc_blocks(mp);
-       return 0;
-
-out_unwind:
-       kmem_free(pag);
-       for (; index > first_initialised; index--) {
-               pag = radix_tree_delete(&mp->m_perag_tree, index);
-               kmem_free(pag);
-       }
-       return error;
-}
-
 static struct xfs_buftarg *
 libxfs_buftarg_alloc(
        struct xfs_mount        *mp,
@@ -1020,8 +966,6 @@ int
 libxfs_umount(
        struct xfs_mount        *mp)
 {
-       struct xfs_perag        *pag;
-       int                     agno;
        int                     error;
 
        libxfs_rtmount_destroy(mp);
@@ -1038,12 +982,8 @@ libxfs_umount(
         * Only try to free the per-AG structures if we set them up in the
         * first place.
         */
-       if (mp->m_flags & LIBXFS_MOUNT_PERAG_DATA_LOADED) {
-               for (agno = 0; agno < mp->m_maxagi; agno++) {
-                       pag = radix_tree_delete(&mp->m_perag_tree, agno);
-                       kmem_free(pag);
-               }
-       }
+       if (mp->m_flags & LIBXFS_MOUNT_PERAG_DATA_LOADED)
+               libxfs_free_perag(mp);
 
        kmem_free(mp->m_attr_geo);
        kmem_free(mp->m_dir_geo);
index d759ff6545c247c555d60785f10e6b061e9c0c81..b76e638098bbb26bde5a7ca97d404fc1daad033c 100644 (file)
 
 #define xfs_finobt_calc_reserves       libxfs_finobt_calc_reserves
 #define xfs_free_extent                        libxfs_free_extent
+#define xfs_free_perag                 libxfs_free_perag
 #define xfs_fs_geometry                        libxfs_fs_geometry
 #define xfs_highbit32                  libxfs_highbit32
 #define xfs_highbit64                  libxfs_highbit64
 #define xfs_idata_realloc              libxfs_idata_realloc
 #define xfs_idestroy_fork              libxfs_idestroy_fork
 #define xfs_iext_lookup_extent         libxfs_iext_lookup_extent
+#define xfs_initialize_perag           libxfs_initialize_perag
 #define xfs_initialize_perag_data      libxfs_initialize_perag_data
 #define xfs_init_local_fork            libxfs_init_local_fork
 
index e1e90268c0b7616da1d8606a01c2e561f6d4f570..7b4e8f55226ec1a78293c9bad738dbf578845b1d 100644 (file)
@@ -546,7 +546,6 @@ int xfs_attr_rmtval_get(struct xfs_da_args *);
 void xfs_bmap_del_free(struct xfs_bmap_free *, struct xfs_bmap_free_item *);
 
 /* xfs_mount.c */
-int xfs_initialize_perag_data(struct xfs_mount *, xfs_agnumber_t);
 void xfs_mount_common(struct xfs_mount *, struct xfs_sb *);
 
 /*
@@ -665,6 +664,15 @@ static inline int test_and_set_bit(int nr, volatile unsigned long *addr)
        return 0;
 }
 
+static inline int xfs_buf_hash_init(struct xfs_perag *pag) { return 0; }
+static inline void xfs_buf_hash_destroy(struct xfs_perag *pag) { }
+
+static inline int xfs_iunlink_init(struct xfs_perag *pag) { return 0; }
+static inline void xfs_iunlink_destroy(struct xfs_perag *pag) { }
+
+xfs_agnumber_t xfs_set_inode_alloc(struct xfs_mount *mp,
+               xfs_agnumber_t agcount);
+
 /* Keep static checkers quiet about nonstatic functions by exporting */
 int xfs_rtbuf_get(struct xfs_mount *mp, struct xfs_trans *tp,
                  xfs_rtblock_t block, int issum, struct xfs_buf **bpp);
index 46e78b0e9ef7298660642fd454eaad6fad33ce5e..1027bc7bd7bd5ee35271fcb82187ac2e8072f2f7 100644 (file)
@@ -27,6 +27,8 @@
 #include "xfs_log_format.h"
 #include "xfs_trans.h"
 #include "xfs_trace.h"
+#include "xfs_inode.h"
+
 
 /*
  * Passive reference counting access wrappers to the perag structures.  If the
@@ -162,6 +164,138 @@ out:
        return error;
 }
 
+STATIC void
+__xfs_free_perag(
+       struct rcu_head *head)
+{
+       struct xfs_perag *pag = container_of(head, struct xfs_perag, rcu_head);
+
+       ASSERT(!delayed_work_pending(&pag->pag_blockgc_work));
+       ASSERT(atomic_read(&pag->pag_ref) == 0);
+       kmem_free(pag);
+}
+
+/*
+ * Free up the per-ag resources associated with the mount structure.
+ */
+void
+xfs_free_perag(
+       struct xfs_mount        *mp)
+{
+       struct xfs_perag        *pag;
+       xfs_agnumber_t          agno;
+
+       for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
+               spin_lock(&mp->m_perag_lock);
+               pag = radix_tree_delete(&mp->m_perag_tree, agno);
+               spin_unlock(&mp->m_perag_lock);
+               ASSERT(pag);
+               ASSERT(atomic_read(&pag->pag_ref) == 0);
+
+               cancel_delayed_work_sync(&pag->pag_blockgc_work);
+               xfs_iunlink_destroy(pag);
+               xfs_buf_hash_destroy(pag);
+
+               call_rcu(&pag->rcu_head, __xfs_free_perag);
+       }
+}
+
+int
+xfs_initialize_perag(
+       struct xfs_mount        *mp,
+       xfs_agnumber_t          agcount,
+       xfs_agnumber_t          *maxagi)
+{
+       struct xfs_perag        *pag;
+       xfs_agnumber_t          index;
+       xfs_agnumber_t          first_initialised = NULLAGNUMBER;
+       int                     error;
+
+       /*
+        * Walk the current per-ag tree so we don't try to initialise AGs
+        * that already exist (growfs case). Allocate and insert all the
+        * AGs we don't find ready for initialisation.
+        */
+       for (index = 0; index < agcount; index++) {
+               pag = xfs_perag_get(mp, index);
+               if (pag) {
+                       xfs_perag_put(pag);
+                       continue;
+               }
+
+               pag = kmem_zalloc(sizeof(*pag), KM_MAYFAIL);
+               if (!pag) {
+                       error = -ENOMEM;
+                       goto out_unwind_new_pags;
+               }
+               pag->pag_agno = index;
+               pag->pag_mount = mp;
+
+               error = radix_tree_preload(GFP_NOFS);
+               if (error)
+                       goto out_free_pag;
+
+               spin_lock(&mp->m_perag_lock);
+               if (radix_tree_insert(&mp->m_perag_tree, index, pag)) {
+                       WARN_ON_ONCE(1);
+                       spin_unlock(&mp->m_perag_lock);
+                       radix_tree_preload_end();
+                       error = -EEXIST;
+                       goto out_free_pag;
+               }
+               spin_unlock(&mp->m_perag_lock);
+               radix_tree_preload_end();
+
+               /* Place kernel structure only init below this point. */
+               spin_lock_init(&pag->pag_ici_lock);
+               spin_lock_init(&pag->pagb_lock);
+               spin_lock_init(&pag->pag_state_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_count = 0;
+               pag->pagb_tree = RB_ROOT;
+
+               error = xfs_buf_hash_init(pag);
+               if (error)
+                       goto out_remove_pag;
+
+               error = xfs_iunlink_init(pag);
+               if (error)
+                       goto out_hash_destroy;
+
+               /* first new pag is fully initialized */
+               if (first_initialised == NULLAGNUMBER)
+                       first_initialised = index;
+       }
+
+       index = xfs_set_inode_alloc(mp, agcount);
+
+       if (maxagi)
+               *maxagi = index;
+
+       mp->m_ag_prealloc_blocks = xfs_prealloc_blocks(mp);
+       return 0;
+
+out_hash_destroy:
+       xfs_buf_hash_destroy(pag);
+out_remove_pag:
+       radix_tree_delete(&mp->m_perag_tree, index);
+out_free_pag:
+       kmem_free(pag);
+out_unwind_new_pags:
+       /* unwind any prior newly initialized pags */
+       for (index = first_initialised; index < agcount; index++) {
+               pag = radix_tree_delete(&mp->m_perag_tree, index);
+               if (!pag)
+                       break;
+               xfs_buf_hash_destroy(pag);
+               xfs_iunlink_destroy(pag);
+               kmem_free(pag);
+       }
+       return error;
+}
+
 static int
 xfs_get_aghdr_buf(
        struct xfs_mount        *mp,
index cb1bd1c03cd7023e3cb19523490578ac78442ee1..f26f72e4735aecd0ace6956af5d5a1ad49237338 100644 (file)
@@ -12,9 +12,103 @@ struct xfs_trans;
 struct xfs_perag;
 
 /*
- * perag get/put wrappers for ref counting
+ * Per-ag infrastructure
  */
-int    xfs_initialize_perag_data(struct xfs_mount *, xfs_agnumber_t);
+
+/* per-AG block reservation data structures*/
+struct xfs_ag_resv {
+       /* number of blocks originally reserved here */
+       xfs_extlen_t                    ar_orig_reserved;
+       /* number of blocks reserved here */
+       xfs_extlen_t                    ar_reserved;
+       /* number of blocks originally asked for */
+       xfs_extlen_t                    ar_asked;
+};
+
+/*
+ * Per-ag incore structure, copies of information in agf and agi, to improve the
+ * performance of allocation group selection.
+ */
+typedef struct xfs_perag {
+       struct xfs_mount *pag_mount;    /* owner filesystem */
+       xfs_agnumber_t  pag_agno;       /* AG this structure belongs to */
+       atomic_t        pag_ref;        /* perag reference count */
+       char            pagf_init;      /* this agf's entry is initialized */
+       char            pagi_init;      /* this agi's entry is initialized */
+       char            pagf_metadata;  /* the agf is preferred to be metadata */
+       char            pagi_inodeok;   /* The agi is ok for inodes */
+       uint8_t         pagf_levels[XFS_BTNUM_AGF];
+                                       /* # of levels in bno & cnt btree */
+       bool            pagf_agflreset; /* agfl requires reset before use */
+       uint32_t        pagf_flcount;   /* count of blocks in freelist */
+       xfs_extlen_t    pagf_freeblks;  /* total free blocks */
+       xfs_extlen_t    pagf_longest;   /* longest free space */
+       uint32_t        pagf_btreeblks; /* # of blocks held in AGF btrees */
+       xfs_agino_t     pagi_freecount; /* number of free inodes */
+       xfs_agino_t     pagi_count;     /* number of allocated inodes */
+
+       /*
+        * Inode allocation search lookup optimisation.
+        * If the pagino matches, the search for new inodes
+        * doesn't need to search the near ones again straight away
+        */
+       xfs_agino_t     pagl_pagino;
+       xfs_agino_t     pagl_leftrec;
+       xfs_agino_t     pagl_rightrec;
+
+       int             pagb_count;     /* pagb slots in use */
+       uint8_t         pagf_refcount_level; /* recount btree height */
+
+       /* Blocks reserved for all kinds of metadata. */
+       struct xfs_ag_resv      pag_meta_resv;
+       /* Blocks reserved for the reverse mapping btree. */
+       struct xfs_ag_resv      pag_rmapbt_resv;
+
+       /* -- kernel only structures below this line -- */
+
+       /*
+        * Bitsets of per-ag metadata that have been checked and/or are sick.
+        * Callers should hold pag_state_lock before accessing this field.
+        */
+       uint16_t        pag_checked;
+       uint16_t        pag_sick;
+       spinlock_t      pag_state_lock;
+
+       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 */
+       struct radix_tree_root pag_ici_root;    /* incore inode cache root */
+       int             pag_ici_reclaimable;    /* reclaimable inodes */
+       unsigned long   pag_ici_reclaim_cursor; /* reclaim restart point */
+
+       /* buffer cache index */
+       spinlock_t      pag_buf_lock;   /* lock for pag_buf_hash */
+       struct rhashtable pag_buf_hash;
+
+       /* for rcu-safe freeing */
+       struct rcu_head rcu_head;
+
+       /* background prealloc block trimming */
+       struct delayed_work     pag_blockgc_work;
+
+       /*
+        * Unlinked inode information.  This incore information reflects
+        * data stored in the AGI, so callers must hold the AGI buffer lock
+        * or have some other means to control concurrency.
+        */
+       struct rhashtable       pagi_unlinked_hash;
+} xfs_perag_t;
+
+int xfs_initialize_perag(struct xfs_mount *mp, xfs_agnumber_t agcount,
+                       xfs_agnumber_t *maxagi);
+int xfs_initialize_perag_data(struct xfs_mount *mp, xfs_agnumber_t agno);
+void xfs_free_perag(struct xfs_mount *mp);
+
 struct xfs_perag *xfs_perag_get(struct xfs_mount *, xfs_agnumber_t);
 struct xfs_perag *xfs_perag_get_tag(struct xfs_mount *, xfs_agnumber_t,
                                   int tag);
index 8a8eb4bc48bb657d4fb812f7c8fbf602cdae1c72..b74b210008ea7e973f296b3fca524887fe27bef3 100644 (file)
@@ -18,6 +18,21 @@ void xfs_ag_resv_alloc_extent(struct xfs_perag *pag, enum xfs_ag_resv_type type,
 void xfs_ag_resv_free_extent(struct xfs_perag *pag, enum xfs_ag_resv_type type,
                struct xfs_trans *tp, xfs_extlen_t len);
 
+static inline struct xfs_ag_resv *
+xfs_perag_resv(
+       struct xfs_perag        *pag,
+       enum xfs_ag_resv_type   type)
+{
+       switch (type) {
+       case XFS_AG_RESV_METADATA:
+               return &pag->pag_meta_resv;
+       case XFS_AG_RESV_RMAPBT:
+               return &pag->pag_rmapbt_resv;
+       default:
+               return NULL;
+       }
+}
+
 /*
  * RMAPBT reservation accounting wrappers. Since rmapbt blocks are sourced from
  * the AGFL, they are allocated one at a time and the reservation updates don't
index d52fdc005fe9e1a7eecb7081b455c6ff7b353671..4faf4a67c38d2dae0b7cff21a0f4c22fb1649082 100644 (file)
@@ -18,6 +18,7 @@
 #include "xfs_trace.h"
 #include "xfs_alloc.h"
 #include "xfs_btree_staging.h"
+#include "xfs_ag.h"
 
 /*
  * Cursor allocation zone.
index 782fdd08f7596199d8ed141d63e71aacf84f1cb2..2d99184d0f8618fbdf965deb4562ba2dd926dd53 100644 (file)
@@ -184,4 +184,24 @@ struct xfs_ino_geometry {
 
 };
 
+/* Faked up kernel bits */
+struct rb_root {
+};
+
+#define RB_ROOT                (struct rb_root) { }
+
+typedef struct wait_queue_head {
+} wait_queue_head_t;
+
+#define init_waitqueue_head(wqh)       do { } while(0)
+
+struct rhashtable {
+};
+
+struct delayed_work {
+};
+
+#define INIT_DELAYED_WORK(work, func)  do { } while(0)
+#define cancel_delayed_work_sync(work) do { } while(0)
+
 #endif /* __XFS_SHARED_H__ */