[XFS] Remove the xlog_ticket allocator
authorDavid Chinner <dgc@sgi.com>
Thu, 10 Apr 2008 02:18:46 +0000 (12:18 +1000)
committerLachlan McIlroy <lachlan@redback.melbourne.sgi.com>
Fri, 18 Apr 2008 01:50:39 +0000 (11:50 +1000)
The ticket allocator is just a simple slab implementation internal to the
log. It requires the icloglock to be held when manipulating it and this
contributes to contention on that lock.

Just kill the entire allocator and use a memory zone instead. While there,
allow us to gracefully fail allocation with ENOMEM.

SGI-PV: 978729
SGI-Modid: xfs-linux-melb:xfs-kern:30771a

Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
fs/xfs/xfs_log.c
fs/xfs/xfs_log_priv.h
fs/xfs/xfs_vfsops.c

index 7a5b12d93537615f20c15fe56c2af798bba9f959..3cf115d8de75f2b412631ac7ff55a267f00dc21e 100644 (file)
@@ -41,6 +41,7 @@
 #include "xfs_inode.h"
 #include "xfs_rw.h"
 
+kmem_zone_t    *xfs_log_ticket_zone;
 
 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
        { (ptr) += (bytes); \
@@ -73,8 +74,6 @@ STATIC int  xlog_state_get_iclog_space(xlog_t         *log,
                                       xlog_ticket_t    *ticket,
                                       int              *continued_write,
                                       int              *logoffsetp);
-STATIC void xlog_state_put_ticket(xlog_t       *log,
-                                 xlog_ticket_t *tic);
 STATIC int  xlog_state_release_iclog(xlog_t            *log,
                                     xlog_in_core_t     *iclog);
 STATIC void xlog_state_switch_iclogs(xlog_t            *log,
@@ -101,7 +100,6 @@ STATIC void xlog_ungrant_log_space(xlog_t    *log,
 
 
 /* local ticket functions */
-STATIC void            xlog_state_ticket_alloc(xlog_t *log);
 STATIC xlog_ticket_t   *xlog_ticket_get(xlog_t *log,
                                         int    unit_bytes,
                                         int    count,
@@ -330,7 +328,7 @@ xfs_log_done(xfs_mount_t    *mp,
                 */
                xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
                xlog_ungrant_log_space(log, ticket);
-               xlog_state_put_ticket(log, ticket);
+               xlog_ticket_put(log, ticket);
        } else {
                xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
                xlog_regrant_reserve_log_space(log, ticket);
@@ -469,6 +467,8 @@ xfs_log_reserve(xfs_mount_t  *mp,
                /* may sleep if need to allocate more tickets */
                internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
                                                  client, flags);
+               if (!internal_ticket)
+                       return XFS_ERROR(ENOMEM);
                internal_ticket->t_trans_type = t_type;
                *ticket = internal_ticket;
                xlog_trace_loggrant(log, internal_ticket, 
@@ -693,7 +693,7 @@ xfs_log_unmount_write(xfs_mount_t *mp)
                if (tic) {
                        xlog_trace_loggrant(log, tic, "unmount rec");
                        xlog_ungrant_log_space(log, tic);
-                       xlog_state_put_ticket(log, tic);
+                       xlog_ticket_put(log, tic);
                }
        } else {
                /*
@@ -1208,7 +1208,6 @@ xlog_alloc_log(xfs_mount_t        *mp,
        spin_lock_init(&log->l_icloglock);
        spin_lock_init(&log->l_grant_lock);
        initnsema(&log->l_flushsema, 0, "ic-flush");
-       xlog_state_ticket_alloc(log);  /* wait until after icloglock inited */
 
        /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
        ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
@@ -1538,7 +1537,6 @@ STATIC void
 xlog_dealloc_log(xlog_t *log)
 {
        xlog_in_core_t  *iclog, *next_iclog;
-       xlog_ticket_t   *tic, *next_tic;
        int             i;
 
        iclog = log->l_iclog;
@@ -1559,22 +1557,6 @@ xlog_dealloc_log(xlog_t *log)
        spinlock_destroy(&log->l_icloglock);
        spinlock_destroy(&log->l_grant_lock);
 
-       /* XXXsup take a look at this again. */
-       if ((log->l_ticket_cnt != log->l_ticket_tcnt)  &&
-           !XLOG_FORCED_SHUTDOWN(log)) {
-               xfs_fs_cmn_err(CE_WARN, log->l_mp,
-                       "xlog_dealloc_log: (cnt: %d, total: %d)",
-                       log->l_ticket_cnt, log->l_ticket_tcnt);
-               /* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
-
-       } else {
-               tic = log->l_unmount_free;
-               while (tic) {
-                       next_tic = tic->t_next;
-                       kmem_free(tic, PAGE_SIZE);
-                       tic = next_tic;
-               }
-       }
        xfs_buf_free(log->l_xbuf);
 #ifdef XFS_LOG_TRACE
        if (log->l_trace != NULL) {
@@ -2794,18 +2776,6 @@ xlog_ungrant_log_space(xlog_t         *log,
 }      /* xlog_ungrant_log_space */
 
 
-/*
- * Atomically put back used ticket.
- */
-STATIC void
-xlog_state_put_ticket(xlog_t       *log,
-                     xlog_ticket_t *tic)
-{
-       spin_lock(&log->l_icloglock);
-       xlog_ticket_put(log, tic);
-       spin_unlock(&log->l_icloglock);
-}      /* xlog_state_put_ticket */
-
 /*
  * Flush iclog to disk if this is the last reference to the given iclog and
  * the WANT_SYNC bit is set.
@@ -3176,92 +3146,19 @@ xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
  */
 
 /*
- *     Algorithm doesn't take into account page size. ;-(
- */
-STATIC void
-xlog_state_ticket_alloc(xlog_t *log)
-{
-       xlog_ticket_t   *t_list;
-       xlog_ticket_t   *next;
-       xfs_caddr_t     buf;
-       uint            i = (PAGE_SIZE / sizeof(xlog_ticket_t)) - 2;
-
-       /*
-        * The kmem_zalloc may sleep, so we shouldn't be holding the
-        * global lock.  XXXmiken: may want to use zone allocator.
-        */
-       buf = (xfs_caddr_t) kmem_zalloc(PAGE_SIZE, KM_SLEEP);
-
-       spin_lock(&log->l_icloglock);
-
-       /* Attach 1st ticket to Q, so we can keep track of allocated memory */
-       t_list = (xlog_ticket_t *)buf;
-       t_list->t_next = log->l_unmount_free;
-       log->l_unmount_free = t_list++;
-       log->l_ticket_cnt++;
-       log->l_ticket_tcnt++;
-
-       /* Next ticket becomes first ticket attached to ticket free list */
-       if (log->l_freelist != NULL) {
-               ASSERT(log->l_tail != NULL);
-               log->l_tail->t_next = t_list;
-       } else {
-               log->l_freelist = t_list;
-       }
-       log->l_ticket_cnt++;
-       log->l_ticket_tcnt++;
-
-       /* Cycle through rest of alloc'ed memory, building up free Q */
-       for ( ; i > 0; i--) {
-               next = t_list + 1;
-               t_list->t_next = next;
-               t_list = next;
-               log->l_ticket_cnt++;
-               log->l_ticket_tcnt++;
-       }
-       t_list->t_next = NULL;
-       log->l_tail = t_list;
-       spin_unlock(&log->l_icloglock);
-}      /* xlog_state_ticket_alloc */
-
-
-/*
- * Put ticket into free list
- *
- * Assumption: log lock is held around this call.
+ * Free a used ticket.
  */
 STATIC void
 xlog_ticket_put(xlog_t         *log,
                xlog_ticket_t   *ticket)
 {
        sv_destroy(&ticket->t_sema);
-
-       /*
-        * Don't think caching will make that much difference.  It's
-        * more important to make debug easier.
-        */
-#if 0
-       /* real code will want to use LIFO for caching */
-       ticket->t_next = log->l_freelist;
-       log->l_freelist = ticket;
-       /* no need to clear fields */
-#else
-       /* When we debug, it is easier if tickets are cycled */
-       ticket->t_next     = NULL;
-       if (log->l_tail) {
-               log->l_tail->t_next = ticket;
-       } else {
-               ASSERT(log->l_freelist == NULL);
-               log->l_freelist = ticket;
-       }
-       log->l_tail         = ticket;
-#endif /* DEBUG */
-       log->l_ticket_cnt++;
+       kmem_zone_free(xfs_log_ticket_zone, ticket);
 }      /* xlog_ticket_put */
 
 
 /*
- * Grab ticket off freelist or allocation some more
+ * Allocate and initialise a new log ticket.
  */
 STATIC xlog_ticket_t *
 xlog_ticket_get(xlog_t         *log,
@@ -3273,21 +3170,9 @@ xlog_ticket_get(xlog_t           *log,
        xlog_ticket_t   *tic;
        uint            num_headers;
 
- alloc:
-       if (log->l_freelist == NULL)
-               xlog_state_ticket_alloc(log);           /* potentially sleep */
-
-       spin_lock(&log->l_icloglock);
-       if (log->l_freelist == NULL) {
-               spin_unlock(&log->l_icloglock);
-               goto alloc;
-       }
-       tic             = log->l_freelist;
-       log->l_freelist = tic->t_next;
-       if (log->l_freelist == NULL)
-               log->l_tail = NULL;
-       log->l_ticket_cnt--;
-       spin_unlock(&log->l_icloglock);
+       tic = kmem_zone_zalloc(xfs_log_ticket_zone, KM_SLEEP|KM_MAYFAIL);
+       if (!tic)
+               return NULL;
 
        /*
         * Permanent reservations have up to 'cnt'-1 active log operations
index 104b623aa082177ae5dfa4c62a36fb27dd912628..c1583960009de8d884725e8a965961ad1775ea42 100644 (file)
@@ -242,7 +242,7 @@ typedef struct xlog_res {
 
 typedef struct xlog_ticket {
        sv_t               t_sema;       /* sleep on this semaphore      : 20 */
-       struct xlog_ticket *t_next;      /*                              :4|8 */
+       struct xlog_ticket *t_next;      /*                              :4|8 */
        struct xlog_ticket *t_prev;      /*                              :4|8 */
        xlog_tid_t         t_tid;        /* transaction identifier       : 4  */
        int                t_curr_res;   /* current reservation in bytes : 4  */
@@ -406,13 +406,8 @@ typedef struct log {
        sema_t                  l_flushsema;    /* iclog flushing semaphore */
        int                     l_flushcnt;     /* # of procs waiting on this
                                                 * sema */
-       int                     l_ticket_cnt;   /* free ticket count */
-       int                     l_ticket_tcnt;  /* total ticket count */
        int                     l_covered_state;/* state of "covering disk
                                                 * log entries" */
-       xlog_ticket_t           *l_freelist;    /* free list of tickets */
-       xlog_ticket_t           *l_unmount_free;/* kmem_free these addresses */
-       xlog_ticket_t           *l_tail;        /* free list of tickets */
        xlog_in_core_t          *l_iclog;       /* head log queue       */
        spinlock_t              l_icloglock;    /* grab to change iclog state */
        xfs_lsn_t               l_tail_lsn;     /* lsn of 1st LR with unflushed
@@ -478,6 +473,8 @@ extern struct xfs_buf *xlog_get_bp(xlog_t *, int);
 extern void     xlog_put_bp(struct xfs_buf *);
 extern int      xlog_bread(xlog_t *, xfs_daddr_t, int, struct xfs_buf *);
 
+extern kmem_zone_t     *xfs_log_ticket_zone;
+
 /* iclog tracing */
 #define XLOG_TRACE_GRAB_FLUSH  1
 #define XLOG_TRACE_REL_FLUSH   2
index c21e4d168297b8fd49aede086456cab48f1321f1..ea94593b5313b7e28e0e869a4baa77b0e2a13b20 100644 (file)
@@ -69,15 +69,17 @@ xfs_init(void)
        /*
         * Initialize all of the zone allocators we use.
         */
+       xfs_log_ticket_zone = kmem_zone_init(sizeof(xlog_ticket_t),
+                                               "xfs_log_ticket");
        xfs_bmap_free_item_zone = kmem_zone_init(sizeof(xfs_bmap_free_item_t),
-                                                "xfs_bmap_free_item");
+                                               "xfs_bmap_free_item");
        xfs_btree_cur_zone = kmem_zone_init(sizeof(xfs_btree_cur_t),
-                                           "xfs_btree_cur");
-       xfs_trans_zone = kmem_zone_init(sizeof(xfs_trans_t), "xfs_trans");
-       xfs_da_state_zone =
-               kmem_zone_init(sizeof(xfs_da_state_t), "xfs_da_state");
+                                               "xfs_btree_cur");
+       xfs_da_state_zone = kmem_zone_init(sizeof(xfs_da_state_t),
+                                               "xfs_da_state");
        xfs_dabuf_zone = kmem_zone_init(sizeof(xfs_dabuf_t), "xfs_dabuf");
        xfs_ifork_zone = kmem_zone_init(sizeof(xfs_ifork_t), "xfs_ifork");
+       xfs_trans_zone = kmem_zone_init(sizeof(xfs_trans_t), "xfs_trans");
        xfs_acl_zone_init(xfs_acl_zone, "xfs_acl");
        xfs_mru_cache_init();
        xfs_filestream_init();
This page took 0.034434 seconds and 5 git commands to generate.