{
struct log *log = mp->m_log;
int log_flags = 0;
+ int push = 0;
if (flags & XFS_TRANS_RELEASE_LOG_RES)
log_flags = XFS_LOG_REL_PERM_RESERV;
xfs_log_done(mp, tp->t_ticket, NULL, log_flags);
xfs_trans_unreserve_and_mod_sb(tp);
- /* background commit is allowed again */
+ /* check for background commit before unlock */
+ if (log->l_cilp->xc_ctx->space_used > XLOG_CIL_SPACE_LIMIT(log))
+ push = 1;
up_read(&log->l_cilp->xc_ctx_lock);
+
+ /*
+ * We need to push CIL every so often so we don't cache more than we
+ * can fit in the log. The limit really is that a checkpoint can't be
+ * more than half the log (the current checkpoint is not allowed to
+ * overwrite the previous checkpoint), but commit latency and memory
+ * usage limit this to a smaller size in most cases.
+ */
+ if (push)
+ xlog_cil_push(log, 0);
return 0;
}
if (!cil)
return 0;
- /* XXX: don't sleep for background? */
new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
new_ctx->ticket = xlog_cil_ticket_alloc(log);
- /* lock out transaction commit */
- down_write(&cil->xc_ctx_lock);
+ /* lock out transaction commit, but don't block on background push */
+ if (!down_write_trylock(&cil->xc_ctx_lock)) {
+ if (!push_now)
+ goto out_free_ticket;
+ down_write(&cil->xc_ctx_lock);
+ }
ctx = cil->xc_ctx;
/* check if we've anything to push */
if (list_empty(&cil->xc_cil))
goto out_skip;
+ /* check for spurious background flush */
+ if (!push_now && cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
+ goto out_skip;
+
/*
* pull all the log vectors off the items in the CIL, and
* remove the items from the CIL. We don't need the CIL lock
out_skip:
up_write(&cil->xc_ctx_lock);
+out_free_ticket:
xfs_log_ticket_put(new_ctx->ticket);
kmem_free(new_ctx);
return 0;
};
/*
+ * The amount of log space we should the CIL to aggregate is difficult to size.
+ * Whatever we chose we have to make we can get a reservation for the log space
+ * effectively, that it is large enough to capture sufficient relogging to
+ * reduce log buffer IO significantly, but it is not too large for the log or
+ * induces too much latency when writing out through the iclogs. We track both
+ * space consumed and the number of vectors in the checkpoint context, so we
+ * need to decide which to use for limiting.
+ *
+ * Every log buffer we write out during a push needs a header reserved, which
+ * is at least one sector and more for v2 logs. Hence we need a reservation of
+ * at least 512 bytes per 32k of log space just for the LR headers. That means
+ * 16KB of reservation per megabyte of delayed logging space we will consume,
+ * plus various headers. The number of headers will vary based on the num of
+ * io vectors, so limiting on a specific number of vectors is going to result
+ * in transactions of varying size. IOWs, it is more consistent to track and
+ * limit space consumed in the log rather than by the number of objects being
+ * logged in order to prevent checkpoint ticket overruns.
+ *
+ * Further, use of static reservations through the log grant mechanism is
+ * problematic. It introduces a lot of complexity (e.g. reserve grant vs write
+ * grant) and a significant deadlock potential because regranting write space
+ * can block on log pushes. Hence if we have to regrant log space during a log
+ * push, we can deadlock.
+ *
+ * However, we can avoid this by use of a dynamic "reservation stealing"
+ * technique during transaction commit whereby unused reservation space in the
+ * transaction ticket is transferred to the CIL ctx commit ticket to cover the
+ * space needed by the checkpoint transaction. This means that we never need to
+ * specifically reserve space for the CIL checkpoint transaction, nor do we
+ * need to regrant space once the checkpoint completes. This also means the
+ * checkpoint transaction ticket is specific to the checkpoint context, rather
+ * than the CIL itself.
+ *
+ * With dynamic reservations, we can basically make up arbitrary limits for the
+ * checkpoint size so long as they don't violate any other size rules. Hence
+ * the initial maximum size for the checkpoint transaction will be set to a
+ * quarter of the log or 8MB, which ever is smaller. 8MB is an arbitrary limit
+ * right now based on the latency of writing out a large amount of data through
+ * the circular iclog buffers.
+ */
+
+#define XLOG_CIL_SPACE_LIMIT(log) \
+ (min((log->l_logsize >> 2), (8 * 1024 * 1024)))
+
+/*
* The reservation head lsn is not made up of a cycle number and block number.
* Instead, it uses a cycle number and byte number. Logs don't expect to
* overflow 31 bits worth of byte offset, so using a byte number will mean
OSDN Git Service
