fs/xfs/xfs_log_priv.h

63  * By covering, we mean changing the h_tail_lsn in the last on-disk
72  * might include space beyond the EOF. So if we just push the EOF a
80  * system is idle. We need two dummy transaction because the h_tail_lsn
92  *		we are done covering previous transactions.
93  *  NEED -- logging has occurred and we need a dummy transaction
95  *  DONE -- we were in the NEED state and have committed a dummy
97  *  NEED2 -- we detected that a dummy transaction has gone to the
99  *  DONE2 -- we committed a dummy transaction when in the NEED2 state.
101  * There are two places where we switch states:
103  * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2.
104  *	We commit the dummy transaction and switch to DONE or DONE2,
105  *	respectively. In all other states, we don't do anything.
107  * 2.) When we finish writing the on-disk log (xlog_state_clean_log).
109  *	No matter what state we are in, if this isn't the dummy
111  *	So, if we aren't in the DONE or DONE2 states, the next state
112  *	is NEED. We can't be finishing a write of the dummy record
115  *	If we are in the DONE state and this was a write of the
116  *		dummy transaction, we move to NEED2.
118  *	If we are in the DONE2 state and this was a write of the
119  *		dummy transaction, we move to IDLE.
125  * This is why we have the NEED2 and DONE2 states before going idle.
142  * we don't care about.
182  * Because of cacheline contention on large machines, we need to separate
191  * so we need to ensure that these fields are located in separate cachelines.
192  * We'll put all the read-only and l_icloglock fields in the first cacheline,
254  * we can avoid the problem of having to hold out new transactions during a
255  * flush until we have a the commit record LSN of the checkpoint. We can
258  * checkpoint is still in the process of committing, we can block waiting for
280  * The amount of log space we allow the CIL to aggregate is difficult to size.
281  * Whatever we choose, we have to make sure we can get a reservation for the
284  * the log or induces too much latency when writing out through the iclogs. We
286  * context, so we need to decide which to use for limiting.
288  * Every log buffer we write out during a push needs a header reserved, which
289  * is at least one sector and more for v2 logs. Hence we need a reservation of
291  * 16KB of reservation per megabyte of delayed logging space we will consume,
301  * can block on log pushes. Hence if we have to regrant log space during a log
302  * push, we can deadlock.
304  * However, we can avoid this by use of a dynamic "reservation stealing"
307  * space needed by the checkpoint transaction. This means that we never need to
308  * specifically reserve space for the CIL checkpoint transaction, nor do we
313  * With dynamic reservations, we can effectively make up arbitrary limits for
315  * Recovery imposes a rule that no transaction exceed half the log, so we are
317  * tries to keep 25% of the log free, so we need to keep below that limit or we
320  * In order to keep background CIL push efficient, we only need to ensure the
322  * repeated physical writes of frequently modified metadata. If we allow the CIL
323  * to grow to a substantial fraction of the log, then we may be pinning hundreds
325  * issues when we are running low on memory - pinned memory cannot be reclaimed,
326  * and the CIL consumes a lot of memory. Hence we need to set an upper physical
342  * To prevent the CIL from overflowing upper commit size bounds, we introduce a
343  * new threshold at which we block committing transactions until the background
348  * the CIL because it is not yielding the CPU. We set the blocking limit at
349  * twice the background push space threshold so we keep in line with the AIL
356  * in the CIL commit fast path, and it's not at all clear that we actually need
357  * such hard limits given the ~7 years we've run without a hard limit before
360  * we've overrun the max size.
489  * When we crack an atomic LSN, we sample it first so that the value will not
490  * change while we are cracking it into the component values. This means we
514  * When we crack the grant head, we sample it first so that the value will not
515  * change while we are cracking it into the component values. This means we
609 	 * a transiently forward state. Instead, we can see the LSN in a  in xlog_valid_lsn()
610 	 * transiently behind state if we happen to race with a cycle wrap.  in xlog_valid_lsn()