10b61f8a4SDave Chinner // SPDX-License-Identifier: GPL-2.0 21da177e4SLinus Torvalds /* 37b718769SNathan Scott * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc. 47b718769SNathan Scott * All Rights Reserved. 51da177e4SLinus Torvalds */ 61da177e4SLinus Torvalds #ifndef __XFS_LOG_PRIV_H__ 71da177e4SLinus Torvalds #define __XFS_LOG_PRIV_H__ 81da177e4SLinus Torvalds 91da177e4SLinus Torvalds struct xfs_buf; 10ad223e60SMark Tinguely struct xlog; 11a844f451SNathan Scott struct xlog_ticket; 121da177e4SLinus Torvalds struct xfs_mount; 131da177e4SLinus Torvalds 141da177e4SLinus Torvalds /* 15fc06c6d0SDave Chinner * Flags for log structure 161da177e4SLinus Torvalds */ 17fc06c6d0SDave Chinner #define XLOG_ACTIVE_RECOVERY 0x2 /* in the middle of recovery */ 18fc06c6d0SDave Chinner #define XLOG_RECOVERY_NEEDED 0x4 /* log was recovered */ 19fc06c6d0SDave Chinner #define XLOG_IO_ERROR 0x8 /* log hit an I/O error, and being 20fc06c6d0SDave Chinner shutdown */ 21fc06c6d0SDave Chinner #define XLOG_TAIL_WARN 0x10 /* log tail verify warning issued */ 221da177e4SLinus Torvalds 231da177e4SLinus Torvalds /* 241da177e4SLinus Torvalds * get client id from packed copy. 251da177e4SLinus Torvalds * 261da177e4SLinus Torvalds * this hack is here because the xlog_pack code copies four bytes 271da177e4SLinus Torvalds * of xlog_op_header containing the fields oh_clientid, oh_flags 281da177e4SLinus Torvalds * and oh_res2 into the packed copy. 291da177e4SLinus Torvalds * 301da177e4SLinus Torvalds * later on this four byte chunk is treated as an int and the 311da177e4SLinus Torvalds * client id is pulled out. 321da177e4SLinus Torvalds * 331da177e4SLinus Torvalds * this has endian issues, of course. 341da177e4SLinus Torvalds */ 35b53e675dSChristoph Hellwig static inline uint xlog_get_client_id(__be32 i) 3603bea6feSChristoph Hellwig { 37b53e675dSChristoph Hellwig return be32_to_cpu(i) >> 24; 3803bea6feSChristoph Hellwig } 391da177e4SLinus Torvalds 401da177e4SLinus Torvalds /* 411da177e4SLinus Torvalds * In core log state 421da177e4SLinus Torvalds */ 431858bb0bSChristoph Hellwig enum xlog_iclog_state { 441858bb0bSChristoph Hellwig XLOG_STATE_ACTIVE, /* Current IC log being written to */ 451858bb0bSChristoph Hellwig XLOG_STATE_WANT_SYNC, /* Want to sync this iclog; no more writes */ 461858bb0bSChristoph Hellwig XLOG_STATE_SYNCING, /* This IC log is syncing */ 471858bb0bSChristoph Hellwig XLOG_STATE_DONE_SYNC, /* Done syncing to disk */ 481858bb0bSChristoph Hellwig XLOG_STATE_CALLBACK, /* Callback functions now */ 491858bb0bSChristoph Hellwig XLOG_STATE_DIRTY, /* Dirty IC log, not ready for ACTIVE status */ 501858bb0bSChristoph Hellwig XLOG_STATE_IOERROR, /* IO error happened in sync'ing log */ 511858bb0bSChristoph Hellwig }; 521da177e4SLinus Torvalds 531da177e4SLinus Torvalds /* 5470e42f2dSDave Chinner * Log ticket flags 551da177e4SLinus Torvalds */ 5670e42f2dSDave Chinner #define XLOG_TIC_PERM_RESERV 0x1 /* permanent reservation */ 570b1b213fSChristoph Hellwig 580b1b213fSChristoph Hellwig #define XLOG_TIC_FLAGS \ 5910547941SDave Chinner { XLOG_TIC_PERM_RESERV, "XLOG_TIC_PERM_RESERV" } 600b1b213fSChristoph Hellwig 611da177e4SLinus Torvalds /* 621da177e4SLinus Torvalds * Below are states for covering allocation transactions. 631da177e4SLinus Torvalds * By covering, we mean changing the h_tail_lsn in the last on-disk 641da177e4SLinus Torvalds * log write such that no allocation transactions will be re-done during 651da177e4SLinus Torvalds * recovery after a system crash. Recovery starts at the last on-disk 661da177e4SLinus Torvalds * log write. 671da177e4SLinus Torvalds * 681da177e4SLinus Torvalds * These states are used to insert dummy log entries to cover 691da177e4SLinus Torvalds * space allocation transactions which can undo non-transactional changes 701da177e4SLinus Torvalds * after a crash. Writes to a file with space 711da177e4SLinus Torvalds * already allocated do not result in any transactions. Allocations 721da177e4SLinus Torvalds * might include space beyond the EOF. So if we just push the EOF a 731da177e4SLinus Torvalds * little, the last transaction for the file could contain the wrong 741da177e4SLinus Torvalds * size. If there is no file system activity, after an allocation 751da177e4SLinus Torvalds * transaction, and the system crashes, the allocation transaction 761da177e4SLinus Torvalds * will get replayed and the file will be truncated. This could 771da177e4SLinus Torvalds * be hours/days/... after the allocation occurred. 781da177e4SLinus Torvalds * 791da177e4SLinus Torvalds * The fix for this is to do two dummy transactions when the 801da177e4SLinus Torvalds * system is idle. We need two dummy transaction because the h_tail_lsn 811da177e4SLinus Torvalds * in the log record header needs to point beyond the last possible 821da177e4SLinus Torvalds * non-dummy transaction. The first dummy changes the h_tail_lsn to 831da177e4SLinus Torvalds * the first transaction before the dummy. The second dummy causes 841da177e4SLinus Torvalds * h_tail_lsn to point to the first dummy. Recovery starts at h_tail_lsn. 851da177e4SLinus Torvalds * 861da177e4SLinus Torvalds * These dummy transactions get committed when everything 871da177e4SLinus Torvalds * is idle (after there has been some activity). 881da177e4SLinus Torvalds * 891da177e4SLinus Torvalds * There are 5 states used to control this. 901da177e4SLinus Torvalds * 911da177e4SLinus Torvalds * IDLE -- no logging has been done on the file system or 921da177e4SLinus Torvalds * we are done covering previous transactions. 931da177e4SLinus Torvalds * NEED -- logging has occurred and we need a dummy transaction 941da177e4SLinus Torvalds * when the log becomes idle. 951da177e4SLinus Torvalds * DONE -- we were in the NEED state and have committed a dummy 961da177e4SLinus Torvalds * transaction. 971da177e4SLinus Torvalds * NEED2 -- we detected that a dummy transaction has gone to the 981da177e4SLinus Torvalds * on disk log with no other transactions. 991da177e4SLinus Torvalds * DONE2 -- we committed a dummy transaction when in the NEED2 state. 1001da177e4SLinus Torvalds * 1011da177e4SLinus Torvalds * There are two places where we switch states: 1021da177e4SLinus Torvalds * 1031da177e4SLinus Torvalds * 1.) In xfs_sync, when we detect an idle log and are in NEED or NEED2. 1041da177e4SLinus Torvalds * We commit the dummy transaction and switch to DONE or DONE2, 1051da177e4SLinus Torvalds * respectively. In all other states, we don't do anything. 1061da177e4SLinus Torvalds * 1071da177e4SLinus Torvalds * 2.) When we finish writing the on-disk log (xlog_state_clean_log). 1081da177e4SLinus Torvalds * 1091da177e4SLinus Torvalds * No matter what state we are in, if this isn't the dummy 1101da177e4SLinus Torvalds * transaction going out, the next state is NEED. 1111da177e4SLinus Torvalds * So, if we aren't in the DONE or DONE2 states, the next state 1121da177e4SLinus Torvalds * is NEED. We can't be finishing a write of the dummy record 1131da177e4SLinus Torvalds * unless it was committed and the state switched to DONE or DONE2. 1141da177e4SLinus Torvalds * 1151da177e4SLinus Torvalds * If we are in the DONE state and this was a write of the 1161da177e4SLinus Torvalds * dummy transaction, we move to NEED2. 1171da177e4SLinus Torvalds * 1181da177e4SLinus Torvalds * If we are in the DONE2 state and this was a write of the 1191da177e4SLinus Torvalds * dummy transaction, we move to IDLE. 1201da177e4SLinus Torvalds * 1211da177e4SLinus Torvalds * 1221da177e4SLinus Torvalds * Writing only one dummy transaction can get appended to 1231da177e4SLinus Torvalds * one file space allocation. When this happens, the log recovery 1241da177e4SLinus Torvalds * code replays the space allocation and a file could be truncated. 1251da177e4SLinus Torvalds * This is why we have the NEED2 and DONE2 states before going idle. 1261da177e4SLinus Torvalds */ 1271da177e4SLinus Torvalds 1281da177e4SLinus Torvalds #define XLOG_STATE_COVER_IDLE 0 1291da177e4SLinus Torvalds #define XLOG_STATE_COVER_NEED 1 1301da177e4SLinus Torvalds #define XLOG_STATE_COVER_DONE 2 1311da177e4SLinus Torvalds #define XLOG_STATE_COVER_NEED2 3 1321da177e4SLinus Torvalds #define XLOG_STATE_COVER_DONE2 4 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds #define XLOG_COVER_OPS 5 1351da177e4SLinus Torvalds 1367e9c6396STim Shimmin /* Ticket reservation region accounting */ 1377e9c6396STim Shimmin #define XLOG_TIC_LEN_MAX 15 1387e9c6396STim Shimmin 1397e9c6396STim Shimmin /* 1407e9c6396STim Shimmin * Reservation region 1417e9c6396STim Shimmin * As would be stored in xfs_log_iovec but without the i_addr which 1427e9c6396STim Shimmin * we don't care about. 1437e9c6396STim Shimmin */ 1447e9c6396STim Shimmin typedef struct xlog_res { 1451259845dSTim Shimmin uint r_len; /* region length :4 */ 1461259845dSTim Shimmin uint r_type; /* region's transaction type :4 */ 1477e9c6396STim Shimmin } xlog_res_t; 1487e9c6396STim Shimmin 1491da177e4SLinus Torvalds typedef struct xlog_ticket { 15010547941SDave Chinner struct list_head t_queue; /* reserve/write queue */ 15114a7235fSChristoph Hellwig struct task_struct *t_task; /* task that owns this ticket */ 1521da177e4SLinus Torvalds xlog_tid_t t_tid; /* transaction identifier : 4 */ 153cc09c0dcSDave Chinner atomic_t t_ref; /* ticket reference count : 4 */ 1541da177e4SLinus Torvalds int t_curr_res; /* current reservation in bytes : 4 */ 1551da177e4SLinus Torvalds int t_unit_res; /* unit reservation in bytes : 4 */ 1567e9c6396STim Shimmin char t_ocnt; /* original count : 1 */ 1577e9c6396STim Shimmin char t_cnt; /* current count : 1 */ 1587e9c6396STim Shimmin char t_clientid; /* who does this belong to; : 1 */ 1597e9c6396STim Shimmin char t_flags; /* properties of reservation : 1 */ 1607e9c6396STim Shimmin 1617e9c6396STim Shimmin /* reservation array fields */ 1627e9c6396STim Shimmin uint t_res_num; /* num in array : 4 */ 1637e9c6396STim Shimmin uint t_res_num_ophdrs; /* num op hdrs : 4 */ 1647e9c6396STim Shimmin uint t_res_arr_sum; /* array sum : 4 */ 1657e9c6396STim Shimmin uint t_res_o_flow; /* sum overflow : 4 */ 1661259845dSTim Shimmin xlog_res_t t_res_arr[XLOG_TIC_LEN_MAX]; /* array of res : 8 * 15 */ 1671da177e4SLinus Torvalds } xlog_ticket_t; 1687e9c6396STim Shimmin 1691da177e4SLinus Torvalds /* 1701da177e4SLinus Torvalds * - A log record header is 512 bytes. There is plenty of room to grow the 1711da177e4SLinus Torvalds * xlog_rec_header_t into the reserved space. 1721da177e4SLinus Torvalds * - ic_data follows, so a write to disk can start at the beginning of 1731da177e4SLinus Torvalds * the iclog. 17412017fafSDavid Chinner * - ic_forcewait is used to implement synchronous forcing of the iclog to disk. 1751da177e4SLinus Torvalds * - ic_next is the pointer to the next iclog in the ring. 1761da177e4SLinus Torvalds * - ic_log is a pointer back to the global log structure. 17779b54d9bSChristoph Hellwig * - ic_size is the full size of the log buffer, minus the cycle headers. 1781da177e4SLinus Torvalds * - ic_offset is the current number of bytes written to in this iclog. 1791da177e4SLinus Torvalds * - ic_refcnt is bumped when someone is writing to the log. 1801da177e4SLinus Torvalds * - ic_state is the state of the iclog. 181114d23aaSDavid Chinner * 182114d23aaSDavid Chinner * Because of cacheline contention on large machines, we need to separate 183114d23aaSDavid Chinner * various resources onto different cachelines. To start with, make the 184114d23aaSDavid Chinner * structure cacheline aligned. The following fields can be contended on 185114d23aaSDavid Chinner * by independent processes: 186114d23aaSDavid Chinner * 18789ae379dSChristoph Hellwig * - ic_callbacks 188114d23aaSDavid Chinner * - ic_refcnt 189114d23aaSDavid Chinner * - fields protected by the global l_icloglock 190114d23aaSDavid Chinner * 191114d23aaSDavid Chinner * so we need to ensure that these fields are located in separate cachelines. 192114d23aaSDavid Chinner * We'll put all the read-only and l_icloglock fields in the first cacheline, 193114d23aaSDavid Chinner * and move everything else out to subsequent cachelines. 1941da177e4SLinus Torvalds */ 195b28708d6SChristoph Hellwig typedef struct xlog_in_core { 196eb40a875SDave Chinner wait_queue_head_t ic_force_wait; 197eb40a875SDave Chinner wait_queue_head_t ic_write_wait; 1981da177e4SLinus Torvalds struct xlog_in_core *ic_next; 1991da177e4SLinus Torvalds struct xlog_in_core *ic_prev; 200ad223e60SMark Tinguely struct xlog *ic_log; 20179b54d9bSChristoph Hellwig u32 ic_size; 20279b54d9bSChristoph Hellwig u32 ic_offset; 2031858bb0bSChristoph Hellwig enum xlog_iclog_state ic_state; 2041da177e4SLinus Torvalds char *ic_datap; /* pointer to iclog data */ 205114d23aaSDavid Chinner 206114d23aaSDavid Chinner /* Callback structures need their own cacheline */ 207114d23aaSDavid Chinner spinlock_t ic_callback_lock ____cacheline_aligned_in_smp; 20889ae379dSChristoph Hellwig struct list_head ic_callbacks; 209114d23aaSDavid Chinner 210114d23aaSDavid Chinner /* reference counts need their own cacheline */ 211114d23aaSDavid Chinner atomic_t ic_refcnt ____cacheline_aligned_in_smp; 212b28708d6SChristoph Hellwig xlog_in_core_2_t *ic_data; 213b28708d6SChristoph Hellwig #define ic_header ic_data->hic_header 214366fc4b8SChristoph Hellwig #ifdef DEBUG 215366fc4b8SChristoph Hellwig bool ic_fail_crc : 1; 216366fc4b8SChristoph Hellwig #endif 21779b54d9bSChristoph Hellwig struct semaphore ic_sema; 21879b54d9bSChristoph Hellwig struct work_struct ic_end_io_work; 21979b54d9bSChristoph Hellwig struct bio ic_bio; 22079b54d9bSChristoph Hellwig struct bio_vec ic_bvec[]; 2211da177e4SLinus Torvalds } xlog_in_core_t; 2221da177e4SLinus Torvalds 2231da177e4SLinus Torvalds /* 22471e330b5SDave Chinner * The CIL context is used to aggregate per-transaction details as well be 22571e330b5SDave Chinner * passed to the iclog for checkpoint post-commit processing. After being 22671e330b5SDave Chinner * passed to the iclog, another context needs to be allocated for tracking the 22771e330b5SDave Chinner * next set of transactions to be aggregated into a checkpoint. 22871e330b5SDave Chinner */ 22971e330b5SDave Chinner struct xfs_cil; 23071e330b5SDave Chinner 23171e330b5SDave Chinner struct xfs_cil_ctx { 23271e330b5SDave Chinner struct xfs_cil *cil; 23371e330b5SDave Chinner xfs_lsn_t sequence; /* chkpt sequence # */ 23471e330b5SDave Chinner xfs_lsn_t start_lsn; /* first LSN of chkpt commit */ 23571e330b5SDave Chinner xfs_lsn_t commit_lsn; /* chkpt commit record lsn */ 23671e330b5SDave Chinner struct xlog_ticket *ticket; /* chkpt ticket */ 23771e330b5SDave Chinner int nvecs; /* number of regions */ 23871e330b5SDave Chinner int space_used; /* aggregate size of regions */ 23971e330b5SDave Chinner struct list_head busy_extents; /* busy extents in chkpt */ 24071e330b5SDave Chinner struct xfs_log_vec *lv_chain; /* logvecs being pushed */ 24189ae379dSChristoph Hellwig struct list_head iclog_entry; 24271e330b5SDave Chinner struct list_head committing; /* ctx committing list */ 2434560e78fSChristoph Hellwig struct work_struct discard_endio_work; 24471e330b5SDave Chinner }; 24571e330b5SDave Chinner 24671e330b5SDave Chinner /* 24771e330b5SDave Chinner * Committed Item List structure 24871e330b5SDave Chinner * 24971e330b5SDave Chinner * This structure is used to track log items that have been committed but not 25071e330b5SDave Chinner * yet written into the log. It is used only when the delayed logging mount 25171e330b5SDave Chinner * option is enabled. 25271e330b5SDave Chinner * 25371e330b5SDave Chinner * This structure tracks the list of committing checkpoint contexts so 25471e330b5SDave Chinner * we can avoid the problem of having to hold out new transactions during a 25571e330b5SDave Chinner * flush until we have a the commit record LSN of the checkpoint. We can 25671e330b5SDave Chinner * traverse the list of committing contexts in xlog_cil_push_lsn() to find a 25771e330b5SDave Chinner * sequence match and extract the commit LSN directly from there. If the 25871e330b5SDave Chinner * checkpoint is still in the process of committing, we can block waiting for 25971e330b5SDave Chinner * the commit LSN to be determined as well. This should make synchronous 26071e330b5SDave Chinner * operations almost as efficient as the old logging methods. 26171e330b5SDave Chinner */ 26271e330b5SDave Chinner struct xfs_cil { 263ad223e60SMark Tinguely struct xlog *xc_log; 26471e330b5SDave Chinner struct list_head xc_cil; 26571e330b5SDave Chinner spinlock_t xc_cil_lock; 2664bb928cdSDave Chinner 2674bb928cdSDave Chinner struct rw_semaphore xc_ctx_lock ____cacheline_aligned_in_smp; 26871e330b5SDave Chinner struct xfs_cil_ctx *xc_ctx; 2694bb928cdSDave Chinner 2704bb928cdSDave Chinner spinlock_t xc_push_lock ____cacheline_aligned_in_smp; 2714bb928cdSDave Chinner xfs_lsn_t xc_push_seq; 27271e330b5SDave Chinner struct list_head xc_committing; 273eb40a875SDave Chinner wait_queue_head_t xc_commit_wait; 274a44f13edSDave Chinner xfs_lsn_t xc_current_sequence; 2754c2d542fSDave Chinner struct work_struct xc_push_work; 2764bb928cdSDave Chinner } ____cacheline_aligned_in_smp; 27771e330b5SDave Chinner 27871e330b5SDave Chinner /* 27980168676SDave Chinner * The amount of log space we allow the CIL to aggregate is difficult to size. 28080168676SDave Chinner * Whatever we choose, we have to make sure we can get a reservation for the 28180168676SDave Chinner * log space effectively, that it is large enough to capture sufficient 28280168676SDave Chinner * relogging to reduce log buffer IO significantly, but it is not too large for 28380168676SDave Chinner * the log or induces too much latency when writing out through the iclogs. We 28480168676SDave Chinner * track both space consumed and the number of vectors in the checkpoint 28580168676SDave Chinner * context, so we need to decide which to use for limiting. 286df806158SDave Chinner * 287df806158SDave Chinner * Every log buffer we write out during a push needs a header reserved, which 288df806158SDave Chinner * is at least one sector and more for v2 logs. Hence we need a reservation of 289df806158SDave Chinner * at least 512 bytes per 32k of log space just for the LR headers. That means 290df806158SDave Chinner * 16KB of reservation per megabyte of delayed logging space we will consume, 291df806158SDave Chinner * plus various headers. The number of headers will vary based on the num of 292df806158SDave Chinner * io vectors, so limiting on a specific number of vectors is going to result 293df806158SDave Chinner * in transactions of varying size. IOWs, it is more consistent to track and 294df806158SDave Chinner * limit space consumed in the log rather than by the number of objects being 295df806158SDave Chinner * logged in order to prevent checkpoint ticket overruns. 296df806158SDave Chinner * 297df806158SDave Chinner * Further, use of static reservations through the log grant mechanism is 298df806158SDave Chinner * problematic. It introduces a lot of complexity (e.g. reserve grant vs write 299df806158SDave Chinner * grant) and a significant deadlock potential because regranting write space 300df806158SDave Chinner * can block on log pushes. Hence if we have to regrant log space during a log 301df806158SDave Chinner * push, we can deadlock. 302df806158SDave Chinner * 303df806158SDave Chinner * However, we can avoid this by use of a dynamic "reservation stealing" 304df806158SDave Chinner * technique during transaction commit whereby unused reservation space in the 305df806158SDave Chinner * transaction ticket is transferred to the CIL ctx commit ticket to cover the 306df806158SDave Chinner * space needed by the checkpoint transaction. This means that we never need to 307df806158SDave Chinner * specifically reserve space for the CIL checkpoint transaction, nor do we 308df806158SDave Chinner * need to regrant space once the checkpoint completes. This also means the 309df806158SDave Chinner * checkpoint transaction ticket is specific to the checkpoint context, rather 310df806158SDave Chinner * than the CIL itself. 311df806158SDave Chinner * 31280168676SDave Chinner * With dynamic reservations, we can effectively make up arbitrary limits for 31380168676SDave Chinner * the checkpoint size so long as they don't violate any other size rules. 31480168676SDave Chinner * Recovery imposes a rule that no transaction exceed half the log, so we are 31580168676SDave Chinner * limited by that. Furthermore, the log transaction reservation subsystem 31680168676SDave Chinner * tries to keep 25% of the log free, so we need to keep below that limit or we 31780168676SDave Chinner * risk running out of free log space to start any new transactions. 31880168676SDave Chinner * 319*108a4235SDave Chinner * In order to keep background CIL push efficient, we only need to ensure the 320*108a4235SDave Chinner * CIL is large enough to maintain sufficient in-memory relogging to avoid 321*108a4235SDave Chinner * repeated physical writes of frequently modified metadata. If we allow the CIL 322*108a4235SDave Chinner * to grow to a substantial fraction of the log, then we may be pinning hundreds 323*108a4235SDave Chinner * of megabytes of metadata in memory until the CIL flushes. This can cause 324*108a4235SDave Chinner * issues when we are running low on memory - pinned memory cannot be reclaimed, 325*108a4235SDave Chinner * and the CIL consumes a lot of memory. Hence we need to set an upper physical 326*108a4235SDave Chinner * size limit for the CIL that limits the maximum amount of memory pinned by the 327*108a4235SDave Chinner * CIL but does not limit performance by reducing relogging efficiency 328*108a4235SDave Chinner * significantly. 329*108a4235SDave Chinner * 330*108a4235SDave Chinner * As such, the CIL push threshold ends up being the smaller of two thresholds: 331*108a4235SDave Chinner * - a threshold large enough that it allows CIL to be pushed and progress to be 332*108a4235SDave Chinner * made without excessive blocking of incoming transaction commits. This is 333*108a4235SDave Chinner * defined to be 12.5% of the log space - half the 25% push threshold of the 334*108a4235SDave Chinner * AIL. 335*108a4235SDave Chinner * - small enough that it doesn't pin excessive amounts of memory but maintains 336*108a4235SDave Chinner * close to peak relogging efficiency. This is defined to be 16x the iclog 337*108a4235SDave Chinner * buffer window (32MB) as measurements have shown this to be roughly the 338*108a4235SDave Chinner * point of diminishing performance increases under highly concurrent 339*108a4235SDave Chinner * modification workloads. 340df806158SDave Chinner */ 341*108a4235SDave Chinner #define XLOG_CIL_SPACE_LIMIT(log) \ 342*108a4235SDave Chinner min_t(int, (log)->l_logsize >> 3, BBTOB(XLOG_TOTAL_REC_SHIFT(log)) << 4) 343df806158SDave Chinner 344df806158SDave Chinner /* 34528496968SChristoph Hellwig * ticket grant locks, queues and accounting have their own cachlines 34628496968SChristoph Hellwig * as these are quite hot and can be operated on concurrently. 34728496968SChristoph Hellwig */ 34828496968SChristoph Hellwig struct xlog_grant_head { 34928496968SChristoph Hellwig spinlock_t lock ____cacheline_aligned_in_smp; 35028496968SChristoph Hellwig struct list_head waiters; 35128496968SChristoph Hellwig atomic64_t grant; 35228496968SChristoph Hellwig }; 35328496968SChristoph Hellwig 35428496968SChristoph Hellwig /* 3551da177e4SLinus Torvalds * The reservation head lsn is not made up of a cycle number and block number. 3561da177e4SLinus Torvalds * Instead, it uses a cycle number and byte number. Logs don't expect to 3571da177e4SLinus Torvalds * overflow 31 bits worth of byte offset, so using a byte number will mean 3581da177e4SLinus Torvalds * that round off problems won't occur when releasing partial reservations. 3591da177e4SLinus Torvalds */ 3609a8d2fdbSMark Tinguely struct xlog { 3614679b2d3SDavid Chinner /* The following fields don't need locking */ 3624679b2d3SDavid Chinner struct xfs_mount *l_mp; /* mount point */ 363a9c21c1bSDavid Chinner struct xfs_ail *l_ailp; /* AIL log is working with */ 36471e330b5SDave Chinner struct xfs_cil *l_cilp; /* CIL log is working with */ 3654679b2d3SDavid Chinner struct xfs_buftarg *l_targ; /* buftarg of log */ 3661058d0f5SChristoph Hellwig struct workqueue_struct *l_ioend_workqueue; /* for I/O completions */ 367f661f1e0SDave Chinner struct delayed_work l_work; /* background flush work */ 3684679b2d3SDavid Chinner uint l_flags; 3694679b2d3SDavid Chinner uint l_quotaoffs_flag; /* XFS_DQ_*, for QUOTAOFFs */ 370d5689eaaSChristoph Hellwig struct list_head *l_buf_cancel_table; 3714679b2d3SDavid Chinner int l_iclog_hsize; /* size of iclog header */ 3724679b2d3SDavid Chinner int l_iclog_heads; /* # of iclog header sectors */ 37348389ef1SAlex Elder uint l_sectBBsize; /* sector size in BBs (2^n) */ 3744679b2d3SDavid Chinner int l_iclog_size; /* size of log in bytes */ 3754679b2d3SDavid Chinner int l_iclog_bufs; /* number of iclog buffers */ 3764679b2d3SDavid Chinner xfs_daddr_t l_logBBstart; /* start block of log */ 3774679b2d3SDavid Chinner int l_logsize; /* size of log in bytes */ 3784679b2d3SDavid Chinner int l_logBBsize; /* size of log in BB chunks */ 3794679b2d3SDavid Chinner 3801da177e4SLinus Torvalds /* The following block of fields are changed while holding icloglock */ 381eb40a875SDave Chinner wait_queue_head_t l_flush_wait ____cacheline_aligned_in_smp; 382d748c623SMatthew Wilcox /* waiting for iclog flush */ 3831da177e4SLinus Torvalds int l_covered_state;/* state of "covering disk 3841da177e4SLinus Torvalds * log entries" */ 3851da177e4SLinus Torvalds xlog_in_core_t *l_iclog; /* head log queue */ 386b22cd72cSEric Sandeen spinlock_t l_icloglock; /* grab to change iclog state */ 3871da177e4SLinus Torvalds int l_curr_cycle; /* Cycle number of log writes */ 3881da177e4SLinus Torvalds int l_prev_cycle; /* Cycle number before last 3891da177e4SLinus Torvalds * block increment */ 3901da177e4SLinus Torvalds int l_curr_block; /* current logical log block */ 3911da177e4SLinus Torvalds int l_prev_block; /* previous logical log block */ 3921da177e4SLinus Torvalds 39384f3c683SDave Chinner /* 3941c3cb9ecSDave Chinner * l_last_sync_lsn and l_tail_lsn are atomics so they can be set and 3951c3cb9ecSDave Chinner * read without needing to hold specific locks. To avoid operations 3961c3cb9ecSDave Chinner * contending with other hot objects, place each of them on a separate 3971c3cb9ecSDave Chinner * cacheline. 39884f3c683SDave Chinner */ 39984f3c683SDave Chinner /* lsn of last LR on disk */ 40084f3c683SDave Chinner atomic64_t l_last_sync_lsn ____cacheline_aligned_in_smp; 4011c3cb9ecSDave Chinner /* lsn of 1st LR with unflushed * buffers */ 4021c3cb9ecSDave Chinner atomic64_t l_tail_lsn ____cacheline_aligned_in_smp; 40384f3c683SDave Chinner 40428496968SChristoph Hellwig struct xlog_grant_head l_reserve_head; 40528496968SChristoph Hellwig struct xlog_grant_head l_write_head; 4063f16b985SDave Chinner 407baff4e44SBrian Foster struct xfs_kobj l_kobj; 408baff4e44SBrian Foster 4094679b2d3SDavid Chinner /* The following field are used for debugging; need to hold icloglock */ 4104679b2d3SDavid Chinner #ifdef DEBUG 4115809d5e0SChristoph Hellwig void *l_iclog_bak[XLOG_MAX_ICLOGS]; 4124679b2d3SDavid Chinner #endif 41312818d24SBrian Foster /* log recovery lsn tracking (for buffer submission */ 41412818d24SBrian Foster xfs_lsn_t l_recovery_lsn; 4159a8d2fdbSMark Tinguely }; 4161da177e4SLinus Torvalds 417d5689eaaSChristoph Hellwig #define XLOG_BUF_CANCEL_BUCKET(log, blkno) \ 418c8ce540dSDarrick J. Wong ((log)->l_buf_cancel_table + ((uint64_t)blkno % XLOG_BC_TABLE_SIZE)) 419d5689eaaSChristoph Hellwig 420b941c719SChristoph Hellwig #define XLOG_FORCED_SHUTDOWN(log) \ 421b941c719SChristoph Hellwig (unlikely((log)->l_flags & XLOG_IO_ERROR)) 422cfcbbbd0SNathan Scott 4231da177e4SLinus Torvalds /* common routines */ 4249a8d2fdbSMark Tinguely extern int 4259a8d2fdbSMark Tinguely xlog_recover( 4269a8d2fdbSMark Tinguely struct xlog *log); 4279a8d2fdbSMark Tinguely extern int 4289a8d2fdbSMark Tinguely xlog_recover_finish( 4299a8d2fdbSMark Tinguely struct xlog *log); 430a7a9250eSHariprasad Kelam extern void 431f0b2efadSBrian Foster xlog_recover_cancel(struct xlog *); 4320e446be4SChristoph Hellwig 433f9668a09SDave Chinner extern __le32 xlog_cksum(struct xlog *log, struct xlog_rec_header *rhead, 4340e446be4SChristoph Hellwig char *dp, int size); 4351da177e4SLinus Torvalds 436eb01c9cdSDavid Chinner extern kmem_zone_t *xfs_log_ticket_zone; 437ad223e60SMark Tinguely struct xlog_ticket * 438ad223e60SMark Tinguely xlog_ticket_alloc( 439ad223e60SMark Tinguely struct xlog *log, 440ad223e60SMark Tinguely int unit_bytes, 441ad223e60SMark Tinguely int count, 442ad223e60SMark Tinguely char client, 443ad223e60SMark Tinguely bool permanent, 44477ba7877SAl Viro xfs_km_flags_t alloc_flags); 44571e330b5SDave Chinner 446eb01c9cdSDavid Chinner 447e6b1f273SChristoph Hellwig static inline void 448e6b1f273SChristoph Hellwig xlog_write_adv_cnt(void **ptr, int *len, int *off, size_t bytes) 449e6b1f273SChristoph Hellwig { 450e6b1f273SChristoph Hellwig *ptr += bytes; 451e6b1f273SChristoph Hellwig *len -= bytes; 452e6b1f273SChristoph Hellwig *off += bytes; 453e6b1f273SChristoph Hellwig } 454e6b1f273SChristoph Hellwig 45571e330b5SDave Chinner void xlog_print_tic_res(struct xfs_mount *mp, struct xlog_ticket *ticket); 456d4ca1d55SBrian Foster void xlog_print_trans(struct xfs_trans *); 4577ec94921SDave Chinner int xlog_write(struct xlog *log, struct xfs_log_vec *log_vector, 4587ec94921SDave Chinner struct xlog_ticket *tic, xfs_lsn_t *start_lsn, 4597ec94921SDave Chinner struct xlog_in_core **commit_iclog, uint flags, 4607ec94921SDave Chinner bool need_start_rec); 461f10e925dSDave Chinner int xlog_commit_record(struct xlog *log, struct xlog_ticket *ticket, 462dd401770SDave Chinner struct xlog_in_core **iclog, xfs_lsn_t *lsn); 4638b41e3f9SChristoph Hellwig void xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket); 4648b41e3f9SChristoph Hellwig void xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket); 46571e330b5SDave Chinner 46671e330b5SDave Chinner /* 4671c3cb9ecSDave Chinner * When we crack an atomic LSN, we sample it first so that the value will not 4681c3cb9ecSDave Chinner * change while we are cracking it into the component values. This means we 4691c3cb9ecSDave Chinner * will always get consistent component values to work from. This should always 47025985edcSLucas De Marchi * be used to sample and crack LSNs that are stored and updated in atomic 4711c3cb9ecSDave Chinner * variables. 4721c3cb9ecSDave Chinner */ 4731c3cb9ecSDave Chinner static inline void 4741c3cb9ecSDave Chinner xlog_crack_atomic_lsn(atomic64_t *lsn, uint *cycle, uint *block) 4751c3cb9ecSDave Chinner { 4761c3cb9ecSDave Chinner xfs_lsn_t val = atomic64_read(lsn); 4771c3cb9ecSDave Chinner 4781c3cb9ecSDave Chinner *cycle = CYCLE_LSN(val); 4791c3cb9ecSDave Chinner *block = BLOCK_LSN(val); 4801c3cb9ecSDave Chinner } 4811c3cb9ecSDave Chinner 4821c3cb9ecSDave Chinner /* 4831c3cb9ecSDave Chinner * Calculate and assign a value to an atomic LSN variable from component pieces. 4841c3cb9ecSDave Chinner */ 4851c3cb9ecSDave Chinner static inline void 4861c3cb9ecSDave Chinner xlog_assign_atomic_lsn(atomic64_t *lsn, uint cycle, uint block) 4871c3cb9ecSDave Chinner { 4881c3cb9ecSDave Chinner atomic64_set(lsn, xlog_assign_lsn(cycle, block)); 4891c3cb9ecSDave Chinner } 4901c3cb9ecSDave Chinner 4911c3cb9ecSDave Chinner /* 492d0eb2f38SDave Chinner * When we crack the grant head, we sample it first so that the value will not 493a69ed03cSDave Chinner * change while we are cracking it into the component values. This means we 494a69ed03cSDave Chinner * will always get consistent component values to work from. 495a69ed03cSDave Chinner */ 496a69ed03cSDave Chinner static inline void 497d0eb2f38SDave Chinner xlog_crack_grant_head_val(int64_t val, int *cycle, int *space) 498a69ed03cSDave Chinner { 499a69ed03cSDave Chinner *cycle = val >> 32; 500a69ed03cSDave Chinner *space = val & 0xffffffff; 501a69ed03cSDave Chinner } 502a69ed03cSDave Chinner 503a69ed03cSDave Chinner static inline void 504d0eb2f38SDave Chinner xlog_crack_grant_head(atomic64_t *head, int *cycle, int *space) 505d0eb2f38SDave Chinner { 506d0eb2f38SDave Chinner xlog_crack_grant_head_val(atomic64_read(head), cycle, space); 507d0eb2f38SDave Chinner } 508d0eb2f38SDave Chinner 509d0eb2f38SDave Chinner static inline int64_t 510d0eb2f38SDave Chinner xlog_assign_grant_head_val(int cycle, int space) 511d0eb2f38SDave Chinner { 512d0eb2f38SDave Chinner return ((int64_t)cycle << 32) | space; 513d0eb2f38SDave Chinner } 514d0eb2f38SDave Chinner 515d0eb2f38SDave Chinner static inline void 516c8a09ff8SDave Chinner xlog_assign_grant_head(atomic64_t *head, int cycle, int space) 517a69ed03cSDave Chinner { 518d0eb2f38SDave Chinner atomic64_set(head, xlog_assign_grant_head_val(cycle, space)); 519a69ed03cSDave Chinner } 520a69ed03cSDave Chinner 521a69ed03cSDave Chinner /* 52271e330b5SDave Chinner * Committed Item List interfaces 52371e330b5SDave Chinner */ 5242c6e24ceSDave Chinner int xlog_cil_init(struct xlog *log); 5252c6e24ceSDave Chinner void xlog_cil_init_post_recovery(struct xlog *log); 5262c6e24ceSDave Chinner void xlog_cil_destroy(struct xlog *log); 5272c6e24ceSDave Chinner bool xlog_cil_empty(struct xlog *log); 52871e330b5SDave Chinner 529a44f13edSDave Chinner /* 530a44f13edSDave Chinner * CIL force routines 531a44f13edSDave Chinner */ 532ad223e60SMark Tinguely xfs_lsn_t 533ad223e60SMark Tinguely xlog_cil_force_lsn( 534ad223e60SMark Tinguely struct xlog *log, 535ad223e60SMark Tinguely xfs_lsn_t sequence); 536a44f13edSDave Chinner 537a44f13edSDave Chinner static inline void 538ad223e60SMark Tinguely xlog_cil_force(struct xlog *log) 539a44f13edSDave Chinner { 540a44f13edSDave Chinner xlog_cil_force_lsn(log, log->l_cilp->xc_current_sequence); 541a44f13edSDave Chinner } 54271e330b5SDave Chinner 543955e47adSTim Shimmin /* 544eb40a875SDave Chinner * Wrapper function for waiting on a wait queue serialised against wakeups 545eb40a875SDave Chinner * by a spinlock. This matches the semantics of all the wait queues used in the 546eb40a875SDave Chinner * log code. 547eb40a875SDave Chinner */ 548f7559793SDarrick J. Wong static inline void 549f7559793SDarrick J. Wong xlog_wait( 550f7559793SDarrick J. Wong struct wait_queue_head *wq, 551f7559793SDarrick J. Wong struct spinlock *lock) 552f7559793SDarrick J. Wong __releases(lock) 553eb40a875SDave Chinner { 554eb40a875SDave Chinner DECLARE_WAITQUEUE(wait, current); 555eb40a875SDave Chinner 556eb40a875SDave Chinner add_wait_queue_exclusive(wq, &wait); 557eb40a875SDave Chinner __set_current_state(TASK_UNINTERRUPTIBLE); 558eb40a875SDave Chinner spin_unlock(lock); 559eb40a875SDave Chinner schedule(); 560eb40a875SDave Chinner remove_wait_queue(wq, &wait); 561eb40a875SDave Chinner } 5621da177e4SLinus Torvalds 563a45086e2SBrian Foster /* 564a45086e2SBrian Foster * The LSN is valid so long as it is behind the current LSN. If it isn't, this 565a45086e2SBrian Foster * means that the next log record that includes this metadata could have a 566a45086e2SBrian Foster * smaller LSN. In turn, this means that the modification in the log would not 567a45086e2SBrian Foster * replay. 568a45086e2SBrian Foster */ 569a45086e2SBrian Foster static inline bool 570a45086e2SBrian Foster xlog_valid_lsn( 571a45086e2SBrian Foster struct xlog *log, 572a45086e2SBrian Foster xfs_lsn_t lsn) 573a45086e2SBrian Foster { 574a45086e2SBrian Foster int cur_cycle; 575a45086e2SBrian Foster int cur_block; 576a45086e2SBrian Foster bool valid = true; 577a45086e2SBrian Foster 578a45086e2SBrian Foster /* 579a45086e2SBrian Foster * First, sample the current lsn without locking to avoid added 580a45086e2SBrian Foster * contention from metadata I/O. The current cycle and block are updated 581a45086e2SBrian Foster * (in xlog_state_switch_iclogs()) and read here in a particular order 582a45086e2SBrian Foster * to avoid false negatives (e.g., thinking the metadata LSN is valid 583a45086e2SBrian Foster * when it is not). 584a45086e2SBrian Foster * 585a45086e2SBrian Foster * The current block is always rewound before the cycle is bumped in 586a45086e2SBrian Foster * xlog_state_switch_iclogs() to ensure the current LSN is never seen in 587a45086e2SBrian Foster * a transiently forward state. Instead, we can see the LSN in a 588a45086e2SBrian Foster * transiently behind state if we happen to race with a cycle wrap. 589a45086e2SBrian Foster */ 5906aa7de05SMark Rutland cur_cycle = READ_ONCE(log->l_curr_cycle); 591a45086e2SBrian Foster smp_rmb(); 5926aa7de05SMark Rutland cur_block = READ_ONCE(log->l_curr_block); 593a45086e2SBrian Foster 594a45086e2SBrian Foster if ((CYCLE_LSN(lsn) > cur_cycle) || 595a45086e2SBrian Foster (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) { 596a45086e2SBrian Foster /* 597a45086e2SBrian Foster * If the metadata LSN appears invalid, it's possible the check 598a45086e2SBrian Foster * above raced with a wrap to the next log cycle. Grab the lock 599a45086e2SBrian Foster * to check for sure. 600a45086e2SBrian Foster */ 601a45086e2SBrian Foster spin_lock(&log->l_icloglock); 602a45086e2SBrian Foster cur_cycle = log->l_curr_cycle; 603a45086e2SBrian Foster cur_block = log->l_curr_block; 604a45086e2SBrian Foster spin_unlock(&log->l_icloglock); 605a45086e2SBrian Foster 606a45086e2SBrian Foster if ((CYCLE_LSN(lsn) > cur_cycle) || 607a45086e2SBrian Foster (CYCLE_LSN(lsn) == cur_cycle && BLOCK_LSN(lsn) > cur_block)) 608a45086e2SBrian Foster valid = false; 609a45086e2SBrian Foster } 610a45086e2SBrian Foster 611a45086e2SBrian Foster return valid; 612a45086e2SBrian Foster } 613a45086e2SBrian Foster 6141da177e4SLinus Torvalds #endif /* __XFS_LOG_PRIV_H__ */ 615