1 /* $NetBSD: rf_raid.h,v 1.8 2000/01/05 02:57:29 oster Exp $ */ 2 /* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: Mark Holland 7 * 8 * Permission to use, copy, modify and distribute this software and 9 * its documentation is hereby granted, provided that both the copyright 10 * notice and this permission notice appear in all copies of the 11 * software, derivative works or modified versions, and any portions 12 * thereof, and that both notices appear in supporting documentation. 13 * 14 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 15 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 16 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 17 * 18 * Carnegie Mellon requests users of this software to return to 19 * 20 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 21 * School of Computer Science 22 * Carnegie Mellon University 23 * Pittsburgh PA 15213-3890 24 * 25 * any improvements or extensions that they make and grant Carnegie the 26 * rights to redistribute these changes. 27 */ 28 29 /********************************************** 30 * rf_raid.h -- main header file for RAID driver 31 **********************************************/ 32 33 34 #ifndef _RF__RF_RAID_H_ 35 #define _RF__RF_RAID_H_ 36 37 #include "rf_archs.h" 38 #include "rf_types.h" 39 #include "rf_threadstuff.h" 40 41 #include "rf_netbsd.h" 42 43 #include <sys/disklabel.h> 44 #include <sys/types.h> 45 46 #include "rf_alloclist.h" 47 #include "rf_stripelocks.h" 48 #include "rf_layout.h" 49 #include "rf_disks.h" 50 #include "rf_debugMem.h" 51 #include "rf_diskqueue.h" 52 #include "rf_reconstruct.h" 53 #include "rf_acctrace.h" 54 55 #if RF_INCLUDE_PARITYLOGGING > 0 56 #include "rf_paritylog.h" 57 #endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 58 59 #define RF_MAX_DISKS 128 /* max disks per array */ 60 #define RF_DEV2RAIDID(_dev) (DISKUNIT(_dev)) 61 62 #define RF_COMPONENT_LABEL_VERSION 1 63 #define RF_RAID_DIRTY 0 64 #define RF_RAID_CLEAN 1 65 66 67 /* 68 * Each row in the array is a distinct parity group, so 69 * each has it's own status, which is one of the following. 70 */ 71 typedef enum RF_RowStatus_e { 72 rf_rs_optimal, 73 rf_rs_degraded, 74 rf_rs_reconstructing, 75 rf_rs_reconfigured 76 } RF_RowStatus_t; 77 78 struct RF_CumulativeStats_s { 79 struct timeval start; /* the time when the stats were last started */ 80 struct timeval stop; /* the time when the stats were last stopped */ 81 long sum_io_us; /* sum of all user response times (us) */ 82 long num_ios; /* total number of I/Os serviced */ 83 long num_sect_moved; /* total number of sectors read or written */ 84 }; 85 86 struct RF_ThroughputStats_s { 87 RF_DECLARE_MUTEX(mutex) /* a mutex used to lock the configuration 88 * stuff */ 89 struct timeval start; /* timer started when numOutstandingRequests 90 * moves from 0 to 1 */ 91 struct timeval stop; /* timer stopped when numOutstandingRequests 92 * moves from 1 to 0 */ 93 RF_uint64 sum_io_us; /* total time timer is enabled */ 94 RF_uint64 num_ios; /* total number of ios processed by RAIDframe */ 95 long num_out_ios; /* number of outstanding ios */ 96 }; 97 98 struct RF_Raid_s { 99 /* This portion never changes, and can be accessed without locking */ 100 /* an exception is Disks[][].status, which requires locking when it is 101 * changed. XXX this is no longer true. numSpare and friends can 102 * change now. 103 */ 104 u_int numRow; /* number of rows of disks, typically == # of 105 * ranks */ 106 u_int numCol; /* number of columns of disks, typically == # 107 * of disks/rank */ 108 u_int numSpare; /* number of spare disks */ 109 int maxQueueDepth; /* max disk queue depth */ 110 RF_SectorCount_t totalSectors; /* total number of sectors in the 111 * array */ 112 RF_SectorCount_t sectorsPerDisk; /* number of sectors on each 113 * disk */ 114 u_int logBytesPerSector; /* base-2 log of the number of bytes 115 * in a sector */ 116 u_int bytesPerSector; /* bytes in a sector */ 117 RF_int32 sectorMask; /* mask of bytes-per-sector */ 118 119 RF_RaidLayout_t Layout; /* all information related to layout */ 120 RF_RaidDisk_t **Disks; /* all information related to physical disks */ 121 RF_DiskQueue_t **Queues;/* all information related to disk queues */ 122 /* NOTE: This is an anchor point via which the queues can be 123 * accessed, but the enqueue/dequeue routines in diskqueue.c use a 124 * local copy of this pointer for the actual accesses. */ 125 /* The remainder of the structure can change, and therefore requires 126 * locking on reads and updates */ 127 RF_DECLARE_MUTEX(mutex) /* mutex used to serialize access to 128 * the fields below */ 129 RF_RowStatus_t *status; /* the status of each row in the array */ 130 int valid; /* indicates successful configuration */ 131 RF_LockTableEntry_t *lockTable; /* stripe-lock table */ 132 RF_LockTableEntry_t *quiesceLock; /* quiesnce table */ 133 int numFailures; /* total number of failures in the array */ 134 135 int parity_good; /* !0 if parity is known to be correct */ 136 int serial_number; /* a "serial number" for this set */ 137 int mod_counter; /* modification counter for component labels */ 138 int clean; /* the clean bit for this array. */ 139 140 int openings; /* Number of IO's which can be scheduled 141 simultaneously (high-level - not a 142 per-component limit)*/ 143 144 /* 145 * Cleanup stuff 146 */ 147 RF_ShutdownList_t *shutdownList; /* shutdown activities */ 148 RF_AllocListElem_t *cleanupList; /* memory to be freed at 149 * shutdown time */ 150 151 /* 152 * Recon stuff 153 */ 154 RF_HeadSepLimit_t headSepLimit; 155 int numFloatingReconBufs; 156 int reconInProgress; 157 RF_DECLARE_COND(waitForReconCond) 158 RF_RaidReconDesc_t *reconDesc; /* reconstruction descriptor */ 159 RF_ReconCtrl_t **reconControl; /* reconstruction control structure 160 * pointers for each row in the array */ 161 162 /* 163 * Array-quiescence stuff 164 */ 165 RF_DECLARE_MUTEX(access_suspend_mutex) 166 RF_DECLARE_COND(quiescent_cond) 167 RF_IoCount_t accesses_suspended; 168 RF_IoCount_t accs_in_flight; 169 int access_suspend_release; 170 int waiting_for_quiescence; 171 RF_CallbackDesc_t *quiesce_wait_list; 172 173 /* 174 * Statistics 175 */ 176 #if !defined(_KERNEL) && !defined(SIMULATE) 177 RF_ThroughputStats_t throughputstats; 178 #endif /* !KERNEL && !SIMULATE */ 179 RF_CumulativeStats_t userstats; 180 int parity_rewrite_stripes_done; 181 int recon_stripes_done; 182 int copyback_stripes_done; 183 184 int recon_in_progress; 185 int parity_rewrite_in_progress; 186 int copyback_in_progress; 187 188 /* 189 * Engine thread control 190 */ 191 RF_DECLARE_MUTEX(node_queue_mutex) 192 RF_DECLARE_COND(node_queue_cond) 193 RF_DagNode_t *node_queue; 194 RF_Thread_t parity_rewrite_thread; 195 RF_Thread_t copyback_thread; 196 RF_Thread_t engine_thread; 197 RF_Thread_t recon_thread; 198 RF_ThreadGroup_t engine_tg; 199 int shutdown_engine; 200 int dags_in_flight; /* debug */ 201 202 /* 203 * PSS (Parity Stripe Status) stuff 204 */ 205 RF_FreeList_t *pss_freelist; 206 long pssTableSize; 207 208 /* 209 * Reconstruction stuff 210 */ 211 int procsInBufWait; 212 int numFullReconBuffers; 213 RF_AccTraceEntry_t *recon_tracerecs; 214 unsigned long accumXorTimeUs; 215 RF_ReconDoneProc_t *recon_done_procs; 216 RF_DECLARE_MUTEX(recon_done_proc_mutex) 217 /* 218 * nAccOutstanding, waitShutdown protected by desc freelist lock 219 * (This may seem strange, since that's a central serialization point 220 * for a per-array piece of data, but otherwise, it'd be an extra 221 * per-array lock, and that'd only be less efficient...) 222 */ 223 RF_DECLARE_COND(outstandingCond) 224 int waitShutdown; 225 int nAccOutstanding; 226 227 RF_DiskId_t **diskids; 228 RF_DiskId_t *sparediskids; 229 230 int raidid; 231 RF_AccTotals_t acc_totals; 232 int keep_acc_totals; 233 234 struct raidcinfo **raid_cinfo; /* array of component info */ 235 236 int terminate_disk_queues; 237 238 /* 239 * XXX 240 * 241 * config-specific information should be moved 242 * somewhere else, or at least hung off this 243 * in some generic way 244 */ 245 246 /* used by rf_compute_workload_shift */ 247 RF_RowCol_t hist_diskreq[RF_MAXROW][RF_MAXCOL]; 248 249 /* used by declustering */ 250 int noRotate; 251 252 #if RF_INCLUDE_PARITYLOGGING > 0 253 /* used by parity logging */ 254 RF_SectorCount_t regionLogCapacity; 255 RF_ParityLogQueue_t parityLogPool; /* pool of unused parity logs */ 256 RF_RegionInfo_t *regionInfo; /* array of region state */ 257 int numParityLogs; 258 int numSectorsPerLog; 259 int regionParityRange; 260 int logsInUse; /* debugging */ 261 RF_ParityLogDiskQueue_t parityLogDiskQueue; /* state of parity 262 * logging disk work */ 263 RF_RegionBufferQueue_t regionBufferPool; /* buffers for holding 264 * region log */ 265 RF_RegionBufferQueue_t parityBufferPool; /* buffers for holding 266 * parity */ 267 caddr_t parityLogBufferHeap; /* pool of unused parity logs */ 268 RF_Thread_t pLogDiskThreadHandle; 269 270 #endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 271 }; 272 #endif /* !_RF__RF_RAID_H_ */ 273