/* $NetBSD: rf_diskqueue.c,v 1.53 2011/05/05 06:04:09 mrg Exp $ */ /* * Copyright (c) 1995 Carnegie-Mellon University. * All rights reserved. * * Author: Mark Holland * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /**************************************************************************** * * rf_diskqueue.c -- higher-level disk queue code * * the routines here are a generic wrapper around the actual queueing * routines. The code here implements thread scheduling, synchronization, * and locking ops (see below) on top of the lower-level queueing code. * * to support atomic RMW, we implement "locking operations". When a * locking op is dispatched to the lower levels of the driver, the * queue is locked, and no further I/Os are dispatched until the queue * receives & completes a corresponding "unlocking operation". This * code relies on the higher layers to guarantee that a locking op * will always be eventually followed by an unlocking op. The model * is that the higher layers are structured so locking and unlocking * ops occur in pairs, i.e. an unlocking op cannot be generated until * after a locking op reports completion. There is no good way to * check to see that an unlocking op "corresponds" to the op that * currently has the queue locked, so we make no such attempt. Since * by definition there can be only one locking op outstanding on a * disk, this should not be a problem. * * In the kernel, we allow multiple I/Os to be concurrently dispatched * to the disk driver. In order to support locking ops in this * environment, when we decide to do a locking op, we stop dispatching * new I/Os and wait until all dispatched I/Os have completed before * dispatching the locking op. * * Unfortunately, the code is different in the 3 different operating * states (user level, kernel, simulator). In the kernel, I/O is * non-blocking, and we have no disk threads to dispatch for us. * Therefore, we have to dispatch new I/Os to the scsi driver at the * time of enqueue, and also at the time of completion. At user * level, I/O is blocking, and so only the disk threads may dispatch * I/Os. Thus at user level, all we can do at enqueue time is enqueue * and wake up the disk thread to do the dispatch. * ****************************************************************************/ #include __KERNEL_RCSID(0, "$NetBSD: rf_diskqueue.c,v 1.53 2011/05/05 06:04:09 mrg Exp $"); #include #include "rf_threadstuff.h" #include "rf_raid.h" #include "rf_diskqueue.h" #include "rf_alloclist.h" #include "rf_acctrace.h" #include "rf_etimer.h" #include "rf_general.h" #include "rf_debugprint.h" #include "rf_shutdown.h" #include "rf_cvscan.h" #include "rf_sstf.h" #include "rf_fifo.h" #include "rf_kintf.h" static void rf_ShutdownDiskQueueSystem(void *); #ifndef RF_DEBUG_DISKQUEUE #define RF_DEBUG_DISKQUEUE 0 #endif #if RF_DEBUG_DISKQUEUE #define Dprintf1(s,a) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL) #define Dprintf2(s,a,b) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),NULL,NULL,NULL,NULL,NULL,NULL) #define Dprintf3(s,a,b,c) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),NULL,NULL,NULL,NULL,NULL) #else #define Dprintf1(s,a) #define Dprintf2(s,a,b) #define Dprintf3(s,a,b,c) #endif /***************************************************************************** * * the disk queue switch defines all the functions used in the * different queueing disciplines queue ID, init routine, enqueue * routine, dequeue routine * ****************************************************************************/ static const RF_DiskQueueSW_t diskqueuesw[] = { {"fifo", /* FIFO */ rf_FifoCreate, rf_FifoEnqueue, rf_FifoDequeue, rf_FifoPeek, rf_FifoPromote}, {"cvscan", /* cvscan */ rf_CvscanCreate, rf_CvscanEnqueue, rf_CvscanDequeue, rf_CvscanPeek, rf_CvscanPromote}, {"sstf", /* shortest seek time first */ rf_SstfCreate, rf_SstfEnqueue, rf_SstfDequeue, rf_SstfPeek, rf_SstfPromote}, {"scan", /* SCAN (two-way elevator) */ rf_ScanCreate, rf_SstfEnqueue, rf_ScanDequeue, rf_ScanPeek, rf_SstfPromote}, {"cscan", /* CSCAN (one-way elevator) */ rf_CscanCreate, rf_SstfEnqueue, rf_CscanDequeue, rf_CscanPeek, rf_SstfPromote}, }; #define NUM_DISK_QUEUE_TYPES (sizeof(diskqueuesw)/sizeof(RF_DiskQueueSW_t)) #define RF_MAX_FREE_DQD 256 #define RF_MIN_FREE_DQD 64 #include /* configures a single disk queue */ static void rf_ShutdownDiskQueue(void *arg) { RF_DiskQueue_t *diskqueue = arg; rf_destroy_mutex2(diskqueue->mutex); } int rf_ConfigureDiskQueue(RF_Raid_t *raidPtr, RF_DiskQueue_t *diskqueue, RF_RowCol_t c, const RF_DiskQueueSW_t *p, RF_SectorCount_t sectPerDisk, dev_t dev, int maxOutstanding, RF_ShutdownList_t **listp, RF_AllocListElem_t *clList) { diskqueue->col = c; diskqueue->qPtr = p; diskqueue->qHdr = (p->Create) (sectPerDisk, clList, listp); diskqueue->dev = dev; diskqueue->numOutstanding = 0; diskqueue->queueLength = 0; diskqueue->maxOutstanding = maxOutstanding; diskqueue->curPriority = RF_IO_NORMAL_PRIORITY; diskqueue->flags = 0; diskqueue->raidPtr = raidPtr; diskqueue->rf_cinfo = &raidPtr->raid_cinfo[c]; rf_init_mutex2(diskqueue->mutex, IPL_VM); rf_ShutdownCreate(listp, rf_ShutdownDiskQueue, diskqueue); return (0); } static void rf_ShutdownDiskQueueSystem(void *ignored) { pool_destroy(&rf_pools.dqd); } int rf_ConfigureDiskQueueSystem(RF_ShutdownList_t **listp) { rf_pool_init(&rf_pools.dqd, sizeof(RF_DiskQueueData_t), "rf_dqd_pl", RF_MIN_FREE_DQD, RF_MAX_FREE_DQD); rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL); return (0); } int rf_ConfigureDiskQueues(RF_ShutdownList_t **listp, RF_Raid_t *raidPtr, RF_Config_t *cfgPtr) { RF_DiskQueue_t *diskQueues, *spareQueues; const RF_DiskQueueSW_t *p; RF_RowCol_t r,c; int rc, i; raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs; for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) { if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) { p = &diskqueuesw[i]; break; } } if (p == NULL) { RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType); p = &diskqueuesw[0]; } raidPtr->qType = p; RF_MallocAndAdd(diskQueues, (raidPtr->numCol + RF_MAXSPARE) * sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *), raidPtr->cleanupList); if (diskQueues == NULL) return (ENOMEM); raidPtr->Queues = diskQueues; for (c = 0; c < raidPtr->numCol; c++) { rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[c], c, p, raidPtr->sectorsPerDisk, raidPtr->Disks[c].dev, cfgPtr->maxOutstandingDiskReqs, listp, raidPtr->cleanupList); if (rc) return (rc); } spareQueues = &raidPtr->Queues[raidPtr->numCol]; for (r = 0; r < raidPtr->numSpare; r++) { rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r], raidPtr->numCol + r, p, raidPtr->sectorsPerDisk, raidPtr->Disks[raidPtr->numCol + r].dev, cfgPtr->maxOutstandingDiskReqs, listp, raidPtr->cleanupList); if (rc) return (rc); } return (0); } /* Enqueue a disk I/O * * In the kernel, I/O is non-blocking and so we'd like to have multiple * I/Os outstanding on the physical disks when possible. * * when any request arrives at a queue, we have two choices: * dispatch it to the lower levels * queue it up * * kernel rules for when to do what: * unlocking req : always dispatch it * normal req : queue empty => dispatch it & set priority * queue not full & priority is ok => dispatch it * else queue it */ void rf_DiskIOEnqueue(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int pri) { RF_ETIMER_START(req->qtime); RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector); req->priority = pri; #if RF_DEBUG_DISKQUEUE if (rf_queueDebug && (req->numSector == 0)) { printf("Warning: Enqueueing zero-sector access\n"); } #endif RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue"); if (RF_OK_TO_DISPATCH(queue, req)) { Dprintf2("Dispatching pri %d regular op to c %d (ok to dispatch)\n", pri, queue->col); rf_DispatchKernelIO(queue, req); } else { queue->queueLength++; /* increment count of number of requests waiting in this queue */ Dprintf2("Enqueueing pri %d regular op to c %d (not ok to dispatch)\n", pri, queue->col); req->queue = (void *) queue; (queue->qPtr->Enqueue) (queue->qHdr, req, pri); } RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue"); } /* get the next set of I/Os started */ void rf_DiskIOComplete(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int status) { int done = 0; RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete"); queue->numOutstanding--; RF_ASSERT(queue->numOutstanding >= 0); /* dispatch requests to the disk until we find one that we can't. */ /* no reason to continue once we've filled up the queue */ /* no reason to even start if the queue is locked */ while (!done && !RF_QUEUE_FULL(queue)) { req = (queue->qPtr->Dequeue) (queue->qHdr); if (req) { Dprintf2("DiskIOComplete: extracting pri %d req from queue at c %d\n", req->priority, queue->col); queue->queueLength--; /* decrement count of number of requests waiting in this queue */ RF_ASSERT(queue->queueLength >= 0); if (RF_OK_TO_DISPATCH(queue, req)) { Dprintf2("DiskIOComplete: dispatching pri %d regular req to c %d (ok to dispatch)\n", req->priority, queue->col); rf_DispatchKernelIO(queue, req); } else { /* we can't dispatch it, so just re-enqueue it. potential trouble here if disk queues batch reqs */ Dprintf2("DiskIOComplete: re-enqueueing pri %d regular req to c %d\n", req->priority, queue->col); queue->queueLength++; (queue->qPtr->Enqueue) (queue->qHdr, req, req->priority); done = 1; } } else { Dprintf1("DiskIOComplete: no more requests to extract.\n", ""); done = 1; } } RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete"); } /* promotes accesses tagged with the given parityStripeID from low priority * to normal priority. This promotion is optional, meaning that a queue * need not implement it. If there is no promotion routine associated with * a queue, this routine does nothing and returns -1. */ int rf_DiskIOPromote(RF_DiskQueue_t *queue, RF_StripeNum_t parityStripeID, RF_ReconUnitNum_t which_ru) { int retval; if (!queue->qPtr->Promote) return (-1); RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote"); retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru); RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote"); return (retval); } RF_DiskQueueData_t * rf_CreateDiskQueueData(RF_IoType_t typ, RF_SectorNum_t ssect, RF_SectorCount_t nsect, void *bf, RF_StripeNum_t parityStripeID, RF_ReconUnitNum_t which_ru, int (*wakeF) (void *, int), void *arg, RF_AccTraceEntry_t *tracerec, RF_Raid_t *raidPtr, RF_DiskQueueDataFlags_t flags, void *kb_proc, int waitflag) { RF_DiskQueueData_t *p; p = pool_get(&rf_pools.dqd, waitflag); if (p == NULL) return (NULL); memset(p, 0, sizeof(RF_DiskQueueData_t)); if (waitflag == PR_WAITOK) { p->bp = getiobuf(NULL, true); } else { p->bp = getiobuf(NULL, false); } if (p->bp == NULL) { pool_put(&rf_pools.dqd, p); return (NULL); } SET(p->bp->b_cflags, BC_BUSY); /* mark buffer busy */ p->sectorOffset = ssect + rf_protectedSectors; p->numSector = nsect; p->type = typ; p->buf = bf; p->parityStripeID = parityStripeID; p->which_ru = which_ru; p->CompleteFunc = wakeF; p->argument = arg; p->next = NULL; p->tracerec = tracerec; p->priority = RF_IO_NORMAL_PRIORITY; p->raidPtr = raidPtr; p->flags = flags; p->b_proc = kb_proc; return (p); } void rf_FreeDiskQueueData(RF_DiskQueueData_t *p) { int s; s = splbio(); /* XXX protect only pool_put, or neither? */ putiobuf(p->bp); pool_put(&rf_pools.dqd, p); splx(s); }