rf_decluster.c

RAIDFrame是个非常好的磁盘阵列RAID仿真工具

/*
 * 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  [e-mail]
 *  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_decluster.c -- code related to the declustered layout
 *
 * Created 10-21-92 (MCH)
 *
 * Nov 93:  adding support for distributed sparing.  This code is a little
 *          complex:  the basic layout used is as follows:
 *          let F = (v-1)/GCD(r,v-1).  The spare space for each set of
 *          F consecutive fulltables is grouped together and placed after
 *          that set of tables.
 *                   +------------------------------+
 *                   |        F fulltables          |
 *                   |        Spare Space           |
 *                   |        F fulltables          |
 *                   |        Spare Space           |
 *                   |            ...               |
 *                   +------------------------------+
 *
 *--------------------------------------------------------------------*/

/*
 * $Locker:  $
 * $Log: rf_decluster.c,v $
 * Revision 1.51  1996/08/21  19:47:10  jimz
 * fix bogus return values from config
 *
 * Revision 1.50  1996/08/20  22:41:42  jimz
 * better diagnostics for bad blockdesigns
 *
 * Revision 1.49  1996/07/31  16:56:18  jimz
 * dataBytesPerStripe, sectorsPerDisk init arch-indep.
 *
 * Revision 1.48  1996/07/29  14:05:12  jimz
 * fix numPUs/numRUs confusion (everything is now numRUs)
 * clean up some commenting, return values
 *
 * Revision 1.47  1996/07/27  23:36:08  jimz
 * Solaris port of simulator
 *
 * Revision 1.46  1996/07/27  18:40:11  jimz
 * cleanup sweep
 *
 * Revision 1.45  1996/07/18  22:57:14  jimz
 * port simulator to AIX
 *
 * Revision 1.44  1996/07/13  00:00:59  jimz
 * sanitized generalized reconstruction architecture
 * cleaned up head sep, rbuf problems
 *
 * Revision 1.43  1996/06/19  17:53:48  jimz
 * move GetNumSparePUs, InstallSpareTable ops into layout switch
 *
 * Revision 1.42  1996/06/17  03:23:48  jimz
 * switch DeclusteredDS typing
 *
 * Revision 1.41  1996/06/11  08:55:15  jimz
 * improved error-checking at configuration time
 *
 * Revision 1.40  1996/06/10  11:55:47  jimz
 * Straightened out some per-array/not-per-array distinctions, fixed
 * a couple bugs related to confusion. Added shutdown lists. Removed
 * layout shutdown function (now subsumed by shutdown lists).
 *
 * Revision 1.39  1996/06/09  02:36:46  jimz
 * lots of little crufty cleanup- fixup whitespace
 * issues, comment #ifdefs, improve typing in some
 * places (esp size-related)
 *
 * Revision 1.38  1996/06/07  22:26:27  jimz
 * type-ify which_ru (RF_ReconUnitNum_t)
 *
 * Revision 1.37  1996/06/07  21:33:04  jimz
 * begin using consistent types for sector numbers,
 * stripe numbers, row+col numbers, recon unit numbers
 *
 * Revision 1.36  1996/06/03  23:28:26  jimz
 * more bugfixes
 * check in tree to sync for IPDS runs with current bugfixes
 * there still may be a problem with threads in the script test
 * getting I/Os stuck- not trivially reproducible (runs ~50 times
 * in a row without getting stuck)
 *
 * Revision 1.35  1996/06/02  17:31:48  jimz
 * Moved a lot of global stuff into array structure, where it belongs.
 * Fixed up paritylogging, pss modules in this manner. Some general
 * code cleanup. Removed lots of dead code, some dead files.
 *
 * Revision 1.34  1996/05/30  23:22:16  jimz
 * bugfixes of serialization, timing problems
 * more cleanup
 *
 * Revision 1.33  1996/05/30  11:29:41  jimz
 * Numerous bug fixes. Stripe lock release code disagreed with the taking code
 * about when stripes should be locked (I made it consistent: no parity, no lock)
 * There was a lot of extra serialization of I/Os which I've removed- a lot of
 * it was to calculate values for the cache code, which is no longer with us.
 * More types, function, macro cleanup. Added code to properly quiesce the array
 * on shutdown. Made a lot of stuff array-specific which was (bogusly) general
 * before. Fixed memory allocation, freeing bugs.
 *
 * Revision 1.32  1996/05/27  18:56:37  jimz
 * more code cleanup
 * better typing
 * compiles in all 3 environments
 *
 * Revision 1.31  1996/05/24  01:59:45  jimz
 * another checkpoint in code cleanup for release
 * time to sync kernel tree
 *
 * Revision 1.30  1996/05/23  00:33:23  jimz
 * code cleanup: move all debug decls to rf_options.c, all extern
 * debug decls to rf_options.h, all debug vars preceded by rf_
 *
 * Revision 1.29  1996/05/18  19:51:34  jimz
 * major code cleanup- fix syntax, make some types consistent,
 * add prototypes, clean out dead code, et cetera
 *
 * Revision 1.28  1995/12/12  18:10:06  jimz
 * MIN -> RF_MIN, MAX -> RF_MAX, ASSERT -> RF_ASSERT
 * fix 80-column brain damage in comments
 *
 * Revision 1.27  1995/12/01  16:00:08  root
 * added copyright info
 *
 * Revision 1.26  1995/11/28  21:35:12  amiri
 * set the RF_BD_DECLUSTERED flag
 *
 * Revision 1.25  1995/11/17  18:56:00  wvcii
 * added prototyping to MapParity
 *
 * Revision 1.24  1995/07/04  22:25:33  holland
 * increased default num bufs
 *
 * Revision 1.23  1995/07/03  20:23:51  holland
 * changed floating recon bufs & head sep yet again
 *
 * Revision 1.22  1995/07/03  18:12:14  holland
 * changed the way the number of floating recon bufs & the head sep
 * limit are set
 *
 * Revision 1.21  1995/07/02  15:07:42  holland
 * bug fixes related to getting distributed sparing numbers
 *
 * Revision 1.20  1995/06/23  13:41:28  robby
 * updeated to prototypes in rf_layout.h
 *
 */

#include "rf_types.h"
#include "rf_raid.h"
#include "rf_raidframe.h"
#include "rf_configure.h"
#include "rf_decluster.h"
#include "rf_debugMem.h"
#include "rf_utils.h"
#include "rf_alloclist.h"
#include "rf_general.h"
#include "rf_shutdown.h"
#include "rf_sys.h"

extern int rf_copyback_in_progress;                /* debug only */

/* configuration code */

int rf_ConfigureDeclustered(
  RF_ShutdownList_t  **listp,
  RF_Raid_t           *raidPtr,
  RF_Config_t         *cfgPtr)
{
    RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
    int b, v, k, r, lambda;				/* block design params */
    int i, j;
    RF_RowCol_t *first_avail_slot;
    RF_StripeCount_t complete_FT_count, numCompleteFullTablesPerDisk;
    RF_DeclusteredConfigInfo_t *info;
    RF_StripeCount_t PUsPerDisk, spareRegionDepthInPUs, numCompleteSpareRegionsPerDisk, extraPUsPerDisk;
    RF_StripeCount_t totSparePUsPerDisk;
    RF_SectorNum_t diskOffsetOfLastFullTableInSUs;
    RF_SectorCount_t SpareSpaceInSUs;
    char *cfgBuf = (char *) (cfgPtr->layoutSpecific);
    RF_StripeNum_t l, SUID;

    SUID = l = 0;
    numCompleteSpareRegionsPerDisk = 0;

    /* 1. create layout specific structure */
    RF_MallocAndAdd(info, sizeof(RF_DeclusteredConfigInfo_t), (RF_DeclusteredConfigInfo_t *), raidPtr->cleanupList);
    if (info == NULL)
      return(ENOMEM);
    layoutPtr->layoutSpecificInfo = (void *) info;
    info->SpareTable = NULL;

    /* 2. extract parameters from the config structure */
    if (layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) {
      (void) bcopy(cfgBuf, info->sparemap_fname, RF_SPAREMAP_NAME_LEN);
    }
    cfgBuf += RF_SPAREMAP_NAME_LEN;
    
    b        = *( (int *) cfgBuf);   cfgBuf += sizeof(int);
    v        = *( (int *) cfgBuf);   cfgBuf += sizeof(int);
    k        = *( (int *) cfgBuf);   cfgBuf += sizeof(int);
    r        = *( (int *) cfgBuf);   cfgBuf += sizeof(int);
    lambda   = *( (int *) cfgBuf);   cfgBuf += sizeof(int);
    raidPtr->noRotate = *( (int *) cfgBuf);   cfgBuf += sizeof(int);

    /* the sparemaps are generated assuming that parity is rotated, so we issue
     * a warning if both distributed sparing and no-rotate are on at the same time
     */
    if ((layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) && raidPtr->noRotate) {
	RF_ERRORMSG("Warning:  distributed sparing specified without parity rotation.\n");
    }

    if (raidPtr->numCol != v) {
        RF_ERRORMSG2("RAID: config error: table element count (%d) not equal to no. of cols (%d)\n", v, raidPtr->numCol);
        return(EINVAL);
    }

    /* 3.  set up the values used in the mapping code */
    info->BlocksPerTable = b;
    info->Lambda = lambda;
    info->NumParityReps = info->groupSize = k;
    info->SUsPerTable = b * (k-1) * layoutPtr->SUsPerPU;/* b blks, k-1 SUs each */
    info->SUsPerFullTable = k * info->SUsPerTable;	/* rot k times */
    info->PUsPerBlock = k-1;
    info->SUsPerBlock = info->PUsPerBlock * layoutPtr->SUsPerPU;
    info->TableDepthInPUs = (b*k) / v;
    info->FullTableDepthInPUs = info->TableDepthInPUs * k;		/* k repetitions */

    /* used only in distributed sparing case */
    info->FullTablesPerSpareRegion = (v-1) / rf_gcd(r, v-1);		/* (v-1)/gcd fulltables */
    info->TablesPerSpareRegion = k * info->FullTablesPerSpareRegion;
    info->SpareSpaceDepthPerRegionInSUs = (r * info->TablesPerSpareRegion / (v-1)) * layoutPtr->SUsPerPU;

    /* check to make sure the block design is sufficiently small */
    if ((raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE)) {
        if (info->FullTableDepthInPUs * layoutPtr->SUsPerPU + info->SpareSpaceDepthPerRegionInSUs > layoutPtr->stripeUnitsPerDisk) {
	    RF_ERRORMSG3("RAID: config error: Full Table depth (%d) + Spare Space (%d) larger than disk size (%d) (BD too big)\n",
		    info->FullTableDepthInPUs, info->SpareSpaceDepthPerRegionInSUs, layoutPtr->stripeUnitsPerDisk);
	    return(EINVAL);
	}
    } else {
	if (info->TableDepthInPUs * layoutPtr->SUsPerPU > layoutPtr->stripeUnitsPerDisk) {
	    RF_ERRORMSG2("RAID: config error: Table depth (%d) larger than disk size (%d) (BD too big)\n",
		    info->TableDepthInPUs * layoutPtr->SUsPerPU, layoutPtr->stripeUnitsPerDisk);
	    return(EINVAL);
	}
    }