void
rf_applyPDA(RF_Raid_t *raidPtr, RF_PhysDiskAddr_t *pda,
RF_PhysDiskAddr_t *ppda, RF_PhysDiskAddr_t *qpda, void *bp)
{
RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
RF_RaidAddr_t s0off = rf_StripeUnitOffset(layoutPtr, ppda->startSector);
RF_SectorCount_t s0len = ppda->numSector, len;
RF_SectorNum_t suoffset;
unsigned coeff;
char *pbuf = ppda->bufPtr;
char *qbuf = qpda->bufPtr;
char *buf;
int delta;
suoffset = rf_StripeUnitOffset(layoutPtr, pda->startSector);
len = pda->numSector;
/* See if pda intersects a recovery pda. */
if ((suoffset < s0off + s0len) && (suoffset + len > s0off)) {
buf = pda->bufPtr;
coeff = rf_RaidAddressToStripeUnitID(&(raidPtr->Layout),
pda->raidAddress);
coeff = (coeff % raidPtr->Layout.numDataCol);
if (suoffset < s0off) {
delta = s0off - suoffset;
buf += rf_RaidAddressToStripeUnitID(&(raidPtr->Layout),
delta);
suoffset = s0off;
len -= delta;
}
if (suoffset > s0off) {
delta = suoffset - s0off;
pbuf += rf_RaidAddressToStripeUnitID(&(raidPtr->Layout),
delta);
qbuf += rf_RaidAddressToStripeUnitID(&(raidPtr->Layout),
delta);
}
if ((suoffset + len) > (s0len + s0off))
len = s0len + s0off - suoffset;
/* Src, dest, len. */
rf_bxor(buf, pbuf, rf_RaidAddressToByte(raidPtr, len), bp);
/* Dest, src, len, coeff. */
rf_IncQ((unsigned long *) qbuf, (unsigned long *) buf,
rf_RaidAddressToByte(raidPtr, len), coeff);
}
}
/* returns an array of ints identifying the disks that comprise the stripe containing the indicated address.
* the caller must _never_ attempt to modify this array.
*/
void
rf_IdentifyStripeDeclustered(RF_Raid_t *raidPtr, RF_RaidAddr_t addr,
RF_RowCol_t **diskids)
{
RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo;
RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable;
RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU;
RF_StripeNum_t base_suid = 0;
RF_StripeNum_t SUID = rf_RaidAddressToStripeUnitID(layoutPtr, addr);
RF_StripeNum_t stripeID, FullTableID;
int tableOffset;
rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid);
FullTableID = SUID / sus_per_fulltable; /* fulltable ID within array
* (across rows) */
stripeID = rf_StripeUnitIDToStripeID(layoutPtr, SUID); /* find stripe offset
* into array */
tableOffset = (stripeID % info->BlocksPerTable); /* find offset into
* block design table */
*diskids = info->LayoutTable[tableOffset];
}
int
rf_PQWriteDoubleRecoveryFunc(RF_DagNode_t *node)
{
/*
* The situation:
*
* We are doing a write that hits only one failed data unit. The other
* failed data unit is not being overwritten, so we need to generate
* it.
*
* For the moment, we assume all the nonfailed data being written is in
* the shadow of the failed data unit. (i.e., either a single data
* unit write or the entire failed stripe unit is being overwritten.)
*
* Recovery strategy: apply the recovery data to the parity and Q.
* Use P & Q to recover the second failed data unit in P. Zero fill
* Q, then apply the recovered data to P. Then apply the data being
* written to the failed drive. Then walk through the surviving drives,
* applying new data when it exists, othewise the recovery data.
* Quite a mess.
*
*
* The params:
*
* read pda0, read pda1, ..., read pda (numDataCol-3),
* write pda0, ..., write pda (numStripeUnitAccess - numDataFailed),
* failed pda, raidPtr, asmap
*/
int np = node->numParams;
RF_AccessStripeMap_t *asmap = (RF_AccessStripeMap_t *)
node->params[np - 1].p;
RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[np - 2].p;
RF_RaidLayout_t *layoutPtr = (RF_RaidLayout_t *) & (raidPtr->Layout);
int i;
RF_RaidAddr_t sosAddr;
unsigned coeff;
RF_StripeCount_t secPerSU = layoutPtr->sectorsPerStripeUnit;
RF_PhysDiskAddr_t *ppda, *qpda, *pda, npda;
int numDataCol = layoutPtr->numDataCol;
RF_Etimer_t timer;
RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
RF_ASSERT(node->numResults == 2);
RF_ASSERT(asmap->failedPDAs[1] == NULL);
RF_ETIMER_START(timer);
ppda = node->results[0];
qpda = node->results[1];
/* apply the recovery data */
for (i = 0; i < numDataCol - 2; i++)
rf_applyPDA(raidPtr, node->params[i].p, ppda, qpda,
node->dagHdr->bp);
/* Determine the other failed data unit. */
pda = asmap->failedPDAs[0];
sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr,
asmap->raidAddress);
/* Need to determine the column of the other failed disk. */
coeff = rf_RaidAddressToStripeUnitID(layoutPtr, pda->raidAddress);
/* Compute the data unit offset within the column. */
coeff = (coeff % raidPtr->Layout.numDataCol);
for (i = 0; i < numDataCol; i++) {
npda.raidAddress = sosAddr + (i * secPerSU);
(raidPtr->Layout.map->MapSector) (raidPtr, npda.raidAddress,
&(npda.row), &(npda.col), &(npda.startSector), 0);
/* Skip over dead disks. */
if (RF_DEAD_DISK(raidPtr->Disks[npda.row][npda.col].status))
if (i != coeff)
break;
}
RF_ASSERT(i < numDataCol);
/*
* Recover the data. The column we want to recover, we write over the
* parity. The column we don't care about, we dump in q.
*/
if (coeff < i) /* Recovering 'a'. */
rf_PQ_recover((unsigned long *) ppda->bufPtr,
(unsigned long *) qpda->bufPtr,
(unsigned long *) ppda->bufPtr,
(unsigned long *) qpda->bufPtr,
rf_RaidAddressToByte(raidPtr, pda->numSector), coeff, i);
else /* Recovering 'b'. */
rf_PQ_recover((unsigned long *) ppda->bufPtr,
(unsigned long *) qpda->bufPtr,
(unsigned long *) qpda->bufPtr,
(unsigned long *) ppda->bufPtr,
rf_RaidAddressToByte(raidPtr, pda->numSector), i, coeff);
/* OK. The valid data is in P. Zero fill Q, then inc it into it. */
bzero(qpda->bufPtr, rf_RaidAddressToByte(raidPtr, qpda->numSector));
rf_IncQ((unsigned long *) qpda->bufPtr, (unsigned long *) ppda->bufPtr,
rf_RaidAddressToByte(raidPtr, qpda->numSector), i);
/* Now apply all the write data to the buffer. */
/*
* Single stripe unit write case: The failed data is the only thing
* we are writing.
*/
RF_ASSERT(asmap->numStripeUnitsAccessed == 1);
/* Dest, src, len, coeff. */
rf_IncQ((unsigned long *) qpda->bufPtr,
(unsigned long *) asmap->failedPDAs[0]->bufPtr,
rf_RaidAddressToByte(raidPtr, qpda->numSector), coeff);
rf_bxor(asmap->failedPDAs[0]->bufPtr, ppda->bufPtr,
rf_RaidAddressToByte(raidPtr, ppda->numSector), node->dagHdr->bp);
/* Now apply all the recovery data. */
for (i = 0; i < numDataCol - 2; i++)
rf_applyPDA(raidPtr, node->params[i].p, ppda, qpda,
node->dagHdr->bp);
RF_ETIMER_STOP(timer);
RF_ETIMER_EVAL(timer);
if (tracerec)
tracerec->q_us += RF_ETIMER_VAL_US(timer);
rf_GenericWakeupFunc(node, 0);
return (0);
}
int
rf_PQDoubleRecoveryFunc(RF_DagNode_t *node)
{
int np = node->numParams;
RF_AccessStripeMap_t *asmap =
(RF_AccessStripeMap_t *) node->params[np - 1].p;
RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[np - 2].p;
RF_RaidLayout_t *layoutPtr = (RF_RaidLayout_t *) & (raidPtr->Layout);
int d, i;
unsigned coeff;
RF_RaidAddr_t sosAddr, suoffset;
RF_SectorCount_t len, secPerSU = layoutPtr->sectorsPerStripeUnit;
int two = 0;
RF_PhysDiskAddr_t *ppda, *ppda2, *qpda, *qpda2, *pda, npda;
char *buf;
int numDataCol = layoutPtr->numDataCol;
RF_Etimer_t timer;
RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
RF_ETIMER_START(timer);
if (asmap->failedPDAs[1] &&
(asmap->failedPDAs[1]->numSector +
asmap->failedPDAs[0]->numSector < secPerSU)) {
RF_ASSERT(0);
ppda = node->params[np - 6].p;
ppda2 = node->params[np - 5].p;
qpda = node->params[np - 4].p;
qpda2 = node->params[np - 3].p;
d = (np - 6);
two = 1;
} else {
ppda = node->params[np - 4].p;
qpda = node->params[np - 3].p;
d = (np - 4);
}
for (i = 0; i < d; i++) {
pda = node->params[i].p;
buf = pda->bufPtr;
suoffset = rf_StripeUnitOffset(layoutPtr, pda->startSector);
len = pda->numSector;
coeff = rf_RaidAddressToStripeUnitID(layoutPtr,
pda->raidAddress);
/* Compute the data unit offset within the column. */
coeff = (coeff % raidPtr->Layout.numDataCol);
/* See if pda intersects a recovery pda. */
rf_applyPDA(raidPtr, pda, ppda, qpda, node->dagHdr->bp);
if (two)
rf_applyPDA(raidPtr, pda, ppda, qpda, node->dagHdr->bp);
}
/*
* Ok, we got the parity back to the point where we can recover. We
* now need to determine the coeff of the columns that need to be
* recovered. We can also only need to recover a single stripe unit.
*/
if (asmap->failedPDAs[1] == NULL) { /*
* Only a single stripe unit
* to recover.
*/
pda = asmap->failedPDAs[0];
sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr,
asmap->raidAddress);
/* Need to determine the column of the other failed disk. */
coeff = rf_RaidAddressToStripeUnitID(layoutPtr,
pda->raidAddress);
/* Compute the data unit offset within the column. */
coeff = (coeff % raidPtr->Layout.numDataCol);
for (i = 0; i < numDataCol; i++) {
npda.raidAddress = sosAddr + (i * secPerSU);
(raidPtr->Layout.map->MapSector) (raidPtr,
npda.raidAddress, &(npda.row), &(npda.col),
&(npda.startSector), 0);
/* Skip over dead disks. */
if (RF_DEAD_DISK(raidPtr->Disks[npda.row][npda.col]
.status))
if (i != coeff)
break;
}
RF_ASSERT(i < numDataCol);
RF_ASSERT(two == 0);
/*
* Recover the data. Since we need only to recover one
* column, we overwrite the parity with the other one.
*/
if (coeff < i) /* Recovering 'a'. */
rf_PQ_recover((unsigned long *) ppda->bufPtr,
(unsigned long *) qpda->bufPtr,
(unsigned long *) pda->bufPtr,
(unsigned long *) ppda->bufPtr,
rf_RaidAddressToByte(raidPtr, pda->numSector),
coeff, i);
else /* Recovering 'b'. */
rf_PQ_recover((unsigned long *) ppda->bufPtr,
(unsigned long *) qpda->bufPtr,
(unsigned long *) ppda->bufPtr,
(unsigned long *) pda->bufPtr,
rf_RaidAddressToByte(raidPtr, pda->numSector),
i, coeff);
} else
RF_PANIC();
RF_ETIMER_STOP(timer);
RF_ETIMER_EVAL(timer);
if (tracerec)
tracerec->q_us += RF_ETIMER_VAL_US(timer);
rf_GenericWakeupFunc(node, 0);
return (0);
}
