#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: rf_stripelocks.c,v 1.35 2021/07/23 00:54:45 oster Exp $");
#include <dev/raidframe/raidframevar.h>
#include "rf_raid.h"
#include "rf_stripelocks.h"
#include "rf_alloclist.h"
#include "rf_debugprint.h"
#include "rf_general.h"
#include "rf_driver.h"
#include "rf_shutdown.h"
#ifdef DEBUG
#define Dprintf1(s,a) rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL)
#define Dprintf2(s,a,b) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),NULL,NULL,NULL,NULL,NULL,NULL)
#define Dprintf3(s,a,b,c) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),NULL,NULL,NULL,NULL,NULL)
#define Dprintf4(s,a,b,c,d) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),NULL,NULL,NULL,NULL)
#define Dprintf5(s,a,b,c,d,e) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),(void *)((unsigned long)e),NULL,NULL,NULL)
#define Dprintf6(s,a,b,c,d,e,f) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),(void *)((unsigned long)e),(void *)((unsigned long)f),NULL,NULL)
#define Dprintf7(s,a,b,c,d,e,f,g) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),(void *)((unsigned long)e),(void *)((unsigned long)f),(void *)((unsigned long)g),NULL)
#define Dprintf8(s,a,b,c,d,e,f,g,h) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),(void *)((unsigned long)e),(void *)((unsigned long)f),(void *)((unsigned long)g),(void *)((unsigned long)h))
#else
#define Dprintf1(s,a) {}
#define Dprintf2(s,a,b) {}
#define Dprintf3(s,a,b,c) {}
#define Dprintf4(s,a,b,c,d) {}
#define Dprintf5(s,a,b,c,d,e) {}
#define Dprintf6(s,a,b,c,d,e,f) {}
#define Dprintf7(s,a,b,c,d,e,f,g) {}
#define Dprintf8(s,a,b,c,d,e,f,g,h) {}
#endif
#define FLUSH
#define HASH_STRIPEID(_sid_) ( (_sid_) & (rf_lockTableSize-1) )
static void AddToWaitersQueue(RF_StripeLockDesc_t * lockDesc,
RF_LockReqDesc_t * lockReqDesc);
static RF_StripeLockDesc_t *AllocStripeLockDesc(RF_Raid_t *raidPtr, RF_StripeNum_t stripeID);
static void FreeStripeLockDesc(RF_Raid_t *raidPtr, RF_StripeLockDesc_t * p);
static RF_LockTableEntry_t *rf_MakeLockTable(void);
#if RF_DEBUG_STRIPELOCK
static void PrintLockedStripes(RF_LockTableEntry_t * lockTable);
#endif
#define SINGLE_RANGE_OVERLAP(_strt1, _stop1, _strt2, _stop2) \
( (_strt1 >= 0) && (_strt2 >= 0) && \
(RF_MAX(_strt1, _strt2) <= RF_MIN(_stop1, _stop2)) )
#define RANGE_OVERLAP(_cand, _pred) \
( SINGLE_RANGE_OVERLAP((_cand)->start, (_cand)->stop, \
(_pred)->start, (_pred)->stop ) || \
SINGLE_RANGE_OVERLAP((_cand)->start2, (_cand)->stop2, \
(_pred)->start, (_pred)->stop ) || \
SINGLE_RANGE_OVERLAP((_cand)->start, (_cand)->stop, \
(_pred)->start2, (_pred)->stop2) || \
SINGLE_RANGE_OVERLAP((_cand)->start2, (_cand)->stop2, \
(_pred)->start2, (_pred)->stop2) )
#define STRIPELOCK_CONFLICT(_cand, _pred) \
RANGE_OVERLAP((_cand), (_pred)) && \
( ( (((_cand)->type == RF_IO_TYPE_READ) && \
((_pred)->type == RF_IO_TYPE_WRITE)) || \
(((_cand)->type == RF_IO_TYPE_WRITE) && \
((_pred)->type == RF_IO_TYPE_READ)) || \
(((_cand)->type == RF_IO_TYPE_WRITE) && \
((_pred)->type == RF_IO_TYPE_WRITE)) \
) \
)
#define RF_MAX_FREE_STRIPELOCK 128
#define RF_MIN_FREE_STRIPELOCK 32
static void rf_ShutdownStripeLocks(RF_LockTableEntry_t * lockTable);
static void rf_ShutdownStripeLockFreeList(void *);
static void rf_RaidShutdownStripeLocks(void *);
static void
rf_ShutdownStripeLockFreeList(void *arg)
{
RF_Raid_t *raidPtr;
raidPtr = (RF_Raid_t *) arg;
pool_destroy(&raidPtr->pools.stripelock);
}
int
rf_ConfigureStripeLockFreeList(RF_ShutdownList_t **listp, RF_Raid_t *raidPtr,
RF_Config_t *cfgPtr)
{
unsigned mask;
rf_pool_init(raidPtr, raidPtr->poolNames.stripelock, &raidPtr->pools.stripelock, sizeof(RF_StripeLockDesc_t),
"strplock", RF_MIN_FREE_STRIPELOCK, RF_MAX_FREE_STRIPELOCK);
rf_ShutdownCreate(listp, rf_ShutdownStripeLockFreeList, raidPtr);
for (mask = 0x1; mask; mask <<= 1)
if (rf_lockTableSize == mask)
break;
if (!mask) {
printf("[WARNING: lock table size must be a power of two. Setting to %d.]\n", RF_DEFAULT_LOCK_TABLE_SIZE);
rf_lockTableSize = RF_DEFAULT_LOCK_TABLE_SIZE;
}
return (0);
}
static void
rf_DestroyLockTable(RF_LockTableEntry_t *lockTable)
{
int i;
for (i = 0; i < rf_lockTableSize; i++) {
rf_destroy_mutex2(lockTable[i].mutex);
}
RF_Free(lockTable, rf_lockTableSize * sizeof(RF_LockTableEntry_t));
}
static RF_LockTableEntry_t *
rf_MakeLockTable(void)
{
RF_LockTableEntry_t *lockTable;
int i;
lockTable = RF_Malloc(rf_lockTableSize * sizeof(*lockTable));
if (lockTable == NULL)
return (NULL);
for (i = 0; i < rf_lockTableSize; i++) {
rf_init_mutex2(lockTable[i].mutex, IPL_VM);
}
return (lockTable);
}
static void
rf_ShutdownStripeLocks(RF_LockTableEntry_t * lockTable)
{
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
PrintLockedStripes(lockTable);
}
#endif
rf_DestroyLockTable(lockTable);
}
static void
rf_RaidShutdownStripeLocks(void *arg)
{
RF_Raid_t *raidPtr = (RF_Raid_t *) arg;
rf_ShutdownStripeLocks(raidPtr->lockTable);
}
int
rf_ConfigureStripeLocks(RF_ShutdownList_t **listp, RF_Raid_t *raidPtr,
RF_Config_t *cfgPtr)
{
raidPtr->lockTable = rf_MakeLockTable();
if (raidPtr->lockTable == NULL)
return (ENOMEM);
rf_ShutdownCreate(listp, rf_RaidShutdownStripeLocks, raidPtr);
return (0);
}
int
rf_AcquireStripeLock(RF_Raid_t *raidPtr, RF_LockTableEntry_t *lockTable, RF_StripeNum_t stripeID,
RF_LockReqDesc_t *lockReqDesc)
{
RF_StripeLockDesc_t *lockDesc;
RF_StripeLockDesc_t *newlockDesc;
RF_LockReqDesc_t *p;
#if defined(DEBUG) && (RF_DEBUG_STRIPELOCK > 0)
int tid = 0;
#endif
int hashval = HASH_STRIPEID(stripeID);
int retcode = 0;
RF_ASSERT(RF_IO_IS_R_OR_W(lockReqDesc->type));
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
if (stripeID == -1) {
Dprintf1("[%d] Lock acquisition suppressed (stripeID == -1)\n", tid);
} else {
Dprintf8("[%d] Trying to acquire stripe lock table 0x%lx SID %ld type %c range %ld-%ld, range2 %ld-%ld hashval %d\n",
tid, (unsigned long) lockTable, stripeID, lockReqDesc->type, lockReqDesc->start,
lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2);
Dprintf3("[%d] lock %ld hashval %d\n", tid, stripeID, hashval);
FLUSH;
}
}
#endif
if (stripeID == -1)
return (0);
lockReqDesc->next = NULL;
newlockDesc = AllocStripeLockDesc(raidPtr, stripeID);
rf_lock_mutex2(lockTable[hashval].mutex);
for (lockDesc = lockTable[hashval].descList; lockDesc;
lockDesc = lockDesc->next) {
if (lockDesc->stripeID == stripeID)
break;
}
if (!lockDesc) {
lockDesc = newlockDesc;
lockDesc->next = lockTable[hashval].descList;
lockTable[hashval].descList = lockDesc;
if (lockReqDesc->type == RF_IO_TYPE_WRITE)
lockDesc->nWriters++;
lockDesc->granted = lockReqDesc;
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf7("[%d] no one waiting: lock %ld %c %ld-%ld %ld-%ld granted\n",
tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2);
FLUSH;
}
#endif
} else {
FreeStripeLockDesc(raidPtr, newlockDesc);
if (lockReqDesc->type == RF_IO_TYPE_WRITE)
lockDesc->nWriters++;
if (lockDesc->nWriters == 0) {
lockReqDesc->next = lockDesc->granted;
lockDesc->granted = lockReqDesc;
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf7("[%d] no writers: lock %ld %c %ld-%ld %ld-%ld granted\n",
tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2);
FLUSH;
}
#endif
} else {
retcode = 0;
for (p = lockDesc->granted; p; p = p->next)
if (STRIPELOCK_CONFLICT(lockReqDesc, p)) {
retcode = 1;
break;
}
if (!retcode)
for (p = lockDesc->waitersH; p; p = p->next)
if (STRIPELOCK_CONFLICT(lockReqDesc, p)) {
retcode = 2;
break;
}
if (!retcode) {
lockReqDesc->next = lockDesc->granted;
lockDesc->granted = lockReqDesc;
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf7("[%d] no conflicts: lock %ld %c %ld-%ld %ld-%ld granted\n",
tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop,
lockReqDesc->start2, lockReqDesc->stop2);
FLUSH;
}
#endif
} else {
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf6("[%d] conflict: lock %ld %c %ld-%ld hashval=%d not granted\n",
tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop,
hashval);
Dprintf3("[%d] lock %ld retcode=%d\n", tid, stripeID, retcode);
FLUSH;
}
#endif
AddToWaitersQueue(lockDesc, lockReqDesc);
}
}
}
rf_unlock_mutex2(lockTable[hashval].mutex);
return (retcode);
}
void
rf_ReleaseStripeLock(RF_Raid_t *raidPtr, RF_LockTableEntry_t *lockTable, RF_StripeNum_t stripeID,
RF_LockReqDesc_t *lockReqDesc)
{
RF_StripeLockDesc_t *lockDesc, *ld_t;
RF_LockReqDesc_t *lr, *lr_t, *callbacklist, *t;
#if defined(DEBUG) && (RF_DEBUG_STRIPELOCK > 0)
int tid = 0;
#endif
int hashval = HASH_STRIPEID(stripeID);
int release_it, consider_it;
RF_LockReqDesc_t *candidate, *candidate_t, *predecessor;
RF_ASSERT(RF_IO_IS_R_OR_W(lockReqDesc->type));
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
if (stripeID == -1) {
Dprintf1("[%d] Lock release suppressed (stripeID == -1)\n", tid);
} else {
Dprintf8("[%d] Releasing stripe lock on stripe ID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2, lockTable);
FLUSH;
}
}
#endif
if (stripeID == -1)
return;
rf_lock_mutex2(lockTable[hashval].mutex);
for (ld_t = NULL, lockDesc = lockTable[hashval].descList;
lockDesc; ld_t = lockDesc, lockDesc = lockDesc->next) {
if (lockDesc->stripeID == stripeID)
break;
}
RF_ASSERT(lockDesc);
for (lr_t = NULL, lr = lockDesc->granted; lr; lr_t = lr, lr = lr->next)
if (lr == lockReqDesc)
break;
RF_ASSERT(lr && (lr == lockReqDesc));
if (lr_t)
lr_t->next = lr->next;
else {
RF_ASSERT(lr == lockDesc->granted);
lockDesc->granted = lr->next;
}
lr->next = NULL;
if (lockReqDesc->type == RF_IO_TYPE_WRITE)
lockDesc->nWriters--;
callbacklist = NULL;
if (lockDesc->nWriters == 0) {
while (lockDesc->waitersH) {
lr = lockDesc->waitersH;
lockDesc->waitersH = lr->next;
RF_ASSERT(lr->type == RF_IO_TYPE_READ);
lr->next = lockDesc->granted;
lockDesc->granted = lr;
RF_ASSERT(!lr->templink);
lr->templink = callbacklist;
callbacklist = lr;
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf8("[%d] No writers: granting lock stripe ID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
tid, stripeID, lr->type, lr->start, lr->stop, lr->start2, lr->stop2, (unsigned long) lockTable);
FLUSH;
}
#endif
}
lockDesc->waitersT = NULL;
} else
for (candidate_t = NULL, candidate = lockDesc->waitersH;
candidate;) {
consider_it = 0;
if (RANGE_OVERLAP(lockReqDesc, candidate))
consider_it = 1;
else
for (t = callbacklist; t; t = t->templink)
if (RANGE_OVERLAP(t, candidate)) {
consider_it = 1;
break;
}
if (!consider_it) {
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf8("[%d] No overlap: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
(unsigned long) lockTable);
FLUSH;
}
#endif
candidate_t = candidate;
candidate = candidate->next;
continue;
}
release_it = 1;
for (predecessor = lockDesc->granted; predecessor;
predecessor = predecessor->next) {
if (STRIPELOCK_CONFLICT(candidate,
predecessor)) {
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf8("[%d] Conflicts with granted lock: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
(unsigned long) lockTable);
FLUSH;
}
#endif
release_it = 0;
break;
}
}
if (release_it)
for (predecessor = lockDesc->waitersH;
predecessor != candidate;
predecessor = predecessor->next) {
if (STRIPELOCK_CONFLICT(candidate,
predecessor)) {
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf8("[%d] Conflicts with waiting lock: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
(unsigned long) lockTable);
FLUSH;
}
#endif
release_it = 0;
break;
}
}
if (release_it) {
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf8("[%d] Granting lock to candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
(unsigned long) lockTable);
FLUSH;
}
#endif
if (candidate_t) {
candidate_t->next = candidate->next;
if (lockDesc->waitersT == candidate)
lockDesc->waitersT = candidate_t;
} else {
RF_ASSERT(candidate == lockDesc->waitersH);
lockDesc->waitersH = lockDesc->waitersH->next;
if (!lockDesc->waitersH)
lockDesc->waitersT = NULL;
}
candidate->next = lockDesc->granted;
lockDesc->granted = candidate;
RF_ASSERT(!candidate->templink);
candidate->templink = callbacklist;
callbacklist = candidate;
if (!candidate_t)
candidate = lockDesc->waitersH;
else
candidate = candidate_t->next;
} else {
candidate_t = candidate;
candidate = candidate->next;
}
}
if (!lockDesc->granted && !lockDesc->waitersH) {
RF_ASSERT(lockDesc->nWriters == 0);
#if RF_DEBUG_STRIPELOCK
if (rf_stripeLockDebug) {
Dprintf3("[%d] Last lock released (table 0x%lx): deleting desc for stripeID %ld\n", tid, (unsigned long) lockTable, stripeID);
FLUSH;
}
#endif
if (ld_t)
ld_t->next = lockDesc->next;
else {
RF_ASSERT(lockDesc == lockTable[hashval].descList);
lockTable[hashval].descList = lockDesc->next;
}
FreeStripeLockDesc(raidPtr, lockDesc);
lockDesc = NULL;
}
rf_unlock_mutex2(lockTable[hashval].mutex);
RF_ASSERT(!((callbacklist) && (!lockDesc)));
for (candidate = callbacklist; candidate;) {
t = candidate;
candidate = candidate->templink;
t->templink = NULL;
(t->cbFunc) (t->cbArg);
}
}
static void
AddToWaitersQueue(RF_StripeLockDesc_t *lockDesc, RF_LockReqDesc_t *lockReqDesc)
{
if (!lockDesc->waitersH) {
lockDesc->waitersH = lockDesc->waitersT = lockReqDesc;
} else {
lockDesc->waitersT->next = lockReqDesc;
lockDesc->waitersT = lockReqDesc;
}
}
static RF_StripeLockDesc_t *
AllocStripeLockDesc(RF_Raid_t *raidPtr, RF_StripeNum_t stripeID)
{
RF_StripeLockDesc_t *p;
p = pool_get(&raidPtr->pools.stripelock, PR_WAITOK);
if (p) {
p->stripeID = stripeID;
p->granted = NULL;
p->waitersH = NULL;
p->waitersT = NULL;
p->nWriters = 0;
p->next = NULL;
}
return (p);
}
static void
FreeStripeLockDesc(RF_Raid_t *raidPtr, RF_StripeLockDesc_t *p)
{
pool_put(&raidPtr->pools.stripelock, p);
}
#if RF_DEBUG_STRIPELOCK
static void
PrintLockedStripes(RF_LockTableEntry_t *lockTable)
{
int i, j, foundone = 0, did;
RF_StripeLockDesc_t *p;
RF_LockReqDesc_t *q;
rf_lock_mutex2(rf_printf_mutex);
printf("Locked stripes:\n");
for (i = 0; i < rf_lockTableSize; i++)
if (lockTable[i].descList) {
foundone = 1;
for (p = lockTable[i].descList; p; p = p->next) {
printf("Stripe ID 0x%lx (%d) nWriters %d\n",
(long) p->stripeID, (int) p->stripeID,
p->nWriters);
if (!(p->granted))
printf("Granted: (none)\n");
else
printf("Granted:\n");
for (did = 1, j = 0, q = p->granted; q;
j++, q = q->next) {
printf(" %c(%ld-%ld", q->type, (long) q->start, (long) q->stop);
if (q->start2 != -1)
printf(",%ld-%ld) ", (long) q->start2,
(long) q->stop2);
else
printf(") ");
if (j && !(j % 4)) {
printf("\n");
did = 1;
} else
did = 0;
}
if (!did)
printf("\n");
if (!(p->waitersH))
printf("Waiting: (none)\n");
else
printf("Waiting:\n");
for (did = 1, j = 0, q = p->waitersH; q;
j++, q = q->next) {
printf("%c(%ld-%ld", q->type, (long) q->start, (long) q->stop);
if (q->start2 != -1)
printf(",%ld-%ld) ", (long) q->start2, (long) q->stop2);
else
printf(") ");
if (j && !(j % 4)) {
printf("\n ");
did = 1;
} else
did = 0;
}
if (!did)
printf("\n");
}
}
if (!foundone)
printf("(none)\n");
else
printf("\n");
rf_unlock_mutex2(rf_printf_mutex);
}
#endif