split-dcache-20050427

split for rw/ro dcaches to allow portion of cache to be dedicated

====================
This delta was composed from multiple commits as part of the CVS->Git migration.
The checkin message with each commit was inconsistent.
The following are the additional commit messages.
====================

update
This commit is contained in:
Derrick Brashear 2005-04-28 23:14:57 +00:00
parent 7a4c7609c8
commit 4fa2511a44
9 changed files with 284 additions and 20 deletions

View File

@ -20,6 +20,11 @@
extern afs_int32 hm_retry_RO;
extern afs_int32 hm_retry_RW;
extern afs_int32 hm_retry_int;
extern afs_int32 afs_blocksUsed_0;
extern afs_int32 afs_blocksUsed_1;
extern afs_int32 afs_blocksUsed_2;
extern afs_int32 afs_pct1;
extern afs_int32 afs_pct2;
#ifdef CONFIG_SYSCTL
static struct ctl_table_header *afs_sysctl = NULL;
@ -49,6 +54,34 @@ static ctl_table afs_sysctl_table[] = {
&afs_bkvolpref, sizeof(afs_int32), 0644, NULL,
&proc_dointvec}
,
{7, "afs_blocksUsed",
&afs_blocksUsed, sizeof(afs_int32), 0444, NULL,
&proc_dointvec}
,
{8, "afs_blocksUsed_0",
&afs_blocksUsed_0, sizeof(afs_int32), 0644, NULL,
&proc_dointvec}
,
{9, "afs_blocksUsed_1",
&afs_blocksUsed_1, sizeof(afs_int32), 0644, NULL,
&proc_dointvec}
,
{10, "afs_blocksUsed_2",
&afs_blocksUsed_2, sizeof(afs_int32), 0644, NULL,
&proc_dointvec}
,
{11, "afs_pct1",
&afs_pct1, sizeof(afs_int32), 0644, NULL,
&proc_dointvec}
,
{12, "afs_pct2",
&afs_pct2, sizeof(afs_int32), 0644, NULL,
&proc_dointvec}
,
{13, "afs_cacheBlocks",
&afs_cacheBlocks, sizeof(afs_int32), 0644, NULL,
&proc_dointvec}
,
{0}
};

View File

@ -828,6 +828,9 @@ struct cm_initparams {
#define NULLIDX (-1) /* null index definition */
/* struct dcache states bits */
#define DRO 1
#define DBackup 2
#define DRW 4
#define DWriting 8 /* file being written (used for cache validation) */
/* dcache data flags */
@ -953,7 +956,7 @@ struct dcache {
char mflags; /* Meta flags */
struct fcache f; /* disk image */
afs_int32 stamp; /* used with vtodc struct for hints */
afs_int32 bucket; /* which bucket these dcache entries are in */
/*
* Locking rules:
*

View File

@ -517,6 +517,18 @@ afs_syscall_call(parm, parm2, parm3, parm4, parm5, parm6)
DInit(temp);
afs_rootFid.Fid.Volume = 0;
code = 0;
} else if (parm == AFSOP_BUCKETPCT) {
/* need to enable this now, will disable again before GO
if we don't have 100% */
splitdcache = 1;
switch (parm2) {
case 1:
afs_tpct1 = parm3;
break;
case 2:
afs_tpct2 = parm3;
break;
}
} else if (parm == AFSOP_ADDCELL) {
/* add a cell. Parameter 2 is 8 hosts (in net order), parm 3 is the null-terminated
* name. Parameter 4 is the length of the name, including the null. Parm 5 is the
@ -693,6 +705,13 @@ afs_syscall_call(parm, parm2, parm3, parm4, parm5, parm6)
afs_osi_Sleep(&afs_initState);
afs_initState = 101;
afs_setTime = parm2;
if (afs_tpct1 + afs_tpct2 != 100) {
afs_tpct1 = 0;
afs_tpct2 = 0;
splitdcache = 0;
} else {
splitdcache = 1;
}
afs_osi_Wakeup(&afs_initState);
#if (!defined(AFS_NONFSTRANS)) || defined(AFS_AIX_IAUTH_ENV)
afs_nfsclient_init();

View File

@ -23,14 +23,31 @@ RCSID
#include "afs/afs_osidnlc.h"
/* Forward declarations. */
static void afs_GetDownD(int anumber, int *aneedSpace);
static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
static void afs_FreeDiscardedDCache(void);
static void afs_DiscardDCache(struct dcache *);
static void afs_FreeDCache(struct dcache *);
/* For split cache */
static afs_int32 afs_DCGetBucket(struct vcache *);
static void afs_DCAdjustSize(struct dcache *, afs_int32, afs_int32);
static void afs_DCMoveBucket(struct dcache *, afs_int32, afs_int32);
static void afs_DCSizeInit(void);
static afs_int32 afs_DCWhichBucket(afs_int32, afs_int32);
/*
* --------------------- Exported definitions ---------------------
*/
/* For split cache */
afs_int32 afs_blocksUsed_0; /*1K blocks in cache - in theory is zero */
afs_int32 afs_blocksUsed_1; /*1K blocks in cache */
afs_int32 afs_blocksUsed_2; /*1K blocks in cache */
afs_int32 afs_pct1 = -1;
afs_int32 afs_pct2 = -1;
afs_uint32 afs_tpct1 = 0;
afs_uint32 afs_tpct2 = 0;
afs_uint32 splitdcache = 0;
afs_lock_t afs_xdcache; /*Lock: alloc new disk cache entries */
afs_int32 afs_freeDCList; /*Free list for disk cache entries */
afs_int32 afs_freeDCCount; /*Count of elts in freeDCList */
@ -119,7 +136,110 @@ struct afs_cacheOps afs_MemCacheOps = {
int cacheDiskType; /*Type of backing disk for cache */
struct afs_cacheOps *afs_cacheType;
static afs_int32
afs_DCGetBucket(struct vcache *avc)
{
/* This should be replaced with some sort of user configurable function */
if (avc->states & CRO) {
return 2;
} else if (avc->states & CBackup) {
return 1;
} else {
/* RW */
}
/* main bucket */
return 1;
}
static void
afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
{
afs_int32 adjustSize = newSize - oldSize;
if (!splitdcache)
return;
switch (adc->bucket)
{
case 0:
afs_blocksUsed_0 += adjustSize;
afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
break;
case 1:
afs_blocksUsed_1 += adjustSize;
afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
break;
case 2:
afs_blocksUsed_2 += adjustSize;
afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
break;
}
return;
}
static void
afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
{
if (!splitdcache)
return;
switch (adc->bucket)
{
case 0:
afs_blocksUsed_0 -= size;
break;
case 1:
afs_blocksUsed_1 -= size;
break;
case 2:
afs_blocksUsed_2 -= size;
break;
}
adc->bucket = newBucket;
switch (adc->bucket)
{
case 0:
afs_blocksUsed_0 += size;
break;
case 1:
afs_blocksUsed_1 += size;
break;
case 2:
afs_blocksUsed_2 += size;
break;
}
return;
}
static void
afs_DCSizeInit(void)
{
afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
}
static afs_int32
afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
{
if (!splitdcache)
return 0;
afs_pct1 = afs_blocksUsed_1*100/afs_cacheBlocks;
afs_pct2 = afs_blocksUsed_2*100/afs_cacheBlocks;
/* Short cut: if we don't know about it, try to kill it */
if (phase < 2 && afs_blocksUsed_0)
return 0;
if (afs_pct1 > afs_tpct1)
return 1;
if (afs_pct2 > afs_tpct2)
return 2;
return 0; /* unlikely */
}
/*
@ -245,7 +365,7 @@ afs_CacheTruncateDaemon(void)
afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
slots_needed =
dc_hiwat - afs_freeDCCount - afs_discardDCCount;
afs_GetDownD(slots_needed, &space_needed);
afs_GetDownD(slots_needed, &space_needed, 0);
if ((space_needed <= 0) && (slots_needed <= 0)) {
break;
}
@ -340,6 +460,7 @@ afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
if (!newSize)
adc->validPos = 0;
newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
afs_DCAdjustSize(adc, oldSize, newSize);
if (newSize > oldSize) {
/* We're growing the file, wakeup the daemon */
afs_MaybeWakeupTruncateDaemon();
@ -376,7 +497,7 @@ afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
#define MAXATONCE 16 /* max we can obtain at once */
static void
afs_GetDownD(int anumber, int *aneedSpace)
afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
{
struct dcache *tdc;
@ -392,6 +513,7 @@ afs_GetDownD(int anumber, int *aneedSpace)
afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
afs_uint32 maxVictimPtr; /* where it is */
int discard;
int curbucket;
AFS_STATCNT(afs_GetDownD);
if (CheckLock(&afs_xdcache) != -1)
@ -408,12 +530,20 @@ afs_GetDownD(int anumber, int *aneedSpace)
if (anumber > MAXATONCE)
anumber = MAXATONCE; /* all we can do */
/* rewrite so phases include a better eligiblity for gc test*/
/*
* The phase variable manages reclaims. Set to 0, the first pass,
* we don't reclaim active entries. Set to 1, we reclaim even active
* ones.
* we don't reclaim active entries, or other than target bucket.
* Set to 1, we reclaim even active ones in target bucket.
* Set to 2, we reclaim any inactive one.
* Set to 3, we reclaim even active ones.
*/
phase = 0;
if (splitdcache) {
phase = 0;
} else {
phase = 4;
}
for (i = 0; i < afs_cacheFiles; i++)
/* turn off all flags */
afs_indexFlags[i] &= ~IFFlag;
@ -422,6 +552,7 @@ afs_GetDownD(int anumber, int *aneedSpace)
/* find oldest entries for reclamation */
maxVictimPtr = victimPtr = 0;
hzero(maxVictimTime);
curbucket = afs_DCWhichBucket(phase, buckethint);
/* select victims from access time array */
for (i = 0; i < afs_cacheFiles; i++) {
if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
@ -429,6 +560,11 @@ afs_GetDownD(int anumber, int *aneedSpace)
continue;
}
tdc = afs_indexTable[i];
if (tdc && (curbucket != tdc->bucket) && (phase < 4))
{
/* Wrong bucket; can't use it! */
continue;
}
if (tdc && (tdc->refCount != 0)) {
/* Referenced; can't use it! */
continue;
@ -510,12 +646,12 @@ afs_GetDownD(int anumber, int *aneedSpace)
if (tvc) {
tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
chunkFlags = afs_indexFlags[tdc->index];
if (phase == 0 && osi_Active(tvc))
skip = 1;
if (phase > 0 && osi_Active(tvc)
&& (tvc->states & CDCLock)
&& (chunkFlags & IFAnyPages))
skip = 1;
if ((((phase / 2) & 1) == 0) && osi_Active(tvc))
skip = 1;
if ((((phase / 2) & 1) == 1) && osi_Active(tvc)
&& (tvc->states & CDCLock)
&& (chunkFlags & IFAnyPages))
skip = 1;
if (chunkFlags & IFDataMod)
skip = 1;
afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
@ -647,17 +783,17 @@ afs_GetDownD(int anumber, int *aneedSpace)
afs_PutDCache(tdc);
}
if (phase == 0) {
if (phase < 5) {
/* Phase is 0 and no one was found, so try phase 1 (ignore
* osi_Active flag) */
if (j == 0) {
phase = 1;
phase++;
for (i = 0; i < afs_cacheFiles; i++)
/* turn off all flags */
afs_indexFlags[i] &= ~IFFlag;
}
} else {
/* found no one in phase 1, we're hosed */
/* found no one in phases 0-5, we're hosed */
if (victimPtr == 0)
break;
}
@ -916,6 +1052,8 @@ afs_FreeDiscardedDCache(void)
afs_CFileTruncate(tfile, 0);
afs_CFileClose(tfile);
afs_AdjustSize(tdc, 0);
tdc->f.states &= ~(DRO|DBackup|DRW);
afs_DCMoveBucket(tdc, 0, 0);
/*
* Free the element we just truncated
@ -1214,6 +1352,7 @@ afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
break; /* leaving refCount high for caller */
}
afs_PutDCache(tdc);
tdc = NULL;
}
index = afs_dcnextTbl[index];
}
@ -1707,7 +1846,8 @@ afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
while (1) {
if (!setLocks)
avc->states |= CDCLock;
afs_GetDownD(5, (int *)0); /* just need slots */
/* just need slots */
afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
if (!setLocks)
avc->states &= ~CDCLock;
if (afs_discardDCList != NULLIDX
@ -1748,6 +1888,8 @@ afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
size =
((tdc->f.chunkBytes +
afs_fsfragsize) ^ afs_fsfragsize) >> 10;
tdc->f.states &= ~(DRO|DBackup|DRW);
afs_DCMoveBucket(tdc, size, 0);
afs_blocksDiscarded -= size;
afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
if (aflags & 2) {
@ -1772,6 +1914,13 @@ afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
*/
afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
tdc->f.fid = avc->fid;
if (avc->states & CRO)
tdc->f.states = DRO;
else if (avc->states & CBackup)
tdc->f.states = DBackup;
else
tdc->f.states = DRW;
afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
hones(tdc->f.versionNo); /* invalid value */
tdc->f.chunk = chunk;
@ -1791,7 +1940,6 @@ afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
afs_dvhashTbl[i] = tdc->index;
tdc->dflags = DFEntryMod;
tdc->mflags = 0;
tdc->f.states = 0;
afs_MaybeWakeupTruncateDaemon();
MReleaseWriteLock(&afs_xdcache);
ConvertWToSLock(&tdc->lock);
@ -2414,6 +2562,8 @@ afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
if (vType(avc) == VDIR) {
DZap(tdc);
}
tdc->f.states &= ~(DRO|DBackup|DRW);
afs_DCMoveBucket(tdc, 0, 0);
ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
ObtainWriteLock(&afs_xcbhash, 454);
@ -2427,6 +2577,8 @@ afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
* avc->lock(W); assert(!setLocks || slowPass)
*/
osi_Assert(!setLocks || slowPass);
tdc->f.states &= ~(DRO|DBackup|DRW);
afs_DCMoveBucket(tdc, 0, 0);
tdc = NULL;
goto done;
}
@ -2837,6 +2989,18 @@ afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
#endif
lasterrtime = osi_Time();
afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
tdc->f.states &= ~(DRO|DBackup|DRW);
afs_DCMoveBucket(tdc, 0, 0);
} else {
if (&tdc->f != 0) {
if (tdc->f.states & DRO) {
afs_DCMoveBucket(tdc, 0, 2);
} else if (tdc->f.states & DBackup) {
afs_DCMoveBucket(tdc, 0, 1);
} else {
afs_DCMoveBucket(tdc, 0, 1);
}
}
}
tdc->refCount = 1;
tdc->index = aslot;
@ -3060,6 +3224,8 @@ afs_InitCacheFile(char *afile, ino_t ainode)
tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
if (tstat.size != 0)
osi_UFSTruncate(tfile, 0);
tdc->f.states &= ~(DRO|DBackup|DRW);
afs_DCMoveBucket(tdc, 0, 0);
/* put entry in free cache slot list */
afs_dvnextTbl[tdc->index] = afs_freeDCList;
afs_freeDCList = index;
@ -3234,6 +3400,10 @@ afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
afs_dcentries = aDentries;
afs_blocksUsed = 0;
afs_stats_cmperf.cacheBucket0_Discarded =
afs_stats_cmperf.cacheBucket1_Discarded =
afs_stats_cmperf.cacheBucket2_Discarded = 0;
afs_DCSizeInit();
QInit(&afs_DLRU);
}
@ -3274,6 +3444,9 @@ shutdown_dcache(void)
afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
afs_blocksUsed = afs_dcentries = 0;
afs_stats_cmperf.cacheBucket0_Discarded =
afs_stats_cmperf.cacheBucket1_Discarded =
afs_stats_cmperf.cacheBucket2_Discarded = 0;
hzero(afs_indexCounter);
afs_freeDCCount = 0;

View File

@ -297,6 +297,7 @@ extern afs_int32 afs_blocksUsed;
extern afs_int32 afs_blocksDiscarded;
extern int afs_WaitForCacheDrain;
extern int cacheDiskType;
extern afs_uint32 afs_tpct1, afs_tpct2, splitdcache;
extern unsigned char *afs_indexFlags;
extern struct afs_cacheOps *afs_cacheType;
extern ino_t cacheInode;

View File

@ -802,10 +802,14 @@ struct afs_stats_CMPerf {
afs_uint32 cbloops;
afs_uint32 osiread_efaults;
afs_int32 cacheBlocksDiscarded; /*# cache blocks free but not truncated */
afs_int32 cacheBucket0_Discarded;
afs_int32 cacheBucket1_Discarded;
afs_int32 cacheBucket2_Discarded;
/*
* Spares for future expansion.
*/
afs_int32 spare[13]; /*Spares */
afs_int32 spare[10]; /*Spares */
};

View File

@ -230,6 +230,8 @@ afs_int32 enable_rxbind = 0;
afs_int32 afs_shutdown = 0;
afs_int32 cacheBlocks; /*Num blocks in the cache */
afs_int32 cacheFiles = 1000; /*Optimal # of files in workstation cache */
afs_int32 rwpct = 0;
afs_int32 ropct = 0;
afs_int32 cacheStatEntries = 300; /*Number of stat cache entries */
char cacheBaseDir[1024]; /*Where the workstation AFS cache lives */
char confDir[1024]; /*Where the workstation AFS configuration lives */
@ -275,6 +277,7 @@ static int enable_dynroot = 0; /* enable dynroot support */
static int enable_fakestat = 0; /* enable fakestat support */
static int enable_backuptree = 0; /* enable backup tree support */
static int enable_nomount = 0; /* do not mount */
static int enable_splitcache = 0;
#ifdef notdef
static int inodes = 60; /* VERY conservative, but has to be */
#endif
@ -1535,7 +1538,21 @@ mainproc(as, arock)
exit(1);
}
}
if (as->parms[34].items) {
char *c;
if (!as->parms[34].items->data ||
((c = strchr(as->parms[34].items->data, '/')) == NULL))
printf("ignoring splitcache (specify as RW/RO percentages: 60/40)\n");
else {
ropct = atoi((char *)c+1);
*c = '\0';
rwpct = atoi((char *)as->parms[30].items->data);
if ((rwpct != 0) && (ropct != 0) && (ropct+rwpct == 100)) {
/* -splitcache */
enable_splitcache = 1;
}
}
}
/*
* Pull out all the configuration info for the workstation's AFS cache and
* the cellular community we're willing to let our users see.
@ -1813,6 +1830,13 @@ mainproc(as, arock)
cparams.inodes = inodes;
#endif
call_syscall(AFSOP_CACHEINIT, &cparams);
/* do it before we init the cache inodes */
if (enable_splitcache) {
call_syscall(AFSOP_BUCKETPCT, 1, rwpct);
call_syscall(AFSOP_BUCKETPCT, 2, ropct);
}
if (afsd_CloseSynch)
call_syscall(AFSOP_CLOSEWAIT);
@ -2178,6 +2202,8 @@ main(argc, argv)
cmd_AddParm(ts, "-settime", CMD_FLAG, CMD_OPTIONAL,
"don't set the time");
cmd_AddParm(ts, "-rxpck", CMD_SINGLE, CMD_OPTIONAL, "set rx_extraPackets to this value");
cmd_AddParm(ts, "-splitcache", CMD_SINGLE, CMD_OPTIONAL, "Percentage RW versus RO in cache (specify as 60/40)");
return (cmd_Dispatch(argc, argv));
}

View File

@ -47,6 +47,7 @@
#define AFSOP_BASIC_INIT 36 /* used to be part of START_AFS */
#define AFSOP_SET_BACKUPTREE 37 /* enable backup tree support */
#define AFSOP_SET_RXPCK 38 /*set rx_extraPackets*/
#define AFSOP_BUCKETPCT 39 /* bucket percentage */
/* The range 20-30 is reserved for AFS system offsets in the afs_syscall */
#define AFSCALL_PIOCTL 20

View File

@ -755,6 +755,10 @@ PrintOverallPerfInfo(a_ovP)
printf("\t%10d srvMaxChainLengthHWM\n", a_ovP->srvMaxChainLengthHWM);
printf("\t%10d srvRecordsHWM\n", a_ovP->srvRecordsHWM);
printf("\t%10d cacheBucket0_Discarded\n", a_ovP->cacheBucket0_Discarded);
printf("\t%10d cacheBucket1_Discarded\n", a_ovP->cacheBucket1_Discarded);
printf("\t%10d cacheBucket2_Discarded\n", a_ovP->cacheBucket2_Discarded);
printf("\t%10d sysName_ID\n", a_ovP->sysName_ID);
printf("\tFile Server up/downtimes, same cell:\n");