NetBSD/sys/dev/raidframe/rf_paritymap.c
jld f1a1ad338d Finally commit the RAIDframe parity map Summer Of Code project.
Drastically reduces the amount of time spent rewriting parity after an
unclean shutdown by keeping better track of which regions might have had
outstanding writes.  Enabled by default; can be disabled on a per-set
basis, or tuned, with the new raidctl(8) commands.

Discussed on tech-kern@ to a general air of approval; exhortations to
commit from mrg@, christos@, and others.

Thanks to Google for their sponsorship, oster@ for mentoring the
project, assorted developers for trying very hard to break it, and
probably more I'm forgetting.
2009-11-17 18:54:26 +00:00

751 lines
20 KiB
C

/* $NetBSD: rf_paritymap.c,v 1.1 2009/11/17 18:54:26 jld Exp $ */
/*-
* Copyright (c) 2009 Jed Davis.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
* ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
* TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
* BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: rf_paritymap.c,v 1.1 2009/11/17 18:54:26 jld Exp $");
#include <sys/callout.h>
#include <sys/kmem.h>
#include <sys/mutex.h>
#include <sys/param.h>
#include <sys/rwlock.h>
#include <sys/systm.h>
#include <sys/types.h>
#include <dev/raidframe/rf_paritymap.h>
#include <dev/raidframe/rf_stripelocks.h>
#include <dev/raidframe/rf_layout.h>
#include <dev/raidframe/rf_raid.h>
#include <dev/raidframe/rf_parityscan.h>
#include <dev/raidframe/rf_kintf.h>
/* Important parameters: */
#define REGION_MINSIZE (25ULL << 20)
#define DFL_TICKMS 40000
#define DFL_COOLDOWN 8 /* 7-8 intervals of 40s = 5min +/- 20s */
/* Internal-use flag bits. */
#define TICKING 1
#define TICKED 2
/* Prototypes! */
static void rf_paritymap_write_locked(struct rf_paritymap *);
static void rf_paritymap_tick(void *);
static u_int rf_paritymap_nreg(RF_Raid_t *);
/* Extract the current status of the parity map. */
void
rf_paritymap_status(struct rf_paritymap *pm, struct rf_pmstat *ps)
{
memset(ps, 0, sizeof(*ps));
if (pm == NULL)
ps->enabled = 0;
else {
ps->enabled = 1;
ps->region_size = pm->region_size;
mutex_enter(&pm->lock);
memcpy(&ps->params, &pm->params, sizeof(ps->params));
memcpy(ps->dirty, pm->disk_now, sizeof(ps->dirty));
memcpy(&ps->ctrs, &pm->ctrs, sizeof(ps->ctrs));
mutex_exit(&pm->lock);
}
}
/*
* Test whether parity in a given sector is suspected of being inconsistent
* on disk (assuming that any pending I/O to it is allowed to complete).
* This may be of interest to future work on parity scrubbing.
*/
int
rf_paritymap_test(struct rf_paritymap *pm, daddr_t sector)
{
unsigned region = sector / pm->region_size;
int retval;
mutex_enter(&pm->lock);
retval = isset(pm->disk_boot->bits, region) ? 1 : 0;
mutex_exit(&pm->lock);
return retval;
}
/* To be called before a write to the RAID is submitted. */
void
rf_paritymap_begin(struct rf_paritymap *pm, daddr_t offset, daddr_t size)
{
unsigned i, b, e;
b = offset / pm->region_size;
e = (offset + size - 1) / pm->region_size;
for (i = b; i <= e; i++)
rf_paritymap_begin_region(pm, i);
}
/* To be called after a write to the RAID completes. */
void
rf_paritymap_end(struct rf_paritymap *pm, daddr_t offset, daddr_t size)
{
unsigned i, b, e;
b = offset / pm->region_size;
e = (offset + size - 1) / pm->region_size;
for (i = b; i <= e; i++)
rf_paritymap_end_region(pm, i);
}
void
rf_paritymap_begin_region(struct rf_paritymap *pm, unsigned region)
{
int needs_write;
KASSERT(region < RF_PARITYMAP_NREG);
pm->ctrs.nwrite++;
/* If it was being kept warm, deal with that. */
mutex_enter(&pm->lock);
if (pm->current->state[region] < 0)
pm->current->state[region] = 0;
/* This shouldn't happen unless RAIDOUTSTANDING is set too high. */
KASSERT(pm->current->state[region] < 127);
pm->current->state[region]++;
needs_write = isclr(pm->disk_now->bits, region);
if (needs_write) {
KASSERT(pm->current->state[region] == 1);
rf_paritymap_write_locked(pm);
}
mutex_exit(&pm->lock);
}
void
rf_paritymap_end_region(struct rf_paritymap *pm, unsigned region)
{
KASSERT(region < RF_PARITYMAP_NREG);
mutex_enter(&pm->lock);
KASSERT(pm->current->state[region] > 0);
--pm->current->state[region];
if (pm->current->state[region] <= 0) {
pm->current->state[region] = -pm->params.cooldown;
KASSERT(pm->current->state[region] <= 0);
mutex_enter(&pm->lk_flags);
if (!(pm->flags & TICKING)) {
pm->flags |= TICKING;
mutex_exit(&pm->lk_flags);
callout_schedule(&pm->ticker,
mstohz(pm->params.tickms));
} else
mutex_exit(&pm->lk_flags);
}
mutex_exit(&pm->lock);
}
/*
* Updates the parity map to account for any changes in current activity
* and/or an ongoing parity scan, then writes it to disk with appropriate
* synchronization.
*/
void
rf_paritymap_write(struct rf_paritymap *pm)
{
mutex_enter(&pm->lock);
rf_paritymap_write_locked(pm);
mutex_exit(&pm->lock);
}
/* As above, but to be used when pm->lock is already held. */
static void
rf_paritymap_write_locked(struct rf_paritymap *pm)
{
char w, w0;
int i, j, setting, clearing;
setting = clearing = 0;
for (i = 0; i < RF_PARITYMAP_NBYTE; i++) {
w0 = pm->disk_now->bits[i];
w = pm->disk_boot->bits[i];
for (j = 0; j < NBBY; j++)
if (pm->current->state[i * NBBY + j] != 0)
w |= 1 << j;
if (w & ~w0)
setting = 1;
if (w0 & ~w)
clearing = 1;
pm->disk_now->bits[i] = w;
}
pm->ctrs.ncachesync += setting + clearing;
pm->ctrs.nclearing += clearing;
/*
* If bits are being set in the parity map, then a sync is
* required afterwards, so that the regions are marked dirty
* on disk before any writes to them take place. If bits are
* being cleared, then a sync is required before the write, so
* that any writes to those regions are processed before the
* region is marked clean. (Synchronization is somewhat
* overkill; a write ordering barrier would suffice, but we
* currently have no way to express that directly.)
*/
if (clearing)
rf_sync_component_caches(pm->raid);
rf_paritymap_kern_write(pm->raid, pm->disk_now);
if (setting)
rf_sync_component_caches(pm->raid);
}
/* Mark all parity as being in need of rewrite. */
void
rf_paritymap_invalidate(struct rf_paritymap *pm)
{
mutex_enter(&pm->lock);
memset(pm->disk_boot, ~(unsigned char)0,
sizeof(struct rf_paritymap_ondisk));
mutex_exit(&pm->lock);
}
/* Mark all parity as being correct. */
void
rf_paritymap_forceclean(struct rf_paritymap *pm)
{
mutex_enter(&pm->lock);
memset(pm->disk_boot, (unsigned char)0,
sizeof(struct rf_paritymap_ondisk));
mutex_exit(&pm->lock);
}
/*
* The cooldown callout routine just defers its work to a thread; it can't do
* the parity map write itself as it would block, and although mutex-induced
* blocking is permitted it seems wise to avoid tying up the softint.
*/
static void
rf_paritymap_tick(void *arg)
{
struct rf_paritymap *pm = arg;
mutex_enter(&pm->lk_flags);
pm->flags |= TICKED;
mutex_exit(&pm->lk_flags);
wakeup(&(pm->raid->iodone)); /* XXX */
}
/*
* This is where the parity cooling work (and rearming the callout if needed)
* is done; the raidio thread calls it when woken up, as by the above.
*/
void
rf_paritymap_checkwork(struct rf_paritymap *pm)
{
int i, zerop, progressp;
mutex_enter(&pm->lk_flags);
if (pm->flags & TICKED) {
zerop = progressp = 0;
pm->flags &= ~TICKED;
mutex_exit(&pm->lk_flags);
mutex_enter(&pm->lock);
for (i = 0; i < RF_PARITYMAP_NREG; i++) {
if (pm->current->state[i] < 0) {
progressp = 1;
pm->current->state[i]++;
if (pm->current->state[i] == 0)
zerop = 1;
}
}
if (progressp)
callout_schedule(&pm->ticker,
mstohz(pm->params.tickms));
else {
mutex_enter(&pm->lk_flags);
pm->flags &= ~TICKING;
mutex_exit(&pm->lk_flags);
}
if (zerop)
rf_paritymap_write_locked(pm);
mutex_exit(&pm->lock);
} else
mutex_exit(&pm->lk_flags);
}
/*
* Set parity map parameters; used both to alter parameters on the fly and to
* establish their initial values. Note that setting a parameter to 0 means
* to leave the previous setting unchanged, and that if this is done for the
* initial setting of "regions", then a default value will be computed based
* on the RAID component size.
*/
int
rf_paritymap_set_params(struct rf_paritymap *pm,
const struct rf_pmparams *params, int todisk)
{
int cooldown, tickms;
u_int regions;
RF_RowCol_t col;
RF_ComponentLabel_t *clabel;
RF_Raid_t *raidPtr;
cooldown = params->cooldown != 0
? params->cooldown : pm->params.cooldown;
tickms = params->tickms != 0
? params->tickms : pm->params.tickms;
regions = params->regions != 0
? params->regions : pm->params.regions;
if (cooldown < 1 || cooldown > 128) {
printf("raid%d: cooldown %d out of range\n", pm->raid->raidid,
cooldown);
return (-1);
}
if (tickms < 10) {
printf("raid%d: tick time %dms out of range\n",
pm->raid->raidid, tickms);
return (-1);
}
if (regions == 0) {
regions = rf_paritymap_nreg(pm->raid);
} else if (regions > RF_PARITYMAP_NREG) {
printf("raid%d: region count %u too large (more than %u)\n",
pm->raid->raidid, regions, RF_PARITYMAP_NREG);
return (-1);
}
/* XXX any currently warm parity will be used with the new tickms! */
pm->params.cooldown = cooldown;
pm->params.tickms = tickms;
/* Apply the initial region count, but do not change it after that. */
if (pm->params.regions == 0)
pm->params.regions = regions;
/* So that the newly set parameters can be tested: */
pm->ctrs.nwrite = pm->ctrs.ncachesync = pm->ctrs.nclearing = 0;
if (todisk) {
raidPtr = pm->raid;
for (col = 0; col < raidPtr->numCol; col++) {
clabel = raidget_component_label(raidPtr, col);
clabel->parity_map_ntick = cooldown;
clabel->parity_map_tickms = tickms;
clabel->parity_map_regions = regions;
raidflush_component_label(raidPtr, col);
}
}
return 0;
}
/*
* The number of regions may not be as many as can fit into the map, because
* when regions are too small, the overhead of setting parity map bits
* becomes significant in comparison to the actual I/O, while the
* corresponding gains in parity verification time become negligible. Thus,
* a minimum region size (defined above) is imposed.
*
* Note that, if the number of regions is less than the maximum, then some of
* the regions will be "fictional", corresponding to no actual disk; some
* parts of the code may process them as normal, but they can not ever be
* written to.
*/
static u_int
rf_paritymap_nreg(RF_Raid_t *raid)
{
daddr_t bytes_per_disk, nreg;
bytes_per_disk = raid->sectorsPerDisk << raid->logBytesPerSector;
nreg = bytes_per_disk / REGION_MINSIZE;
if (nreg > RF_PARITYMAP_NREG)
nreg = RF_PARITYMAP_NREG;
return (u_int)nreg;
}
/*
* Initialize a parity map given specific parameters. This neither reads nor
* writes the parity map config in the component labels; for that, see below.
*/
int
rf_paritymap_init(struct rf_paritymap *pm, RF_Raid_t *raid,
const struct rf_pmparams *params)
{
daddr_t rstripes;
struct rf_pmparams safe;
pm->raid = raid;
pm->params.regions = 0;
if (0 != rf_paritymap_set_params(pm, params, 0)) {
/*
* If the parameters are out-of-range, then bring the
* parity map up with something reasonable, so that
* the admin can at least go and fix it (or ignore it
* entirely).
*/
safe.cooldown = DFL_COOLDOWN;
safe.tickms = DFL_TICKMS;
safe.regions = 0;
if (0 != rf_paritymap_set_params(pm, &safe, 0))
return (-1);
}
rstripes = howmany(raid->Layout.numStripe, pm->params.regions);
pm->region_size = rstripes * raid->Layout.dataSectorsPerStripe;
callout_init(&pm->ticker, CALLOUT_MPSAFE);
callout_setfunc(&pm->ticker, rf_paritymap_tick, pm);
pm->flags = 0;
pm->disk_boot = kmem_alloc(sizeof(struct rf_paritymap_ondisk),
KM_SLEEP);
pm->disk_now = kmem_alloc(sizeof(struct rf_paritymap_ondisk),
KM_SLEEP);
pm->current = kmem_zalloc(sizeof(struct rf_paritymap_current),
KM_SLEEP);
rf_paritymap_kern_read(pm->raid, pm->disk_boot);
memcpy(pm->disk_now, pm->disk_boot, sizeof(*pm->disk_now));
mutex_init(&pm->lock, MUTEX_DEFAULT, IPL_NONE);
mutex_init(&pm->lk_flags, MUTEX_DEFAULT, IPL_SOFTCLOCK);
return 0;
}
/*
* Destroys a parity map; unless "force" is set, also cleans parity for any
* regions which were still in cooldown (but are not dirty on disk).
*/
void
rf_paritymap_destroy(struct rf_paritymap *pm, int force)
{
int i;
callout_halt(&pm->ticker, NULL); /* XXX stop? halt? */
callout_destroy(&pm->ticker);
if (!force) {
for (i = 0; i < RF_PARITYMAP_NREG; i++) {
/* XXX check for > 0 ? */
if (pm->current->state[i] < 0)
pm->current->state[i] = 0;
}
rf_paritymap_write_locked(pm);
}
mutex_destroy(&pm->lock);
mutex_destroy(&pm->lk_flags);
kmem_free(pm->disk_boot, sizeof(struct rf_paritymap_ondisk));
kmem_free(pm->disk_now, sizeof(struct rf_paritymap_ondisk));
kmem_free(pm->current, sizeof(struct rf_paritymap_current));
}
/*
* Rewrite parity, taking parity map into account; this is the equivalent of
* the old rf_RewriteParity, and is likewise to be called from a suitable
* thread and shouldn't have multiple copies running in parallel and so on.
*
* Note that the fictional regions are "cleaned" in one shot, so that very
* small RAIDs (useful for testing) will not experience potentially severe
* regressions in rewrite time.
*/
int
rf_paritymap_rewrite(struct rf_paritymap *pm)
{
int i, ret_val = 0;
daddr_t reg_b, reg_e;
/* Process only the actual regions. */
for (i = 0; i < pm->params.regions; i++) {
mutex_enter(&pm->lock);
if (isset(pm->disk_boot->bits, i)) {
mutex_exit(&pm->lock);
reg_b = i * pm->region_size;
reg_e = reg_b + pm->region_size;
if (reg_e > pm->raid->totalSectors)
reg_e = pm->raid->totalSectors;
if (rf_RewriteParityRange(pm->raid, reg_b,
reg_e - reg_b)) {
ret_val = 1;
if (pm->raid->waitShutdown)
return ret_val;
} else {
mutex_enter(&pm->lock);
clrbit(pm->disk_boot->bits, i);
rf_paritymap_write_locked(pm);
mutex_exit(&pm->lock);
}
} else {
mutex_exit(&pm->lock);
}
}
/* Now, clear the fictional regions, if any. */
rf_paritymap_forceclean(pm);
rf_paritymap_write(pm);
return ret_val;
}
/*
* How to merge the on-disk parity maps when reading them in from the
* various components; returns whether they differ. In the case that
* they do differ, sets *dst to the union of *dst and *src.
*
* In theory, it should be safe to take the intersection (or just pick
* a single component arbitrarily), but the paranoid approach costs
* little.
*
* Appropriate locking, if any, is the responsibility of the caller.
*/
int
rf_paritymap_merge(struct rf_paritymap_ondisk *dst,
struct rf_paritymap_ondisk *src)
{
int i, discrep = 0;
for (i = 0; i < RF_PARITYMAP_NBYTE; i++) {
if (dst->bits[i] != src->bits[i])
discrep = 1;
dst->bits[i] |= src->bits[i];
}
return discrep;
}
/*
* Detach a parity map from its RAID. This is not meant to be applied except
* when unconfiguring the RAID after all I/O has been resolved, as otherwise
* an out-of-date parity map could be treated as current.
*/
void
rf_paritymap_detach(RF_Raid_t *raidPtr)
{
if (raidPtr->parity_map == NULL)
return;
simple_lock(&(raidPtr->iodone_lock));
struct rf_paritymap *pm = raidPtr->parity_map;
raidPtr->parity_map = NULL;
simple_unlock(&(raidPtr->iodone_lock));
/* XXXjld is that enough locking? Or too much? */
rf_paritymap_destroy(pm, 0);
kmem_free(pm, sizeof(*pm));
}
/*
* Attach a parity map to a RAID set if appropriate. Includes
* configure-time processing of parity-map fields of component label.
*/
void
rf_paritymap_attach(RF_Raid_t *raidPtr, int force)
{
RF_RowCol_t col;
int pm_use, pm_zap;
int g_tickms, g_ntick, g_regions;
int good;
RF_ComponentLabel_t *clabel;
u_int flags, regions;
struct rf_pmparams params;
if (raidPtr->Layout.map->faultsTolerated == 0) {
/* There isn't any parity. */
return;
}
pm_use = 1;
pm_zap = 0;
g_tickms = DFL_TICKMS;
g_ntick = DFL_COOLDOWN;
g_regions = 0;
/*
* Collect opinions on the set config. If this is the initial
* config (raidctl -C), treat all labels as invalid, since
* there may be random data present.
*/
if (!force) {
for (col = 0; col < raidPtr->numCol; col++) {
clabel = raidget_component_label(raidPtr, col);
flags = clabel->parity_map_flags;
/* Check for use by non-parity-map kernel. */
if (clabel->parity_map_modcount
!= clabel->mod_counter) {
flags &= ~RF_PMLABEL_WASUSED;
}
if (flags & RF_PMLABEL_VALID) {
g_tickms = clabel->parity_map_tickms;
g_ntick = clabel->parity_map_ntick;
regions = clabel->parity_map_regions;
if (g_regions == 0)
g_regions = regions;
else if (g_regions != regions) {
pm_zap = 1; /* important! */
}
if (flags & RF_PMLABEL_DISABLE) {
pm_use = 0;
}
if (!(flags & RF_PMLABEL_WASUSED)) {
pm_zap = 1;
}
} else {
pm_zap = 1;
}
}
} else {
pm_zap = 1;
}
/* Finally, create and attach the parity map. */
if (pm_use) {
params.cooldown = g_ntick;
params.tickms = g_tickms;
params.regions = g_regions;
raidPtr->parity_map = kmem_alloc(sizeof(struct rf_paritymap),
KM_SLEEP);
if (0 != rf_paritymap_init(raidPtr->parity_map, raidPtr,
&params)) {
/* It failed; do without. */
kmem_free(raidPtr->parity_map,
sizeof(struct rf_paritymap));
raidPtr->parity_map = NULL;
return;
}
if (g_regions == 0)
/* Pick up the autoconfigured region count. */
g_regions = raidPtr->parity_map->params.regions;
if (pm_zap) {
good = raidPtr->parity_good && !force;
if (good)
rf_paritymap_forceclean(raidPtr->parity_map);
else
rf_paritymap_invalidate(raidPtr->parity_map);
/* This needs to be on disk before WASUSED is set. */
rf_paritymap_write(raidPtr->parity_map);
}
}
/* Alter labels in-core to reflect the current view of things. */
for (col = 0; col < raidPtr->numCol; col++) {
clabel = raidget_component_label(raidPtr, col);
if (pm_use)
flags = RF_PMLABEL_VALID | RF_PMLABEL_WASUSED;
else
flags = RF_PMLABEL_VALID | RF_PMLABEL_DISABLE;
clabel->parity_map_flags = flags;
clabel->parity_map_tickms = g_tickms;
clabel->parity_map_ntick = g_ntick;
clabel->parity_map_regions = g_regions;
raidflush_component_label(raidPtr, col);
}
}
/*
* For initializing the parity-map fields of a component label, both on
* initial creation and on reconstruct/copyback/etc.
*/
void
rf_paritymap_init_label(struct rf_paritymap *pm, RF_ComponentLabel_t *clabel)
{
if (pm != NULL) {
clabel->parity_map_flags =
RF_PMLABEL_VALID | RF_PMLABEL_WASUSED;
clabel->parity_map_tickms = pm->params.tickms;
clabel->parity_map_ntick = pm->params.cooldown;
/*
* XXXjld: If the number of regions is changed on disk, and
* then a new component is labeled before the next configure,
* then it will get the old value and they will conflict on
* the next boot (and the default will be used instead).
*/
clabel->parity_map_regions = pm->params.regions;
} else {
/*
* XXXjld: if the map is disabled, and all the components are
* replaced without an intervening unconfigure/reconfigure,
* then it will become enabled on the next unconfig/reconfig.
*/
}
}
/* Will the parity map be disabled next time? */
int
rf_paritymap_get_disable(RF_Raid_t *raidPtr)
{
RF_ComponentLabel_t *clabel;
RF_RowCol_t col;
int dis;
dis = 0;
for (col = 0; col < raidPtr->numCol; col++) {
clabel = raidget_component_label(raidPtr, col);
if (clabel->parity_map_flags & RF_PMLABEL_DISABLE)
dis = 1;
}
return dis;
}
/* Set whether the parity map will be disabled next time. */
void
rf_paritymap_set_disable(RF_Raid_t *raidPtr, int dis)
{
RF_ComponentLabel_t *clabel;
RF_RowCol_t col;
for (col = 0; col < raidPtr->numCol; col++) {
clabel = raidget_component_label(raidPtr, col);
if (dis)
clabel->parity_map_flags |= RF_PMLABEL_DISABLE;
else
clabel->parity_map_flags &= ~RF_PMLABEL_DISABLE;
raidflush_component_label(raidPtr, col);
}
}