mirror_zfs/module/zfs/spa_history.c
Richard Yao b8d06fca08 Switch KM_SLEEP to KM_PUSHPAGE
Differences between how paging is done on Solaris and Linux can cause
deadlocks if KM_SLEEP is used in any the following contexts.

  * The txg_sync thread
  * The zvol write/discard threads
  * The zpl_putpage() VFS callback

This is because KM_SLEEP will allow for direct reclaim which may result
in the VM calling back in to the filesystem or block layer to write out
pages.  If a lock is held over this operation the potential exists to
deadlock the system.  To ensure forward progress all memory allocations
in these contexts must us KM_PUSHPAGE which disables performing any I/O
to accomplish the memory allocation.

Previously, this behavior was acheived by setting PF_MEMALLOC on the
thread.  However, that resulted in unexpected side effects such as the
exhaustion of pages in ZONE_DMA.  This approach touchs more of the zfs
code, but it is more consistent with the right way to handle these cases
under Linux.

This is patch lays the ground work for being able to safely revert the
following commits which used PF_MEMALLOC:

  21ade34 Disable direct reclaim for z_wr_* threads
  cfc9a5c Fix zpl_writepage() deadlock
  eec8164 Fix ASSERTION(!dsl_pool_sync_context(tx->tx_pool))

Signed-off-by: Richard Yao <ryao@cs.stonybrook.edu>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #726
2012-08-27 12:01:37 -07:00

515 lines
14 KiB
C

/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License (the "License").
* You may not use this file except in compliance with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* or http://www.opensolaris.org/os/licensing.
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
/*
* Copyright (c) 2006, 2010, Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2011 by Delphix. All rights reserved.
*/
#include <sys/spa.h>
#include <sys/spa_impl.h>
#include <sys/zap.h>
#include <sys/dsl_synctask.h>
#include <sys/dmu_tx.h>
#include <sys/dmu_objset.h>
#include <sys/utsname.h>
#include <sys/cmn_err.h>
#include <sys/sunddi.h>
#include "zfs_comutil.h"
#ifdef _KERNEL
#include <sys/zone.h>
#endif
/*
* Routines to manage the on-disk history log.
*
* The history log is stored as a dmu object containing
* <packed record length, record nvlist> tuples.
*
* Where "record nvlist" is a nvlist containing uint64_ts and strings, and
* "packed record length" is the packed length of the "record nvlist" stored
* as a little endian uint64_t.
*
* The log is implemented as a ring buffer, though the original creation
* of the pool ('zpool create') is never overwritten.
*
* The history log is tracked as object 'spa_t::spa_history'. The bonus buffer
* of 'spa_history' stores the offsets for logging/retrieving history as
* 'spa_history_phys_t'. 'sh_pool_create_len' is the ending offset in bytes of
* where the 'zpool create' record is stored. This allows us to never
* overwrite the original creation of the pool. 'sh_phys_max_off' is the
* physical ending offset in bytes of the log. This tells you the length of
* the buffer. 'sh_eof' is the logical EOF (in bytes). Whenever a record
* is added, 'sh_eof' is incremented by the the size of the record.
* 'sh_eof' is never decremented. 'sh_bof' is the logical BOF (in bytes).
* This is where the consumer should start reading from after reading in
* the 'zpool create' portion of the log.
*
* 'sh_records_lost' keeps track of how many records have been overwritten
* and permanently lost.
*/
/* convert a logical offset to physical */
static uint64_t
spa_history_log_to_phys(uint64_t log_off, spa_history_phys_t *shpp)
{
uint64_t phys_len;
phys_len = shpp->sh_phys_max_off - shpp->sh_pool_create_len;
return ((log_off - shpp->sh_pool_create_len) % phys_len
+ shpp->sh_pool_create_len);
}
void
spa_history_create_obj(spa_t *spa, dmu_tx_t *tx)
{
dmu_buf_t *dbp;
spa_history_phys_t *shpp;
objset_t *mos = spa->spa_meta_objset;
ASSERT(spa->spa_history == 0);
spa->spa_history = dmu_object_alloc(mos, DMU_OT_SPA_HISTORY,
SPA_MAXBLOCKSIZE, DMU_OT_SPA_HISTORY_OFFSETS,
sizeof (spa_history_phys_t), tx);
VERIFY(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT,
DMU_POOL_HISTORY, sizeof (uint64_t), 1,
&spa->spa_history, tx) == 0);
VERIFY(0 == dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp));
ASSERT(dbp->db_size >= sizeof (spa_history_phys_t));
shpp = dbp->db_data;
dmu_buf_will_dirty(dbp, tx);
/*
* Figure out maximum size of history log. We set it at
* 0.1% of pool size, with a max of 1G and min of 128KB.
*/
shpp->sh_phys_max_off =
metaslab_class_get_dspace(spa_normal_class(spa)) / 1000;
shpp->sh_phys_max_off = MIN(shpp->sh_phys_max_off, 1<<30);
shpp->sh_phys_max_off = MAX(shpp->sh_phys_max_off, 128<<10);
dmu_buf_rele(dbp, FTAG);
}
/*
* Change 'sh_bof' to the beginning of the next record.
*/
static int
spa_history_advance_bof(spa_t *spa, spa_history_phys_t *shpp)
{
objset_t *mos = spa->spa_meta_objset;
uint64_t firstread, reclen, phys_bof;
char buf[sizeof (reclen)];
int err;
phys_bof = spa_history_log_to_phys(shpp->sh_bof, shpp);
firstread = MIN(sizeof (reclen), shpp->sh_phys_max_off - phys_bof);
if ((err = dmu_read(mos, spa->spa_history, phys_bof, firstread,
buf, DMU_READ_PREFETCH)) != 0)
return (err);
if (firstread != sizeof (reclen)) {
if ((err = dmu_read(mos, spa->spa_history,
shpp->sh_pool_create_len, sizeof (reclen) - firstread,
buf + firstread, DMU_READ_PREFETCH)) != 0)
return (err);
}
reclen = LE_64(*((uint64_t *)buf));
shpp->sh_bof += reclen + sizeof (reclen);
shpp->sh_records_lost++;
return (0);
}
static int
spa_history_write(spa_t *spa, void *buf, uint64_t len, spa_history_phys_t *shpp,
dmu_tx_t *tx)
{
uint64_t firstwrite, phys_eof;
objset_t *mos = spa->spa_meta_objset;
int err;
ASSERT(MUTEX_HELD(&spa->spa_history_lock));
/* see if we need to reset logical BOF */
while (shpp->sh_phys_max_off - shpp->sh_pool_create_len -
(shpp->sh_eof - shpp->sh_bof) <= len) {
if ((err = spa_history_advance_bof(spa, shpp)) != 0) {
return (err);
}
}
phys_eof = spa_history_log_to_phys(shpp->sh_eof, shpp);
firstwrite = MIN(len, shpp->sh_phys_max_off - phys_eof);
shpp->sh_eof += len;
dmu_write(mos, spa->spa_history, phys_eof, firstwrite, buf, tx);
len -= firstwrite;
if (len > 0) {
/* write out the rest at the beginning of physical file */
dmu_write(mos, spa->spa_history, shpp->sh_pool_create_len,
len, (char *)buf + firstwrite, tx);
}
return (0);
}
static char *
spa_history_zone(void)
{
#ifdef _KERNEL
#ifdef HAVE_SPL
return ("linux");
#else
return (curproc->p_zone->zone_name);
#endif
#else
return ("global");
#endif
}
/*
* Write out a history event.
*/
/*ARGSUSED*/
static void
spa_history_log_sync(void *arg1, void *arg2, dmu_tx_t *tx)
{
spa_t *spa = arg1;
history_arg_t *hap = arg2;
const char *history_str = hap->ha_history_str;
objset_t *mos = spa->spa_meta_objset;
dmu_buf_t *dbp;
spa_history_phys_t *shpp;
size_t reclen;
uint64_t le_len;
nvlist_t *nvrecord;
char *record_packed = NULL;
int ret;
/*
* If we have an older pool that doesn't have a command
* history object, create it now.
*/
mutex_enter(&spa->spa_history_lock);
if (!spa->spa_history)
spa_history_create_obj(spa, tx);
mutex_exit(&spa->spa_history_lock);
/*
* Get the offset of where we need to write via the bonus buffer.
* Update the offset when the write completes.
*/
VERIFY(0 == dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp));
shpp = dbp->db_data;
dmu_buf_will_dirty(dbp, tx);
#ifdef ZFS_DEBUG
{
dmu_object_info_t doi;
dmu_object_info_from_db(dbp, &doi);
ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_SPA_HISTORY_OFFSETS);
}
#endif
VERIFY(nvlist_alloc(&nvrecord, NV_UNIQUE_NAME, KM_PUSHPAGE) == 0);
VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_TIME,
gethrestime_sec()) == 0);
VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_WHO, hap->ha_uid) == 0);
if (hap->ha_zone != NULL)
VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_ZONE,
hap->ha_zone) == 0);
#ifdef _KERNEL
VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_HOST,
utsname.nodename) == 0);
#endif
if (hap->ha_log_type == LOG_CMD_POOL_CREATE ||
hap->ha_log_type == LOG_CMD_NORMAL) {
VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_CMD,
history_str) == 0);
zfs_dbgmsg("command: %s", history_str);
} else {
VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_INT_EVENT,
hap->ha_event) == 0);
VERIFY(nvlist_add_uint64(nvrecord, ZPOOL_HIST_TXG,
tx->tx_txg) == 0);
VERIFY(nvlist_add_string(nvrecord, ZPOOL_HIST_INT_STR,
history_str) == 0);
zfs_dbgmsg("internal %s pool:%s txg:%llu %s",
zfs_history_event_names[hap->ha_event], spa_name(spa),
(longlong_t)tx->tx_txg, history_str);
}
VERIFY(nvlist_size(nvrecord, &reclen, NV_ENCODE_XDR) == 0);
record_packed = kmem_alloc(reclen, KM_PUSHPAGE);
VERIFY(nvlist_pack(nvrecord, &record_packed, &reclen,
NV_ENCODE_XDR, KM_PUSHPAGE) == 0);
mutex_enter(&spa->spa_history_lock);
if (hap->ha_log_type == LOG_CMD_POOL_CREATE)
VERIFY(shpp->sh_eof == shpp->sh_pool_create_len);
/* write out the packed length as little endian */
le_len = LE_64((uint64_t)reclen);
ret = spa_history_write(spa, &le_len, sizeof (le_len), shpp, tx);
if (!ret)
ret = spa_history_write(spa, record_packed, reclen, shpp, tx);
if (!ret && hap->ha_log_type == LOG_CMD_POOL_CREATE) {
shpp->sh_pool_create_len += sizeof (le_len) + reclen;
shpp->sh_bof = shpp->sh_pool_create_len;
}
mutex_exit(&spa->spa_history_lock);
nvlist_free(nvrecord);
kmem_free(record_packed, reclen);
dmu_buf_rele(dbp, FTAG);
strfree(hap->ha_history_str);
if (hap->ha_zone != NULL)
strfree(hap->ha_zone);
kmem_free(hap, sizeof (history_arg_t));
}
/*
* Write out a history event.
*/
int
spa_history_log(spa_t *spa, const char *history_str, history_log_type_t what)
{
history_arg_t *ha;
int err = 0;
dmu_tx_t *tx;
ASSERT(what != LOG_INTERNAL);
tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
err = dmu_tx_assign(tx, TXG_WAIT);
if (err) {
dmu_tx_abort(tx);
return (err);
}
ha = kmem_alloc(sizeof (history_arg_t), KM_PUSHPAGE);
ha->ha_history_str = strdup(history_str);
ha->ha_zone = strdup(spa_history_zone());
ha->ha_log_type = what;
ha->ha_uid = crgetuid(CRED());
/* Kick this off asynchronously; errors are ignored. */
dsl_sync_task_do_nowait(spa_get_dsl(spa), NULL,
spa_history_log_sync, spa, ha, 0, tx);
dmu_tx_commit(tx);
/* spa_history_log_sync will free ha and strings */
return (err);
}
/*
* Read out the command history.
*/
int
spa_history_get(spa_t *spa, uint64_t *offp, uint64_t *len, char *buf)
{
objset_t *mos = spa->spa_meta_objset;
dmu_buf_t *dbp;
uint64_t read_len, phys_read_off, phys_eof;
uint64_t leftover = 0;
spa_history_phys_t *shpp;
int err;
/*
* If the command history doesn't exist (older pool),
* that's ok, just return ENOENT.
*/
if (!spa->spa_history)
return (ENOENT);
/*
* The history is logged asynchronously, so when they request
* the first chunk of history, make sure everything has been
* synced to disk so that we get it.
*/
if (*offp == 0 && spa_writeable(spa))
txg_wait_synced(spa_get_dsl(spa), 0);
if ((err = dmu_bonus_hold(mos, spa->spa_history, FTAG, &dbp)) != 0)
return (err);
shpp = dbp->db_data;
#ifdef ZFS_DEBUG
{
dmu_object_info_t doi;
dmu_object_info_from_db(dbp, &doi);
ASSERT3U(doi.doi_bonus_type, ==, DMU_OT_SPA_HISTORY_OFFSETS);
}
#endif
mutex_enter(&spa->spa_history_lock);
phys_eof = spa_history_log_to_phys(shpp->sh_eof, shpp);
if (*offp < shpp->sh_pool_create_len) {
/* read in just the zpool create history */
phys_read_off = *offp;
read_len = MIN(*len, shpp->sh_pool_create_len -
phys_read_off);
} else {
/*
* Need to reset passed in offset to BOF if the passed in
* offset has since been overwritten.
*/
*offp = MAX(*offp, shpp->sh_bof);
phys_read_off = spa_history_log_to_phys(*offp, shpp);
/*
* Read up to the minimum of what the user passed down or
* the EOF (physical or logical). If we hit physical EOF,
* use 'leftover' to read from the physical BOF.
*/
if (phys_read_off <= phys_eof) {
read_len = MIN(*len, phys_eof - phys_read_off);
} else {
read_len = MIN(*len,
shpp->sh_phys_max_off - phys_read_off);
if (phys_read_off + *len > shpp->sh_phys_max_off) {
leftover = MIN(*len - read_len,
phys_eof - shpp->sh_pool_create_len);
}
}
}
/* offset for consumer to use next */
*offp += read_len + leftover;
/* tell the consumer how much you actually read */
*len = read_len + leftover;
if (read_len == 0) {
mutex_exit(&spa->spa_history_lock);
dmu_buf_rele(dbp, FTAG);
return (0);
}
err = dmu_read(mos, spa->spa_history, phys_read_off, read_len, buf,
DMU_READ_PREFETCH);
if (leftover && err == 0) {
err = dmu_read(mos, spa->spa_history, shpp->sh_pool_create_len,
leftover, buf + read_len, DMU_READ_PREFETCH);
}
mutex_exit(&spa->spa_history_lock);
dmu_buf_rele(dbp, FTAG);
return (err);
}
static void
log_internal(history_internal_events_t event, spa_t *spa,
dmu_tx_t *tx, const char *fmt, va_list adx)
{
history_arg_t *ha;
va_list adx_copy;
/*
* If this is part of creating a pool, not everything is
* initialized yet, so don't bother logging the internal events.
*/
if (tx->tx_txg == TXG_INITIAL)
return;
ha = kmem_alloc(sizeof (history_arg_t), KM_PUSHPAGE);
va_copy(adx_copy, adx);
ha->ha_history_str = kmem_vasprintf(fmt, adx_copy);
va_end(adx_copy);
ha->ha_log_type = LOG_INTERNAL;
ha->ha_event = event;
ha->ha_zone = NULL;
ha->ha_uid = 0;
if (dmu_tx_is_syncing(tx)) {
spa_history_log_sync(spa, ha, tx);
} else {
dsl_sync_task_do_nowait(spa_get_dsl(spa), NULL,
spa_history_log_sync, spa, ha, 0, tx);
}
/* spa_history_log_sync() will free ha and strings */
}
void
spa_history_log_internal(history_internal_events_t event, spa_t *spa,
dmu_tx_t *tx, const char *fmt, ...)
{
dmu_tx_t *htx = tx;
va_list adx;
/* create a tx if we didn't get one */
if (tx == NULL) {
htx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
if (dmu_tx_assign(htx, TXG_WAIT) != 0) {
dmu_tx_abort(htx);
return;
}
}
va_start(adx, fmt);
log_internal(event, spa, htx, fmt, adx);
va_end(adx);
/* if we didn't get a tx from the caller, commit the one we made */
if (tx == NULL)
dmu_tx_commit(htx);
}
void
spa_history_log_version(spa_t *spa, history_internal_events_t event)
{
#ifdef _KERNEL
uint64_t current_vers = spa_version(spa);
if (current_vers >= SPA_VERSION_ZPOOL_HISTORY) {
spa_history_log_internal(event, spa, NULL,
"pool spa %llu; zfs spa %llu; zpl %d; uts %s %s %s %s",
(u_longlong_t)current_vers, SPA_VERSION, ZPL_VERSION,
utsname.nodename, utsname.release, utsname.version,
utsname.machine);
}
cmn_err(CE_CONT, "!%s version %llu pool %s using %llu",
event == LOG_POOL_IMPORT ? "imported" :
event == LOG_POOL_CREATE ? "created" : "accessed",
(u_longlong_t)current_vers, spa_name(spa), SPA_VERSION);
#endif
}
#if defined(_KERNEL) && defined(HAVE_SPL)
EXPORT_SYMBOL(spa_history_create_obj);
EXPORT_SYMBOL(spa_history_get);
EXPORT_SYMBOL(spa_history_log);
EXPORT_SYMBOL(spa_history_log_internal);
EXPORT_SYMBOL(spa_history_log_version);
#endif