zfs_fm.c revision bf82a41b568b2bd31bf9814587eb25ee2e7b05ff
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * CDDL HEADER START
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * The contents of this file are subject to the terms of the
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Common Development and Distribution License (the "License").
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * You may not use this file except in compliance with the License.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * See the License for the specific language governing permissions
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * and limitations under the License.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * When distributing Covered Code, include this CDDL HEADER in each
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * If applicable, add the following below this CDDL HEADER, with the
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * fields enclosed by brackets "[]" replaced with your own identifying
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * information: Portions Copyright [yyyy] [name of copyright owner]
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * CDDL HEADER END
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Use is subject to license terms.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz#pragma ident "%Z%%M% %I% %E% SMI"
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * This general routine is responsible for generating all the different ZFS
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * ereports. The payload is dependent on the class, and which arguments are
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * supplied to the function:
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * EREPORT POOL VDEV IO
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * block X X X
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * device X X
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * If we are in a loading state, all errors are chained together by the same
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * SPA-wide ENA (Error Numeric Association).
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * For isolated I/O requests, we get the ENA from the zio_t. The propagation
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * gets very complicated due to RAID-Z, gang blocks, and vdev caching. We want
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * to chain together all ereports associated with a logical piece of data. For
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * read I/Os, there are basically three 'types' of I/O, which form a roughly
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * layered diagram:
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * +---------------+
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * | Aggregate I/O | No associated logical data or device
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * +---------------+
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * +---------------+ Reads associated with a piece of logical data.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * | Read I/O | This includes reads on behalf of RAID-Z,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * +---------------+ mirrors, gang blocks, retries, etc.
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * +---------------+ Reads associated with a particular device, but
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * | Physical I/O | no logical data. Issued as part of vdev caching
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * +---------------+ and I/O aggregation.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Note that 'physical I/O' here is not the same terminology as used in the rest
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * of ZIO. Typically, 'physical I/O' simply means that there is no attached
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * blockpointer. But I/O with no associated block pointer can still be related
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * to a logical piece of data (i.e. RAID-Z requests).
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * Purely physical I/O always have unique ENAs. They are not related to a
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * particular piece of logical data, and therefore cannot be chained together.
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * We still generate an ereport, but the DE doesn't correlate it with any
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * logical piece of data. When such an I/O fails, the delegated I/O requests
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * will issue a retry, which will trigger the 'real' ereport with the correct
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * We keep track of the ENA for a ZIO chain through the 'io_logical' member.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * When a new logical I/O is issued, we set this to point to itself. Child I/Os
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * then inherit this pointer, so that when it is first set subsequent failures
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * will use the same ENA. If a physical I/O is issued (by passing the
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * ZIO_FLAG_NOBOOKMARK flag), then this pointer is reset, guaranteeing that a
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * unique ENA will be generated. For an aggregate I/O, this pointer is set to
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * NULL, and no ereport will be generated (since it doesn't actually correspond
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner * to any particular device or piece of data).
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulknerzfs_ereport_post(const char *subclass, spa_t *spa, vdev_t *vd, zio_t *zio,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * If we are doing a spa_tryimport(), ignore errors.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * If we are in the middle of opening a pool, and the previous attempt
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * failed, don't bother logging any new ereports - we're just going to
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * get the same diagnosis anyway.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Ignore any errors from I/Os that we are going to retry
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * anyway - we only generate errors from the final failure.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Checksum errors are generated after the pipeline stage
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * responsible for retrying the I/O (VDEV_IO_ASSESS), so this
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * only applies to standard I/O errors.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz if (zio_should_retry(zio) && zio->io_error != ECKSUM)
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * If this is not a read or write zio, ignore the error. This
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * can occur if the DKIOCFLUSHWRITECACHE ioctl fails.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Ignore any errors from speculative I/Os, as failure is an
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * expected result.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz if ((ereport = fm_nvlist_create(NULL)) == NULL)
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz if ((detector = fm_nvlist_create(NULL)) == NULL) {
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Serialize ereport generation
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Determine the ENA to use for this event. If we are in a loading
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * state, use a SPA-wide ENA. Otherwise, if we are in an I/O state, use
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * a root zio-wide ENA. Otherwise, simply use a unique ENA.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz spa->spa_ena = fm_ena_generate(0, FM_ENA_FMT1);
3e14f97f673e8a630f076077de35afdd43dc1587Roger A. Faulkner } else if (zio != NULL && zio->io_logical != NULL) {
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Construct the full class, detector, and other standard FMA fields.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz (void) snprintf(class, sizeof (class), "%s.%s",
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz fm_fmri_zfs_set(detector, FM_ZFS_SCHEME_VERSION, spa_guid(spa),
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz fm_ereport_set(ereport, FM_EREPORT_VERSION, class, ena, detector, NULL);
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Construct the per-ereport payload, depending on which parameters are
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * passed in.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * If we are importing a faulted pool, then we treat it like an open,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * not an import. Otherwise, the DE will ignore all faults during
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * import, since the default behavior is to mark the devices as
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * persistently unavailable, not leave them in the faulted state.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz state = spa->spa_import_faulted ? SPA_LOAD_OPEN : spa->spa_load_state;
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * Generic payload members common to all ereports.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * The direct reference to spa_name is used rather than spa_name()
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * because of the asynchronous nature of the zio pipeline. spa_name()
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * asserts that the config lock is held in some form. This is always
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * the case in I/O context, but because the check for RW_WRITER compares
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * against 'curthread', we may be in an asynchronous context and blow
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * this assert. Rather than loosen this assert, we acknowledge that all
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * contexts in which this function is called (pool open, I/O) are safe,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz * and dereference the name directly.
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz DATA_TYPE_STRING, spa->spa_name, FM_EREPORT_PAYLOAD_ZFS_POOL_GUID,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz FM_EREPORT_PAYLOAD_ZFS_POOL_CONTEXT, DATA_TYPE_INT32,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_POOL_FAILMODE,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz spa_get_failmode(spa) == ZIO_FAILURE_MODE_WAIT ?
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE ?
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz FM_EREPORT_FAILMODE_CONTINUE : FM_EREPORT_FAILMODE_PANIC,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz fm_payload_set(ereport, FM_EREPORT_PAYLOAD_ZFS_VDEV_GUID,
34f9b3eef6fdadbda0a846aa4d68691ac40eace5Roland Mainz DATA_TYPE_STRING, vd->vdev_ops->vdev_op_type, NULL);
NULL);
if (size)
#ifdef _KERNEL
if (vd)