-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* Author: Phil Schwan <phil@clusterfs.com>
*/
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
#define DEBUG_SUBSYSTEM S_OST
#include <linux/module.h>
CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444,
"number of OSS create threads to start");
+static char *oss_cpts;
+CFS_MODULE_PARM(oss_cpts, "s", charp, 0444,
+ "CPU partitions OSS threads should run on");
+
+static char *oss_io_cpts;
+CFS_MODULE_PARM(oss_io_cpts, "s", charp, 0444,
+ "CPU partitions OSS IO threads should run on");
+
+/*
+ * this page is allocated statically when module is initializing
+ * it is used to simulate data corruptions, see ost_checksum_bulk()
+ * for details. as the original pages provided by the layers below
+ * can be remain in the internal cache, we do not want to modify
+ * them.
+ */
+static struct page *ost_page_to_corrupt = NULL;
+
/**
* Do not return server-side uid/gid to remote client
*/
-static void ost_drop_id(struct obd_export *exp, struct obdo *oa)
+static void ost_drop_id(struct obd_export *exp, struct obdo *oa)
{
if (exp_connect_rmtclient(exp)) {
oa->o_uid = -1;
}
}
+/**
+ * Validate oa from client.
+ * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
+ * req are valid.
+ * a. for single MDS seq = FID_SEQ_OST_MDT0,
+ * b. for CMD, seq = FID_SEQ_OST_MDT0, FID_SEQ_OST_MDT1 - FID_SEQ_OST_MAX
+ */
+static int ost_validate_obdo(struct obd_export *exp, struct obdo *oa,
+ struct obd_ioobj *ioobj)
+{
+ if (oa != NULL && !(oa->o_valid & OBD_MD_FLGROUP)) {
+ oa->o_seq = FID_SEQ_OST_MDT0;
+ if (ioobj)
+ ioobj->ioo_seq = FID_SEQ_OST_MDT0;
+ /* remove fid_seq_is_rsvd() after FID-on-OST allows SEQ > 9 */
+ } else if (oa == NULL || !(fid_seq_is_rsvd(oa->o_seq) ||
+ fid_seq_is_mdt0(oa->o_seq))) {
+ CERROR("%s: client %s sent invalid object "POSTID"\n",
+ exp->exp_obd->obd_name, obd_export_nid2str(exp),
+ oa ? oa->o_id : -1, oa ? oa->o_seq : -1);
+ return -EPROTO;
+ }
+ obdo_from_ostid(oa, &oa->o_oi);
+ if (ioobj)
+ ioobj_from_obdo(ioobj, oa);
+ return 0;
+}
+
void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
{
struct oti_req_ack_lock *ack_lock;
if (oti == NULL)
return;
- if (req->rq_repmsg)
+ if (req->rq_repmsg) {
+ __u64 versions[PTLRPC_NUM_VERSIONS] = { 0 };
lustre_msg_set_transno(req->rq_repmsg, oti->oti_transno);
+ versions[0] = oti->oti_pre_version;
+ lustre_msg_set_versions(req->rq_repmsg, versions);
+ }
req->rq_transno = oti->oti_transno;
/* XXX 4 == entries in oti_ack_locks??? */
if (body->oa.o_id == 0)
RETURN(-EPROTO);
+ rc = ost_validate_obdo(exp, &body->oa, NULL);
+ if (rc)
+ RETURN(rc);
+
/* If there's a DLM request, cancel the locks mentioned in it*/
if (req_capsule_field_present(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT)) {
struct ldlm_request *dlm;
memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
/* Do the destroy and set the reply status accordingly */
- req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL, capa);
+ req->rq_status = obd_destroy(req->rq_svc_thread->t_env, exp,
+ &repbody->oa, NULL, oti, NULL, capa);
RETURN(0);
}
__u64 start, __u64 count, struct lustre_handle *lh,
int mode, int flags)
{
- struct ldlm_res_id res_id = { .name = { oa->o_id, 0, oa->o_gr, 0} };
+ struct ldlm_res_id res_id;
ldlm_policy_data_t policy;
__u64 end = start + count;
ENTRY;
LASSERT(!lustre_handle_is_used(lh));
- LASSERT((oa->o_valid & (OBD_MD_FLID | OBD_MD_FLGROUP)) ==
- (OBD_MD_FLID | OBD_MD_FLGROUP));
+ /* o_id and o_gr are used for localizing resource, if client miss to set
+ * them, do not trigger ASSERTION. */
+ if (unlikely((oa->o_valid & (OBD_MD_FLID | OBD_MD_FLGROUP)) !=
+ (OBD_MD_FLID | OBD_MD_FLGROUP)))
+ RETURN(-EPROTO);
if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
!(oa->o_flags & OBD_FL_SRVLOCK))
RETURN(0);
+ osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
CDEBUG(D_INODE, "OST-side extent lock.\n");
policy.l_extent.start = start & CFS_PAGE_MASK;
static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
{
struct ost_body *body, *repbody;
- struct obd_info oinfo = { { { 0 } } };
+ struct obd_info *oinfo;
struct lustre_handle lh = { 0 };
+ struct lustre_capa *capa = NULL;
int rc;
ENTRY;
if (body == NULL)
RETURN(-EFAULT);
+ rc = ost_validate_obdo(exp, &body->oa, NULL);
+ if (rc)
+ RETURN(rc);
+
+ if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
+ capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
+ if (capa == NULL) {
+ CERROR("Missing capability for OST GETATTR");
+ RETURN(-EFAULT);
+ }
+ }
+
rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
RETURN(rc);
repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
repbody->oa = body->oa;
- rc = ost_lock_get(exp, &body->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0);
+ rc = ost_lock_get(exp, &repbody->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0);
if (rc)
RETURN(rc);
- oinfo.oi_oa = &repbody->oa;
- if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) {
- oinfo.oi_capa = req_capsule_client_get(&req->rq_pill,
- &RMF_CAPA1);
- if (oinfo.oi_capa == NULL) {
- CERROR("Missing capability for OST GETATTR");
- RETURN (-EFAULT);
- }
- }
+ OBD_ALLOC_PTR(oinfo);
+ if (!oinfo)
+ GOTO(unlock, rc = -ENOMEM);
+ oinfo->oi_oa = &repbody->oa;
+ oinfo->oi_capa = capa;
- req->rq_status = obd_getattr(exp, &oinfo);
- ost_lock_put(exp, &lh, LCK_PR);
+ req->rq_status = obd_getattr(req->rq_svc_thread->t_env, exp, oinfo);
+
+ OBD_FREE_PTR(oinfo);
ost_drop_id(exp, &repbody->oa);
- RETURN(0);
+
+unlock:
+ ost_lock_put(exp, &lh, LCK_PR);
+ RETURN(rc);
}
static int ost_statfs(struct ptlrpc_request *req)
osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
- req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs,
- cfs_time_current_64() - CFS_HZ, 0);
+ req->rq_status = obd_statfs(req->rq_svc_thread->t_env, req->rq_export,
+ osfs,
+ cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
+ 0);
if (req->rq_status != 0)
CERROR("ost: statfs failed: rc %d\n", req->rq_status);
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_EINPROGRESS))
+ req->rq_status = -EINPROGRESS;
+
RETURN(0);
}
if (body == NULL)
RETURN(-EFAULT);
+ rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
+ if (rc)
+ RETURN(rc);
+
rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
RETURN(rc);
repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
- memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
- oti->oti_logcookies = &repbody->oa.o_lcookie;
+ repbody->oa = body->oa;
+ oti->oti_logcookies = &body->oa.o_lcookie;
- req->rq_status = obd_create(exp, &repbody->oa, NULL, oti);
+ req->rq_status = obd_create(req->rq_svc_thread->t_env, exp,
+ &repbody->oa, NULL, oti);
//obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
RETURN(0);
}
static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
struct obd_trans_info *oti)
{
- struct obd_info oinfo = { { { 0 } } };
struct ost_body *body, *repbody;
int rc, flags = 0;
struct lustre_handle lh = {0,};
if (body == NULL)
RETURN(-EFAULT);
- oinfo.oi_oa = &body->oa;
- oinfo.oi_policy.l_extent.start = oinfo.oi_oa->o_size;
- oinfo.oi_policy.l_extent.end = oinfo.oi_oa->o_blocks;
+ rc = ost_validate_obdo(exp, &body->oa, NULL);
+ if (rc)
+ RETURN(rc);
- if ((oinfo.oi_oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
+ if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
RETURN(-EPROTO);
/* standard truncate optimization: if file body is completely
* destroyed, don't send data back to the server. */
- if (oinfo.oi_oa->o_size == 0)
+ if (body->oa.o_size == 0)
flags |= LDLM_AST_DISCARD_DATA;
repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
- rc = ost_lock_get(exp, oinfo.oi_oa, oinfo.oi_oa->o_size,
- oinfo.oi_oa->o_blocks, &lh, LCK_PW, flags);
+ repbody->oa = body->oa;
+
+ rc = ost_lock_get(exp, &repbody->oa, repbody->oa.o_size,
+ repbody->oa.o_blocks, &lh, LCK_PW, flags);
if (rc == 0) {
- if (oinfo.oi_oa->o_valid & OBD_MD_FLFLAGS &&
- oinfo.oi_oa->o_flags == OBD_FL_SRVLOCK)
+ struct obd_info *oinfo;
+ struct lustre_capa *capa = NULL;
+
+ if (repbody->oa.o_valid & OBD_MD_FLFLAGS &&
+ repbody->oa.o_flags == OBD_FL_SRVLOCK)
/*
* If OBD_FL_SRVLOCK is the only bit set in
* ->o_flags, clear OBD_MD_FLFLAGS to avoid falling
* through filter_setattr() to filter_iocontrol().
*/
- oinfo.oi_oa->o_valid &= ~OBD_MD_FLFLAGS;
+ repbody->oa.o_valid &= ~OBD_MD_FLFLAGS;
- if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) {
- oinfo.oi_capa = req_capsule_client_get(&req->rq_pill,
- &RMF_CAPA1);
- if (oinfo.oi_capa == NULL) {
+ if (repbody->oa.o_valid & OBD_MD_FLOSSCAPA) {
+ capa = req_capsule_client_get(&req->rq_pill,
+ &RMF_CAPA1);
+ if (capa == NULL) {
CERROR("Missing capability for OST PUNCH");
- RETURN (-EFAULT);
+ GOTO(unlock, rc = -EFAULT);
}
}
- req->rq_status = obd_punch(exp, &oinfo, oti, NULL);
+
+ OBD_ALLOC_PTR(oinfo);
+ if (!oinfo)
+ GOTO(unlock, rc = -ENOMEM);
+ oinfo->oi_oa = &repbody->oa;
+ oinfo->oi_policy.l_extent.start = oinfo->oi_oa->o_size;
+ oinfo->oi_policy.l_extent.end = oinfo->oi_oa->o_blocks;
+ oinfo->oi_capa = capa;
+ oinfo->oi_flags = OBD_FL_PUNCH;
+
+ req->rq_status = obd_punch(req->rq_svc_thread->t_env, exp,
+ oinfo, oti, NULL);
+ OBD_FREE_PTR(oinfo);
+unlock:
ost_lock_put(exp, &lh, LCK_PW);
}
- repbody->oa = *oinfo.oi_oa;
+
ost_drop_id(exp, &repbody->oa);
RETURN(rc);
}
-static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
+static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req,
+ struct obd_trans_info *oti)
{
struct ost_body *body, *repbody;
+ struct obd_info *oinfo;
struct lustre_capa *capa = NULL;
int rc;
ENTRY;
if (body == NULL)
RETURN(-EFAULT);
+ rc = ost_validate_obdo(exp, &body->oa, NULL);
+ if (rc)
+ RETURN(rc);
+
if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
if (capa == NULL) {
RETURN(rc);
repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
- memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
- req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
- repbody->oa.o_blocks, capa);
+ repbody->oa = body->oa;
+
+ OBD_ALLOC_PTR(oinfo);
+ if (!oinfo)
+ RETURN(-ENOMEM);
+
+ oinfo->oi_oa = &repbody->oa;
+ oinfo->oi_capa = capa;
+ oinfo->oi_jobid = oti->oti_jobid;
+ req->rq_status = obd_sync(req->rq_svc_thread->t_env, exp, oinfo,
+ repbody->oa.o_size, repbody->oa.o_blocks,
+ NULL);
+ OBD_FREE_PTR(oinfo);
+
ost_drop_id(exp, &repbody->oa);
RETURN(0);
}
struct obd_trans_info *oti)
{
struct ost_body *body, *repbody;
+ struct obd_info *oinfo;
+ struct lustre_capa *capa = NULL;
int rc;
- struct obd_info oinfo = { { { 0 } } };
ENTRY;
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
RETURN(-EFAULT);
- rc = req_capsule_server_pack(&req->rq_pill);
+ rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
if (rc)
RETURN(rc);
- repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
- repbody->oa = body->oa;
+ rc = req_capsule_server_pack(&req->rq_pill);
+ if (rc)
+ RETURN(rc);
- oinfo.oi_oa = &repbody->oa;
- if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) {
- oinfo.oi_capa = req_capsule_client_get(&req->rq_pill,
- &RMF_CAPA1);
- if (oinfo.oi_capa == NULL) {
+ if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
+ capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
+ if (capa == NULL) {
CERROR("Missing capability for OST SETATTR");
RETURN (-EFAULT);
}
}
- req->rq_status = obd_setattr(exp, &oinfo, oti);
+
+ repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
+ repbody->oa = body->oa;
+
+ OBD_ALLOC_PTR(oinfo);
+ if (!oinfo)
+ RETURN(-ENOMEM);
+ oinfo->oi_oa = &repbody->oa;
+ oinfo->oi_capa = capa;
+
+ req->rq_status = obd_setattr(req->rq_svc_thread->t_env, exp, oinfo,
+ oti);
+
+ OBD_FREE_PTR(oinfo);
+
ost_drop_id(exp, &repbody->oa);
RETURN(0);
}
-static int ost_bulk_timeout(void *data)
-{
- ENTRY;
- /* We don't fail the connection here, because having the export
- * killed makes the (vital) call to commitrw very sad.
- */
- RETURN(1);
-}
-
static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc,
- cksum_type_t cksum_type)
+ cksum_type_t cksum_type)
{
- __u32 cksum;
- int i;
-
- cksum = init_checksum(cksum_type);
- for (i = 0; i < desc->bd_iov_count; i++) {
- struct page *page = desc->bd_iov[i].kiov_page;
- int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
- char *ptr = kmap(page) + off;
- int len = desc->bd_iov[i].kiov_len;
-
- /* corrupt the data before we compute the checksum, to
- * simulate a client->OST data error */
- if (i == 0 && opc == OST_WRITE &&
- OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE))
- memcpy(ptr, "bad3", min(4, len));
- cksum = compute_checksum(cksum, ptr, len, cksum_type);
- /* corrupt the data after we compute the checksum, to
- * simulate an OST->client data error */
- if (i == 0 && opc == OST_READ &&
- OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
- memcpy(ptr, "bad4", min(4, len));
- /* nobody should use corrupted page again */
- ClearPageUptodate(page);
- }
- kunmap(page);
- }
-
- return cksum;
+ struct cfs_crypto_hash_desc *hdesc;
+ unsigned int bufsize;
+ int i, err;
+ unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
+ __u32 cksum;
+
+ hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
+ if (IS_ERR(hdesc)) {
+ CERROR("Unable to initialize checksum hash %s\n",
+ cfs_crypto_hash_name(cfs_alg));
+ return PTR_ERR(hdesc);
+ }
+ CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
+ for (i = 0; i < desc->bd_iov_count; i++) {
+
+ /* corrupt the data before we compute the checksum, to
+ * simulate a client->OST data error */
+ if (i == 0 && opc == OST_WRITE &&
+ OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
+ int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
+ int len = desc->bd_iov[i].kiov_len;
+ struct page *np = ost_page_to_corrupt;
+ char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
+
+ if (np) {
+ char *ptr2 = kmap(np) + off;
+
+ memcpy(ptr2, ptr, len);
+ memcpy(ptr2, "bad3", min(4, len));
+ kunmap(np);
+ desc->bd_iov[i].kiov_page = np;
+ } else {
+ CERROR("can't alloc page for corruption\n");
+ }
+ }
+ cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
+ desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
+ desc->bd_iov[i].kiov_len);
+
+ /* corrupt the data after we compute the checksum, to
+ * simulate an OST->client data error */
+ if (i == 0 && opc == OST_READ &&
+ OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
+ int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
+ int len = desc->bd_iov[i].kiov_len;
+ struct page *np = ost_page_to_corrupt;
+ char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
+
+ if (np) {
+ char *ptr2 = kmap(np) + off;
+
+ memcpy(ptr2, ptr, len);
+ memcpy(ptr2, "bad4", min(4, len));
+ kunmap(np);
+ desc->bd_iov[i].kiov_page = np;
+ } else {
+ CERROR("can't alloc page for corruption\n");
+ }
+ }
+ }
+
+ bufsize = 4;
+ err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
+ if (err)
+ cfs_crypto_hash_final(hdesc, NULL, NULL);
+
+ return cksum;
}
static int ost_brw_lock_get(int mode, struct obd_export *exp,
int i;
ENTRY;
- osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
+ osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
LASSERT(mode == LCK_PR || mode == LCK_PW);
LASSERT(!lustre_handle_is_used(lh));
EXIT;
}
-struct ost_prolong_data {
- struct obd_export *opd_exp;
- ldlm_policy_data_t opd_policy;
- struct obdo *opd_oa;
- ldlm_mode_t opd_mode;
- int opd_lock_match;
- int opd_timeout;
-};
-
-static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data)
+/* Allocate thread local buffers if needed */
+static struct ost_thread_local_cache *ost_tls_get(struct ptlrpc_request *r)
{
- struct ost_prolong_data *opd = data;
-
- LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
-
- if (lock->l_req_mode != lock->l_granted_mode) {
- /* scan granted locks only */
- return LDLM_ITER_STOP;
- }
-
- if (lock->l_export != opd->opd_exp) {
- /* prolong locks only for given client */
- return LDLM_ITER_CONTINUE;
- }
-
- if (!(lock->l_granted_mode & opd->opd_mode)) {
- /* we aren't interesting in all type of locks */
- return LDLM_ITER_CONTINUE;
- }
-
- if (lock->l_policy_data.l_extent.end < opd->opd_policy.l_extent.start ||
- lock->l_policy_data.l_extent.start > opd->opd_policy.l_extent.end) {
- /* the request doesn't cross the lock, skip it */
- return LDLM_ITER_CONTINUE;
- }
-
- /* Fill the obdo with the matched lock handle.
- * XXX: it is possible in some cases the IO RPC is covered by several
- * locks, even for the write case, so it may need to be a lock list. */
- if (opd->opd_oa && !(opd->opd_oa->o_valid & OBD_MD_FLHANDLE)) {
- opd->opd_oa->o_handle.cookie = lock->l_handle.h_cookie;
- opd->opd_oa->o_valid |= OBD_MD_FLHANDLE;
- }
+ struct ost_thread_local_cache *tls =
+ (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
- if (!(lock->l_flags & LDLM_FL_AST_SENT)) {
- /* ignore locks not being cancelled */
- return LDLM_ITER_CONTINUE;
+ /* In normal mode of operation an I/O request is serviced only
+ * by ll_ost_io threads each of them has own tls buffers allocated by
+ * ost_thread_init().
+ * During recovery, an I/O request may be queued until any of the ost
+ * service threads process it. Not necessary it should be one of
+ * ll_ost_io threads. In that case we dynamically allocating tls
+ * buffers for the request service time. */
+ if (unlikely(tls == NULL)) {
+ LASSERT(r->rq_export->exp_in_recovery);
+ OBD_ALLOC_PTR(tls);
+ if (tls != NULL) {
+ tls->temporary = 1;
+ r->rq_svc_thread->t_data = tls;
+ }
}
-
- CDEBUG(D_DLMTRACE,"refresh lock: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
- lock->l_resource->lr_name.name[0],
- lock->l_resource->lr_name.name[1],
- opd->opd_policy.l_extent.start, opd->opd_policy.l_extent.end);
- /* OK. this is a possible lock the user holds doing I/O
- * let's refresh eviction timer for it */
- ldlm_refresh_waiting_lock(lock, opd->opd_timeout);
- opd->opd_lock_match = 1;
-
- return LDLM_ITER_CONTINUE;
+ return tls;
}
-static int ost_rw_prolong_locks(struct ptlrpc_request *req, struct obd_ioobj *obj,
- struct niobuf_remote *nb, struct obdo *oa,
- ldlm_mode_t mode)
+/* Free thread local buffers if they were allocated only for servicing
+ * this one request */
+static void ost_tls_put(struct ptlrpc_request *r)
{
- struct ldlm_res_id res_id;
- int nrbufs = obj->ioo_bufcnt;
- struct ost_prolong_data opd = { 0 };
- ENTRY;
-
- osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
+ struct ost_thread_local_cache *tls =
+ (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
- opd.opd_mode = mode;
- opd.opd_exp = req->rq_export;
- opd.opd_policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
- opd.opd_policy.l_extent.end = (nb[nrbufs - 1].offset +
- nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
-
- /* prolong locks for the current service time of the corresponding
- * portal (= OST_IO_PORTAL) */
- opd.opd_timeout = AT_OFF ? obd_timeout / 2:
- max(at_est2timeout(at_get(&req->rq_rqbd->
- rqbd_service->srv_at_estimate)), ldlm_timeout);
-
- CDEBUG(D_INFO,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
- res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
- opd.opd_policy.l_extent.end);
-
- if (oa->o_valid & OBD_MD_FLHANDLE) {
- struct ldlm_lock *lock;
-
- lock = ldlm_handle2lock(&oa->o_handle);
- if (lock != NULL) {
- ost_prolong_locks_iter(lock, &opd);
- if (opd.opd_lock_match) {
- LDLM_LOCK_PUT(lock);
- RETURN(1);
- }
-
- /* Check if the lock covers the whole IO region,
- * otherwise iterate through the resource. */
- if (lock->l_policy_data.l_extent.end >=
- opd.opd_policy.l_extent.end &&
- lock->l_policy_data.l_extent.start <=
- opd.opd_policy.l_extent.start) {
- LDLM_LOCK_PUT(lock);
- RETURN(0);
- }
- LDLM_LOCK_PUT(lock);
- }
+ if (unlikely(tls->temporary)) {
+ OBD_FREE_PTR(tls);
+ r->rq_svc_thread->t_data = NULL;
}
-
- opd.opd_oa = oa;
- ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id,
- ost_prolong_locks_iter, &opd);
- RETURN(opd.opd_lock_match);
}
static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti)
struct lustre_handle lockh = { 0 };
int niocount, npages, nob = 0, rc, i;
int no_reply = 0;
+ struct ost_thread_local_cache *tls;
ENTRY;
req->rq_bulk_read = 1;
if (ioo == NULL)
GOTO(out, rc = -EFAULT);
+ rc = ost_validate_obdo(exp, &body->oa, ioo);
+ if (rc)
+ RETURN(rc);
+
niocount = ioo->ioo_bufcnt;
remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
if (remote_nb == NULL)
if (rc)
GOTO(out, rc);
- /*
- * Per-thread array of struct niobuf_{local,remote}'s was allocated by
- * ost_thread_init().
- */
- local_nb = ost_tls(req)->local;
+ tls = ost_tls_get(req);
+ if (tls == NULL)
+ GOTO(out_bulk, rc = -ENOMEM);
+ local_nb = tls->local;
rc = ost_brw_lock_get(LCK_PR, exp, ioo, remote_nb, &lockh);
if (rc != 0)
- GOTO(out_bulk, rc);
+ GOTO(out_tls, rc);
/*
* If getting the lock took more time than
GOTO(out_lock, rc = -ETIMEDOUT);
}
+ repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
+ memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
+
npages = OST_THREAD_POOL_SIZE;
- rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1, ioo,
- remote_nb, &npages, local_nb, oti, capa);
+ rc = obd_preprw(req->rq_svc_thread->t_env, OBD_BRW_READ, exp,
+ &repbody->oa, 1, ioo, remote_nb, &npages, local_nb,
+ oti, capa);
if (rc != 0)
GOTO(out_lock, rc);
desc = ptlrpc_prep_bulk_exp(req, npages,
BULK_PUT_SOURCE, OST_BULK_PORTAL);
if (desc == NULL)
- GOTO(out_lock, rc = -ENOMEM);
-
- if (!lustre_handle_is_used(&lockh))
- /* no needs to try to prolong lock if server is asked
- * to handle locking (= OBD_BRW_SRVLOCK) */
- ost_rw_prolong_locks(req, ioo, remote_nb, &body->oa,
- LCK_PW | LCK_PR);
+ GOTO(out_commitrw, rc = -ENOMEM);
nob = 0;
for (i = 0; i < npages; i++) {
nob += page_rc;
if (page_rc != 0) { /* some data! */
LASSERT (local_nb[i].page != NULL);
- ptlrpc_prep_bulk_page(desc, local_nb[i].page,
- local_nb[i].offset & ~CFS_PAGE_MASK,
- page_rc);
+ ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].page,
+ local_nb[i].lnb_page_offset,
+ page_rc);
}
if (page_rc != local_nb[i].len) { /* short read */
}
if (body->oa.o_valid & OBD_MD_FLCKSUM) {
- cksum_type_t cksum_type = OBD_CKSUM_CRC32;
-
- if (body->oa.o_valid & OBD_MD_FLFLAGS)
- cksum_type = cksum_type_unpack(body->oa.o_flags);
- body->oa.o_flags = cksum_type_pack(cksum_type);
- body->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
- body->oa.o_cksum = ost_checksum_bulk(desc, OST_READ, cksum_type);
- CDEBUG(D_PAGE,"checksum at read origin: %x\n",body->oa.o_cksum);
+ cksum_type_t cksum_type =
+ cksum_type_unpack(repbody->oa.o_valid & OBD_MD_FLFLAGS ?
+ repbody->oa.o_flags : 0);
+ repbody->oa.o_flags = cksum_type_pack(cksum_type);
+ repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
+ repbody->oa.o_cksum = ost_checksum_bulk(desc, OST_READ,cksum_type);
+ CDEBUG(D_PAGE, "checksum at read origin: %x\n",
+ repbody->oa.o_cksum);
} else {
- body->oa.o_valid = 0;
+ repbody->oa.o_valid = 0;
}
/* We're finishing using body->oa as an input variable */
/* Check if client was evicted while we were doing i/o before touching
network */
if (rc == 0) {
- /* Check if there is eviction in progress, and if so, wait for
- * it to finish */
- if (unlikely(cfs_atomic_read(&exp->exp_obd->
- obd_evict_inprogress))) {
- lwi = LWI_INTR(NULL, NULL);
- rc = l_wait_event(exp->exp_obd->
- obd_evict_inprogress_waitq,
- !cfs_atomic_read(&exp->exp_obd->
- obd_evict_inprogress),
- &lwi);
- }
- /* Check if client was evicted or tried to reconnect already */
- if (exp->exp_failed || exp->exp_abort_active_req)
- rc = -ENOTCONN;
- else {
- rc = sptlrpc_svc_wrap_bulk(req, desc);
- if (rc == 0)
- rc = ptlrpc_start_bulk_transfer(desc);
- }
-
- if (rc == 0) {
- time_t start = cfs_time_current_sec();
- do {
- long timeoutl = req->rq_deadline -
- cfs_time_current_sec();
- cfs_duration_t timeout = timeoutl <= 0 ?
- CFS_TICK : cfs_time_seconds(timeoutl);
- lwi = LWI_TIMEOUT_INTERVAL(timeout,
- cfs_time_seconds(1),
- ost_bulk_timeout,
- desc);
- rc = l_wait_event(desc->bd_waitq,
- !ptlrpc_server_bulk_active(desc) ||
- exp->exp_failed ||
- exp->exp_abort_active_req,
- &lwi);
- LASSERT(rc == 0 || rc == -ETIMEDOUT);
- /* Wait again if we changed deadline */
- } while ((rc == -ETIMEDOUT) &&
- (req->rq_deadline > cfs_time_current_sec()));
-
- if (rc == -ETIMEDOUT) {
- DEBUG_REQ(D_ERROR, req,
- "timeout on bulk PUT after %ld%+lds",
- req->rq_deadline - start,
- cfs_time_current_sec() -
- req->rq_deadline);
- ptlrpc_abort_bulk(desc);
- } else if (exp->exp_failed) {
- DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT");
- rc = -ENOTCONN;
- ptlrpc_abort_bulk(desc);
- } else if (exp->exp_abort_active_req) {
- DEBUG_REQ(D_ERROR, req, "Reconnect on bulk PUT");
- /* we don't reply anyway */
- rc = -ETIMEDOUT;
- ptlrpc_abort_bulk(desc);
- } else if (!desc->bd_success ||
- desc->bd_nob_transferred != desc->bd_nob) {
- DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
- desc->bd_success ?
- "truncated" : "network error on",
- desc->bd_nob_transferred,
- desc->bd_nob);
- /* XXX should this be a different errno? */
- rc = -ETIMEDOUT;
- }
- } else {
- DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d", rc);
- }
+ if (likely(!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)))
+ rc = target_bulk_io(exp, desc, &lwi);
no_reply = rc != 0;
}
+out_commitrw:
/* Must commit after prep above in all cases */
- rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1, ioo,
- remote_nb, npages, local_nb, oti, rc);
+ rc = obd_commitrw(req->rq_svc_thread->t_env, OBD_BRW_READ, exp,
+ &repbody->oa, 1, ioo, remote_nb, npages, local_nb,
+ oti, rc);
- if (rc == 0) {
- repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
- memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
+ if (rc == 0)
ost_drop_id(exp, &repbody->oa);
- }
out_lock:
ost_brw_lock_put(LCK_PR, ioo, remote_nb, &lockh);
+out_tls:
+ ost_tls_put(req);
out_bulk:
- if (desc)
- ptlrpc_free_bulk(desc);
+ if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
+ ptlrpc_free_bulk_nopin(desc);
out:
LASSERT(rc <= 0);
if (rc == 0) {
} else {
/* reply out callback would free */
ptlrpc_req_drop_rs(req);
- CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
- "client will retry\n",
- exp->exp_obd->obd_name,
- exp->exp_client_uuid.uuid,
- exp->exp_connection->c_remote_uuid.uuid,
- libcfs_id2str(req->rq_peer));
+ LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
+ "client will retry: rc %d\n",
+ exp->exp_obd->obd_name,
+ obd_uuid2str(&exp->exp_client_uuid),
+ obd_export_nid2str(exp), rc);
+ }
+ /* send a bulk after reply to simulate a network delay or reordering
+ * by a router */
+ if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) {
+ cfs_waitq_t waitq;
+ struct l_wait_info lwi1;
+
+ CDEBUG(D_INFO, "reorder BULK\n");
+ cfs_waitq_init(&waitq);
+
+ lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
+ l_wait_event(waitq, 0, &lwi1);
+ rc = target_bulk_io(exp, desc, &lwi);
+ ptlrpc_free_bulk(desc);
}
RETURN(rc);
int rc, i, j;
obd_count client_cksum = 0, server_cksum = 0;
cksum_type_t cksum_type = OBD_CKSUM_CRC32;
- int no_reply = 0;
+ int no_reply = 0, mmap = 0;
__u32 o_uid = 0, o_gid = 0;
+ struct ost_thread_local_cache *tls;
ENTRY;
req->rq_bulk_write = 1;
/* pause before transaction has been started */
OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
- /* Check if there is eviction in progress, and if so, wait for it to
- * finish */
- if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
- lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
- rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
- !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
- &lwi);
- }
- if (exp->exp_failed)
- GOTO(out, rc = -ENOTCONN);
-
/* ost_body, ioobj & noibuf_remote are verified and swabbed in
* ost_rw_hpreq_check(). */
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
GOTO(out, rc = -EFAULT);
- if ((body->oa.o_flags & OBD_BRW_MEMALLOC) &&
- (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
- libcfs_memory_pressure_set();
-
objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
RCL_CLIENT) / sizeof(*ioo);
ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
if (ioo == NULL)
GOTO(out, rc = -EFAULT);
+
+ rc = ost_validate_obdo(exp, &body->oa, ioo);
+ if (rc)
+ RETURN(rc);
+
for (niocount = i = 0; i < objcount; i++)
niocount += ioo[i].ioo_bufcnt;
&RMF_NIOBUF_REMOTE, RCL_CLIENT) / sizeof(*remote_nb)))
GOTO(out, rc = -EFAULT);
+ if ((remote_nb[0].flags & OBD_BRW_MEMALLOC) &&
+ (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
+ cfs_memory_pressure_set();
+
if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
if (capa == NULL) {
rc = req_capsule_server_pack(&req->rq_pill);
if (rc != 0)
GOTO(out, rc);
- OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, obd_fail_val);
+ CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
- /*
- * Per-thread array of struct niobuf_{local,remote}'s was allocated by
- * ost_thread_init().
- */
- local_nb = ost_tls(req)->local;
+ tls = ost_tls_get(req);
+ if (tls == NULL)
+ GOTO(out_bulk, rc = -ENOMEM);
+ local_nb = tls->local;
rc = ost_brw_lock_get(LCK_PW, exp, ioo, remote_nb, &lockh);
if (rc != 0)
- GOTO(out_bulk, rc);
+ GOTO(out_tls, rc);
/*
* If getting the lock took more time than
GOTO(out_lock, rc = -ETIMEDOUT);
}
- if (!lustre_handle_is_used(&lockh))
- /* no needs to try to prolong lock if server is asked
- * to handle locking (= OBD_BRW_SRVLOCK) */
- ost_rw_prolong_locks(req, ioo, remote_nb,&body->oa, LCK_PW);
-
/* obd_preprw clobbers oa->valid, so save what we need */
if (body->oa.o_valid & OBD_MD_FLCKSUM) {
client_cksum = body->oa.o_cksum;
if (body->oa.o_valid & OBD_MD_FLFLAGS)
cksum_type = cksum_type_unpack(body->oa.o_flags);
}
+ if (body->oa.o_valid & OBD_MD_FLFLAGS && body->oa.o_flags & OBD_FL_MMAP)
+ mmap = 1;
/* Because we already sync grant info with client when reconnect,
* grant info will be cleared for resent req, then fed_grant and
o_uid = body->oa.o_uid;
o_gid = body->oa.o_gid;
}
+
+ repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
+ memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
+
npages = OST_THREAD_POOL_SIZE;
- rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount,
- ioo, remote_nb, &npages, local_nb, oti, capa);
+ rc = obd_preprw(req->rq_svc_thread->t_env, OBD_BRW_WRITE, exp,
+ &repbody->oa, objcount, ioo, remote_nb, &npages,
+ local_nb, oti, capa);
if (rc != 0)
GOTO(out_lock, rc);
desc = ptlrpc_prep_bulk_exp(req, npages,
BULK_GET_SINK, OST_BULK_PORTAL);
if (desc == NULL)
- GOTO(out_lock, rc = -ENOMEM);
+ GOTO(skip_transfer, rc = -ENOMEM);
/* NB Having prepped, we must commit... */
for (i = 0; i < npages; i++)
- ptlrpc_prep_bulk_page(desc, local_nb[i].page,
- local_nb[i].offset & ~CFS_PAGE_MASK,
- local_nb[i].len);
+ ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].page,
+ local_nb[i].lnb_page_offset,
+ local_nb[i].len);
rc = sptlrpc_svc_prep_bulk(req, desc);
if (rc != 0)
GOTO(out_lock, rc);
- /* Check if client was evicted or tried to reconnect while we
- * were doing i/o before touching network */
- if (desc->bd_export->exp_failed ||
- desc->bd_export->exp_abort_active_req)
- rc = -ENOTCONN;
- else
- rc = ptlrpc_start_bulk_transfer(desc);
- if (rc == 0) {
- time_t start = cfs_time_current_sec();
- do {
- long timeoutl = req->rq_deadline -
- cfs_time_current_sec();
- cfs_duration_t timeout = timeoutl <= 0 ?
- CFS_TICK : cfs_time_seconds(timeoutl);
- lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
- ost_bulk_timeout, desc);
- rc = l_wait_event(desc->bd_waitq,
- !ptlrpc_server_bulk_active(desc) ||
- desc->bd_export->exp_failed ||
- desc->bd_export->exp_abort_active_req,
- &lwi);
- LASSERT(rc == 0 || rc == -ETIMEDOUT);
- /* Wait again if we changed deadline */
- } while ((rc == -ETIMEDOUT) &&
- (req->rq_deadline > cfs_time_current_sec()));
-
- if (rc == -ETIMEDOUT) {
- DEBUG_REQ(D_ERROR, req,
- "timeout on bulk GET after %ld%+lds",
- req->rq_deadline - start,
- cfs_time_current_sec() -
- req->rq_deadline);
- ptlrpc_abort_bulk(desc);
- } else if (desc->bd_export->exp_failed) {
- DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET");
- rc = -ENOTCONN;
- ptlrpc_abort_bulk(desc);
- } else if (desc->bd_export->exp_abort_active_req) {
- DEBUG_REQ(D_ERROR, req, "Reconnect on bulk GET");
- /* we don't reply anyway */
- rc = -ETIMEDOUT;
- ptlrpc_abort_bulk(desc);
- } else if (!desc->bd_success) {
- DEBUG_REQ(D_ERROR, req, "network error on bulk GET");
- /* XXX should this be a different errno? */
- rc = -ETIMEDOUT;
- } else {
- rc = sptlrpc_svc_unwrap_bulk(req, desc);
- }
- } else {
- DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc);
- }
+ rc = target_bulk_io(exp, desc, &lwi);
no_reply = rc != 0;
- repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
- memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
-
- if (unlikely(client_cksum != 0 && rc == 0)) {
+skip_transfer:
+ if (client_cksum != 0 && rc == 0) {
static int cksum_counter;
repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
repbody->oa.o_cksum = server_cksum;
cksum_counter++;
if (unlikely(client_cksum != server_cksum)) {
- CERROR("client csum %x, server csum %x\n",
- client_cksum, server_cksum);
+ CDEBUG_LIMIT(mmap ? D_INFO : D_ERROR,
+ "client csum %x, server csum %x\n",
+ client_cksum, server_cksum);
cksum_counter = 0;
} else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
}
/* Must commit after prep above in all cases */
- rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa, objcount, ioo,
- remote_nb, npages, local_nb, oti, rc);
+ rc = obd_commitrw(req->rq_svc_thread->t_env, OBD_BRW_WRITE, exp,
+ &repbody->oa, objcount, ioo, remote_nb, npages,
+ local_nb, oti, rc);
if (rc == -ENOTCONN)
/* quota acquire process has been given up because
* either the client has been evicted or the client
*/
repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
- if (unlikely(client_cksum != server_cksum && rc == 0)) {
+ if (unlikely(client_cksum != server_cksum && rc == 0 && !mmap)) {
int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
char *msg;
char *via;
}
LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from "
- "%s%s%s inum "LPU64"/"LPU64" object "
+ "%s%s%s inode "DFID" object "
LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
exp->exp_obd->obd_name, msg,
libcfs_id2str(req->rq_peer),
via, router,
body->oa.o_valid & OBD_MD_FLFID ?
- body->oa.o_fid : (__u64)0,
+ body->oa.o_parent_seq : (__u64)0,
+ body->oa.o_valid & OBD_MD_FLFID ?
+ body->oa.o_parent_oid : 0,
body->oa.o_valid & OBD_MD_FLFID ?
- body->oa.o_generation :(__u64)0,
+ body->oa.o_parent_ver : 0,
body->oa.o_id,
body->oa.o_valid & OBD_MD_FLGROUP ?
- body->oa.o_gr : (__u64)0,
- local_nb[0].offset,
- local_nb[npages-1].offset +
+ body->oa.o_seq : (__u64)0,
+ local_nb[0].lnb_file_offset,
+ local_nb[npages-1].lnb_file_offset +
local_nb[npages-1].len - 1 );
CERROR("client csum %x, original server csum %x, "
"server csum now %x\n",
out_lock:
ost_brw_lock_put(LCK_PW, ioo, remote_nb, &lockh);
+out_tls:
+ ost_tls_put(req);
out_bulk:
if (desc)
- ptlrpc_free_bulk(desc);
+ ptlrpc_free_bulk_nopin(desc);
out:
if (rc == 0) {
oti_to_request(oti, req);
} else {
/* reply out callback would free */
ptlrpc_req_drop_rs(req);
- CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
- "client will retry\n",
- exp->exp_obd->obd_name,
- exp->exp_client_uuid.uuid,
- exp->exp_connection->c_remote_uuid.uuid,
- libcfs_id2str(req->rq_peer));
+ LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s), "
+ "client will retry: rc %d\n",
+ exp->exp_obd->obd_name,
+ obd_uuid2str(&exp->exp_client_uuid),
+ obd_export_nid2str(exp), rc);
}
- libcfs_memory_pressure_clr();
+ cfs_memory_pressure_clr();
RETURN(rc);
}
/* OBD will also check if KEY_IS(KEY_GRANT_SHRINK), and will cast val to
* a struct ost_body * value */
- rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL);
+ rc = obd_set_info_async(req->rq_svc_thread->t_env, exp, keylen,
+ key, vallen, val, NULL);
out:
lustre_msg_set_status(req->rq_repmsg, 0);
RETURN(rc);
}
keylen = req_capsule_get_size(pill, &RMF_SETINFO_KEY, RCL_CLIENT);
- rc = obd_get_info(exp, keylen, key, &replylen, NULL, NULL);
+ if (KEY_IS(KEY_FIEMAP)) {
+ struct ll_fiemap_info_key *fm_key = key;
+ int rc;
+
+ rc = ost_validate_obdo(exp, &fm_key->oa, NULL);
+ if (rc)
+ RETURN(rc);
+ }
+
+ rc = obd_get_info(req->rq_svc_thread->t_env, exp, keylen, key,
+ &replylen, NULL, NULL);
if (rc)
RETURN(rc);
RETURN(-ENOMEM);
/* call again to fill in the reply buffer */
- rc = obd_get_info(exp, keylen, key, &replylen, reply, NULL);
+ rc = obd_get_info(req->rq_svc_thread->t_env, exp, keylen, key,
+ &replylen, reply, NULL);
lustre_msg_set_status(req->rq_repmsg, 0);
RETURN(rc);
}
-#ifdef HAVE_QUOTA_SUPPORT
static int ost_handle_quotactl(struct ptlrpc_request *req)
{
struct obd_quotactl *oqctl, *repoqc;
if (rc)
RETURN(-ENOMEM);
- req->rq_status = obd_quotacheck(req->rq_export, oqctl);
- RETURN(0);
-}
-
-static int ost_handle_quota_adjust_qunit(struct ptlrpc_request *req)
-{
- struct quota_adjust_qunit *oqaq, *repoqa;
- struct lustre_quota_ctxt *qctxt;
- int rc;
- ENTRY;
-
- qctxt = &req->rq_export->exp_obd->u.obt.obt_qctxt;
- oqaq = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
- if (oqaq == NULL)
- GOTO(out, rc = -EPROTO);
-
- rc = req_capsule_server_pack(&req->rq_pill);
- if (rc)
- GOTO(out, rc);
-
- repoqa = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
- req->rq_status = obd_quota_adjust_qunit(req->rq_export, oqaq, qctxt);
- *repoqa = *oqaq;
-
- out:
- RETURN(rc);
+ /* deprecated, not used any more */
+ req->rq_status = -EOPNOTSUPP;
+ RETURN(-EOPNOTSUPP);
}
-#endif
static int ost_llog_handle_connect(struct obd_export *exp,
struct ptlrpc_request *req)
return rc;
}
+/* Ensure that data and metadata are synced to the disk when lock is cancelled
+ * (if requested) */
+int ost_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
+ void *data, int flag)
+{
+ struct lu_env env;
+ __u32 sync_lock_cancel = 0;
+ __u32 len = sizeof(sync_lock_cancel);
+ int rc = 0;
+
+ ENTRY;
+
+ rc = lu_env_init(&env, LCT_DT_THREAD);
+ if (unlikely(rc != 0))
+ RETURN(rc);
+
+ rc = obd_get_info(&env, lock->l_export, sizeof(KEY_SYNC_LOCK_CANCEL),
+ KEY_SYNC_LOCK_CANCEL, &len, &sync_lock_cancel, NULL);
+ if (rc == 0 && flag == LDLM_CB_CANCELING &&
+ (lock->l_granted_mode & (LCK_PW|LCK_GROUP)) &&
+ (sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
+ (sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
+ lock->l_flags & LDLM_FL_CBPENDING))) {
+ struct obd_info *oinfo;
+ struct obdo *oa;
+ int rc;
+
+ OBD_ALLOC_PTR(oinfo);
+ if (!oinfo)
+ GOTO(out_env, rc = -ENOMEM);
+ OBDO_ALLOC(oa);
+ if (!oa) {
+ OBD_FREE_PTR(oinfo);
+ GOTO(out_env, rc = -ENOMEM);
+ }
+ oa->o_id = lock->l_resource->lr_name.name[0];
+ oa->o_seq = lock->l_resource->lr_name.name[1];
+ oa->o_valid = OBD_MD_FLID|OBD_MD_FLGROUP;
+ oinfo->oi_oa = oa;
+ oinfo->oi_capa = BYPASS_CAPA;
+
+ rc = obd_sync(&env, lock->l_export, oinfo,
+ lock->l_policy_data.l_extent.start,
+ lock->l_policy_data.l_extent.end, NULL);
+ if (rc)
+ CERROR("Error %d syncing data on lock cancel\n", rc);
+
+ OBDO_FREE(oa);
+ OBD_FREE_PTR(oinfo);
+ }
+
+ rc = ldlm_server_blocking_ast(lock, desc, data, flag);
+out_env:
+ lu_env_fini(&env);
+ RETURN(rc);
+}
+
static int ost_filter_recovery_request(struct ptlrpc_request *req,
struct obd_device *obd, int *process)
{
RETURN(0);
default:
- DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
+ DEBUG_REQ(D_WARNING, req, "not permitted during recovery");
*process = -EAGAIN;
RETURN(0);
}
case OST_SYNC:
case OST_SET_INFO:
case OST_GET_INFO:
-#ifdef HAVE_QUOTA_SUPPORT
case OST_QUOTACHECK:
case OST_QUOTACTL:
case OST_QUOTA_ADJUST_QUNIT:
-#endif
rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION);
if (rc)
CERROR("bad opc %u version %08x, expecting %08x\n",
return rc;
}
+struct ost_prolong_data {
+ struct ptlrpc_request *opd_req;
+ struct obd_export *opd_exp;
+ struct obdo *opd_oa;
+ struct ldlm_res_id opd_resid;
+ struct ldlm_extent opd_extent;
+ ldlm_mode_t opd_mode;
+ unsigned int opd_locks;
+ int opd_timeout;
+};
+
+/* prolong locks for the current service time of the corresponding
+ * portal (= OST_IO_PORTAL)
+ */
+static inline int prolong_timeout(struct ptlrpc_request *req)
+{
+ struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
+
+ if (AT_OFF)
+ return obd_timeout / 2;
+
+ return max(at_est2timeout(at_get(&svcpt->scp_at_estimate)),
+ ldlm_timeout);
+}
+
+static void ost_prolong_lock_one(struct ost_prolong_data *opd,
+ struct ldlm_lock *lock)
+{
+ LASSERT(lock->l_export == opd->opd_exp);
+
+ if (lock->l_destroyed) /* lock already cancelled */
+ return;
+
+ /* XXX: never try to grab resource lock here because we're inside
+ * exp_bl_list_lock; in ldlm_lockd.c to handle waiting list we take
+ * res lock and then exp_bl_list_lock. */
+
+ if (!(lock->l_flags & LDLM_FL_AST_SENT))
+ /* ignore locks not being cancelled */
+ return;
+
+ LDLM_DEBUG(lock,
+ "refreshed for req x"LPU64" ext("LPU64"->"LPU64") to %ds.\n",
+ opd->opd_req->rq_xid, opd->opd_extent.start,
+ opd->opd_extent.end, opd->opd_timeout);
+
+ /* OK. this is a possible lock the user holds doing I/O
+ * let's refresh eviction timer for it */
+ ldlm_refresh_waiting_lock(lock, opd->opd_timeout);
+ ++opd->opd_locks;
+}
+
+static void ost_prolong_locks(struct ost_prolong_data *data)
+{
+ struct obd_export *exp = data->opd_exp;
+ struct obdo *oa = data->opd_oa;
+ struct ldlm_lock *lock;
+ ENTRY;
+
+ if (oa->o_valid & OBD_MD_FLHANDLE) {
+ /* mostly a request should be covered by only one lock, try
+ * fast path. */
+ lock = ldlm_handle2lock(&oa->o_handle);
+ if (lock != NULL) {
+ /* Fast path to check if the lock covers the whole IO
+ * region exclusively. */
+ if (lock->l_granted_mode == LCK_PW &&
+ ldlm_extent_contain(&lock->l_policy_data.l_extent,
+ &data->opd_extent)) {
+ /* bingo */
+ ost_prolong_lock_one(data, lock);
+ LDLM_LOCK_PUT(lock);
+ RETURN_EXIT;
+ }
+ LDLM_LOCK_PUT(lock);
+ }
+ }
+
+
+ cfs_spin_lock_bh(&exp->exp_bl_list_lock);
+ cfs_list_for_each_entry(lock, &exp->exp_bl_list, l_exp_list) {
+ LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
+ LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
+
+ if (!ldlm_res_eq(&data->opd_resid, &lock->l_resource->lr_name))
+ continue;
+
+ if (!ldlm_extent_overlap(&lock->l_policy_data.l_extent,
+ &data->opd_extent))
+ continue;
+
+ ost_prolong_lock_one(data, lock);
+ }
+ cfs_spin_unlock_bh(&exp->exp_bl_list_lock);
+
+ EXIT;
+}
+
/**
* Returns 1 if the given PTLRPC matches the given LDLM locks, or 0 if it does
* not.
{
struct niobuf_remote *nb;
struct obd_ioobj *ioo;
- struct ost_body *body;
- int objcount, niocount;
- int mode, opc, i;
- __u64 start, end;
+ int mode, opc;
+ struct ldlm_extent ext;
ENTRY;
opc = lustre_msg_get_opc(req->rq_reqmsg);
LASSERT(opc == OST_READ || opc == OST_WRITE);
- /* As the request may be covered by several locks, do not look at
- * o_handle, look at the RPC IO region. */
- body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
- if (body == NULL)
- RETURN(0);
-
- objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
- RCL_CLIENT) / sizeof(*ioo);
ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
- if (ioo == NULL)
- RETURN(0);
-
- for (niocount = i = 0; i < objcount; i++)
- niocount += ioo[i].ioo_bufcnt;
+ LASSERT(ioo != NULL);
nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
- if (nb == NULL ||
- niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE,
- RCL_CLIENT) / sizeof(*nb)))
- RETURN(0);
+ LASSERT(nb != NULL);
- mode = LCK_PW;
- if (opc == OST_READ)
- mode |= LCK_PR;
-
- start = nb[0].offset & CFS_PAGE_MASK;
- end = (nb[ioo->ioo_bufcnt - 1].offset +
- nb[ioo->ioo_bufcnt - 1].len - 1) | ~CFS_PAGE_MASK;
+ ext.start = nb->offset;
+ nb += ioo->ioo_bufcnt - 1;
+ ext.end = nb->offset + nb->len - 1;
LASSERT(lock->l_resource != NULL);
- if (!osc_res_name_eq(ioo->ioo_id, ioo->ioo_gr,
+ if (!osc_res_name_eq(ioo->ioo_id, ioo->ioo_seq,
&lock->l_resource->lr_name))
RETURN(0);
+ mode = LCK_PW;
+ if (opc == OST_READ)
+ mode |= LCK_PR;
if (!(lock->l_granted_mode & mode))
RETURN(0);
- if (lock->l_policy_data.l_extent.end < start ||
- lock->l_policy_data.l_extent.start > end)
- RETURN(0);
-
- RETURN(1);
+ RETURN(ldlm_extent_overlap(&lock->l_policy_data.l_extent, &ext));
}
/**
*/
static int ost_rw_hpreq_check(struct ptlrpc_request *req)
{
- struct niobuf_remote *nb;
- struct obd_ioobj *ioo;
+ struct obd_device *obd = req->rq_export->exp_obd;
struct ost_body *body;
- int objcount, niocount;
- int mode, opc, i;
+ struct obd_ioobj *ioo;
+ struct niobuf_remote *nb;
+ struct ost_prolong_data opd = { 0 };
+ int mode, opc;
ENTRY;
+ /*
+ * Use LASSERT to do sanity check because malformed RPCs should have
+ * been filtered out in ost_hpreq_handler().
+ */
opc = lustre_msg_get_opc(req->rq_reqmsg);
LASSERT(opc == OST_READ || opc == OST_WRITE);
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
- if (body == NULL)
- RETURN(-EFAULT);
+ LASSERT(body != NULL);
- objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
- RCL_CLIENT) / sizeof(*ioo);
ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
- if (ioo == NULL)
- RETURN(-EFAULT);
+ LASSERT(ioo != NULL);
- for (niocount = i = 0; i < objcount; i++)
- niocount += ioo[i].ioo_bufcnt;
nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
- if (nb == NULL ||
- niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE,
- RCL_CLIENT) / sizeof(*nb)))
- RETURN(-EFAULT);
- if (niocount != 0 && (nb[0].flags & OBD_BRW_SRVLOCK))
- RETURN(-EFAULT);
+ LASSERT(nb != NULL);
+ LASSERT(!(nb->flags & OBD_BRW_SRVLOCK));
+
+ osc_build_res_name(ioo->ioo_id, ioo->ioo_seq, &opd.opd_resid);
+ opd.opd_req = req;
mode = LCK_PW;
if (opc == OST_READ)
mode |= LCK_PR;
- RETURN(ost_rw_prolong_locks(req, ioo, nb, &body->oa, mode));
-}
+ opd.opd_mode = mode;
+ opd.opd_exp = req->rq_export;
+ opd.opd_oa = &body->oa;
+ opd.opd_extent.start = nb->offset;
+ nb += ioo->ioo_bufcnt - 1;
+ opd.opd_extent.end = nb->offset + nb->len - 1;
+ opd.opd_timeout = prolong_timeout(req);
-static int ost_punch_prolong_locks(struct ptlrpc_request *req, struct obdo *oa)
-{
- struct ldlm_res_id res_id = { .name = { oa->o_id } };
- struct ost_prolong_data opd = { 0 };
- __u64 start, end;
- ENTRY;
+ DEBUG_REQ(D_RPCTRACE, req,
+ "%s %s: refresh rw locks: " LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
+ obd->obd_name, cfs_current()->comm,
+ opd.opd_resid.name[0], opd.opd_resid.name[1],
+ opd.opd_extent.start, opd.opd_extent.end);
- start = oa->o_size;
- end = start + oa->o_blocks;
+ ost_prolong_locks(&opd);
- opd.opd_mode = LCK_PW;
- opd.opd_exp = req->rq_export;
- opd.opd_policy.l_extent.start = start & CFS_PAGE_MASK;
- if (oa->o_blocks == OBD_OBJECT_EOF || end < start)
- opd.opd_policy.l_extent.end = OBD_OBJECT_EOF;
- else
- opd.opd_policy.l_extent.end = end | ~CFS_PAGE_MASK;
-
- /* prolong locks for the current service time of the corresponding
- * portal (= OST_IO_PORTAL) */
- opd.opd_timeout = AT_OFF ? obd_timeout / 2:
- max(at_est2timeout(at_get(&req->rq_rqbd->
- rqbd_service->srv_at_estimate)), ldlm_timeout);
-
- CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
- res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
- opd.opd_policy.l_extent.end);
-
- opd.opd_oa = oa;
- ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id,
- ost_prolong_locks_iter, &opd);
- RETURN(opd.opd_lock_match);
+ CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
+ obd->obd_name, opd.opd_locks, req);
+
+ RETURN(opd.opd_locks);
+}
+
+static void ost_rw_hpreq_fini(struct ptlrpc_request *req)
+{
+ (void)ost_rw_hpreq_check(req);
}
/**
ENTRY;
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
- if (body == NULL)
- RETURN(0); /* can't return -EFAULT here */
+ LASSERT(body != NULL);
if (body->oa.o_valid & OBD_MD_FLHANDLE &&
body->oa.o_handle.cookie == lock->l_handle.h_cookie)
RETURN(1);
+
RETURN(0);
}
*/
static int ost_punch_hpreq_check(struct ptlrpc_request *req)
{
+ struct obd_device *obd = req->rq_export->exp_obd;
struct ost_body *body;
+ struct obdo *oa;
+ struct ost_prolong_data opd = { 0 };
+ __u64 start, end;
+ ENTRY;
body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
- if (body == NULL)
- RETURN(-EFAULT);
+ LASSERT(body != NULL);
+
+ oa = &body->oa;
+ LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS) ||
+ !(oa->o_flags & OBD_FL_SRVLOCK));
+
+ start = oa->o_size;
+ end = start + oa->o_blocks;
+
+ opd.opd_req = req;
+ opd.opd_mode = LCK_PW;
+ opd.opd_exp = req->rq_export;
+ opd.opd_oa = oa;
+ opd.opd_extent.start = start;
+ opd.opd_extent.end = end;
+ if (oa->o_blocks == OBD_OBJECT_EOF)
+ opd.opd_extent.end = OBD_OBJECT_EOF;
+ opd.opd_timeout = prolong_timeout(req);
+
+ osc_build_res_name(oa->o_id, oa->o_seq, &opd.opd_resid);
+
+ CDEBUG(D_DLMTRACE,
+ "%s: refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
+ obd->obd_name,
+ opd.opd_resid.name[0], opd.opd_resid.name[1],
+ opd.opd_extent.start, opd.opd_extent.end);
+
+ ost_prolong_locks(&opd);
- LASSERT(!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
- !(body->oa.o_flags & OBD_FL_SRVLOCK));
+ CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
+ obd->obd_name, opd.opd_locks, req);
- RETURN(ost_punch_prolong_locks(req, &body->oa));
+ RETURN(opd.opd_locks > 0);
+}
+
+static void ost_punch_hpreq_fini(struct ptlrpc_request *req)
+{
+ (void)ost_punch_hpreq_check(req);
}
struct ptlrpc_hpreq_ops ost_hpreq_rw = {
- .hpreq_lock_match = ost_rw_hpreq_lock_match,
- .hpreq_check = ost_rw_hpreq_check,
+ .hpreq_lock_match = ost_rw_hpreq_lock_match,
+ .hpreq_check = ost_rw_hpreq_check,
+ .hpreq_fini = ost_rw_hpreq_fini
};
struct ptlrpc_hpreq_ops ost_hpreq_punch = {
- .hpreq_lock_match = ost_punch_hpreq_lock_match,
- .hpreq_check = ost_punch_hpreq_check,
+ .hpreq_lock_match = ost_punch_hpreq_lock_match,
+ .hpreq_check = ost_punch_hpreq_check,
+ .hpreq_fini = ost_punch_hpreq_fini
};
/** Assign high priority operations to the request if needed. */
struct niobuf_remote *nb;
struct obd_ioobj *ioo;
int objcount, niocount;
+ int rc;
int i;
/* RPCs on the H-P queue can be inspected before
CERROR("Missing/short ost_body\n");
RETURN(-EFAULT);
}
+
objcount = req_capsule_get_size(&req->rq_pill,
&RMF_OBD_IOOBJ,
RCL_CLIENT) /
RETURN(-EFAULT);
}
+ rc = ost_validate_obdo(req->rq_export, &body->oa, ioo);
+ if (rc) {
+ CERROR("invalid object ids\n");
+ RETURN(rc);
+ }
+
for (niocount = i = 0; i < objcount; i++) {
if (ioo[i].ioo_bufcnt == 0) {
CERROR("ioo[%d] has zero bufcnt\n", i);
struct obd_device *obd = NULL;
ENTRY;
+ /* OST module is kept between remounts, but the last reference
+ * to specific module (say, osd or ofd) kills all related keys
+ * from the environment. so we have to refill it until the root
+ * cause is fixed properly */
+ lu_env_refill(req->rq_svc_thread->t_env);
+
LASSERT(current->journal_info == NULL);
/* primordial rpcs don't affect server recovery */
if (rc)
RETURN(rc);
+ if (req && req->rq_reqmsg && req->rq_export &&
+ (req->rq_export->exp_connect_flags & OBD_CONNECT_JOBSTATS))
+ oti->oti_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
+
switch (lustre_msg_get_opc(req->rq_reqmsg)) {
case OST_CONNECT: {
CDEBUG(D_INODE, "connect\n");
req_capsule_set(&req->rq_pill, &RQF_OST_BRW_WRITE);
CDEBUG(D_INODE, "write\n");
/* req->rq_request_portal would be nice, if it was set */
- if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
- CERROR("%s: deny write request from %s to portal %u\n",
- req->rq_export->exp_obd->obd_name,
- obd_export_nid2str(req->rq_export),
- req->rq_rqbd->rqbd_service->srv_req_portal);
+ if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
+ CERROR("%s: deny write request from %s to portal %u\n",
+ req->rq_export->exp_obd->obd_name,
+ obd_export_nid2str(req->rq_export),
+ ptlrpc_req2svc(req)->srv_req_portal);
GOTO(out, rc = -EPROTO);
}
if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
req_capsule_set(&req->rq_pill, &RQF_OST_BRW_READ);
CDEBUG(D_INODE, "read\n");
/* req->rq_request_portal would be nice, if it was set */
- if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
- CERROR("%s: deny read request from %s to portal %u\n",
- req->rq_export->exp_obd->obd_name,
- obd_export_nid2str(req->rq_export),
- req->rq_rqbd->rqbd_service->srv_req_portal);
+ if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
+ CERROR("%s: deny read request from %s to portal %u\n",
+ req->rq_export->exp_obd->obd_name,
+ obd_export_nid2str(req->rq_export),
+ ptlrpc_req2svc(req)->srv_req_portal);
GOTO(out, rc = -EPROTO);
}
if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
req_capsule_set(&req->rq_pill, &RQF_OST_SYNC);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET))
RETURN(0);
- rc = ost_sync(req->rq_export, req);
+ rc = ost_sync(req->rq_export, req, oti);
break;
case OST_SET_INFO:
DEBUG_REQ(D_INODE, req, "set_info");
req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC);
rc = ost_get_info(req->rq_export, req);
break;
-#ifdef HAVE_QUOTA_SUPPORT
case OST_QUOTACHECK:
CDEBUG(D_INODE, "quotacheck\n");
req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK);
RETURN(0);
rc = ost_handle_quotactl(req);
break;
- case OST_QUOTA_ADJUST_QUNIT:
- CDEBUG(D_INODE, "quota_adjust_qunit\n");
- req_capsule_set(&req->rq_pill, &RQF_OST_QUOTA_ADJUST_QUNIT);
- rc = ost_handle_quota_adjust_qunit(req);
- break;
-#endif
case OBD_PING:
DEBUG_REQ(D_INODE, req, "ping");
req_capsule_set(&req->rq_pill, &RQF_OBD_PING);
if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE))
RETURN(0);
rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
- ldlm_server_blocking_ast,
+ ost_blocking_ast,
ldlm_server_glimpse_ast);
fail = OBD_FAIL_OST_LDLM_REPLY_NET;
break;
LASSERT(thread != NULL);
LASSERT(thread->t_data == NULL);
- LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
OBD_ALLOC_PTR(tls);
if (tls == NULL)
#define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
+static struct cfs_cpt_table *ost_io_cptable;
+
/* Sigh - really, this is an OSS, the _server_, not the _target_ */
static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
{
- struct ost_obd *ost = &obd->u.ost;
- struct lprocfs_static_vars lvars;
- int oss_min_threads;
- int oss_max_threads;
- int oss_min_create_threads;
- int oss_max_create_threads;
- int rc;
- ENTRY;
+ static struct ptlrpc_service_conf svc_conf;
+ struct ost_obd *ost = &obd->u.ost;
+ struct lprocfs_static_vars lvars;
+ nodemask_t *mask;
+ int rc;
+ ENTRY;
rc = cfs_cleanup_group_info();
if (rc)
lprocfs_ost_init_vars(&lvars);
lprocfs_obd_setup(obd, lvars.obd_vars);
- cfs_sema_init(&ost->ost_health_sem, 1);
-
- if (oss_num_threads) {
- /* If oss_num_threads is set, it is the min and the max. */
- if (oss_num_threads > OSS_THREADS_MAX)
- oss_num_threads = OSS_THREADS_MAX;
- if (oss_num_threads < OSS_THREADS_MIN)
- oss_num_threads = OSS_THREADS_MIN;
- oss_max_threads = oss_min_threads = oss_num_threads;
- } else {
- /* Base min threads on memory and cpus */
- oss_min_threads =
- cfs_num_possible_cpus() * CFS_NUM_CACHEPAGES >>
- (27 - CFS_PAGE_SHIFT);
- if (oss_min_threads < OSS_THREADS_MIN)
- oss_min_threads = OSS_THREADS_MIN;
- /* Insure a 4x range for dynamic threads */
- if (oss_min_threads > OSS_THREADS_MAX / 4)
- oss_min_threads = OSS_THREADS_MAX / 4;
- oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4 + 1);
- }
-
- ost->ost_service =
- ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
- OST_MAXREPSIZE, OST_REQUEST_PORTAL,
- OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
- ost_handle, LUSTRE_OSS_NAME,
- obd->obd_proc_entry, target_print_req,
- oss_min_threads, oss_max_threads,
- "ll_ost", LCT_DT_THREAD, NULL);
- if (ost->ost_service == NULL) {
- CERROR("failed to start service\n");
- GOTO(out_lprocfs, rc = -ENOMEM);
+ cfs_mutex_init(&ost->ost_health_mutex);
+
+ svc_conf = (typeof(svc_conf)) {
+ .psc_name = LUSTRE_OSS_NAME,
+ .psc_watchdog_factor = OSS_SERVICE_WATCHDOG_FACTOR,
+ .psc_buf = {
+ .bc_nbufs = OST_NBUFS,
+ .bc_buf_size = OST_BUFSIZE,
+ .bc_req_max_size = OST_MAXREQSIZE,
+ .bc_rep_max_size = OST_MAXREPSIZE,
+ .bc_req_portal = OST_REQUEST_PORTAL,
+ .bc_rep_portal = OSC_REPLY_PORTAL,
+ },
+ .psc_thr = {
+ .tc_thr_name = "ll_ost",
+ .tc_thr_factor = OSS_THR_FACTOR,
+ .tc_nthrs_init = OSS_NTHRS_INIT,
+ .tc_nthrs_base = OSS_NTHRS_BASE,
+ .tc_nthrs_max = OSS_NTHRS_MAX,
+ .tc_nthrs_user = oss_num_threads,
+ .tc_cpu_affinity = 1,
+ .tc_ctx_tags = LCT_DT_THREAD,
+ },
+ .psc_cpt = {
+ .cc_pattern = oss_cpts,
+ },
+ .psc_ops = {
+ .so_req_handler = ost_handle,
+ .so_req_printer = target_print_req,
+ .so_hpreq_handler = ptlrpc_hpreq_handler,
+ },
+ };
+ ost->ost_service = ptlrpc_register_service(&svc_conf,
+ obd->obd_proc_entry);
+ if (IS_ERR(ost->ost_service)) {
+ rc = PTR_ERR(ost->ost_service);
+ CERROR("failed to start service: %d\n", rc);
+ GOTO(out_lprocfs, rc);
}
- rc = ptlrpc_start_threads(obd, ost->ost_service);
- if (rc)
- GOTO(out_service, rc = -EINVAL);
-
- if (oss_num_create_threads) {
- if (oss_num_create_threads > OSS_MAX_CREATE_THREADS)
- oss_num_create_threads = OSS_MAX_CREATE_THREADS;
- if (oss_num_create_threads < OSS_MIN_CREATE_THREADS)
- oss_num_create_threads = OSS_MIN_CREATE_THREADS;
- oss_min_create_threads = oss_max_create_threads =
- oss_num_create_threads;
- } else {
- oss_min_create_threads = OSS_MIN_CREATE_THREADS;
- oss_max_create_threads = OSS_MAX_CREATE_THREADS;
+ memset(&svc_conf, 0, sizeof(svc_conf));
+ svc_conf = (typeof(svc_conf)) {
+ .psc_name = "ost_create",
+ .psc_watchdog_factor = OSS_SERVICE_WATCHDOG_FACTOR,
+ .psc_buf = {
+ .bc_nbufs = OST_NBUFS,
+ .bc_buf_size = OST_BUFSIZE,
+ .bc_req_max_size = OST_MAXREQSIZE,
+ .bc_rep_max_size = OST_MAXREPSIZE,
+ .bc_req_portal = OST_CREATE_PORTAL,
+ .bc_rep_portal = OSC_REPLY_PORTAL,
+ },
+ .psc_thr = {
+ .tc_thr_name = "ll_ost_create",
+ .tc_thr_factor = OSS_CR_THR_FACTOR,
+ .tc_nthrs_init = OSS_CR_NTHRS_INIT,
+ .tc_nthrs_base = OSS_CR_NTHRS_BASE,
+ .tc_nthrs_max = OSS_CR_NTHRS_MAX,
+ .tc_nthrs_user = oss_num_create_threads,
+ .tc_cpu_affinity = 1,
+ .tc_ctx_tags = LCT_DT_THREAD,
+ },
+ .psc_cpt = {
+ .cc_pattern = oss_cpts,
+ },
+ .psc_ops = {
+ .so_req_handler = ost_handle,
+ .so_req_printer = target_print_req,
+ .so_hpreq_handler = NULL,
+ },
+ };
+ ost->ost_create_service = ptlrpc_register_service(&svc_conf,
+ obd->obd_proc_entry);
+ if (IS_ERR(ost->ost_create_service)) {
+ rc = PTR_ERR(ost->ost_create_service);
+ CERROR("failed to start OST create service: %d\n", rc);
+ GOTO(out_service, rc);
}
- ost->ost_create_service =
- ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
- OST_MAXREPSIZE, OST_CREATE_PORTAL,
- OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
- ost_handle, "ost_create",
- obd->obd_proc_entry, target_print_req,
- oss_min_create_threads, oss_max_create_threads,
- "ll_ost_creat", LCT_DT_THREAD, NULL);
- if (ost->ost_create_service == NULL) {
- CERROR("failed to start OST create service\n");
- GOTO(out_service, rc = -ENOMEM);
+ mask = cfs_cpt_table->ctb_nodemask;
+ /* event CPT feature is disabled in libcfs level by set partition
+ * number to 1, we still want to set node affinity for io service */
+ if (cfs_cpt_number(cfs_cpt_table) == 1 && nodes_weight(*mask) > 1) {
+ int cpt = 0;
+ int i;
+
+ ost_io_cptable = cfs_cpt_table_alloc(nodes_weight(*mask));
+ for_each_node_mask(i, *mask) {
+ if (ost_io_cptable == NULL) {
+ CWARN("OSS failed to create CPT table\n");
+ break;
+ }
+
+ rc = cfs_cpt_set_node(ost_io_cptable, cpt++, i);
+ if (!rc) {
+ CWARN("OSS Failed to set node %d for"
+ "IO CPT table\n", i);
+ cfs_cpt_table_free(ost_io_cptable);
+ ost_io_cptable = NULL;
+ break;
+ }
+ }
+ }
+
+ memset(&svc_conf, 0, sizeof(svc_conf));
+ svc_conf = (typeof(svc_conf)) {
+ .psc_name = "ost_io",
+ .psc_watchdog_factor = OSS_SERVICE_WATCHDOG_FACTOR,
+ .psc_buf = {
+ .bc_nbufs = OST_NBUFS,
+ .bc_buf_size = OST_BUFSIZE,
+ .bc_req_max_size = OST_MAXREQSIZE,
+ .bc_rep_max_size = OST_MAXREPSIZE,
+ .bc_req_portal = OST_IO_PORTAL,
+ .bc_rep_portal = OSC_REPLY_PORTAL,
+ },
+ .psc_thr = {
+ .tc_thr_name = "ll_ost_io",
+ .tc_thr_factor = OSS_THR_FACTOR,
+ .tc_nthrs_init = OSS_NTHRS_INIT,
+ .tc_nthrs_base = OSS_NTHRS_BASE,
+ .tc_nthrs_max = OSS_NTHRS_MAX,
+ .tc_nthrs_user = oss_num_threads,
+ .tc_cpu_affinity = 1,
+ .tc_ctx_tags = LCT_DT_THREAD,
+ },
+ .psc_cpt = {
+ .cc_cptable = ost_io_cptable,
+ .cc_pattern = ost_io_cptable == NULL ?
+ oss_io_cpts : NULL,
+ },
+ .psc_ops = {
+ .so_thr_init = ost_thread_init,
+ .so_thr_done = ost_thread_done,
+ .so_req_handler = ost_handle,
+ .so_hpreq_handler = ost_hpreq_handler,
+ .so_req_printer = target_print_req,
+ .so_hpreq_handler = NULL,
+ },
+ };
+ ost->ost_io_service = ptlrpc_register_service(&svc_conf,
+ obd->obd_proc_entry);
+ if (IS_ERR(ost->ost_io_service)) {
+ rc = PTR_ERR(ost->ost_io_service);
+ CERROR("failed to start OST I/O service: %d\n", rc);
+ ost->ost_io_service = NULL;
+ GOTO(out_create, rc);
}
- rc = ptlrpc_start_threads(obd, ost->ost_create_service);
- if (rc)
- GOTO(out_create, rc = -EINVAL);
-
- ost->ost_io_service =
- ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
- OST_MAXREPSIZE, OST_IO_PORTAL,
- OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
- ost_handle, "ost_io",
- obd->obd_proc_entry, target_print_req,
- oss_min_threads, oss_max_threads,
- "ll_ost_io", LCT_DT_THREAD, ost_hpreq_handler);
- if (ost->ost_io_service == NULL) {
- CERROR("failed to start OST I/O service\n");
- GOTO(out_create, rc = -ENOMEM);
- }
-
- ost->ost_io_service->srv_init = ost_thread_init;
- ost->ost_io_service->srv_done = ost_thread_done;
- ost->ost_io_service->srv_cpu_affinity = 1;
- rc = ptlrpc_start_threads(obd, ost->ost_io_service);
- if (rc)
- GOTO(out_io, rc = -EINVAL);
-
ping_evictor_start();
RETURN(0);
-out_io:
- ptlrpc_unregister_service(ost->ost_io_service);
- ost->ost_io_service = NULL;
out_create:
ptlrpc_unregister_service(ost->ost_create_service);
ost->ost_create_service = NULL;
ping_evictor_stop();
- cfs_spin_lock_bh(&obd->obd_processing_task_lock);
- if (obd->obd_recovering) {
- target_cancel_recovery_timer(obd);
- obd->obd_recovering = 0;
- }
- cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
-
- cfs_down(&ost->ost_health_sem);
+ /* there is no recovery for OST OBD, all recovery is controlled by
+ * obdfilter OBD */
+ LASSERT(obd->obd_recovering == 0);
+ cfs_mutex_lock(&ost->ost_health_mutex);
ptlrpc_unregister_service(ost->ost_service);
ptlrpc_unregister_service(ost->ost_create_service);
ptlrpc_unregister_service(ost->ost_io_service);
ost->ost_service = NULL;
ost->ost_create_service = NULL;
- cfs_up(&ost->ost_health_sem);
+ ost->ost_io_service = NULL;
- lprocfs_obd_cleanup(obd);
+ cfs_mutex_unlock(&ost->ost_health_mutex);
+
+ lprocfs_obd_cleanup(obd);
+
+ if (ost_io_cptable != NULL) {
+ cfs_cpt_table_free(ost_io_cptable);
+ ost_io_cptable = NULL;
+ }
- RETURN(err);
+ RETURN(err);
}
-static int ost_health_check(struct obd_device *obd)
+static int ost_health_check(const struct lu_env *env, struct obd_device *obd)
{
struct ost_obd *ost = &obd->u.ost;
int rc = 0;
- cfs_down(&ost->ost_health_sem);
+ cfs_mutex_lock(&ost->ost_health_mutex);
rc |= ptlrpc_service_health_check(ost->ost_service);
rc |= ptlrpc_service_health_check(ost->ost_create_service);
rc |= ptlrpc_service_health_check(ost->ost_io_service);
- cfs_up(&ost->ost_health_sem);
+ cfs_mutex_unlock(&ost->ost_health_mutex);
/*
* health_check to return 0 on healthy
int rc;
ENTRY;
+ ost_page_to_corrupt = cfs_alloc_page(CFS_ALLOC_STD);
+
lprocfs_ost_init_vars(&lvars);
rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
LUSTRE_OSS_NAME, NULL);
static void /*__exit*/ ost_exit(void)
{
+ if (ost_page_to_corrupt)
+ page_cache_release(ost_page_to_corrupt);
+
class_unregister_type(LUSTRE_OSS_NAME);
}