-/* -*- 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.
* in the LICENSE file that accompanied this code).
*
* You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see [sun.com URL with a
- * copy of GPLv2].
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
*
* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
* CA 95054 USA or visit www.sun.com if you need additional information or
* 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>
#include <lustre_debug.h>
#include <linux/init.h>
#include <lprocfs_status.h>
-#include <lustre_commit_confd.h>
#include <libcfs/list.h>
#include <lustre_quota.h>
#include "ost_internal.h"
CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444,
"number of OSS create threads to start");
+/**
+ * Do not return server-side uid/gid to remote client
+ */
+static void ost_drop_id(struct obd_export *exp, struct obdo *oa)
+{
+ if (exp_connect_rmtclient(exp)) {
+ oa->o_uid = -1;
+ oa->o_gid = -1;
+ oa->o_valid &= ~(OBD_MD_FLUID | OBD_MD_FLGID);
+ }
+}
+
+/**
+ * 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 (!ack_lock->mode)
break;
/* XXX not even calling target_send_reply in some cases... */
- ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode);
+ ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode, 0);
}
}
struct obd_trans_info *oti)
{
struct ost_body *body, *repbody;
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
+ struct lustre_capa *capa = NULL;
int rc;
ENTRY;
- body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
- lustre_swab_ost_body);
+ /* Get the request body */
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
RETURN(-EFAULT);
if (body->oa.o_id == 0)
RETURN(-EPROTO);
- if (lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1)) {
+ 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;
- dlm = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*dlm),
- lustre_swab_ldlm_request);
+
+ dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
if (dlm == NULL)
RETURN (-EFAULT);
ldlm_request_cancel(req, dlm, 0);
}
-
- rc = lustre_pack_reply(req, 2, size, NULL);
+
+ /* If there's a capability, get it */
+ 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 DESTROY");
+ RETURN (-EFAULT);
+ }
+ }
+
+ /* Prepare the reply */
+ rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
RETURN(rc);
+ /* Get the log cancellation cookie */
if (body->oa.o_valid & OBD_MD_FLCOOKIE)
- oti->oti_logcookies = obdo_logcookie(&body->oa);
- repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
- sizeof(*repbody));
+ oti->oti_logcookies = &body->oa.o_lcookie;
+
+ /* Finish the reply */
+ repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
- req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL);
+
+ /* Do the destroy and set the reply status accordingly */
+ req->rq_status = obd_destroy(req->rq_svc_thread->t_env, exp,
+ &repbody->oa, NULL, oti, NULL, capa);
RETURN(0);
}
+/**
+ * Helper function for getting server side [start, start+count] DLM lock
+ * if asked by client.
+ */
+static int ost_lock_get(struct obd_export *exp, struct obdo *oa,
+ __u64 start, __u64 count, struct lustre_handle *lh,
+ int mode, int flags)
+{
+ struct ldlm_res_id res_id;
+ ldlm_policy_data_t policy;
+ __u64 end = start + count;
+
+ ENTRY;
+
+ LASSERT(!lustre_handle_is_used(lh));
+ /* 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;
+
+ /* If ->o_blocks is EOF it means "lock till the end of the
+ * file". Otherwise, it's size of a hole being punched (in bytes) */
+ if (count == OBD_OBJECT_EOF || end < start)
+ policy.l_extent.end = OBD_OBJECT_EOF;
+ else
+ policy.l_extent.end = end | ~CFS_PAGE_MASK;
+
+ RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
+ LDLM_EXTENT, &policy, mode, &flags,
+ ldlm_blocking_ast, ldlm_completion_ast,
+ ldlm_glimpse_ast, NULL, 0, NULL, lh));
+}
+
+/* Helper function: release lock, if any. */
+static void ost_lock_put(struct obd_export *exp,
+ struct lustre_handle *lh, int mode)
+{
+ ENTRY;
+ if (lustre_handle_is_used(lh))
+ ldlm_lock_decref(lh, mode);
+ EXIT;
+}
+
static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
{
struct ost_body *body, *repbody;
- struct obd_info oinfo = { { { 0 } } };
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
+ struct obd_info *oinfo;
+ struct lustre_handle lh = { 0 };
+ struct lustre_capa *capa = NULL;
int rc;
ENTRY;
- body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
- lustre_swab_ost_body);
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
RETURN(-EFAULT);
- rc = lustre_pack_reply(req, 2, size, NULL);
+ 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 = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
- sizeof(*repbody));
+ repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
repbody->oa = body->oa;
- oinfo.oi_oa = &repbody->oa;
- if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
- oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
- REQ_REC_OFF + 1);
- req->rq_status = obd_getattr(exp, &oinfo);
- RETURN(0);
+ rc = ost_lock_get(exp, &repbody->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0);
+ if (rc)
+ RETURN(rc);
+
+ OBD_ALLOC_PTR(oinfo);
+ if (!oinfo)
+ GOTO(unlock, rc = -ENOMEM);
+ oinfo->oi_oa = &repbody->oa;
+ oinfo->oi_capa = capa;
+
+ req->rq_status = obd_getattr(req->rq_svc_thread->t_env, exp, oinfo);
+
+ OBD_FREE_PTR(oinfo);
+
+ ost_drop_id(exp, &repbody->oa);
+
+unlock:
+ ost_lock_put(exp, &lh, LCK_PR);
+ RETURN(rc);
}
static int ost_statfs(struct ptlrpc_request *req)
{
struct obd_statfs *osfs;
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*osfs) };
int rc;
ENTRY;
- rc = lustre_pack_reply(req, 2, size, NULL);
+ rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
RETURN(rc);
- osfs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*osfs));
+ 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() - HZ, 0);
- if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
- osfs->os_bfree = osfs->os_bavail = 64;
+ 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);
struct obd_trans_info *oti)
{
struct ost_body *body, *repbody;
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
int rc;
ENTRY;
- body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
- lustre_swab_ost_body);
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
RETURN(-EFAULT);
- rc = lustre_pack_reply(req, 2, size, NULL);
+ rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
if (rc)
RETURN(rc);
- repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
- sizeof(*repbody));
- memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
- oti->oti_logcookies = obdo_logcookie(&repbody->oa);
- req->rq_status = obd_create(exp, &repbody->oa, NULL, oti);
- //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
- RETURN(0);
-}
-
-/*
- * Helper function for ost_punch(): if asked by client, acquire [size, EOF]
- * lock on the file being truncated.
- */
-static int ost_punch_lock_get(struct obd_export *exp, struct obdo *oa,
- struct lustre_handle *lh)
-{
- int flags;
- struct ldlm_res_id res_id;
- ldlm_policy_data_t policy;
- __u64 start;
- __u64 finis;
-
- ENTRY;
-
- osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
- LASSERT(!lustre_handle_is_used(lh));
-
- if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
- !(oa->o_flags & OBD_FL_TRUNCLOCK))
- RETURN(0);
-
- CDEBUG(D_INODE, "OST-side truncate lock.\n");
-
- start = oa->o_size;
- finis = start + oa->o_blocks;
-
- /*
- * standard truncate optimization: if file body is completely
- * destroyed, don't send data back to the server.
- */
- flags = (start == 0) ? LDLM_AST_DISCARD_DATA : 0;
-
- policy.l_extent.start = start & CFS_PAGE_MASK;
-
- /*
- * If ->o_blocks is EOF it means "lock till the end of the
- * file". Otherwise, it's size of a hole being punched (in bytes)
- */
- if (oa->o_blocks == OBD_OBJECT_EOF || finis < start)
- policy.l_extent.end = OBD_OBJECT_EOF;
- else
- policy.l_extent.end = finis | ~CFS_PAGE_MASK;
+ rc = req_capsule_server_pack(&req->rq_pill);
+ if (rc)
+ RETURN(rc);
- RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
- LDLM_EXTENT, &policy, LCK_PW, &flags,
- ldlm_blocking_ast, ldlm_completion_ast,
- ldlm_glimpse_ast, NULL, 0, NULL, lh));
-}
+ repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
+ repbody->oa = body->oa;
+ oti->oti_logcookies = &body->oa.o_lcookie;
-/*
- * Helper function for ost_punch(): release lock acquired by
- * ost_punch_lock_get(), if any.
- */
-static void ost_punch_lock_put(struct obd_export *exp, struct obdo *oa,
- struct lustre_handle *lh)
-{
- ENTRY;
- if (lustre_handle_is_used(lh))
- ldlm_lock_decref(lh, LCK_PW);
- EXIT;
+ 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;
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
- int rc;
+ int rc, flags = 0;
struct lustre_handle lh = {0,};
ENTRY;
/* check that we do support OBD_CONNECT_TRUNCLOCK. */
CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
- body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
- lustre_swab_ost_body);
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
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(-EINVAL);
+ RETURN(-EPROTO);
- rc = lustre_pack_reply(req, 2, size, NULL);
+ rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
RETURN(rc);
- repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
- sizeof(*repbody));
- rc = ost_punch_lock_get(exp, oinfo.oi_oa, &lh);
+ /* standard truncate optimization: if file body is completely
+ * destroyed, don't send data back to the server. */
+ if (body->oa.o_size == 0)
+ flags |= LDLM_AST_DISCARD_DATA;
+
+ repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
+ 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_TRUNCLOCK)
+ 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_TRUNCLOCK is the only bit set in
+ * 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 (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");
+ GOTO(unlock, rc = -EFAULT);
+ }
+ }
- if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
- oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
- REQ_REC_OFF + 1);
- req->rq_status = obd_punch(exp, &oinfo, oti, NULL);
- ost_punch_lock_put(exp, oinfo.oi_oa, &lh);
+ 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;
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
int rc;
ENTRY;
- body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
- lustre_swab_ost_body);
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
RETURN(-EFAULT);
- if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
- capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 1);
+ 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 SYNC");
+ RETURN (-EFAULT);
+ }
+ }
- rc = lustre_pack_reply(req, 2, size, NULL);
+ rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
RETURN(rc);
- repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
- sizeof(*repbody));
- 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 = 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;
+ 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;
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
+ struct obd_info *oinfo;
+ struct lustre_capa *capa = NULL;
int rc;
- struct obd_info oinfo = { { { 0 } } };
ENTRY;
- body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
- lustre_swab_ost_body);
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
if (body == NULL)
RETURN(-EFAULT);
- rc = lustre_pack_reply(req, 2, size, NULL);
+ rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
if (rc)
RETURN(rc);
- repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
- sizeof(*repbody));
- repbody->oa = body->oa;
-
- oinfo.oi_oa = &repbody->oa;
- if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
- oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
- REQ_REC_OFF + 1);
- req->rq_status = obd_setattr(exp, &oinfo, oti);
- 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 int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo,
- struct niobuf_remote *rnb, int nrnb,
- struct niobuf_remote **pp_rnbp)
-{
- /* Copy a remote niobuf, splitting it into page-sized chunks
- * and setting ioo[i].ioo_bufcnt accordingly */
- struct niobuf_remote *pp_rnb;
- int i;
- int j;
- int page;
- int rnbidx = 0;
- int npages = 0;
-
- /*
- * array of sufficient size already preallocated by caller
- */
- LASSERT(pp_rnbp != NULL);
- LASSERT(*pp_rnbp != NULL);
-
- /* first count and check the number of pages required */
- for (i = 0; i < nioo; i++)
- for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) {
- obd_off offset = rnb[rnbidx].offset;
- obd_off p0 = offset >> CFS_PAGE_SHIFT;
- obd_off pn = (offset + rnb[rnbidx].len - 1) >>
- CFS_PAGE_SHIFT;
-
- LASSERT(rnbidx < nrnb);
-
- npages += (pn + 1 - p0);
+ rc = req_capsule_server_pack(&req->rq_pill);
+ if (rc)
+ RETURN(rc);
- if (rnb[rnbidx].len == 0) {
- CERROR("zero len BRW: obj %d objid "LPX64
- " buf %u\n", i, ioo[i].ioo_id, j);
- return -EINVAL;
- }
- if (j > 0 &&
- rnb[rnbidx].offset <= rnb[rnbidx-1].offset) {
- CERROR("unordered BRW: obj %d objid "LPX64
- " buf %u offset "LPX64" <= "LPX64"\n",
- i, ioo[i].ioo_id, j, rnb[rnbidx].offset,
- rnb[rnbidx].offset);
- return -EINVAL;
- }
+ 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);
}
-
- LASSERT(rnbidx == nrnb);
-
- if (npages == nrnb) { /* all niobufs are for single pages */
- *pp_rnbp = rnb;
- return npages;
}
- pp_rnb = *pp_rnbp;
+ repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
+ repbody->oa = body->oa;
- /* now do the actual split */
- page = rnbidx = 0;
- for (i = 0; i < nioo; i++) {
- int obj_pages = 0;
+ OBD_ALLOC_PTR(oinfo);
+ if (!oinfo)
+ RETURN(-ENOMEM);
+ oinfo->oi_oa = &repbody->oa;
+ oinfo->oi_capa = capa;
- for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) {
- obd_off off = rnb[rnbidx].offset;
- int nob = rnb[rnbidx].len;
+ req->rq_status = obd_setattr(req->rq_svc_thread->t_env, exp, oinfo,
+ oti);
- LASSERT(rnbidx < nrnb);
- do {
- obd_off poff = off & ~CFS_PAGE_MASK;
- int pnob = (poff + nob > CFS_PAGE_SIZE) ?
- CFS_PAGE_SIZE - poff : nob;
-
- LASSERT(page < npages);
- pp_rnb[page].len = pnob;
- pp_rnb[page].offset = off;
- pp_rnb[page].flags = rnb[rnbidx].flags;
-
- CDEBUG(0, " obj %d id "LPX64
- "page %d(%d) "LPX64" for %d, flg %x\n",
- i, ioo[i].ioo_id, obj_pages, page,
- pp_rnb[page].offset, pp_rnb[page].len,
- pp_rnb[page].flags);
- page++;
- obj_pages++;
-
- off += pnob;
- nob -= pnob;
- } while (nob > 0);
- LASSERT(nob == 0);
- }
- ioo[i].ioo_bufcnt = obj_pages;
- }
- LASSERT(page == npages);
+ OBD_FREE_PTR(oinfo);
- return npages;
+ ost_drop_id(exp, &repbody->oa);
+ RETURN(0);
}
static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc,
- 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));
- kunmap(page);
- }
-
- return cksum;
-}
-
-/*
- * populate @nio by @nrpages pages from per-thread page pool
- */
-static void ost_nio_pages_get(struct ptlrpc_request *req,
- struct niobuf_local *nio, int nrpages)
-{
- int i;
- struct ost_thread_local_cache *tls;
-
- ENTRY;
-
- LASSERT(nrpages <= OST_THREAD_POOL_SIZE);
- LASSERT(req != NULL);
- LASSERT(req->rq_svc_thread != NULL);
-
- tls = ost_tls(req);
- LASSERT(tls != NULL);
-
- memset(nio, 0, nrpages * sizeof *nio);
- for (i = 0; i < nrpages; ++ i) {
- struct page *page;
-
- page = tls->page[i];
- LASSERT(page != NULL);
- POISON_PAGE(page, 0xf1);
- nio[i].page = page;
- LL_CDEBUG_PAGE(D_INFO, page, "%d\n", i);
- }
- EXIT;
-}
-
-/*
- * Dual for ost_nio_pages_get(). Poison pages in pool for debugging
- */
-static void ost_nio_pages_put(struct ptlrpc_request *req,
- struct niobuf_local *nio, int nrpages)
+ cksum_type_t cksum_type)
{
- int i;
-
- ENTRY;
-
- LASSERT(nrpages <= OST_THREAD_POOL_SIZE);
-
- for (i = 0; i < nrpages; ++ i)
- POISON_PAGE(nio[i].page, 0xf2);
- EXIT;
+ 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;
+ char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
+ memcpy(ptr, "bad3", min(4, len));
+ kunmap(desc->bd_iov[i].kiov_page);
+ }
+ 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;
+ char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
+ memcpy(ptr, "bad4", min(4, len));
+ kunmap(desc->bd_iov[i].kiov_page);
+ /* nobody should use corrupted page again */
+ ClearPageUptodate(desc->bd_iov[i].kiov_page);
+ }
+ }
+
+ 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));
if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
RETURN(0);
- /* EXPENSIVE ASSERTION */
for (i = 1; i < nrbufs; i ++)
- LASSERT((nb[0].flags & OBD_BRW_SRVLOCK) ==
- (nb[i].flags & OBD_BRW_SRVLOCK));
+ if ((nb[0].flags & OBD_BRW_SRVLOCK) !=
+ (nb[i].flags & OBD_BRW_SRVLOCK))
+ RETURN(-EFAULT);
policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
policy.l_extent.end = (nb[nrbufs - 1].offset +
EXIT;
}
-struct ost_prolong_data {
- struct obd_export *opd_exp;
- ldlm_policy_data_t opd_policy;
- ldlm_mode_t opd_mode;
-};
-
-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;
+ struct ost_thread_local_cache *tls =
+ (struct ost_thread_local_cache *)(r->rq_svc_thread->t_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;
- }
-
- 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;
+ }
}
-
- /* OK. this is a possible lock the user holds doing I/O
- * let's refresh eviction timer for it */
- ldlm_refresh_waiting_lock(lock);
-
- return LDLM_ITER_CONTINUE;
+ return tls;
}
-static void ost_prolong_locks(struct obd_export *exp, struct obd_ioobj *obj,
- struct niobuf_remote *nb, 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;
- 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 = exp;
- 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;
-
- 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);
- ldlm_resource_iterate(exp->exp_obd->obd_namespace, &res_id,
- ost_prolong_locks_iter, &opd);
+ if (unlikely(tls->temporary)) {
+ OBD_FREE_PTR(tls);
+ r->rq_svc_thread->t_data = NULL;
+ }
}
static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti)
{
- struct ptlrpc_bulk_desc *desc;
+ struct ptlrpc_bulk_desc *desc = NULL;
struct obd_export *exp = req->rq_export;
struct niobuf_remote *remote_nb;
- struct niobuf_remote *pp_rnb = NULL;
struct niobuf_local *local_nb;
struct obd_ioobj *ioo;
struct ost_body *body, *repbody;
struct lustre_capa *capa = NULL;
struct l_wait_info lwi;
struct lustre_handle lockh = { 0 };
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
int niocount, npages, nob = 0, rc, i;
int no_reply = 0;
+ struct ost_thread_local_cache *tls;
ENTRY;
req->rq_bulk_read = 1;
/* Check if there is eviction in progress, and if so, wait for it to
* finish */
- if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
+ 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,
- !atomic_read(&exp->exp_obd->obd_evict_inprogress),
+ !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
&lwi);
}
if (exp->exp_failed)
GOTO(out, rc = -ENOTCONN);
- body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
- lustre_swab_ost_body);
- if (body == NULL) {
- CERROR("Missing/short ost_body\n");
+ /* 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);
- }
- ioo = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*ioo),
- lustre_swab_obd_ioobj);
- if (ioo == NULL) {
- CERROR("Missing/short ioobj\n");
+ /*
+ * A req_capsule_X_get_array(pill, field, ptr_to_element_count) function
+ * would be useful here and wherever we get &RMF_OBD_IOOBJ and
+ * &RMF_NIOBUF_REMOTE.
+ */
+ 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);
niocount = ioo->ioo_bufcnt;
- if (niocount > PTLRPC_MAX_BRW_PAGES) {
- DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)",
- niocount);
+ remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
+ if (remote_nb == NULL)
GOTO(out, rc = -EFAULT);
- }
- remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2,
- niocount * sizeof(*remote_nb),
- lustre_swab_niobuf_remote);
- if (remote_nb == NULL) {
- CERROR("Missing/short niobuf\n");
- GOTO(out, rc = -EFAULT);
- }
- if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */
- for (i = 1; i < niocount; i++)
- lustre_swab_niobuf_remote (&remote_nb[i]);
+ 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 BRW READ");
+ GOTO(out, rc = -EFAULT);
+ }
}
- if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
- capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3);
-
- rc = lustre_pack_reply(req, 2, size, NULL);
+ rc = req_capsule_server_pack(&req->rq_pill);
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;
- pp_rnb = ost_tls(req)->remote;
-
- /* FIXME all niobuf splitting should be done in obdfilter if needed */
- /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
- npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb);
- if (npages < 0)
- GOTO(out, rc = npages);
-
- LASSERT(npages <= OST_THREAD_POOL_SIZE);
-
- ost_nio_pages_get(req, local_nb, npages);
-
- desc = ptlrpc_prep_bulk_exp(req, npages,
- BULK_PUT_SOURCE, OST_BULK_PORTAL);
- if (desc == NULL)
- GOTO(out, rc = -ENOMEM);
+ 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, pp_rnb, &lockh);
+ 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
* client was willing to wait, drop it. b=11330
*/
GOTO(out_lock, rc = -ETIMEDOUT);
}
- rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1,
- ioo, npages, pp_rnb, local_nb, oti, capa);
+ 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(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);
- ost_prolong_locks(exp, ioo, pp_rnb, LCK_PW | LCK_PR);
+ desc = ptlrpc_prep_bulk_exp(req, npages,
+ BULK_PUT_SOURCE, OST_BULK_PORTAL);
+ if (desc == NULL)
+ GOTO(out_commitrw, rc = -ENOMEM);
nob = 0;
for (i = 0; i < npages; i++) {
break;
}
- LASSERTF(page_rc <= pp_rnb[i].len, "page_rc (%d) > "
- "pp_rnb[%d].len (%d)\n", page_rc, i, pp_rnb[i].len);
nob += page_rc;
if (page_rc != 0) { /* some data! */
LASSERT (local_nb[i].page != NULL);
ptlrpc_prep_bulk_page(desc, local_nb[i].page,
- pp_rnb[i].offset & ~CFS_PAGE_MASK,
+ local_nb[i].offset & ~CFS_PAGE_MASK,
page_rc);
}
- if (page_rc != pp_rnb[i].len) { /* short read */
+ if (page_rc != local_nb[i].len) { /* short read */
/* All subsequent pages should be 0 */
while(++i < npages)
LASSERT(local_nb[i].rc == 0);
}
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(atomic_read(&exp->exp_obd->
- obd_evict_inprogress))) {
- lwi = LWI_INTR(NULL, NULL);
- rc = l_wait_event(exp->exp_obd->
- obd_evict_inprogress_waitq,
- !atomic_read(&exp->exp_obd->
- obd_evict_inprogress),
- &lwi);
- }
- if (exp->exp_failed)
- rc = -ENOTCONN;
- else {
- sptlrpc_svc_wrap_bulk(req, desc);
-
- 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 || rc) ?
- 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_bulk_active(desc) ||
- exp->exp_failed, &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 (!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, 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);
- ost_nio_pages_put(req, local_nb, npages);
-
- if (rc == 0) {
- repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
- sizeof(*repbody));
- 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, pp_rnb, &lockh);
+ ost_brw_lock_put(LCK_PR, ioo, remote_nb, &lockh);
+out_tls:
+ ost_tls_put(req);
out_bulk:
- ptlrpc_free_bulk(desc);
+ if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
+ ptlrpc_free_bulk(desc);
out:
LASSERT(rc <= 0);
if (rc == 0) {
req->rq_status = nob;
+ ptlrpc_lprocfs_brw(req, nob);
target_committed_to_req(req);
ptlrpc_reply(req);
} else if (!no_reply) {
} 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);
static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
{
- struct ptlrpc_bulk_desc *desc;
+ struct ptlrpc_bulk_desc *desc = NULL;
struct obd_export *exp = req->rq_export;
struct niobuf_remote *remote_nb;
- struct niobuf_remote *pp_rnb;
struct niobuf_local *local_nb;
struct obd_ioobj *ioo;
struct ost_body *body, *repbody;
struct lustre_handle lockh = {0};
struct lustre_capa *capa = NULL;
__u32 *rcs;
- __u32 size[3] = { sizeof(struct ptlrpc_body), sizeof(*body) };
int objcount, niocount, npages;
- int rc, swab, i, j;
+ 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(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,
- !atomic_read(&exp->exp_obd->obd_evict_inprogress),
- &lwi);
- }
- if (exp->exp_failed)
- GOTO(out, rc = -ENOTCONN);
-
- swab = lustre_msg_swabbed(req->rq_reqmsg);
- body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
- lustre_swab_ost_body);
- if (body == NULL) {
- CERROR("Missing/short ost_body\n");
+ /* 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);
- }
- lustre_set_req_swabbed(req, REQ_REC_OFF + 1);
- objcount = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1) /
- sizeof(*ioo);
- if (objcount == 0) {
- CERROR("Missing/short ioobj\n");
+ 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);
- }
- if (objcount > 1) {
- CERROR("too many ioobjs (%d)\n", objcount);
- GOTO(out, rc = -EFAULT);
- }
- ioo = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1,
- objcount * sizeof(*ioo));
- LASSERT (ioo != NULL);
- for (niocount = i = 0; i < objcount; i++) {
- if (swab)
- lustre_swab_obd_ioobj(&ioo[i]);
- if (ioo[i].ioo_bufcnt == 0) {
- CERROR("ioo[%d] has zero bufcnt\n", i);
- 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;
- }
- if (niocount > PTLRPC_MAX_BRW_PAGES) {
- DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)",
- niocount);
+ /*
+ * It'd be nice to have a capsule function to indicate how many elements
+ * there were in a buffer for an RMF that's declared to be an array.
+ * It's easy enough to compute the number of elements here though.
+ */
+ remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
+ if (remote_nb == NULL || niocount != (req_capsule_get_size(&req->rq_pill,
+ &RMF_NIOBUF_REMOTE, RCL_CLIENT) / sizeof(*remote_nb)))
GOTO(out, rc = -EFAULT);
- }
- remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2,
- niocount * sizeof(*remote_nb),
- lustre_swab_niobuf_remote);
- if (remote_nb == NULL) {
- CERROR("Missing/short niobuf\n");
- GOTO(out, rc = -EFAULT);
- }
- if (swab) { /* swab the remaining niobufs */
- for (i = 1; i < niocount; i++)
- lustre_swab_niobuf_remote (&remote_nb[i]);
- }
+ 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 = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3);
+ 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 BRW WRITE");
+ GOTO(out, rc = -EFAULT);
+ }
+ }
- size[REPLY_REC_OFF + 1] = niocount * sizeof(*rcs);
- rc = lustre_pack_reply(req, 3, size, NULL);
+ req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
+ niocount * sizeof(*rcs));
+ 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);
- rcs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1,
- niocount * sizeof(*rcs));
-
- /*
- * Per-thread array of struct niobuf_{local,remote}'s was allocated by
- * ost_thread_init().
- */
- local_nb = ost_tls(req)->local;
- pp_rnb = ost_tls(req)->remote;
+ CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
+ rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
- /* FIXME all niobuf splitting should be done in obdfilter if needed */
- /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
- npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
- if (npages < 0)
- GOTO(out, rc = npages);
-
- LASSERT(npages <= OST_THREAD_POOL_SIZE);
-
- ost_nio_pages_get(req, local_nb, npages);
-
- desc = ptlrpc_prep_bulk_exp(req, npages,
- BULK_GET_SINK, OST_BULK_PORTAL);
- if (desc == NULL)
- GOTO(out, rc = -ENOMEM);
+ 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, pp_rnb, &lockh);
+ 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
* client was willing to wait, drop it. b=11330
*/
GOTO(out_lock, rc = -ETIMEDOUT);
}
- ost_prolong_locks(exp, ioo, pp_rnb, 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
- * total_grant will not be modified in following preprw_write */
+ * grant info will be cleared for resent req, then fed_grant and
+ * total_grant will not be modified in following preprw_write */
if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
body->oa.o_valid &= ~OBD_MD_FLGRANT;
}
- rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount,
- ioo, npages, pp_rnb, local_nb, oti, capa);
+ if (exp_connect_rmtclient(exp)) {
+ 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(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(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,
- pp_rnb[i].offset & ~CFS_PAGE_MASK,
- pp_rnb[i].len);
+ local_nb[i].offset & ~CFS_PAGE_MASK,
+ local_nb[i].len);
- /* Check if client was evicted while we were doing i/o before touching
- network */
- if (desc->bd_export->exp_failed)
- 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 || rc) ?
- 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_bulk_active(desc) ||
- desc->bd_export->exp_failed, &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_success ||
- desc->bd_nob_transferred != desc->bd_nob) {
- DEBUG_REQ(D_ERROR, req, "%s bulk GET %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, "ptlrpc_bulk_get failed: rc %d", rc);
- }
- no_reply = rc != 0;
-
- if (rc == 0)
- sptlrpc_svc_unwrap_bulk(req, desc);
+ rc = sptlrpc_svc_prep_bulk(req, desc);
+ if (rc != 0)
+ GOTO(out_lock, rc);
- repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
- sizeof(*repbody));
- memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
+ rc = target_bulk_io(exp, desc, &lwi);
+ no_reply = rc != 0;
- 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, 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
+ * has timed out the request already */
+ no_reply = 1;
+
+ if (exp_connect_rmtclient(exp)) {
+ repbody->oa.o_uid = o_uid;
+ repbody->oa.o_gid = o_gid;
+ }
- if (unlikely(client_cksum != server_cksum && rc == 0)) {
+ /*
+ * Disable sending mtime back to the client. If the client locked the
+ * whole object, then it has already updated the mtime on its side,
+ * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
+ */
+ repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
+
+ if (unlikely(client_cksum != server_cksum && rc == 0 && !mmap)) {
int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
char *msg;
char *via;
via = " via ";
router = libcfs_nid2str(desc->bd_sender);
}
-
+
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,
- pp_rnb[0].offset,
- pp_rnb[npages-1].offset+pp_rnb[npages-1].len
- - 1 );
+ body->oa.o_seq : (__u64)0,
+ local_nb[0].offset,
+ local_nb[npages-1].offset +
+ local_nb[npages-1].len - 1 );
CERROR("client csum %x, original server csum %x, "
"server csum now %x\n",
client_cksum, server_cksum, new_cksum);
}
- ost_nio_pages_put(req, local_nb, npages);
-
if (rc == 0) {
+ int nob = 0;
+
/* set per-requested niobuf return codes */
for (i = j = 0; i < niocount; i++) {
- int nob = remote_nb[i].len;
+ int len = remote_nb[i].len;
+ nob += len;
rcs[i] = 0;
do {
LASSERT(j < npages);
if (local_nb[j].rc < 0)
rcs[i] = local_nb[j].rc;
- nob -= pp_rnb[j].len;
+ len -= local_nb[j].len;
j++;
- } while (nob > 0);
- LASSERT(nob == 0);
+ } while (len > 0);
+ LASSERT(len == 0);
}
LASSERT(j == npages);
+ ptlrpc_lprocfs_brw(req, nob);
}
out_lock:
- ost_brw_lock_put(LCK_PW, ioo, pp_rnb, &lockh);
+ ost_brw_lock_put(LCK_PW, ioo, remote_nb, &lockh);
+out_tls:
+ ost_tls_put(req);
out_bulk:
- ptlrpc_free_bulk(desc);
+ if (desc)
+ ptlrpc_free_bulk(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);
}
+ cfs_memory_pressure_clr();
RETURN(rc);
}
+/**
+ * Implementation of OST_SET_INFO.
+ *
+ * OST_SET_INFO is like ioctl(): heavily overloaded. Specifically, it takes a
+ * "key" and a value RPC buffers as arguments, with the value's contents
+ * interpreted according to the key.
+ *
+ * Value types that need swabbing have swabbing done explicitly, either here or
+ * in functions called from here. This should be corrected: all swabbing should
+ * be done in the capsule abstraction, as that will then allow us to move
+ * swabbing exclusively to the client without having to modify server code
+ * outside the capsule abstraction's implementation itself. To correct this
+ * will require minor changes to the capsule abstraction; see the comments for
+ * req_capsule_extend() in layout.c.
+ */
static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
{
+ struct ost_body *body = NULL, *repbody;
char *key, *val = NULL;
int keylen, vallen, rc = 0;
+ int is_grant_shrink = 0;
ENTRY;
- key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
+ key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
if (key == NULL) {
DEBUG_REQ(D_HA, req, "no set_info key");
RETURN(-EFAULT);
}
- keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
+ keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
+ RCL_CLIENT);
+
+ vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
+ RCL_CLIENT);
- rc = lustre_pack_reply(req, 1, NULL, NULL);
+ if ((is_grant_shrink = KEY_IS(KEY_GRANT_SHRINK)))
+ /* In this case the value is actually an RMF_OST_BODY, so we
+ * transmutate the type of this PTLRPC */
+ req_capsule_extend(&req->rq_pill, &RQF_OST_SET_GRANT_INFO);
+
+ rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
RETURN(rc);
- vallen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1);
- if (vallen)
- val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, 0);
+ if (vallen) {
+ if (is_grant_shrink) {
+ body = req_capsule_client_get(&req->rq_pill,
+ &RMF_OST_BODY);
+ if (!body)
+ RETURN(-EFAULT);
+
+ repbody = req_capsule_server_get(&req->rq_pill,
+ &RMF_OST_BODY);
+ memcpy(repbody, body, sizeof(*body));
+ val = (char*)repbody;
+ } else {
+ val = req_capsule_client_get(&req->rq_pill,
+ &RMF_SETINFO_VAL);
+ }
+ }
if (KEY_IS(KEY_EVICT_BY_NID)) {
if (val && vallen)
obd_export_evict_by_nid(exp->exp_obd, val);
-
GOTO(out, rc = 0);
+ } else if (KEY_IS(KEY_MDS_CONN) && ptlrpc_req_need_swab(req)) {
+ if (vallen < sizeof(__u32))
+ RETURN(-EFAULT);
+ __swab32s((__u32 *)val);
}
- rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL);
+ /* 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(req->rq_svc_thread->t_env, exp, keylen,
+ key, vallen, val, NULL);
out:
lustre_msg_set_status(req->rq_repmsg, 0);
RETURN(rc);
struct req_capsule *pill = &req->rq_pill;
ENTRY;
- req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC);
-
/* this common part for get_info rpc */
key = req_capsule_client_get(pill, &RMF_SETINFO_KEY);
if (key == NULL) {
}
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;
- __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
int rc;
ENTRY;
- oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
- lustre_swab_obd_quotactl);
+ oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
if (oqctl == NULL)
GOTO(out, rc = -EPROTO);
- rc = lustre_pack_reply(req, 2, size, NULL);
+ rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
GOTO(out, rc);
- repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
-
+ repoqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
req->rq_status = obd_quotactl(req->rq_export, oqctl);
*repoqc = *oqctl;
+
out:
RETURN(rc);
}
RETURN(-EPROTO);
rc = req_capsule_server_pack(&req->rq_pill);
- if (rc) {
- CERROR("ost: out of memory while packing quotacheck reply\n");
+ 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, NULL);
+ *repoqa = *oqaq;
+
+ out:
+ RETURN(rc);
+}
+#endif
+
static int ost_llog_handle_connect(struct obd_export *exp,
struct ptlrpc_request *req)
{
int rc;
ENTRY;
- body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*body));
+ body = req_capsule_client_get(&req->rq_pill, &RMF_LLOGD_CONN_BODY);
rc = obd_llog_connect(exp, body);
RETURN(rc);
}
-static int filter_export_check_flavor(struct filter_obd *filter,
- struct obd_export *exp,
- struct ptlrpc_request *req)
+#define ost_init_sec_none(reply, exp) \
+do { \
+ reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \
+ OBD_CONNECT_RMT_CLIENT_FORCE | \
+ OBD_CONNECT_OSS_CAPA); \
+ cfs_spin_lock(&exp->exp_lock); \
+ exp->exp_connect_flags = reply->ocd_connect_flags; \
+ cfs_spin_unlock(&exp->exp_lock); \
+} while (0)
+
+static int ost_init_sec_level(struct ptlrpc_request *req)
{
- int rc = 0;
-
- /* FIXME
- * this should be done in filter_connect()/filter_reconnect(), but
- * we can't obtain information like NID, which stored in incoming
- * request, thus can't decide what flavor to use. so we do it here.
- *
- * This hack should be removed after the OST stack be rewritten, just
- * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect().
- */
- if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_INVALID)
- return 0;
+ struct obd_export *exp = req->rq_export;
+ struct req_capsule *pill = &req->rq_pill;
+ struct obd_device *obd = exp->exp_obd;
+ struct filter_obd *filter = &obd->u.filter;
+ char *client = libcfs_nid2str(req->rq_peer.nid);
+ struct obd_connect_data *data, *reply;
+ int rc = 0, remote;
+ ENTRY;
+
+ data = req_capsule_client_get(pill, &RMF_CONNECT_DATA);
+ reply = req_capsule_server_get(pill, &RMF_CONNECT_DATA);
+ if (data == NULL || reply == NULL)
+ RETURN(-EFAULT);
+
+ /* connection from MDT is always trusted */
+ if (req->rq_auth_usr_mdt) {
+ ost_init_sec_none(reply, exp);
+ RETURN(0);
+ }
+
+ /* no GSS support case */
+ if (!req->rq_auth_gss) {
+ if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
+ CWARN("client %s -> target %s does not user GSS, "
+ "can not run under security level %d.\n",
+ client, obd->obd_name, filter->fo_sec_level);
+ RETURN(-EACCES);
+ } else {
+ ost_init_sec_none(reply, exp);
+ RETURN(0);
+ }
+ }
+
+ /* old version case */
+ if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) ||
+ !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) {
+ if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
+ CWARN("client %s -> target %s uses old version, "
+ "can not run under security level %d.\n",
+ client, obd->obd_name, filter->fo_sec_level);
+ RETURN(-EACCES);
+ } else {
+ CWARN("client %s -> target %s uses old version, "
+ "run under security level %d.\n",
+ client, obd->obd_name, filter->fo_sec_level);
+ ost_init_sec_none(reply, exp);
+ RETURN(0);
+ }
+ }
+
+ remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE;
+ if (remote) {
+ if (!req->rq_auth_remote)
+ CDEBUG(D_SEC, "client (local realm) %s -> target %s "
+ "asked to be remote.\n", client, obd->obd_name);
+ } else if (req->rq_auth_remote) {
+ remote = 1;
+ CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set "
+ "as remote by default.\n", client, obd->obd_name);
+ }
+
+ if (remote) {
+ if (!filter->fo_fl_oss_capa) {
+ CDEBUG(D_SEC, "client %s -> target %s is set as remote,"
+ " but OSS capabilities are not enabled: %d.\n",
+ client, obd->obd_name, filter->fo_fl_oss_capa);
+ RETURN(-EACCES);
+ }
+ }
- CDEBUG(D_SEC, "from %s\n", sptlrpc_part2name(req->rq_sp_from));
- spin_lock(&exp->exp_lock);
- exp->exp_sp_peer = req->rq_sp_from;
-
- read_lock(&filter->fo_sptlrpc_lock);
- sptlrpc_rule_set_choose(&filter->fo_sptlrpc_rset, exp->exp_sp_peer,
- req->rq_peer.nid, &exp->exp_flvr);
- read_unlock(&filter->fo_sptlrpc_lock);
-
- if (exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
- CERROR("invalid rpc flavor %x, expect %x, from %s\n",
- req->rq_flvr.sf_rpc, exp->exp_flvr.sf_rpc,
- libcfs_nid2str(req->rq_peer.nid));
- exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_INVALID;
- rc = -EACCES;
+ switch (filter->fo_sec_level) {
+ case LUSTRE_SEC_NONE:
+ if (!remote) {
+ ost_init_sec_none(reply, exp);
+ break;
+ } else {
+ CDEBUG(D_SEC, "client %s -> target %s is set as remote, "
+ "can not run under security level %d.\n",
+ client, obd->obd_name, filter->fo_sec_level);
+ RETURN(-EACCES);
+ }
+ case LUSTRE_SEC_REMOTE:
+ if (!remote)
+ ost_init_sec_none(reply, exp);
+ break;
+ case LUSTRE_SEC_ALL:
+ if (!remote) {
+ reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
+ OBD_CONNECT_RMT_CLIENT_FORCE);
+ if (!filter->fo_fl_oss_capa)
+ reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA;
+
+ cfs_spin_lock(&exp->exp_lock);
+ exp->exp_connect_flags = reply->ocd_connect_flags;
+ cfs_spin_unlock(&exp->exp_lock);
+ }
+ break;
+ default:
+ RETURN(-EINVAL);
+ }
+
+ RETURN(rc);
+}
+
+/*
+ * FIXME
+ * this should be done in filter_connect()/filter_reconnect(), but
+ * we can't obtain information like NID, which stored in incoming
+ * request, thus can't decide what flavor to use. so we do it here.
+ *
+ * This hack should be removed after the OST stack be rewritten, just
+ * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect().
+ */
+static int ost_connect_check_sptlrpc(struct ptlrpc_request *req)
+{
+ struct obd_export *exp = req->rq_export;
+ struct filter_obd *filter = &exp->exp_obd->u.filter;
+ struct sptlrpc_flavor flvr;
+ int rc = 0;
+
+ if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
+ LUSTRE_ECHO_NAME) == 0)) {
+ exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
+ return 0;
}
- spin_unlock(&exp->exp_lock);
+ if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
+ cfs_read_lock(&filter->fo_sptlrpc_lock);
+ sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset,
+ req->rq_sp_from,
+ req->rq_peer.nid,
+ &flvr);
+ cfs_read_unlock(&filter->fo_sptlrpc_lock);
+
+ cfs_spin_lock(&exp->exp_lock);
+
+ exp->exp_sp_peer = req->rq_sp_from;
+ exp->exp_flvr = flvr;
+
+ if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
+ exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
+ CERROR("unauthorized rpc flavor %x from %s, "
+ "expect %x\n", req->rq_flvr.sf_rpc,
+ libcfs_nid2str(req->rq_peer.nid),
+ exp->exp_flvr.sf_rpc);
+ rc = -EACCES;
+ }
+
+ cfs_spin_unlock(&exp->exp_lock);
+ } else {
+ if (exp->exp_sp_peer != req->rq_sp_from) {
+ CERROR("RPC source %s doesn't match %s\n",
+ sptlrpc_part2name(req->rq_sp_from),
+ sptlrpc_part2name(exp->exp_sp_peer));
+ rc = -EACCES;
+ } else {
+ rc = sptlrpc_target_export_check(exp, 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)
+{
+ __u32 sync_lock_cancel = 0;
+ __u32 len = sizeof(sync_lock_cancel);
+ int rc = 0;
+ ENTRY;
+
+ rc = obd_get_info(NULL, lock->l_export, sizeof(KEY_SYNC_LOCK_CANCEL),
+ KEY_SYNC_LOCK_CANCEL, &len, &sync_lock_cancel, NULL);
+
+ if (!rc && 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)
+ RETURN(-ENOMEM);
+ OBDO_ALLOC(oa);
+ if (!oa) {
+ OBD_FREE_PTR(oinfo);
+ RETURN(-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;
+
+ rc = obd_sync(NULL, 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);
+ 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 *svc = req->rq_rqbd->rqbd_service;
+
+ if (AT_OFF)
+ return obd_timeout / 2;
+
+ return max(at_est2timeout(at_get(&svc->srv_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.
+ */
+static int ost_rw_hpreq_lock_match(struct ptlrpc_request *req,
+ struct ldlm_lock *lock)
+{
+ struct niobuf_remote *nb;
+ struct obd_ioobj *ioo;
+ int mode, opc;
+ struct ldlm_extent ext;
+ ENTRY;
+
+ opc = lustre_msg_get_opc(req->rq_reqmsg);
+ LASSERT(opc == OST_READ || opc == OST_WRITE);
+
+ ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
+ LASSERT(ioo != NULL);
+
+ nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
+ LASSERT(nb != NULL);
+
+ 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_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);
+
+ RETURN(ldlm_extent_overlap(&lock->l_policy_data.l_extent, &ext));
+}
+
+/**
+ * High-priority queue request check for whether the given PTLRPC request (\a
+ * req) is blocking an LDLM lock cancel.
+ *
+ * Returns 1 if the given given PTLRPC request (\a req) is blocking an LDLM lock
+ * cancel, 0 if it is not, and -EFAULT if the request is malformed.
+ *
+ * Only OST_READs, OST_WRITEs and OST_PUNCHes go on the h-p RPC queue. This
+ * function looks only at OST_READs and OST_WRITEs.
+ */
+static int ost_rw_hpreq_check(struct ptlrpc_request *req)
+{
+ struct obd_device *obd = req->rq_export->exp_obd;
+ struct ost_body *body;
+ 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);
+ LASSERT(body != NULL);
+
+ ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
+ LASSERT(ioo != NULL);
+
+ nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
+ 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;
+ 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);
+
+ 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);
+
+ ost_prolong_locks(&opd);
+
+ 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);
+}
+
+/**
+ * Like ost_rw_hpreq_lock_match(), but for OST_PUNCH RPCs.
+ */
+static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req,
+ struct ldlm_lock *lock)
+{
+ struct ost_body *body;
+ ENTRY;
+
+ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
+ 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);
+}
+
+/**
+ * Like ost_rw_hpreq_check(), but for OST_PUNCH RPCs.
+ */
+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);
+ 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);
+
+ CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
+ obd->obd_name, opd.opd_locks, req);
+
+ 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_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_fini = ost_punch_hpreq_fini
+};
+
+/** Assign high priority operations to the request if needed. */
+static int ost_hpreq_handler(struct ptlrpc_request *req)
+{
+ ENTRY;
+ if (req->rq_export) {
+ int opc = lustre_msg_get_opc(req->rq_reqmsg);
+ struct ost_body *body;
+
+ if (opc == OST_READ || opc == OST_WRITE) {
+ 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
+ * ost_handler() initializes their pills, so we
+ * initialize that here. Capsule initialization is
+ * idempotent, as is setting the pill's format (provided
+ * it doesn't change).
+ */
+ req_capsule_init(&req->rq_pill, req, RCL_SERVER);
+ if (opc == OST_READ)
+ req_capsule_set(&req->rq_pill,
+ &RQF_OST_BRW_READ);
+ else
+ req_capsule_set(&req->rq_pill,
+ &RQF_OST_BRW_WRITE);
+
+ body = req_capsule_client_get(&req->rq_pill,
+ &RMF_OST_BODY);
+ if (body == NULL) {
+ CERROR("Missing/short ost_body\n");
+ RETURN(-EFAULT);
+ }
+
+ objcount = req_capsule_get_size(&req->rq_pill,
+ &RMF_OBD_IOOBJ,
+ RCL_CLIENT) /
+ sizeof(*ioo);
+ if (objcount == 0) {
+ CERROR("Missing/short ioobj\n");
+ RETURN(-EFAULT);
+ }
+ if (objcount > 1) {
+ CERROR("too many ioobjs (%d)\n", objcount);
+ RETURN(-EFAULT);
+ }
+
+ ioo = req_capsule_client_get(&req->rq_pill,
+ &RMF_OBD_IOOBJ);
+ if (ioo == NULL) {
+ CERROR("Missing/short ioobj\n");
+ 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);
+ RETURN(-EFAULT);
+ }
+ niocount += ioo[i].ioo_bufcnt;
+ }
+ if (niocount > PTLRPC_MAX_BRW_PAGES) {
+ DEBUG_REQ(D_RPCTRACE, req,
+ "bulk has too many pages (%d)",
+ niocount);
+ RETURN(-EFAULT);
+ }
+
+ nb = req_capsule_client_get(&req->rq_pill,
+ &RMF_NIOBUF_REMOTE);
+ if (nb == NULL) {
+ CERROR("Missing/short niobuf\n");
+ RETURN(-EFAULT);
+ }
+
+ if (niocount == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
+ req->rq_ops = &ost_hpreq_rw;
+ } else if (opc == OST_PUNCH) {
+ req_capsule_init(&req->rq_pill, req, RCL_SERVER);
+ req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
+
+ body = req_capsule_client_get(&req->rq_pill,
+ &RMF_OST_BODY);
+ if (body == NULL) {
+ CERROR("Missing/short ost_body\n");
+ RETURN(-EFAULT);
+ }
+
+ if (!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
+ !(body->oa.o_flags & OBD_FL_SRVLOCK))
+ req->rq_ops = &ost_hpreq_punch;
+ }
+ }
+ RETURN(0);
+}
+
/* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
int ost_handle(struct ptlrpc_request *req)
{
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 */
req_capsule_init(&req->rq_pill, req, RCL_SERVER);
- /* XXX identical to MDS */
if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) {
- int recovering;
-
- if (req->rq_export == NULL) {
+ if (!class_connected_export(req->rq_export)) {
CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
lustre_msg_get_opc(req->rq_reqmsg),
libcfs_id2str(req->rq_peer));
obd = req->rq_export->exp_obd;
/* Check for aborted recovery. */
- spin_lock_bh(&obd->obd_processing_task_lock);
- recovering = obd->obd_recovering;
- spin_unlock_bh(&obd->obd_processing_task_lock);
- if (recovering) {
+ if (obd->obd_recovering) {
rc = ost_filter_recovery_request(req, obd,
&should_process);
if (rc || !should_process)
}
oti_init(oti, req);
-
+
rc = ost_msg_check_version(req->rq_reqmsg);
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");
if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2))
RETURN(0);
if (!rc) {
- struct obd_export *exp = req->rq_export;
-
- obd = exp->exp_obd;
-
- rc = filter_export_check_flavor(&obd->u.filter,
- exp, req);
+ rc = ost_init_sec_level(req);
+ if (!rc)
+ rc = ost_connect_check_sptlrpc(req);
}
break;
}
break;
case OST_CREATE:
CDEBUG(D_INODE, "create\n");
+ req_capsule_set(&req->rq_pill, &RQF_OST_CREATE);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET))
RETURN(0);
- if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
- GOTO(out, rc = -ENOSPC);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
GOTO(out, rc = -EROFS);
rc = ost_create(req->rq_export, req, oti);
break;
case OST_DESTROY:
CDEBUG(D_INODE, "destroy\n");
+ req_capsule_set(&req->rq_pill, &RQF_OST_DESTROY);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET))
RETURN(0);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
break;
case OST_GETATTR:
CDEBUG(D_INODE, "getattr\n");
+ req_capsule_set(&req->rq_pill, &RQF_OST_GETATTR);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET))
RETURN(0);
rc = ost_getattr(req->rq_export, req);
break;
case OST_SETATTR:
CDEBUG(D_INODE, "setattr\n");
+ req_capsule_set(&req->rq_pill, &RQF_OST_SETATTR);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET))
RETURN(0);
rc = ost_setattr(req->rq_export, req, oti);
break;
case OST_WRITE:
+ 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){
/* ost_brw_write sends its own replies */
RETURN(rc);
case OST_READ:
+ 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){
RETURN(rc);
case OST_PUNCH:
CDEBUG(D_INODE, "punch\n");
+ req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET))
RETURN(0);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
break;
case OST_STATFS:
CDEBUG(D_INODE, "statfs\n");
+ req_capsule_set(&req->rq_pill, &RQF_OST_STATFS);
if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET))
RETURN(0);
rc = ost_statfs(req);
break;
case OST_SYNC:
CDEBUG(D_INODE, "sync\n");
+ 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_SET_INFO);
+ req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
rc = ost_set_info(req->rq_export, req);
break;
case OST_GET_INFO:
DEBUG_REQ(D_INODE, req, "get_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_OBD_LOG_CANCEL_NET))
RETURN(0);
rc = llog_origin_handle_cancel(req);
+ if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
+ RETURN(0);
req->rq_status = rc;
rc = req_capsule_server_pack(&req->rq_pill);
if (rc)
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;
*/
static void ost_thread_done(struct ptlrpc_thread *thread)
{
- int i;
struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local
* Storage */
*/
tls = thread->t_data;
if (tls != NULL) {
- for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) {
- if (tls->page[i] != NULL)
- OBD_PAGE_FREE(tls->page[i]);
- }
OBD_FREE_PTR(tls);
thread->t_data = NULL;
}
*/
static int ost_thread_init(struct ptlrpc_thread *thread)
{
- int result;
- int i;
struct ost_thread_local_cache *tls;
ENTRY;
LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
OBD_ALLOC_PTR(tls);
- if (tls != NULL) {
- result = 0;
- thread->t_data = tls;
- /*
- * populate pool
- */
- for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) {
- OBD_PAGE_ALLOC(tls->page[i], OST_THREAD_POOL_GFP);
- if (tls->page[i] == NULL) {
- ost_thread_done(thread);
- result = -ENOMEM;
- break;
- }
- }
- } else
- result = -ENOMEM;
- RETURN(result);
+ if (tls == NULL)
+ RETURN(-ENOMEM);
+ thread->t_data = tls;
+ RETURN(0);
}
#define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
int rc;
ENTRY;
- rc = cleanup_group_info();
+ rc = cfs_cleanup_group_info();
if (rc)
RETURN(rc);
lprocfs_ost_init_vars(&lvars);
lprocfs_obd_setup(obd, lvars.obd_vars);
- sema_init(&ost->ost_health_sem, 1);
+ cfs_mutex_init(&ost->ost_health_mutex);
if (oss_num_threads) {
/* If oss_num_threads is set, it is the min and the max. */
- if (oss_num_threads > OSS_THREADS_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 = num_possible_cpus() * num_physpages >>
- (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);
- }
+ } else {
+ /* Base min threads on memory and cpus */
+ oss_min_threads =
+ cfs_num_online_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_handle, LUSTRE_OSS_NAME,
obd->obd_proc_entry, target_print_req,
oss_min_threads, oss_max_threads,
- "ll_ost", LCT_DT_THREAD);
+ "ll_ost", LCT_DT_THREAD, NULL);
if (ost->ost_service == NULL) {
CERROR("failed to start service\n");
GOTO(out_lprocfs, rc = -ENOMEM);
}
- rc = ptlrpc_start_threads(obd, ost->ost_service);
+ rc = ptlrpc_start_threads(ost->ost_service);
if (rc)
GOTO(out_service, rc = -EINVAL);
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);
+ "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);
}
- rc = ptlrpc_start_threads(obd, ost->ost_create_service);
+ rc = ptlrpc_start_threads(ost->ost_create_service);
if (rc)
GOTO(out_create, rc = -EINVAL);
ost_handle, "ost_io",
obd->obd_proc_entry, target_print_req,
oss_min_threads, oss_max_threads,
- "ll_ost_io", LCT_DT_THREAD);
+ "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);
+ rc = ptlrpc_start_threads(ost->ost_io_service);
if (rc)
GOTO(out_io, rc = -EINVAL);
ping_evictor_stop();
- spin_lock_bh(&obd->obd_processing_task_lock);
- if (obd->obd_recovering) {
- target_cancel_recovery_timer(obd);
- obd->obd_recovering = 0;
- }
- spin_unlock_bh(&obd->obd_processing_task_lock);
-
- 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;
- up(&ost->ost_health_sem);
+ cfs_mutex_unlock(&ost->ost_health_mutex);
lprocfs_obd_cleanup(obd);
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;
- 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);
- up(&ost->ost_health_sem);
+ cfs_mutex_unlock(&ost->ost_health_mutex);
/*
* health_check to return 0 on healthy