X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fofd%2Fofd_dev.c;h=c2cda399c16be858884403393fd6f5e73a7acbd5;hb=6115eb7fd5;hp=b66f850d5cb9c44e27e5b4901c074aeea74ac144;hpb=bafa12c06d87c05bb263eed37ad8af1a2df99894;p=fs%2Flustre-release.git diff --git a/lustre/ofd/ofd_dev.c b/lustre/ofd/ofd_dev.c index b66f850..c2cda39 100644 --- a/lustre/ofd/ofd_dev.c +++ b/lustre/ofd/ofd_dev.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * 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 - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,23 +23,26 @@ * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2013, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ * Lustre is a trademark of Sun Microsystems, Inc. * - * lustre/ofd/ofd.c + * lustre/ofd/ofd_dev.c + * + * This file contains OSD API methods for OBD Filter Device (OFD), + * request handlers and supplemental functions to set OFD up and clean it up. * - * Author: Alex Zhuravlev - * Author: Mike Pershin - * Author: Johann Lombardi + * Author: Alex Zhuravlev + * Author: Mike Pershin + * Author: Johann Lombardi */ /* * The OBD Filter Device (OFD) module belongs to the Object Storage * Server stack and connects the RPC oriented Unified Target (TGT) * layer (see lustre/include/lu_target.h) to the storage oriented OSD - * layer (see lustre/doc/osd-api.txt). + * layer (see Documentation/osd-api.txt). * * TGT * | DT and OBD APIs @@ -70,12 +69,14 @@ #define DEBUG_SUBSYSTEM S_FILTER #include -#include +#include +#include #include #include -#include #include #include +#include +#include #include "ofd_internal.h" @@ -93,6 +94,20 @@ static struct lu_kmem_descr ofd_caches[] = { } }; +/** + * Connect OFD to the next device in the stack. + * + * This function is used for device stack configuration and links OFD + * device with bottom OSD device. + * + * \param[in] env execution environment + * \param[in] m OFD device + * \param[in] next name of next device in the stack + * \param[out] exp export to return + * + * \retval 0 and export in \a exp if successful + * \retval negative value on error + */ static int ofd_connect_to_next(const struct lu_env *env, struct ofd_device *m, const char *next, struct obd_export **exp) { @@ -134,6 +149,18 @@ out: RETURN(rc); } +/** + * Initialize stack of devices. + * + * This function initializes OFD-OSD device stack to serve OST requests + * + * \param[in] env execution environment + * \param[in] m OFD device + * \param[in] cfg Lustre config for this server + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_stack_init(const struct lu_env *env, struct ofd_device *m, struct lustre_cfg *cfg) { @@ -141,6 +168,7 @@ static int ofd_stack_init(const struct lu_env *env, struct lu_device *d; struct ofd_thread_info *info = ofd_info(env); struct lustre_mount_info *lmi; + struct lustre_mount_data *lmd; int rc; char *osdname; @@ -152,6 +180,10 @@ static int ofd_stack_init(const struct lu_env *env, RETURN(-ENODEV); } + lmd = s2lsi(lmi->lmi_sb)->lsi_lmd; + if (lmd != NULL && lmd->lmd_flags & LMD_FLG_SKIP_LFSCK) + m->ofd_skip_lfsck = 1; + /* find bottom osd */ OBD_ALLOC(osdname, MTI_NAME_MAXLEN); if (osdname == NULL) @@ -173,6 +205,19 @@ static int ofd_stack_init(const struct lu_env *env, RETURN(rc); } +/** + * Finalize the device stack OFD-OSD. + * + * This function cleans OFD-OSD device stack and + * disconnects OFD from the OSD. + * + * \param[in] env execution environment + * \param[in] m OFD device + * \param[in] top top device of stack + * + * \retval 0 if successful + * \retval negative value on error + */ static void ofd_stack_fini(const struct lu_env *env, struct ofd_device *m, struct lu_device *top) { @@ -191,19 +236,23 @@ static void ofd_stack_fini(const struct lu_env *env, struct ofd_device *m, if (obd->obd_fail) strcat(flags, "A"); lustre_cfg_bufs_set_string(&bufs, 1, flags); - lcfg = lustre_cfg_new(LCFG_CLEANUP, &bufs); - if (!lcfg) { - CERROR("Cannot alloc lcfg!\n"); + OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount, bufs.lcfg_buflen)); + if (!lcfg) RETURN_EXIT; - } + lustre_cfg_init(lcfg, LCFG_CLEANUP, &bufs); LASSERT(top); top->ld_ops->ldo_process_config(env, top, lcfg); - lustre_cfg_free(lcfg); + OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, lcfg->lcfg_buflens)); + + if (m->ofd_los != NULL) { + local_oid_storage_fini(env, m->ofd_los); + m->ofd_los = NULL; + } lu_site_purge(env, top->ld_site, ~0); if (!cfs_hash_is_empty(top->ld_site->ls_obj_hash)) { - LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL); + LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_OTHER, NULL); lu_site_print(env, top->ld_site, &msgdata, lu_cdebug_printer); } @@ -213,22 +262,54 @@ static void ofd_stack_fini(const struct lu_env *env, struct ofd_device *m, EXIT; } +static void ofd_stack_pre_fini(const struct lu_env *env, struct ofd_device *m, + struct lu_device *top) +{ + struct lustre_cfg_bufs bufs; + struct lustre_cfg *lcfg; + ENTRY; + + LASSERT(top); + + lustre_cfg_bufs_reset(&bufs, ofd_name(m)); + lustre_cfg_bufs_set_string(&bufs, 1, NULL); + OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount, bufs.lcfg_buflen)); + if (!lcfg) { + CERROR("%s: failed to trigger LCFG_PRE_CLEANUP\n", ofd_name(m)); + } else { + lustre_cfg_init(lcfg, LCFG_PRE_CLEANUP, &bufs); + top->ld_ops->ldo_process_config(env, top, lcfg); + OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount, + lcfg->lcfg_buflens)); + } + + EXIT; +} + /* For interoperability, see mdt_interop_param[]. */ static struct cfg_interop_param ofd_interop_param[] = { { "ost.quota_type", NULL }, { NULL } }; -/* Some parameters were moved from ofd to osd and only their +/** + * Check if parameters are symlinks to the OSD. + * + * Some parameters were moved from ofd to osd and only their * symlinks were kept in ofd by LU-3106. They are: * -writehthrough_cache_enable - * -readcache_max_filese + * -readcache_max_filesize * -read_cache_enable * -brw_stats - * Since they are not included by the static lprocfs var list, - * a pre-check is added for them to avoid "unknown param" error - * message confuses the customer. If they are matched in this - * check, they will be passed to the osd directly. + * + * Since they are not included by the static lprocfs var list, a pre-check + * is added for them to avoid "unknown param" errors. If they are matched + * in this check, they will be passed to the OSD directly. + * + * \param[in] param parameters to check + * + * \retval true if param is symlink to OSD param + * false otherwise */ static bool match_symlink_param(char *param) { @@ -253,7 +334,19 @@ static bool match_symlink_param(char *param) return false; } -/* used by MGS to process specific configurations */ +/** + * Process various configuration parameters. + * + * This function is used by MGS to process specific configurations and + * pass them through to the next device in server stack, i.e. the OSD. + * + * \param[in] env execution environment + * \param[in] d LU device of OFD + * \param[in] cfg parameters to process + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_process_config(const struct lu_env *env, struct lu_device *d, struct lustre_cfg *cfg) { @@ -326,6 +419,18 @@ static int ofd_process_config(const struct lu_env *env, struct lu_device *d, RETURN(rc); } +/** + * Implementation of lu_object_operations::loo_object_init for OFD + * + * Allocate just the next object (OSD) in stack. + * + * \param[in] env execution environment + * \param[in] o lu_object of OFD object + * \param[in] conf additional configuration parameters, not used here + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_object_init(const struct lu_env *env, struct lu_object *o, const struct lu_object_conf *conf) { @@ -349,6 +454,14 @@ static int ofd_object_init(const struct lu_env *env, struct lu_object *o, RETURN(rc); } +/** + * Implementation of lu_object_operations::loo_object_free. + * + * Finish OFD object lifecycle and free its memory. + * + * \param[in] env execution environment + * \param[in] o LU object of OFD object + */ static void ofd_object_free(const struct lu_env *env, struct lu_object *o) { struct ofd_object *of = ofd_obj(o); @@ -366,18 +479,45 @@ static void ofd_object_free(const struct lu_env *env, struct lu_object *o) EXIT; } +/** + * Implementation of lu_object_operations::loo_object_print. + * + * Print OFD part of compound OFD-OSD object. See lu_object_print() and + * LU_OBJECT_DEBUG() for more details about the compound object printing. + * + * \param[in] env execution environment + * \param[in] cookie opaque data passed to the printer function + * \param[in] p printer function to use + * \param[in] o LU object of OFD object + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_object_print(const struct lu_env *env, void *cookie, lu_printer_t p, const struct lu_object *o) { return (*p)(env, cookie, LUSTRE_OST_NAME"-object@%p", o); } -struct lu_object_operations ofd_obj_ops = { +static struct lu_object_operations ofd_obj_ops = { .loo_object_init = ofd_object_init, .loo_object_free = ofd_object_free, .loo_object_print = ofd_object_print }; +/** + * Implementation of lu_device_operations::lod_object_alloc. + * + * This function allocates OFD part of compound OFD-OSD object and + * initializes its header, because OFD is the top device in stack + * + * \param[in] env execution environment + * \param[in] hdr object header, NULL for OFD + * \param[in] d lu_device + * + * \retval allocated object if successful + * \retval NULL value on failed allocation + */ static struct lu_object *ofd_object_alloc(const struct lu_env *env, const struct lu_object_header *hdr, struct lu_device *d) @@ -403,8 +543,19 @@ static struct lu_object *ofd_object_alloc(const struct lu_env *env, } } -extern int ost_handle(struct ptlrpc_request *req); - +/** + * Return the result of LFSCK run to the OFD. + * + * Notify OFD about result of LFSCK run. That may block the new object + * creation until problem is fixed by LFSCK. + * + * \param[in] env execution environment + * \param[in] data pointer to the OFD device + * \param[in] event LFSCK event type + * + * \retval 0 if successful + * \retval negative value on unknown event + */ static int ofd_lfsck_out_notify(const struct lu_env *env, void *data, enum lfsck_events event) { @@ -439,6 +590,20 @@ static int ofd_lfsck_out_notify(const struct lu_env *env, void *data, return 0; } +/** + * Implementation of lu_device_operations::ldo_prepare. + * + * This method is called after layer has been initialized and before it starts + * serving user requests. In OFD it starts lfsk check routines and initializes + * recovery. + * + * \param[in] env execution environment + * \param[in] pdev higher device in stack, NULL for OFD + * \param[in] dev lu_device of OFD device + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_prepare(const struct lu_env *env, struct lu_device *pdev, struct lu_device *dev) { @@ -484,27 +649,49 @@ static int ofd_prepare(const struct lu_env *env, struct lu_device *pdev, RETURN(rc); } +/** + * Implementation of lu_device_operations::ldo_recovery_complete. + * + * This method notifies all layers about 'recovery complete' event. That means + * device is in full state and consistent. An OFD calculates available grant + * space upon this event. + * + * \param[in] env execution environment + * \param[in] dev lu_device of OFD device + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_recovery_complete(const struct lu_env *env, struct lu_device *dev) { + struct ofd_thread_info *oti = ofd_info(env); struct ofd_device *ofd = ofd_dev(dev); struct lu_device *next = &ofd->ofd_osd->dd_lu_dev; - int rc = 0, max_precreate; + int rc = 0; ENTRY; - /* Grant space for object precreation on the self export. - * This initial reserved space (i.e. 10MB for zfs and 280KB for ldiskfs) + /* + * Grant space for object precreation on the self export. + * The initial reserved space (i.e. 10MB for zfs and 280KB for ldiskfs) * is enough to create 10k objects. More space is then acquired for - * precreation in ofd_grant_create(). + * precreation in tgt_grant_create(). */ - max_precreate = OST_MAX_PRECREATE * ofd->ofd_dt_conf.ddp_inodespace / 2; - ofd_grant_connect(env, dev->ld_obd->obd_self_export, max_precreate, - false); + memset(&oti->fti_ocd, 0, sizeof(oti->fti_ocd)); + oti->fti_ocd.ocd_grant = OST_MAX_PRECREATE / 2; + oti->fti_ocd.ocd_grant *= ofd->ofd_lut.lut_dt_conf.ddp_inodespace; + oti->fti_ocd.ocd_connect_flags = OBD_CONNECT_GRANT | + OBD_CONNECT_GRANT_PARAM; + tgt_grant_connect(env, dev->ld_obd->obd_self_export, &oti->fti_ocd, + true); rc = next->ld_ops->ldo_recovery_complete(env, next); RETURN(rc); } +/** + * lu_device_operations matrix for OFD device. + */ static struct lu_device_operations ofd_lu_ops = { .ldo_object_alloc = ofd_object_alloc, .ldo_process_config = ofd_process_config, @@ -514,10 +701,18 @@ static struct lu_device_operations ofd_lu_ops = { LPROC_SEQ_FOPS(lprocfs_nid_stats_clear); +/** + * Initialize all needed procfs entries for OFD device. + * + * \param[in] ofd OFD device + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_procfs_init(struct ofd_device *ofd) { struct obd_device *obd = ofd_obd(ofd); - cfs_proc_dir_entry_t *entry; + struct proc_dir_entry *entry; int rc = 0; ENTRY; @@ -525,7 +720,7 @@ static int ofd_procfs_init(struct ofd_device *ofd) /* lprocfs must be setup before the ofd so state can be safely added * to /proc incrementally as the ofd is setup */ obd->obd_vars = lprocfs_ofd_obd_vars; - rc = lprocfs_obd_setup(obd); + rc = lprocfs_obd_setup(obd, false); if (rc) { CERROR("%s: lprocfs_obd_setup failed: %d.\n", obd->obd_name, rc); @@ -541,8 +736,7 @@ static int ofd_procfs_init(struct ofd_device *ofd) obd->obd_uses_nid_stats = 1; - entry = lprocfs_seq_register("exports", obd->obd_proc_entry, NULL, - NULL); + entry = lprocfs_register("exports", obd->obd_proc_entry, NULL, NULL); if (IS_ERR(entry)) { rc = PTR_ERR(entry); CERROR("%s: error %d setting up lprocfs for %s\n", @@ -552,9 +746,6 @@ static int ofd_procfs_init(struct ofd_device *ofd) obd->obd_proc_exports_entry = entry; entry = lprocfs_add_simple(obd->obd_proc_exports_entry, "clear", -#ifndef HAVE_ONLY_PROCFS_SEQ - NULL, NULL, -#endif obd, &lprocfs_nid_stats_clear_fops); if (IS_ERR(entry)) { rc = PTR_ERR(entry); @@ -578,13 +769,15 @@ obd_cleanup: } /** - * ofd_procfs_add_brw_stats_symlink - expose osd stats to ofd layer + * Expose OSD statistics to OFD layer. * * The osd interfaces to the backend file system exposes useful data * such as brw_stats and read or write cache states. This same data * needs to be exposed into the obdfilter (ofd) layer to maintain * backwards compatibility. This function creates the symlinks in the * proc layer to enable this. + * + * \param[in] ofd OFD device */ static void ofd_procfs_add_brw_stats_symlink(struct ofd_device *ofd) { @@ -613,6 +806,11 @@ static void ofd_procfs_add_brw_stats_symlink(struct ofd_device *ofd) osd_obd->obd_type->typ_name, obd->obd_name); } +/** + * Cleanup all procfs entries in OFD. + * + * \param[in] ofd OFD device + */ static void ofd_procfs_fini(struct ofd_device *ofd) { struct obd_device *obd = ofd_obd(ofd); @@ -623,13 +821,37 @@ static void ofd_procfs_fini(struct ofd_device *ofd) lprocfs_job_stats_fini(obd); } -extern int ost_handle(struct ptlrpc_request *req); - +/** + * Stop SEQ/FID server on OFD. + * + * \param[in] env execution environment + * \param[in] ofd OFD device + * + * \retval 0 if successful + * \retval negative value on error + */ int ofd_fid_fini(const struct lu_env *env, struct ofd_device *ofd) { return seq_site_fini(env, &ofd->ofd_seq_site); } +/** + * Start SEQ/FID server on OFD. + * + * The SEQ/FID server on OFD is needed to allocate FIDs for new objects. + * It also connects to the master server to get own FID sequence (SEQ) range + * to this particular OFD. Typically that happens when the OST is first + * formatted or in the rare case that it exhausts the local sequence range. + * + * The sequence range is allocated out to the MDTs for OST object allocations, + * and not directly to the clients. + * + * \param[in] env execution environment + * \param[in] ofd OFD device + * + * \retval 0 if successful + * \retval negative value on error + */ int ofd_fid_init(const struct lu_env *env, struct ofd_device *ofd) { struct seq_server_site *ss = &ofd->ofd_seq_site; @@ -643,72 +865,73 @@ int ofd_fid_init(const struct lu_env *env, struct ofd_device *ofd) ss->ss_lu = lu->ld_site; ss->ss_node_id = ofd->ofd_lut.lut_lsd.lsd_osd_index; + OBD_ALLOC(name, sizeof(obd_name) * 2 + 10); + if (name == NULL) + return -ENOMEM; + OBD_ALLOC_PTR(ss->ss_server_seq); if (ss->ss_server_seq == NULL) - GOTO(out_free, rc = -ENOMEM); - - OBD_ALLOC(name, strlen(obd_name) + 10); - if (!name) { - OBD_FREE_PTR(ss->ss_server_seq); - ss->ss_server_seq = NULL; - GOTO(out_free, rc = -ENOMEM); - } + GOTO(out_name, rc = -ENOMEM); rc = seq_server_init(env, ss->ss_server_seq, ofd->ofd_osd, obd_name, LUSTRE_SEQ_SERVER, ss); if (rc) { CERROR("%s : seq server init error %d\n", obd_name, rc); - GOTO(out_free, rc); + GOTO(out_server, rc); } ss->ss_server_seq->lss_space.lsr_index = ss->ss_node_id; OBD_ALLOC_PTR(ss->ss_client_seq); if (ss->ss_client_seq == NULL) - GOTO(out_free, rc = -ENOMEM); + GOTO(out_server, rc = -ENOMEM); - snprintf(name, strlen(obd_name) + 6, "%p-super", obd_name); + /* + * It always printed as "%p", so that the name is unique in the kernel, + * even if the filesystem is mounted twice. So sizeof(.) * 2 is enough. + */ + snprintf(name, sizeof(obd_name) * 2 + 7, "%p-super", obd_name); rc = seq_client_init(ss->ss_client_seq, NULL, LUSTRE_SEQ_DATA, name, NULL); if (rc) { CERROR("%s : seq client init error %d\n", obd_name, rc); - GOTO(out_free, rc); + GOTO(out_client, rc); } - OBD_FREE(name, strlen(obd_name) + 10); - name = NULL; rc = seq_server_set_cli(env, ss->ss_server_seq, ss->ss_client_seq); -out_free: if (rc) { - if (ss->ss_server_seq) { - seq_server_fini(ss->ss_server_seq, env); - OBD_FREE_PTR(ss->ss_server_seq); - ss->ss_server_seq = NULL; - } - - if (ss->ss_client_seq) { - seq_client_fini(ss->ss_client_seq); - OBD_FREE_PTR(ss->ss_client_seq); - ss->ss_client_seq = NULL; - } - - if (name) { - OBD_FREE(name, strlen(obd_name) + 10); - name = NULL; - } +out_client: + seq_client_fini(ss->ss_client_seq); + OBD_FREE_PTR(ss->ss_client_seq); + ss->ss_client_seq = NULL; +out_server: + seq_server_fini(ss->ss_server_seq, env); + OBD_FREE_PTR(ss->ss_server_seq); + ss->ss_server_seq = NULL; } +out_name: + OBD_FREE(name, sizeof(obd_name) * 2 + 10); return rc; } -int ofd_set_info_hdl(struct tgt_session_info *tsi) +/** + * OFD request handler for OST_SET_INFO RPC. + * + * This is OFD-specific part of request handling + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ +static int ofd_set_info_hdl(struct tgt_session_info *tsi) { struct ptlrpc_request *req = tgt_ses_req(tsi); struct ost_body *body = NULL, *repbody; void *key, *val = NULL; int keylen, vallen, rc = 0; bool is_grant_shrink; - struct ofd_device *ofd = ofd_exp(tsi->tsi_exp); ENTRY; @@ -745,16 +968,12 @@ int ofd_set_info_hdl(struct tgt_session_info *tsi) *repbody = *body; /** handle grant shrink, similar to a read request */ - ofd_grant_prepare_read(tsi->tsi_env, tsi->tsi_exp, + tgt_grant_prepare_read(tsi->tsi_env, tsi->tsi_exp, &repbody->oa); } else if (KEY_IS(KEY_EVICT_BY_NID)) { if (vallen > 0) obd_export_evict_by_nid(tsi->tsi_exp->exp_obd, val); rc = 0; - } else if (KEY_IS(KEY_CAPA_KEY)) { - rc = ofd_update_capa_key(ofd, val); - } else if (KEY_IS(KEY_SPTLRPC_CONF)) { - rc = tgt_adapt_sptlrpc_conf(tsi->tsi_tgt, 0); } else { CERROR("%s: Unsupported key %s\n", tgt_name(tsi->tsi_tgt), (char *)key); @@ -766,8 +985,22 @@ int ofd_set_info_hdl(struct tgt_session_info *tsi) RETURN(rc); } +/** + * Get FIEMAP (FIle Extent MAPping) for object with the given FID. + * + * This function returns a list of extents which describes how a file's + * blocks are laid out on the disk. + * + * \param[in] env execution environment + * \param[in] ofd OFD device + * \param[in] fid FID of object + * \param[in] fiemap fiemap structure to fill with data + * + * \retval 0 if \a fiemap is filled with data successfully + * \retval negative value on error + */ int ofd_fiemap_get(const struct lu_env *env, struct ofd_device *ofd, - struct lu_fid *fid, struct ll_user_fiemap *fiemap) + struct lu_fid *fid, struct fiemap *fiemap) { struct ofd_object *fo; int rc; @@ -789,45 +1022,54 @@ int ofd_fiemap_get(const struct lu_env *env, struct ofd_device *ofd, return rc; } -struct locked_region { - struct list_head list; - struct lustre_handle lh; -}; -static int lock_region(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, - unsigned long long begin, unsigned long long end, - struct list_head *locked) +static int ofd_lock_unlock_region(struct ldlm_namespace *ns, + struct ldlm_res_id *res_id, + unsigned long long begin, + unsigned long long end) { - struct locked_region *region = NULL; __u64 flags = 0; int rc; + struct lustre_handle lh = { 0 }; LASSERT(begin <= end); - OBD_ALLOC_PTR(region); - if (region == NULL) - return -ENOMEM; - rc = tgt_extent_lock(ns, res_id, begin, end, ®ion->lh, - LCK_PR, &flags); + rc = tgt_extent_lock(ns, res_id, begin, end, &lh, LCK_PR, &flags); if (rc != 0) return rc; - CDEBUG(D_OTHER, "ost lock [%llu,%llu], lh=%p\n", begin, end, - ®ion->lh); - list_add(®ion->list, locked); + CDEBUG(D_OTHER, "ost lock [%llu,%llu], lh=%p\n", begin, end, &lh); + tgt_extent_unlock(&lh, LCK_PR); return 0; } +/** + * Lock the sparse areas of given resource. + * + * The locking of sparse areas will cause dirty data to be flushed back from + * clients. This is used when getting the FIEMAP of an object to make sure + * there is no unaccounted cached data on clients. + * + * This function goes through \a fiemap list of extents and locks only sparse + * areas between extents. + * + * \param[in] ns LDLM namespace + * \param[in] res_id resource ID + * \param[in] fiemap file extents mapping on disk + * \param[in] locked list head of regions list + * + * \retval 0 if successful + * \retval negative value on error + */ static int lock_zero_regions(struct ldlm_namespace *ns, struct ldlm_res_id *res_id, - struct ll_user_fiemap *fiemap, - struct list_head *locked) + struct fiemap *fiemap) { __u64 begin = fiemap->fm_start; unsigned int i; int rc = 0; - struct ll_fiemap_extent *fiemap_start = fiemap->fm_extents; + struct fiemap_extent *fiemap_start = fiemap->fm_extents; ENTRY; @@ -836,8 +1078,8 @@ static int lock_zero_regions(struct ldlm_namespace *ns, if (fiemap_start[i].fe_logical > begin) { CDEBUG(D_OTHER, "ost lock [%llu,%llu]\n", begin, fiemap_start[i].fe_logical); - rc = lock_region(ns, res_id, begin, - fiemap_start[i].fe_logical, locked); + rc = ofd_lock_unlock_region(ns, res_id, begin, + fiemap_start[i].fe_logical); if (rc) RETURN(rc); } @@ -848,27 +1090,33 @@ static int lock_zero_regions(struct ldlm_namespace *ns, if (begin < (fiemap->fm_start + fiemap->fm_length)) { CDEBUG(D_OTHER, "ost lock [%llu,%llu]\n", begin, fiemap->fm_start + fiemap->fm_length); - rc = lock_region(ns, res_id, begin, - fiemap->fm_start + fiemap->fm_length, locked); + rc = ofd_lock_unlock_region(ns, res_id, begin, + fiemap->fm_start + fiemap->fm_length); } RETURN(rc); } -static void -unlock_zero_regions(struct ldlm_namespace *ns, struct list_head *locked) -{ - struct locked_region *entry, *temp; - list_for_each_entry_safe(entry, temp, locked, list) { - CDEBUG(D_OTHER, "ost unlock lh=%p\n", &entry->lh); - tgt_extent_unlock(&entry->lh, LCK_PR); - list_del(&entry->list); - OBD_FREE_PTR(entry); - } -} - -int ofd_get_info_hdl(struct tgt_session_info *tsi) +/** + * OFD request handler for OST_GET_INFO RPC. + * + * This is OFD-specific part of request handling. The OFD-specific keys are: + * - KEY_LAST_ID (obsolete) + * - KEY_FIEMAP + * - KEY_LAST_FID + * + * This function reads needed data from storage and fills reply with it. + * + * Note: the KEY_LAST_ID is obsolete, replaced by KEY_LAST_FID on newer MDTs, + * and is kept for compatibility. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ +static int ofd_get_info_hdl(struct tgt_session_info *tsi) { struct obd_export *exp = tsi->tsi_exp; struct ofd_device *ofd = ofd_exp(exp); @@ -889,7 +1137,7 @@ int ofd_get_info_hdl(struct tgt_session_info *tsi) RCL_CLIENT); if (KEY_IS(KEY_LAST_ID)) { - obd_id *last_id; + u64 *last_id; struct ofd_seq *oseq; req_capsule_extend(tsi->tsi_pill, &RQF_OST_GET_INFO_LAST_ID); @@ -900,7 +1148,7 @@ int ofd_get_info_hdl(struct tgt_session_info *tsi) last_id = req_capsule_server_get(tsi->tsi_pill, &RMF_OBD_ID); oseq = ofd_seq_load(tsi->tsi_env, ofd, - (obd_seq)exp->exp_filter_data.fed_group); + (u64)exp->exp_filter_data.fed_group); if (IS_ERR(oseq)) rc = -EFAULT; else @@ -908,21 +1156,22 @@ int ofd_get_info_hdl(struct tgt_session_info *tsi) ofd_seq_put(tsi->tsi_env, oseq); } else if (KEY_IS(KEY_FIEMAP)) { struct ll_fiemap_info_key *fm_key; - struct ll_user_fiemap *fiemap; + struct fiemap *fiemap; struct lu_fid *fid; req_capsule_extend(tsi->tsi_pill, &RQF_OST_GET_INFO_FIEMAP); fm_key = req_capsule_client_get(tsi->tsi_pill, &RMF_FIEMAP_KEY); - rc = tgt_validate_obdo(tsi, &fm_key->oa); + rc = tgt_validate_obdo(tsi, &fm_key->lfik_oa); if (rc) RETURN(err_serious(rc)); - fid = &fm_key->oa.o_oi.oi_fid; + fid = &fm_key->lfik_oa.o_oi.oi_fid; CDEBUG(D_INODE, "get FIEMAP of object "DFID"\n", PFID(fid)); - replylen = fiemap_count_to_size(fm_key->fiemap.fm_extent_count); + replylen = fiemap_count_to_size( + fm_key->lfik_fiemap.fm_extent_count); req_capsule_set_size(tsi->tsi_pill, &RMF_FIEMAP_VAL, RCL_SERVER, replylen); @@ -934,26 +1183,19 @@ int ofd_get_info_hdl(struct tgt_session_info *tsi) if (fiemap == NULL) RETURN(-ENOMEM); - *fiemap = fm_key->fiemap; + *fiemap = fm_key->lfik_fiemap; rc = ofd_fiemap_get(tsi->tsi_env, ofd, fid, fiemap); /* LU-3219: Lock the sparse areas to make sure dirty * flushed back from client, then call fiemap again. */ - if (fm_key->oa.o_valid & OBD_MD_FLFLAGS && - fm_key->oa.o_flags & OBD_FL_SRVLOCK) { - struct list_head locked; - - INIT_LIST_HEAD(&locked); + if (fm_key->lfik_oa.o_valid & OBD_MD_FLFLAGS && + fm_key->lfik_oa.o_flags & OBD_FL_SRVLOCK) { ost_fid_build_resid(fid, &fti->fti_resid); rc = lock_zero_regions(ofd->ofd_namespace, - &fti->fti_resid, fiemap, - &locked); - if (rc == 0 && !list_empty(&locked)) { + &fti->fti_resid, fiemap); + if (rc == 0) rc = ofd_fiemap_get(tsi->tsi_env, ofd, fid, fiemap); - unlock_zero_regions(ofd->ofd_namespace, - &locked); - } } } else if (KEY_IS(KEY_LAST_FID)) { struct ofd_device *ofd = ofd_exp(exp); @@ -1001,6 +1243,17 @@ out_put: RETURN(rc); } +/** + * OFD request handler for OST_GETATTR RPC. + * + * This is OFD-specific part of request handling. It finds the OFD object + * by its FID, gets attributes from storage and packs result to the reply. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_getattr_hdl(struct tgt_session_info *tsi) { struct ofd_thread_info *fti = tsi2ofd_info(tsi); @@ -1009,7 +1262,7 @@ static int ofd_getattr_hdl(struct tgt_session_info *tsi) struct lustre_handle lh = { 0 }; struct ofd_object *fo; __u64 flags = 0; - ldlm_mode_t lock_mode = LCK_PR; + enum ldlm_mode lock_mode = LCK_PR; bool srvlock; int rc; ENTRY; @@ -1046,8 +1299,7 @@ static int ofd_getattr_hdl(struct tgt_session_info *tsi) __u64 curr_version; obdo_from_la(&repbody->oa, &fti->fti_attr, - OFD_VALID_FLAGS | LA_UID | LA_GID); - tgt_drop_id(tsi->tsi_exp, &repbody->oa); + OFD_VALID_FLAGS | LA_UID | LA_GID | LA_PROJID); /* Store object version in reply */ curr_version = dt_version_get(tsi->tsi_env, @@ -1056,6 +1308,16 @@ static int ofd_getattr_hdl(struct tgt_session_info *tsi) repbody->oa.o_valid |= OBD_MD_FLDATAVERSION; repbody->oa.o_data_version = curr_version; } + + if (fo->ofo_ff.ff_layout_version > 0) { + repbody->oa.o_valid |= OBD_MD_LAYOUT_VERSION; + repbody->oa.o_layout_version = + fo->ofo_ff.ff_layout_version + fo->ofo_ff.ff_range; + + CDEBUG(D_INODE, DFID": get layout version: %u\n", + PFID(&tsi->tsi_fid), + repbody->oa.o_layout_version); + } } ofd_object_put(tsi->tsi_env, fo); @@ -1072,6 +1334,17 @@ out: RETURN(rc); } +/** + * OFD request handler for OST_SETATTR RPC. + * + * This is OFD-specific part of request handling. It finds the OFD object + * by its FID, sets attributes from request and packs result to the reply. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_setattr_hdl(struct tgt_session_info *tsi) { struct ofd_thread_info *fti = tsi2ofd_info(tsi); @@ -1080,7 +1353,6 @@ static int ofd_setattr_hdl(struct tgt_session_info *tsi) struct ost_body *repbody; struct ldlm_resource *res; struct ofd_object *fo; - struct filter_fid *ff = NULL; int rc = 0; ENTRY; @@ -1120,19 +1392,13 @@ static int ofd_setattr_hdl(struct tgt_session_info *tsi) la_from_obdo(&fti->fti_attr, &body->oa, body->oa.o_valid); fti->fti_attr.la_valid &= ~LA_TYPE; - if (body->oa.o_valid & OBD_MD_FLFID) { - ff = &fti->fti_mds_fid; - ofd_prepare_fidea(ff, &body->oa); - } - /* setting objects attributes (including owner/group) */ - rc = ofd_attr_set(tsi->tsi_env, fo, &fti->fti_attr, ff); + rc = ofd_attr_set(tsi->tsi_env, fo, &fti->fti_attr, &body->oa); if (rc != 0) GOTO(out_put, rc); obdo_from_la(&repbody->oa, &fti->fti_attr, - OFD_VALID_FLAGS | LA_UID | LA_GID); - tgt_drop_id(tsi->tsi_exp, &repbody->oa); + OFD_VALID_FLAGS | LA_UID | LA_GID | LA_PROJID); ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_SETATTR, tsi->tsi_jobid, 1); @@ -1156,6 +1422,20 @@ out: return rc; } +/** + * Destroy OST orphans. + * + * This is part of OST_CREATE RPC handling. If there is flag OBD_FL_DELORPHAN + * set then we must destroy possible orphaned objects. + * + * \param[in] env execution environment + * \param[in] exp OBD export + * \param[in] ofd OFD device + * \param[in] oa obdo structure for reply + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_orphans_destroy(const struct lu_env *env, struct obd_export *exp, struct ofd_device *ofd, struct obdo *oa) @@ -1164,10 +1444,10 @@ static int ofd_orphans_destroy(const struct lu_env *env, struct lu_fid *fid = &info->fti_fid; struct ost_id *oi = &oa->o_oi; struct ofd_seq *oseq; - obd_seq seq = ostid_seq(oi); - obd_id end_id = ostid_id(oi); - obd_id last; - obd_id oid; + u64 seq = ostid_seq(oi); + u64 end_id = ostid_id(oi); + u64 last; + u64 oid; int skip_orphan; int rc = 0; @@ -1226,8 +1506,7 @@ done: rc = ofd_seq_last_oid_write(env, ofd, oseq); } else { /* don't reuse orphan object, return last used objid */ - ostid_set_id(oi, last); - rc = 0; + rc = ostid_set_id(oi, last); } GOTO(out_put, rc); @@ -1237,6 +1516,17 @@ out_put: return rc; } +/** + * OFD request handler for OST_CREATE RPC. + * + * This is OFD-specific part of request handling. Its main purpose is to + * create new data objects on OST, but it also used to destroy orphans. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_create_hdl(struct tgt_session_info *tsi) { struct ptlrpc_request *req = tgt_ses_req(tsi); @@ -1245,11 +1535,13 @@ static int ofd_create_hdl(struct tgt_session_info *tsi) struct obdo *rep_oa; struct obd_export *exp = tsi->tsi_exp; struct ofd_device *ofd = ofd_exp(exp); - obd_seq seq = ostid_seq(&oa->o_oi); - obd_id oid = ostid_id(&oa->o_oi); + u64 seq = ostid_seq(&oa->o_oi); + u64 oid = ostid_id(&oa->o_oi); struct ofd_seq *oseq; - int rc = 0, diff; + s64 diff; + int rc = 0; int sync_trans = 0; + long granted = 0; ENTRY; @@ -1270,14 +1562,13 @@ static int ofd_create_hdl(struct tgt_session_info *tsi) rep_oa = &repbody->oa; rep_oa->o_oi = oa->o_oi; - LASSERT(seq >= FID_SEQ_OST_MDT0); LASSERT(oa->o_valid & OBD_MD_FLGROUP); CDEBUG(D_INFO, "ofd_create("DOSTID")\n", POSTID(&oa->o_oi)); oseq = ofd_seq_load(tsi->tsi_env, ofd, seq); if (IS_ERR(oseq)) { - CERROR("%s: Can't find FID Sequence "LPX64": rc = %ld\n", + CERROR("%s: Can't find FID Sequence %#llx: rc = %ld\n", ofd_name(ofd), seq, PTR_ERR(oseq)); GOTO(out_sem, rc = -EINVAL); } @@ -1286,7 +1577,7 @@ static int ofd_create_hdl(struct tgt_session_info *tsi) (oa->o_flags & OBD_FL_RECREATE_OBJS)) { if (!ofd_obd(ofd)->obd_recovering || oid > ofd_seq_last_oid(oseq)) { - CERROR("%s: recreate objid "DOSTID" > last id "LPU64 + CERROR("%s: recreate objid "DOSTID" > last id %llu" "\n", ofd_name(ofd), POSTID(&oa->o_oi), ofd_seq_last_oid(oseq)); GOTO(out_nolock, rc = -EINVAL); @@ -1311,17 +1602,26 @@ static int ofd_create_hdl(struct tgt_session_info *tsi) oseq->os_destroys_in_progress = 1; mutex_lock(&oseq->os_create_lock); if (!oseq->os_destroys_in_progress) { - CERROR("%s:["LPU64"] destroys_in_progress already" + CERROR("%s:[%llu] destroys_in_progress already" " cleared\n", ofd_name(ofd), seq); - ostid_set_id(&rep_oa->o_oi, ofd_seq_last_oid(oseq)); - GOTO(out, rc = 0); + rc = ostid_set_id(&rep_oa->o_oi, + ofd_seq_last_oid(oseq)); + GOTO(out, rc); } diff = oid - ofd_seq_last_oid(oseq); - CDEBUG(D_HA, "ofd_last_id() = "LPU64" -> diff = %d\n", - ofd_seq_last_oid(oseq), diff); + CDEBUG(D_HA, "ofd_last_id() = %llu -> diff = %lld\n", + ofd_seq_last_oid(oseq), diff); if (-diff > OST_MAX_PRECREATE) { - /* FIXME: should reset precreate_next_id on MDS */ - rc = 0; + LCONSOLE(D_INFO, "%s: too large difference between MDS " + "LAST_ID "DFID" (%llu) and OST LAST_ID "DFID" " + "(%llu), trust the OST\n", + ofd_name(ofd), PFID(&oa->o_oi.oi_fid), oid, + PFID(&oseq->os_oi.oi_fid), + ofd_seq_last_oid(oseq)); + + /* Let MDS know that we are so far ahead. */ + rc = ostid_set_id(&rep_oa->o_oi, + ofd_seq_last_oid(oseq) + 1); } else if (diff < 0) { rc = ofd_orphans_destroy(tsi->tsi_env, exp, ofd, rep_oa); @@ -1365,23 +1665,35 @@ static int ofd_create_hdl(struct tgt_session_info *tsi) ofd_name(ofd), POSTID(&oa->o_oi)); GOTO(out, rc = -EINVAL); } + + if (diff < 0) { + /* LU-5648 */ + CERROR("%s: invalid precreate request for " + DOSTID", last_id %llu. " + "Likely MDS last_id corruption\n", + ofd_name(ofd), POSTID(&oa->o_oi), + ofd_seq_last_oid(oseq)); + GOTO(out, rc = -EINVAL); + } } } if (diff > 0) { - cfs_time_t enough_time = cfs_time_shift(DISK_TIMEOUT); - obd_id next_id; - int created = 0; - int count; + time64_t enough_time = ktime_get_seconds() + DISK_TIMEOUT; + u64 next_id; + int created = 0; + int count; if (!(oa->o_valid & OBD_MD_FLFLAGS) || !(oa->o_flags & OBD_FL_DELORPHAN)) { /* don't enforce grant during orphan recovery */ - rc = ofd_grant_create(tsi->tsi_env, - ofd_obd(ofd)->obd_self_export, - &diff); - if (rc) { + granted = tgt_grant_create(tsi->tsi_env, + ofd_obd(ofd)->obd_self_export, + &diff); + if (granted < 0) { + rc = granted; + granted = 0; CDEBUG(D_HA, "%s: failed to acquire grant " - "space for precreate (%d): rc = %d\n", + "space for precreate (%lld): rc = %d\n", ofd_name(ofd), diff, rc); diff = 0; } @@ -1394,26 +1706,31 @@ static int ofd_create_hdl(struct tgt_session_info *tsi) * LFSCK will eventually clean up any orphans. LU-14 */ if (diff > 5 * OST_MAX_PRECREATE) { diff = OST_MAX_PRECREATE / 2; - LCONSOLE_WARN("%s: precreate FID "DOSTID" is over %u " - "larger than the LAST_ID "DOSTID", only " - "precreating the last %u objects.\n", - ofd_name(ofd), POSTID(&oa->o_oi), - 5 * OST_MAX_PRECREATE, - POSTID(&oseq->os_oi), diff); + LCONSOLE_WARN("%s: Too many FIDs to precreate " + "OST replaced or reformatted: " + "LFSCK will clean up", + ofd_name(ofd)); + + CDEBUG(D_HA, "%s: precreate FID "DOSTID" is over " + "%u larger than the LAST_ID "DOSTID", only " + "precreating the last %lld objects.\n", + ofd_name(ofd), POSTID(&oa->o_oi), + 5 * OST_MAX_PRECREATE, + POSTID(&oseq->os_oi), diff); ofd_seq_last_oid_set(oseq, ostid_id(&oa->o_oi) - diff); } while (diff > 0) { next_id = ofd_seq_last_oid(oseq) + 1; - count = ofd_precreate_batch(ofd, diff); + count = ofd_precreate_batch(ofd, (int)diff); - CDEBUG(D_HA, "%s: reserve %d objects in group "LPX64 - " at "LPU64"\n", ofd_name(ofd), + CDEBUG(D_HA, "%s: reserve %d objects in group %#llx" + " at %llu\n", ofd_name(ofd), count, seq, next_id); if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) - && cfs_time_after(jiffies, enough_time)) { - CDEBUG(D_HA, "%s: Slow creates, %d/%d objects" + && ktime_get_seconds() > enough_time) { + CDEBUG(D_HA, "%s: Slow creates, %d/%lld objects" " created at a rate of %d/s\n", ofd_name(ofd), created, diff + created, created / DISK_TIMEOUT); @@ -1434,7 +1751,7 @@ static int ofd_create_hdl(struct tgt_session_info *tsi) lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) LCONSOLE_WARN("%s: can't create the same count of" " objects when replaying the request" - " (diff is %d). see LU-4621\n", + " (diff is %lld). see LU-4621\n", ofd_name(ofd), diff); if (created > 0) @@ -1446,11 +1763,13 @@ static int ofd_create_hdl(struct tgt_session_info *tsi) ofd_name(ofd), rc); if (!(oa->o_valid & OBD_MD_FLFLAGS) || - !(oa->o_flags & OBD_FL_DELORPHAN)) - ofd_grant_commit(tsi->tsi_env, - ofd_obd(ofd)->obd_self_export, rc); + !(oa->o_flags & OBD_FL_DELORPHAN)) { + tgt_grant_commit(ofd_obd(ofd)->obd_self_export, + granted, rc); + granted = 0; + } - ostid_set_id(&rep_oa->o_oi, ofd_seq_last_oid(oseq)); + rc = ostid_set_id(&rep_oa->o_oi, ofd_seq_last_oid(oseq)); } EXIT; ofd_counter_incr(exp, LPROC_OFD_STATS_CREATE, @@ -1477,6 +1796,17 @@ out_sem: return rc; } +/** + * OFD request handler for OST_DESTROY RPC. + * + * This is OFD-specific part of request handling. It destroys data objects + * related to destroyed object on MDT. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_destroy_hdl(struct tgt_session_info *tsi) { const struct ost_body *body = tsi->tsi_ost_body; @@ -1484,8 +1814,8 @@ static int ofd_destroy_hdl(struct tgt_session_info *tsi) struct ofd_device *ofd = ofd_exp(tsi->tsi_exp); struct ofd_thread_info *fti = tsi2ofd_info(tsi); struct lu_fid *fid = &fti->fti_fid; - obd_id oid; - obd_count count; + u64 oid; + u32 count; int rc = 0; ENTRY; @@ -1502,7 +1832,7 @@ static int ofd_destroy_hdl(struct tgt_session_info *tsi) dlm = req_capsule_client_get(tsi->tsi_pill, &RMF_DLM_REQ); if (dlm == NULL) RETURN(-EFAULT); - ldlm_request_cancel(tgt_ses_req(tsi), dlm, 0); + ldlm_request_cancel(tgt_ses_req(tsi), dlm, 0, LATF_SKIP); } *fid = body->oa.o_oi.oi_fid; @@ -1554,6 +1884,17 @@ out: return rc; } +/** + * OFD request handler for OST_STATFS RPC. + * + * This function gets statfs data from storage as part of request + * processing. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_statfs_hdl(struct tgt_session_info *tsi) { struct obd_statfs *osfs; @@ -1561,10 +1902,12 @@ static int ofd_statfs_hdl(struct tgt_session_info *tsi) ENTRY; + OBD_FAIL_TIMEOUT(OBD_FAIL_OST_STATFS_DELAY, 10); + osfs = req_capsule_server_get(tsi->tsi_pill, &RMF_OBD_STATFS); rc = ofd_statfs(tsi->tsi_env, tsi->tsi_exp, osfs, - cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS), 0); + ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS, 0); if (rc != 0) CERROR("%s: statfs failed: rc = %d\n", tgt_name(tsi->tsi_tgt), rc); @@ -1578,6 +1921,17 @@ static int ofd_statfs_hdl(struct tgt_session_info *tsi) RETURN(rc); } +/** + * OFD request handler for OST_SYNC RPC. + * + * Sync object data or all filesystem data to the disk and pack the + * result in reply. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_sync_hdl(struct tgt_session_info *tsi) { struct ost_body *body = tsi->tsi_ost_body; @@ -1626,6 +1980,17 @@ put: return rc; } +/** + * OFD request handler for OST_PUNCH RPC. + * + * This is part of request processing. Validate request fields, + * punch (truncate) the given OFD object and pack reply. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_punch_hdl(struct tgt_session_info *tsi) { const struct obdo *oa = &tsi->tsi_ost_body->oa; @@ -1634,7 +1999,6 @@ static int ofd_punch_hdl(struct tgt_session_info *tsi) struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace; struct ldlm_resource *res; struct ofd_object *fo; - struct filter_fid *ff = NULL; __u64 flags = 0; struct lustre_handle lh = { 0, }; int rc; @@ -1643,6 +2007,8 @@ static int ofd_punch_hdl(struct tgt_session_info *tsi) ENTRY; + OBD_FAIL_TIMEOUT(OBD_FAIL_OST_PAUSE_PUNCH, cfs_fail_val); + /* check that we do support OBD_CONNECT_TRUNCLOCK. */ CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK); @@ -1679,8 +2045,8 @@ static int ofd_punch_hdl(struct tgt_session_info *tsi) RETURN(rc); } - CDEBUG(D_INODE, "calling punch for object "DFID", valid = "LPX64 - ", start = "LPD64", end = "LPD64"\n", PFID(&tsi->tsi_fid), + CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx" + ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid), oa->o_valid, start, end); fo = ofd_object_find_exists(tsi->tsi_env, ofd_exp(tsi->tsi_exp), @@ -1693,13 +2059,8 @@ static int ofd_punch_hdl(struct tgt_session_info *tsi) info->fti_attr.la_size = start; info->fti_attr.la_valid |= LA_SIZE; - if (oa->o_valid & OBD_MD_FLFID) { - ff = &info->fti_mds_fid; - ofd_prepare_fidea(ff, oa); - } - rc = ofd_object_punch(tsi->tsi_env, fo, start, end, &info->fti_attr, - ff, (struct obdo *)oa); + (struct obdo *)oa); if (rc) GOTO(out_put, rc); @@ -1720,17 +2081,217 @@ out: res = ldlm_resource_get(ns, NULL, &tsi->tsi_resid, LDLM_EXTENT, 0); if (!IS_ERR(res)) { + struct ost_lvb *res_lvb; + ldlm_res_lvbo_update(res, NULL, 0); + res_lvb = res->lr_lvb_data; + repbody->oa.o_valid |= OBD_MD_FLBLOCKS; + repbody->oa.o_blocks = res_lvb->lvb_blocks; ldlm_resource_putref(res); } } return rc; } +static int ofd_ladvise_prefetch(const struct lu_env *env, + struct ofd_object *fo, + struct niobuf_local *lnb, + __u64 start, __u64 end, enum dt_bufs_type dbt) +{ + struct ofd_thread_info *info = ofd_info(env); + pgoff_t start_index, end_index, pages; + struct niobuf_remote rnb; + unsigned long nr_local; + int rc = 0; + + if (end <= start) + RETURN(-EINVAL); + + ofd_read_lock(env, fo); + if (!ofd_object_exists(fo)) + GOTO(out_unlock, rc = -ENOENT); + + rc = ofd_attr_get(env, fo, &info->fti_attr); + if (rc) + GOTO(out_unlock, rc); + + if (end > info->fti_attr.la_size) + end = info->fti_attr.la_size; + + if (end <= start) + GOTO(out_unlock, rc); + + /* We need page aligned offset and length */ + start_index = start >> PAGE_SHIFT; + end_index = (end - 1) >> PAGE_SHIFT; + pages = end_index - start_index + 1; + while (pages > 0) { + nr_local = pages <= PTLRPC_MAX_BRW_PAGES ? pages : + PTLRPC_MAX_BRW_PAGES; + rnb.rnb_offset = start_index << PAGE_SHIFT; + rnb.rnb_len = nr_local << PAGE_SHIFT; + rc = dt_bufs_get(env, ofd_object_child(fo), &rnb, lnb, dbt); + if (unlikely(rc < 0)) + break; + nr_local = rc; + rc = dt_read_prep(env, ofd_object_child(fo), lnb, nr_local); + dt_bufs_put(env, ofd_object_child(fo), lnb, nr_local); + if (unlikely(rc)) + break; + start_index += nr_local; + pages -= nr_local; + } + +out_unlock: + ofd_read_unlock(env, fo); + RETURN(rc); +} + +/** + * OFD request handler for OST_LADVISE RPC. + * + * Tune cache or perfetch policies according to advices. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative errno on error + */ +static int ofd_ladvise_hdl(struct tgt_session_info *tsi) +{ + struct ptlrpc_request *req = tgt_ses_req(tsi); + struct obd_export *exp = tsi->tsi_exp; + struct ofd_device *ofd = ofd_exp(exp); + struct ost_body *body, *repbody; + struct ofd_thread_info *info; + struct ofd_object *fo; + struct ptlrpc_thread *svc_thread = req->rq_svc_thread; + const struct lu_env *env = svc_thread->t_env; + struct tgt_thread_big_cache *tbc = svc_thread->t_data; + enum dt_bufs_type dbt = DT_BUFS_TYPE_READAHEAD; + struct lu_ladvise *ladvise; + int num_advise; + struct ladvise_hdr *ladvise_hdr; + struct obd_ioobj ioo; + struct lustre_handle lockh = { 0 }; + __u64 flags = 0; + int i; + struct dt_object *dob; + __u64 start; + __u64 end; + int rc = 0; + ENTRY; + + CFS_FAIL_TIMEOUT(OBD_FAIL_OST_LADVISE_PAUSE, cfs_fail_val); + body = tsi->tsi_ost_body; + + if ((body->oa.o_valid & OBD_MD_FLID) != OBD_MD_FLID) + RETURN(err_serious(-EPROTO)); + + ladvise_hdr = req_capsule_client_get(tsi->tsi_pill, + &RMF_OST_LADVISE_HDR); + if (ladvise_hdr == NULL) + RETURN(err_serious(-EPROTO)); + + if (ladvise_hdr->lah_magic != LADVISE_MAGIC || + ladvise_hdr->lah_count < 1) + RETURN(err_serious(-EPROTO)); + + if ((ladvise_hdr->lah_flags & (~LF_MASK)) != 0) + RETURN(err_serious(-EPROTO)); + + ladvise = req_capsule_client_get(tsi->tsi_pill, &RMF_OST_LADVISE); + if (ladvise == NULL) + RETURN(err_serious(-EPROTO)); + + num_advise = req_capsule_get_size(&req->rq_pill, + &RMF_OST_LADVISE, RCL_CLIENT) / + sizeof(*ladvise); + if (num_advise < ladvise_hdr->lah_count) + RETURN(err_serious(-EPROTO)); + + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); + repbody->oa = body->oa; + + info = ofd_info_init(env, exp); + + rc = ostid_to_fid(&info->fti_fid, &body->oa.o_oi, + ofd->ofd_lut.lut_lsd.lsd_osd_index); + if (rc != 0) + RETURN(rc); + + fo = ofd_object_find(env, ofd, &info->fti_fid); + if (IS_ERR(fo)) { + rc = PTR_ERR(fo); + RETURN(rc); + } + LASSERT(fo != NULL); + dob = ofd_object_child(fo); + + if (ptlrpc_connection_is_local(exp->exp_connection)) + dbt |= DT_BUFS_TYPE_LOCAL; + + for (i = 0; i < num_advise; i++, ladvise++) { + start = ladvise->lla_start; + end = ladvise->lla_end; + if (end <= start) { + rc = err_serious(-EPROTO); + break; + } + + /* Handle different advice types */ + switch (ladvise->lla_advice) { + default: + rc = -ENOTSUPP; + break; + case LU_LADVISE_WILLREAD: + if (tbc == NULL) + RETURN(-ENOMEM); + + ioo.ioo_oid = body->oa.o_oi; + ioo.ioo_bufcnt = 1; + rc = tgt_extent_lock(exp->exp_obd->obd_namespace, + &tsi->tsi_resid, start, end - 1, + &lockh, LCK_PR, &flags); + if (rc != 0) + break; + + req->rq_status = ofd_ladvise_prefetch(env, fo, + tbc->local, + start, end, dbt); + tgt_extent_unlock(&lockh, LCK_PR); + break; + case LU_LADVISE_DONTNEED: + rc = dt_ladvise(env, dob, ladvise->lla_start, + ladvise->lla_end, LU_LADVISE_DONTNEED); + break; + } + if (rc != 0) + break; + } + + ofd_object_put(env, fo); + req->rq_status = rc; + RETURN(rc); +} + +/** + * OFD request handler for OST_QUOTACTL RPC. + * + * This is part of request processing to validate incoming request fields, + * get the requested data from OSD and pack reply. + * + * \param[in] tsi target session environment for this request + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_quotactl(struct tgt_session_info *tsi) { - struct obd_quotactl *oqctl, *repoqc; - int rc; + struct obd_quotactl *oqctl, *repoqc; + struct lu_nodemap *nodemap; + int id; + int rc; ENTRY; @@ -1742,77 +2303,101 @@ static int ofd_quotactl(struct tgt_session_info *tsi) if (repoqc == NULL) RETURN(err_serious(-ENOMEM)); - /* report success for quota on/off for interoperability with current MDT - * stack */ - if (oqctl->qc_cmd == Q_QUOTAON || oqctl->qc_cmd == Q_QUOTAOFF) - RETURN(0); - *repoqc = *oqctl; + + nodemap = nodemap_get_from_exp(tsi->tsi_exp); + if (IS_ERR(nodemap)) + RETURN(PTR_ERR(nodemap)); + + id = repoqc->qc_id; + if (oqctl->qc_type == USRQUOTA) + id = nodemap_map_id(nodemap, NODEMAP_UID, + NODEMAP_CLIENT_TO_FS, + repoqc->qc_id); + else if (oqctl->qc_type == GRPQUOTA) + id = nodemap_map_id(nodemap, NODEMAP_GID, + NODEMAP_CLIENT_TO_FS, + repoqc->qc_id); + + nodemap_putref(nodemap); + + if (repoqc->qc_id != id) + swap(repoqc->qc_id, id); + rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, repoqc); ofd_counter_incr(tsi->tsi_exp, LPROC_OFD_STATS_QUOTACTL, tsi->tsi_jobid, 1); + if (repoqc->qc_id != id) + swap(repoqc->qc_id, id); + RETURN(rc); } -/* High priority request handlers for OFD */ - -/* prolong locks for the current service time of the corresponding - * portal (= OST_IO_PORTAL) +/** + * Calculate the amount of time for lock prolongation. + * + * This is helper for ofd_prolong_extent_locks() function to get + * the timeout extra time. + * + * \param[in] req current request + * + * \retval amount of time to extend the timeout with */ -static inline int prolong_timeout(struct ptlrpc_request *req) +static inline time64_t prolong_timeout(struct ptlrpc_request *req) { struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; + time64_t req_timeout; if (AT_OFF) return obd_timeout / 2; - return max(at_est2timeout(at_get(&svcpt->scp_at_estimate)), - ldlm_timeout); + req_timeout = req->rq_deadline - req->rq_arrival_time.tv_sec; + return max_t(time64_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)), + req_timeout); } -static int ofd_prolong_one_lock(struct tgt_session_info *tsi, - struct ldlm_lock *lock, - struct ldlm_extent *extent, int timeout) -{ - - if (lock->l_flags & LDLM_FL_DESTROYED) /* lock already cancelled */ - return 0; - - /* 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 0; - - LDLM_DEBUG(lock, "refreshed for req x"LPU64" ext("LPU64"->"LPU64") " - "to %ds.\n", tgt_ses_req(tsi)->rq_xid, extent->start, - extent->end, 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, timeout); - return 1; -} - -static int ofd_prolong_extent_locks(struct tgt_session_info *tsi, - __u64 start, __u64 end) +/** + * Prolong lock timeout for the given extent. + * + * This function finds all locks related with incoming request and + * prolongs their timeout. + * + * If a client is holding a lock for a long time while it sends + * read or write RPCs to the OST for the object under this lock, + * then we don't want the OST to evict the client. Otherwise, + * if the network or disk is very busy then the client may not + * be able to make any progress to clear out dirty pages under + * the lock and the application will fail. + * + * Every time a Bulk Read/Write (BRW) request arrives for the object + * covered by the lock, extend the timeout on that lock. The RPC should + * contain a lock handle for the lock it is using, but this + * isn't handled correctly by all client versions, and the + * request may cover multiple locks. + * + * \param[in] tsi target session environment for this request + * \param[in] data struct of data to prolong locks + * + */ +static void ofd_prolong_extent_locks(struct tgt_session_info *tsi, + struct ldlm_prolong_args *data) { - struct obd_export *exp = tsi->tsi_exp; struct obdo *oa = &tsi->tsi_ost_body->oa; - struct ldlm_extent extent = { - .start = start, - .end = end - }; struct ldlm_lock *lock; - int timeout = prolong_timeout(tgt_ses_req(tsi)); - int lock_count = 0; ENTRY; + data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi)); + data->lpa_export = tsi->tsi_exp; + data->lpa_resid = tsi->tsi_resid; + + CDEBUG(D_RPCTRACE, "Prolong locks for req %p with x%llu" + " ext(%llu->%llu)\n", tgt_ses_req(tsi), + tgt_ses_req(tsi)->rq_xid, data->lpa_extent.start, + data->lpa_extent.end); + if (oa->o_valid & OBD_MD_FLHANDLE) { /* mostly a request should be covered by only one lock, try * fast path. */ @@ -1820,51 +2405,54 @@ static int ofd_prolong_extent_locks(struct tgt_session_info *tsi, 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, - &extent)) { + if (ldlm_extent_contain(&lock->l_policy_data.l_extent, + &data->lpa_extent)) { /* bingo */ - LASSERT(lock->l_export == exp); - lock_count = ofd_prolong_one_lock(tsi, lock, - &extent, timeout); + LASSERT(lock->l_export == data->lpa_export); + ldlm_lock_prolong_one(lock, data); + LDLM_LOCK_PUT(lock); + if (data->lpa_locks_cnt > 0) + RETURN_EXIT; + /* The lock was destroyed probably lets try + * resource tree. */ + } else { + lock->l_last_used = ktime_get(); LDLM_LOCK_PUT(lock); - RETURN(lock_count); } - LDLM_LOCK_PUT(lock); } } - spin_lock_bh(&exp->exp_bl_list_lock); - 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(&tsi->tsi_resid, &lock->l_resource->lr_name)) - continue; - - if (!ldlm_extent_overlap(&lock->l_policy_data.l_extent, - &extent)) - continue; - - lock_count += ofd_prolong_one_lock(tsi, lock, &extent, timeout); - } - spin_unlock_bh(&exp->exp_bl_list_lock); - - RETURN(lock_count); + ldlm_resource_prolong(data); + EXIT; } /** - * Returns 1 if the given PTLRPC matches the given LDLM lock, or 0 if it does - * not. + * Implementation of ptlrpc_hpreq_ops::hpreq_lock_match for OFD RW requests. + * + * Determine if \a lock and the lock from request \a req are equivalent + * by comparing their resource names, modes, and extents. + * + * It is used to give priority to read and write RPCs being done + * under this lock so that the client can drop the contended + * lock more quickly and let other clients use it. This improves + * overall performance in the case where the first client gets a + * very large lock extent that prevents other clients from + * submitting their writes. + * + * \param[in] req ptlrpc_request being processed + * \param[in] lock contended lock to match + * + * \retval 1 if lock is matched + * \retval 0 otherwise */ static int ofd_rw_hpreq_lock_match(struct ptlrpc_request *req, struct ldlm_lock *lock) { - struct niobuf_remote *rnb; - struct obd_ioobj *ioo; - ldlm_mode_t mode; - struct ldlm_extent ext; - __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); + struct niobuf_remote *rnb; + struct obd_ioobj *ioo; + enum ldlm_mode mode; + struct ldlm_extent ext; + __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); ENTRY; @@ -1882,8 +2470,13 @@ static int ofd_rw_hpreq_lock_match(struct ptlrpc_request *req, if (!ostid_res_name_eq(&ioo->ioo_oid, &lock->l_resource->lr_name)) RETURN(0); - mode = LCK_PW; + /* a bulk write can only hold a reference on a PW extent lock + * or GROUP lock. + */ + mode = LCK_PW | LCK_GROUP; if (opc == OST_READ) + /* whereas a bulk read can be protected by either a PR or PW + * extent lock */ mode |= LCK_PR; if (!(lock->l_granted_mode & mode)) @@ -1893,34 +2486,39 @@ static int ofd_rw_hpreq_lock_match(struct ptlrpc_request *req, } /** - * High-priority queue request check for whether the given PTLRPC request - * (\a req) is blocking an LDLM lock cancel. + * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OFD RW requests. * - * 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. + * Check for whether the given PTLRPC request (\a req) is blocking + * an LDLM lock cancel. Also checks whether the request is covered by an LDLM + * lock. * - * 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. + * \param[in] req the incoming request + * + * \retval 1 if \a req is blocking an LDLM lock cancel + * \retval 0 if it is not + * \retval -ESTALE if lock is not found */ static int ofd_rw_hpreq_check(struct ptlrpc_request *req) { struct tgt_session_info *tsi; struct obd_ioobj *ioo; struct niobuf_remote *rnb; - __u64 start, end; - int lock_count; + int opc; + struct ldlm_prolong_args pa = { 0 }; ENTRY; /* Don't use tgt_ses_info() to get session info, because lock_match() * can be called while request has no processing thread yet. */ tsi = lu_context_key_get(&req->rq_session, &tgt_session_key); - LASSERT(tsi != NULL); /* * Use LASSERT below because malformed RPCs should have * been filtered out in tgt_hpreq_handler(). */ + 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); @@ -1928,57 +2526,113 @@ static int ofd_rw_hpreq_check(struct ptlrpc_request *req) LASSERT(rnb != NULL); LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK)); - start = rnb->rnb_offset; + pa.lpa_mode = LCK_PW | LCK_GROUP; + if (opc == OST_READ) + pa.lpa_mode |= LCK_PR; + + pa.lpa_extent.start = rnb->rnb_offset; rnb += ioo->ioo_bufcnt - 1; - end = rnb->rnb_offset + rnb->rnb_len - 1; + pa.lpa_extent.end = rnb->rnb_offset + rnb->rnb_len - 1; DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks: "DFID - " ("LPU64"->"LPU64")\n", - tgt_name(tsi->tsi_tgt), current->comm, - PFID(&tsi->tsi_fid), start, end); + " (%llu->%llu)\n", tgt_name(tsi->tsi_tgt), + current->comm, PFID(&tsi->tsi_fid), pa.lpa_extent.start, + pa.lpa_extent.end); - lock_count = ofd_prolong_extent_locks(tsi, start, end); + ofd_prolong_extent_locks(tsi, &pa); CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n", - tgt_name(tsi->tsi_tgt), lock_count, req); + tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req); - RETURN(lock_count > 0); + if (pa.lpa_blocks_cnt > 0) + RETURN(1); + + RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE); } +/** + * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OFD RW requests. + * + * Called after the request has been handled. It refreshes lock timeout again + * so that client has more time to send lock cancel RPC. + * + * \param[in] req request which is being processed. + */ static void ofd_rw_hpreq_fini(struct ptlrpc_request *req) { ofd_rw_hpreq_check(req); } /** - * Like tgt_rw_hpreq_lock_match(), but for OST_PUNCH RPCs. + * Implementation of ptlrpc_hpreq_ops::hpreq_lock_match for OST_PUNCH request. + * + * This function checks if the given lock is the same by its resname, mode + * and extent as one taken from the request. + * It is used to give priority to punch/truncate RPCs that might lead to + * the fastest release of that lock when a lock is contended. + * + * \param[in] req ptlrpc_request being processed + * \param[in] lock contended lock to match + * + * \retval 1 if lock is matched + * \retval 0 otherwise */ static int ofd_punch_hpreq_lock_match(struct ptlrpc_request *req, struct ldlm_lock *lock) { struct tgt_session_info *tsi; + struct obdo *oa; + struct ldlm_extent ext; + + ENTRY; /* Don't use tgt_ses_info() to get session info, because lock_match() * can be called while request has no processing thread yet. */ tsi = lu_context_key_get(&req->rq_session, &tgt_session_key); - LASSERT(tsi != NULL); + /* + * Use LASSERT below because malformed RPCs should have + * been filtered out in tgt_hpreq_handler(). + */ LASSERT(tsi->tsi_ost_body != NULL); if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE && tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie) - return 1; + RETURN(1); - return 0; + oa = &tsi->tsi_ost_body->oa; + ext.start = oa->o_size; + ext.end = oa->o_blocks; + + LASSERT(lock->l_resource != NULL); + if (!ostid_res_name_eq(&oa->o_oi, &lock->l_resource->lr_name)) + RETURN(0); + + if (!(lock->l_granted_mode & (LCK_PW | LCK_GROUP))) + RETURN(0); + + RETURN(ldlm_extent_overlap(&lock->l_policy_data.l_extent, &ext)); } /** - * Like ost_rw_hpreq_check(), but for OST_PUNCH RPCs. + * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request. + * + * High-priority queue request check for whether the given punch request + * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is + * covered by an LDLM lock. + * + + * + * \param[in] req the incoming request + * + * \retval 1 if \a req is blocking an LDLM lock cancel + * \retval 0 if it is not + * \retval -ESTALE if lock is not found */ static int ofd_punch_hpreq_check(struct ptlrpc_request *req) { struct tgt_session_info *tsi; struct obdo *oa; - int lock_count; + struct ldlm_prolong_args pa = { 0 }; ENTRY; @@ -1991,37 +2645,60 @@ static int ofd_punch_hpreq_check(struct ptlrpc_request *req) LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & OBD_FL_SRVLOCK)); + pa.lpa_mode = LCK_PW | LCK_GROUP; + pa.lpa_extent.start = oa->o_size; + pa.lpa_extent.end = oa->o_blocks; + CDEBUG(D_DLMTRACE, - "%s: refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", + "%s: refresh locks: %llu/%llu (%llu->%llu)\n", tgt_name(tsi->tsi_tgt), tsi->tsi_resid.name[0], - tsi->tsi_resid.name[1], oa->o_size, oa->o_blocks); + tsi->tsi_resid.name[1], pa.lpa_extent.start, pa.lpa_extent.end); - lock_count = ofd_prolong_extent_locks(tsi, oa->o_size, oa->o_blocks); + ofd_prolong_extent_locks(tsi, &pa); CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n", - tgt_name(tsi->tsi_tgt), lock_count, req); + tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req); + + if (pa.lpa_blocks_cnt > 0) + RETURN(1); - RETURN(lock_count > 0); + RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE); } +/** + * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request. + * + * Called after the request has been handled. It refreshes lock timeout again + * so that client has more time to send lock cancel RPC. + * + * \param[in] req request which is being processed. + */ static void ofd_punch_hpreq_fini(struct ptlrpc_request *req) { ofd_punch_hpreq_check(req); } -struct ptlrpc_hpreq_ops ofd_hpreq_rw = { +static struct ptlrpc_hpreq_ops ofd_hpreq_rw = { .hpreq_lock_match = ofd_rw_hpreq_lock_match, .hpreq_check = ofd_rw_hpreq_check, .hpreq_fini = ofd_rw_hpreq_fini }; -struct ptlrpc_hpreq_ops ofd_hpreq_punch = { +static struct ptlrpc_hpreq_ops ofd_hpreq_punch = { .hpreq_lock_match = ofd_punch_hpreq_lock_match, .hpreq_check = ofd_punch_hpreq_check, .hpreq_fini = ofd_punch_hpreq_fini }; -/** Assign high priority operations to the IO requests */ +/** + * Assign high priority operations to an IO request. + * + * Check if the incoming request is a candidate for + * high-priority processing. If it is, assign it a high + * priority operations table. + * + * \param[in] tsi target session environment for this request + */ static void ofd_hp_brw(struct tgt_session_info *tsi) { struct niobuf_remote *rnb; @@ -2036,18 +2713,31 @@ static void ofd_hp_brw(struct tgt_session_info *tsi) LASSERT(rnb != NULL); /* must exist after request preprocessing */ /* no high priority if server lock is needed */ - if (rnb->rnb_flags & OBD_BRW_SRVLOCK) + if (rnb->rnb_flags & OBD_BRW_SRVLOCK || + (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) + & MSG_REPLAY)) return; } tgt_ses_req(tsi)->rq_ops = &ofd_hpreq_rw; } +/** + * Assign high priority operations to an punch request. + * + * Check if the incoming request is a candidate for + * high-priority processing. If it is, assign it a high + * priority operations table. + * + * \param[in] tsi target session environment for this request + */ static void ofd_hp_punch(struct tgt_session_info *tsi) { LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */ /* no high-priority if server lock is needed */ - if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS && - tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) + if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS && + tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) || + tgt_conn_flags(tsi) & OBD_CONNECT_MDS || + lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY) return; tgt_ses_req(tsi)->rq_ops = &ofd_hpreq_punch; } @@ -2057,6 +2747,14 @@ static void ofd_hp_punch(struct tgt_session_info *tsi) #define OST_BRW_READ OST_READ #define OST_BRW_WRITE OST_WRITE +/** + * Table of OFD-specific request handlers + * + * This table contains all opcodes accepted by OFD and + * specifies handlers for them. The tgt_request_handler() + * uses such table from each target to process incoming + * requests. + */ static struct tgt_handler ofd_tgt_handlers[] = { TGT_RPC_HANDLER(OST_FIRST_OPC, 0, OST_CONNECT, tgt_connect, @@ -2087,6 +2785,7 @@ TGT_OST_HDL_HP(HABEO_CORPUS| HABEO_REFERO | MUTABOR, ofd_hp_punch), TGT_OST_HDL(HABEO_CORPUS| HABEO_REFERO, OST_SYNC, ofd_sync_hdl), TGT_OST_HDL(0 | HABEO_REFERO, OST_QUOTACTL, ofd_quotactl), +TGT_OST_HDL(HABEO_CORPUS | HABEO_REFERO, OST_LADVISE, ofd_ladvise_hdl), }; static struct tgt_opc_slice ofd_common_slice[] = { @@ -2121,18 +2820,77 @@ static struct tgt_opc_slice ofd_common_slice[] = { .tos_hs = tgt_lfsck_handlers }, { + .tos_opc_start = SEC_FIRST_OPC, + .tos_opc_end = SEC_LAST_OPC, + .tos_hs = tgt_sec_ctx_handlers + }, + { .tos_hs = NULL } }; +/* context key constructor/destructor: ofd_key_init(), ofd_key_fini() */ +LU_KEY_INIT_FINI(ofd, struct ofd_thread_info); + +/** + * Implementation of lu_context_key::lct_key_exit. + * + * Optional method called on lu_context_exit() for all allocated + * keys. + * It is used in OFD to sanitize context values which may be re-used + * during another request processing by the same thread. + * + * \param[in] ctx execution context + * \param[in] key context key + * \param[in] data ofd_thread_info + */ +static void ofd_key_exit(const struct lu_context *ctx, + struct lu_context_key *key, void *data) +{ + struct ofd_thread_info *info = data; + + info->fti_env = NULL; + info->fti_exp = NULL; + + info->fti_xid = 0; + info->fti_pre_version = 0; + + memset(&info->fti_attr, 0, sizeof info->fti_attr); +} + +struct lu_context_key ofd_thread_key = { + .lct_tags = LCT_DT_THREAD, + .lct_init = ofd_key_init, + .lct_fini = ofd_key_fini, + .lct_exit = ofd_key_exit +}; + +/** + * Initialize OFD device according to parameters in the config log \a cfg. + * + * This is the main starting point of OFD initialization. It fills all OFD + * parameters with their initial values and calls other initializing functions + * to set up all OFD subsystems. + * + * \param[in] env execution environment + * \param[in] m OFD device + * \param[in] ldt LU device type of OFD + * \param[in] cfg configuration log + * + * \retval 0 if successful + * \retval negative value on error + */ static int ofd_init0(const struct lu_env *env, struct ofd_device *m, struct lu_device_type *ldt, struct lustre_cfg *cfg) { - const char *dev = lustre_cfg_string(cfg, 0); - struct ofd_thread_info *info = NULL; - struct obd_device *obd; - struct obd_statfs *osfs; - int rc; + const char *dev = lustre_cfg_string(cfg, 0); + struct ofd_thread_info *info = NULL; + struct obd_device *obd; + struct tg_grants_data *tgd = &m->ofd_lut.lut_tgd; + struct lu_fid fid; + struct nm_config_file *nodemap_config; + struct obd_device_target *obt; + int rc; ENTRY; @@ -2146,30 +2904,19 @@ static int ofd_init0(const struct lu_env *env, struct ofd_device *m, if (rc != 0) RETURN(rc); - obd->u.obt.obt_magic = OBT_MAGIC; + obt = &obd->u.obt; + obt->obt_magic = OBT_MAGIC; m->ofd_fmd_max_num = OFD_FMD_MAX_NUM_DEFAULT; m->ofd_fmd_max_age = OFD_FMD_MAX_AGE_DEFAULT; spin_lock_init(&m->ofd_flags_lock); m->ofd_raid_degraded = 0; + m->ofd_checksum_t10pi_enforce = 0; m->ofd_syncjournal = 0; ofd_slc_set(m); - m->ofd_grant_compat_disable = 0; m->ofd_soft_sync_limit = OFD_SOFT_SYNC_LIMIT_DEFAULT; - /* statfs data */ - spin_lock_init(&m->ofd_osfs_lock); - m->ofd_osfs_age = cfs_time_shift_64(-1000); - m->ofd_osfs_unstable = 0; - m->ofd_statfs_inflight = 0; - m->ofd_osfs_inflight = 0; - - /* grant data */ - spin_lock_init(&m->ofd_grant_lock); - m->ofd_tot_dirty = 0; - m->ofd_tot_granted = 0; - m->ofd_tot_pending = 0; m->ofd_seq_count = 0; init_waitqueue_head(&m->ofd_inconsistency_thread.t_ctl_waitq); INIT_LIST_HEAD(&m->ofd_inconsistency_list); @@ -2178,12 +2925,6 @@ static int ofd_init0(const struct lu_env *env, struct ofd_device *m, spin_lock_init(&m->ofd_batch_lock); init_rwsem(&m->ofd_lastid_rwsem); - obd->u.filter.fo_fl_oss_capa = 0; - INIT_LIST_HEAD(&obd->u.filter.fo_capa_keys); - obd->u.filter.fo_capa_hash = init_capa_hash(); - if (obd->u.filter.fo_capa_hash == NULL) - RETURN(-ENOMEM); - m->ofd_dt_dev.dd_lu_dev.ld_ops = &ofd_lu_ops; m->ofd_dt_dev.dd_lu_dev.ld_obd = obd; /* set this lu_device to obd, because error handling need it */ @@ -2211,7 +2952,7 @@ static int ofd_init0(const struct lu_env *env, struct ofd_device *m, info = ofd_info_init(env, NULL); if (info == NULL) - RETURN(-EFAULT); + GOTO(err_fini_proc, rc = -EFAULT); rc = ofd_stack_init(env, m, cfg); if (rc) { @@ -2221,24 +2962,6 @@ static int ofd_init0(const struct lu_env *env, struct ofd_device *m, ofd_procfs_add_brw_stats_symlink(m); - /* populate cached statfs data */ - osfs = &ofd_info(env)->fti_u.osfs; - rc = ofd_statfs_internal(env, m, osfs, 0, NULL); - if (rc != 0) { - CERROR("%s: can't get statfs data, rc %d\n", obd->obd_name, rc); - GOTO(err_fini_stack, rc); - } - if (!IS_PO2(osfs->os_bsize)) { - CERROR("%s: blocksize (%d) is not a power of 2\n", - obd->obd_name, osfs->os_bsize); - GOTO(err_fini_stack, rc = -EPROTO); - } - m->ofd_blockbits = fls(osfs->os_bsize) - 1; - - m->ofd_precreate_batch = OFD_PRECREATE_BATCH_DEFAULT; - if (osfs->os_bsize * osfs->os_blocks < OFD_PRECREATE_SMALL_FS) - m->ofd_precreate_batch = OFD_PRECREATE_BATCH_SMALL; - snprintf(info->fti_u.name, sizeof(info->fti_u.name), "%s-%s", "filter"/*LUSTRE_OST_NAME*/, obd->obd_uuid.uuid); m->ofd_namespace = ldlm_namespace_new(obd, info->fti_u.name, @@ -2256,38 +2979,64 @@ static int ofd_init0(const struct lu_env *env, struct ofd_device *m, ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL, "filter_ldlm_cb_client", &obd->obd_ldlm_client); - dt_conf_get(env, m->ofd_osd, &m->ofd_dt_conf); - - /* Allow at most ddp_grant_reserved% of the available filesystem space - * to be granted to clients, so that any errors in the grant overhead - * calculations do not allow granting more space to clients than can be - * written. Assumes that in aggregate the grant overhead calculations do - * not have more than ddp_grant_reserved% estimation error in them. */ - m->ofd_grant_ratio = - ofd_grant_ratio_conv(m->ofd_dt_conf.ddp_grant_reserved); - rc = tgt_init(env, &m->ofd_lut, obd, m->ofd_osd, ofd_common_slice, OBD_FAIL_OST_ALL_REQUEST_NET, OBD_FAIL_OST_ALL_REPLY_NET); if (rc) GOTO(err_free_ns, rc); + tgd->tgd_reserved_pcnt = 0; + + m->ofd_brw_size = m->ofd_lut.lut_dt_conf.ddp_brw_size; + m->ofd_cksum_types_supported = + obd_cksum_types_supported_server(obd->obd_name); + m->ofd_precreate_batch = OFD_PRECREATE_BATCH_DEFAULT; + if (tgd->tgd_osfs.os_bsize * tgd->tgd_osfs.os_blocks < + OFD_PRECREATE_SMALL_FS) + m->ofd_precreate_batch = OFD_PRECREATE_BATCH_SMALL; + rc = ofd_fs_setup(env, m, obd); if (rc) GOTO(err_fini_lut, rc); - rc = ofd_start_inconsistency_verification_thread(m); + fid.f_seq = FID_SEQ_LOCAL_NAME; + fid.f_oid = 1; + fid.f_ver = 0; + rc = local_oid_storage_init(env, m->ofd_osd, &fid, + &m->ofd_los); if (rc != 0) GOTO(err_fini_fs, rc); + nodemap_config = nm_config_file_register_tgt(env, m->ofd_osd, + m->ofd_los); + if (IS_ERR(nodemap_config)) { + rc = PTR_ERR(nodemap_config); + if (rc != -EROFS) + GOTO(err_fini_los, rc); + } else { + obt->obt_nodemap_config_file = nodemap_config; + } + + rc = ofd_start_inconsistency_verification_thread(m); + if (rc != 0) + GOTO(err_fini_nm, rc); + + tgt_adapt_sptlrpc_conf(&m->ofd_lut); + RETURN(0); +err_fini_nm: + nm_config_file_deregister_tgt(env, obt->obt_nodemap_config_file); + obt->obt_nodemap_config_file = NULL; +err_fini_los: + local_oid_storage_fini(env, m->ofd_los); + m->ofd_los = NULL; err_fini_fs: ofd_fs_cleanup(env, m); err_fini_lut: tgt_fini(env, &m->ofd_lut); err_free_ns: - ldlm_namespace_free(m->ofd_namespace, 0, obd->obd_force); + ldlm_namespace_free(m->ofd_namespace, NULL, obd->obd_force); obd->obd_namespace = m->ofd_namespace = NULL; err_fini_stack: ofd_stack_fini(env, m, &m->ofd_osd->dd_lu_dev); @@ -2296,6 +3045,15 @@ err_fini_proc: return rc; } +/** + * Stop the OFD device + * + * This function stops the OFD device and all its subsystems. + * This is the end of OFD lifecycle. + * + * \param[in] env execution environment + * \param[in] m OFD device + */ static void ofd_fini(const struct lu_env *env, struct ofd_device *m) { struct obd_device *obd = ofd_obd(m); @@ -2305,7 +3063,12 @@ static void ofd_fini(const struct lu_env *env, struct ofd_device *m) stop.ls_status = LS_PAUSED; stop.ls_flags = 0; lfsck_stop(env, m->ofd_osd, &stop); + ofd_stack_pre_fini(env, m, &m->ofd_dt_dev.dd_lu_dev); target_recovery_fini(obd); + if (m->ofd_namespace != NULL) + ldlm_namespace_free_prior(m->ofd_namespace, NULL, + d->ld_obd->obd_force); + obd_exports_barrier(obd); obd_zombie_barrier(); @@ -2313,13 +3076,11 @@ static void ofd_fini(const struct lu_env *env, struct ofd_device *m) ofd_stop_inconsistency_verification_thread(m); lfsck_degister(env, m->ofd_osd); ofd_fs_cleanup(env, m); - - ofd_free_capa_keys(m); - cleanup_capa_hash(obd->u.filter.fo_capa_hash); + nm_config_file_deregister_tgt(env, obd->u.obt.obt_nodemap_config_file); + obd->u.obt.obt_nodemap_config_file = NULL; if (m->ofd_namespace != NULL) { - ldlm_namespace_free(m->ofd_namespace, NULL, - d->ld_obd->obd_force); + ldlm_namespace_free_post(m->ofd_namespace); d->ld_obd->obd_namespace = m->ofd_namespace = NULL; } @@ -2330,6 +3091,17 @@ static void ofd_fini(const struct lu_env *env, struct ofd_device *m) EXIT; } +/** + * Implementation of lu_device_type_operations::ldto_device_fini. + * + * Finalize device. Dual to ofd_device_init(). It is called from + * obd_precleanup() and stops the current device. + * + * \param[in] env execution environment + * \param[in] d LU device of OFD + * + * \retval NULL + */ static struct lu_device *ofd_device_fini(const struct lu_env *env, struct lu_device *d) { @@ -2338,6 +3110,16 @@ static struct lu_device *ofd_device_fini(const struct lu_env *env, RETURN(NULL); } +/** + * Implementation of lu_device_type_operations::ldto_device_free. + * + * Free OFD device. Dual to ofd_device_alloc(). + * + * \param[in] env execution environment + * \param[in] d LU device of OFD + * + * \retval NULL + */ static struct lu_device *ofd_device_free(const struct lu_env *env, struct lu_device *d) { @@ -2348,6 +3130,19 @@ static struct lu_device *ofd_device_free(const struct lu_env *env, RETURN(NULL); } +/** + * Implementation of lu_device_type_operations::ldto_device_alloc. + * + * This function allocates the new OFD device. It is called from + * obd_setup() if OBD device had lu_device_type defined. + * + * \param[in] env execution environment + * \param[in] t lu_device_type of OFD device + * \param[in] cfg configuration log + * + * \retval pointer to the lu_device of just allocated OFD + * \retval ERR_PTR of return value on error + */ static struct lu_device *ofd_device_alloc(const struct lu_env *env, struct lu_device_type *t, struct lustre_cfg *cfg) @@ -2371,32 +3166,7 @@ static struct lu_device *ofd_device_alloc(const struct lu_env *env, return l; } -/* thread context key constructor/destructor */ -LU_KEY_INIT_FINI(ofd, struct ofd_thread_info); - -static void ofd_key_exit(const struct lu_context *ctx, - struct lu_context_key *key, void *data) -{ - struct ofd_thread_info *info = data; - - info->fti_env = NULL; - info->fti_exp = NULL; - - info->fti_xid = 0; - info->fti_pre_version = 0; - info->fti_used = 0; - - memset(&info->fti_attr, 0, sizeof info->fti_attr); -} - -struct lu_context_key ofd_thread_key = { - .lct_tags = LCT_DT_THREAD, - .lct_init = ofd_key_init, - .lct_fini = ofd_key_fini, - .lct_exit = ofd_key_exit -}; - -/* type constructor/destructor: mdt_type_init, mdt_type_fini */ +/* type constructor/destructor: ofd_type_init(), ofd_type_fini() */ LU_TYPE_INIT_FINI(ofd, &ofd_thread_key); static struct lu_device_type_operations ofd_device_type_ops = { @@ -2418,7 +3188,16 @@ static struct lu_device_type ofd_device_type = { .ldt_ctx_tags = LCT_DT_THREAD }; -int __init ofd_init(void) +/** + * Initialize OFD module. + * + * This function is called upon module loading. It registers OFD device type + * and prepares all in-memory structures used by all OFD devices. + * + * \retval 0 if successful + * \retval negative value on error + */ +static int __init ofd_init(void) { int rc; @@ -2433,22 +3212,26 @@ int __init ofd_init(void) } rc = class_register_type(&ofd_obd_ops, NULL, true, NULL, -#ifndef HAVE_ONLY_PROCFS_SEQ - NULL, -#endif LUSTRE_OST_NAME, &ofd_device_type); return rc; } -void __exit ofd_exit(void) +/** + * Stop OFD module. + * + * This function is called upon OFD module unloading. + * It frees all related structures and unregisters OFD device type. + */ +static void __exit ofd_exit(void) { ofd_fmd_exit(); lu_kmem_fini(ofd_caches); class_unregister_type(LUSTRE_OST_NAME); } -MODULE_AUTHOR("Whamcloud, Inc. "); +MODULE_AUTHOR("OpenSFS, Inc. "); MODULE_DESCRIPTION("Lustre Object Filtering Device"); +MODULE_VERSION(LUSTRE_VERSION_STRING); MODULE_LICENSE("GPL"); module_init(ofd_init);