X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosp%2Fosp_dev.c;h=2bb469e9a0a945f75801eded9354b2603389ae8d;hp=c86efd820f62f533e6a29d8af32a095027794d10;hb=713174908cb8e5e3ceadd3ca1cb42a88b200e576;hpb=b33cba87140d6801368d68772c864f49178720eb diff --git a/lustre/osp/osp_dev.c b/lustre/osp/osp_dev.c index c86efd8..2bb469e 100644 --- a/lustre/osp/osp_dev.c +++ b/lustre/osp/osp_dev.c @@ -27,7 +27,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Intel, Inc. + * Copyright (c) 2012, 2015, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -35,24 +35,57 @@ * * lustre/osp/osp_dev.c * - * Lustre OST Proxy Device - * * Author: Alex Zhuravlev * Author: Mikhail Pershin + * Author: Di Wang + */ +/* + * The Object Storage Proxy (OSP) module provides an implementation of + * the DT API for remote MDTs and OSTs. Every local OSP device (or + * object) is a proxy for a remote OSD device (or object). Thus OSP + * converts DT operations into RPCs, which are sent to the OUT service + * on a remote target, converted back to DT operations, and + * executed. Of course there are many ways in which this description + * is inaccurate but it's a good enough mental model. OSP is used by + * the MDT stack in several ways: + * + * - OSP devices allocate FIDs for the stripe sub-objects of a striped + * file or directory. + * + * - OSP objects represent the remote MDT and OST objects that are + * the stripes of a striped object. + * + * - OSP devices log, send, and track synchronous operations (setattr + * and unlink) to remote targets. + * + * - OSP objects are the bottom slice of the compound LU object + * representing a remote MDT object: MDT/MDD/LOD/OSP. + * + * - OSP objects are used by LFSCK to represent remote OST objects + * during the verification of MDT-OST consistency. + * + * - OSP devices batch idempotent requests (declare_attr_get() and + * declare_xattr_get()) to the remote target and cache their results. + * + * In addition the OSP layer implements a subset of the OBD device API + * to support being a client of a remote target, connecting to other + * layers, and FID allocation. */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_MDS -#include +#include + +#include +#include +#include #include +#include #include "osp_internal.h" /* Slab for OSP object allocation */ -cfs_mem_cache_t *osp_object_kmem; +struct kmem_cache *osp_object_kmem; static struct lu_kmem_descr osp_caches[] = { { @@ -65,24 +98,45 @@ static struct lu_kmem_descr osp_caches[] = { } }; -struct lu_object *osp_object_alloc(const struct lu_env *env, - const struct lu_object_header *hdr, - struct lu_device *d) +/** + * Implementation of lu_device_operations::ldo_object_alloc + * + * Allocates an OSP object in memory, whose FID is on the remote target. + * + * \param[in] env execution environment + * \param[in] hdr The header of the object stack. If it is NULL, it + * means the object is not built from top device, i.e. + * it is a sub-stripe object of striped directory or + * an OST object. + * \param[in] d OSP device + * + * \retval object object being created if the creation succeed. + * \retval NULL NULL if the creation failed. + */ +static struct lu_object *osp_object_alloc(const struct lu_env *env, + const struct lu_object_header *hdr, + struct lu_device *d) { - struct lu_object_header *h; + struct lu_object_header *h = NULL; struct osp_object *o; struct lu_object *l; - LASSERT(hdr == NULL); - - OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, CFS_ALLOC_IO); + OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, GFP_NOFS); if (o != NULL) { l = &o->opo_obj.do_lu; - h = &o->opo_header; - lu_object_header_init(h); - dt_object_init(&o->opo_obj, h, d); - lu_object_add_top(h, l); + /* If hdr is NULL, it means the object is not built + * from the top dev(MDT/OST), usually it happens when + * building striped object, like data object on MDT or + * striped object for directory */ + if (hdr == NULL) { + h = &o->opo_header; + lu_object_header_init(h); + dt_object_init(&o->opo_obj, h, d); + lu_object_add_top(h, l); + } else { + dt_object_init(&o->opo_obj, h, d); + } l->lo_ops = &osp_lu_obj_ops; @@ -92,106 +146,314 @@ struct lu_object *osp_object_alloc(const struct lu_env *env, } } -/* Update opd_last_used_id along with checking for gap in objid sequence */ -void osp_update_last_id(struct osp_device *d, obd_id objid) +/** + * Find or create the local object + * + * Finds or creates the local file referenced by \a reg_id and return the + * attributes of the local file. + * + * \param[in] env execution environment + * \param[in] osp OSP device + * \param[out] attr attributes of the object + * \param[in] reg_id the local object ID of the file. It will be used + * to compose a local FID{FID_SEQ_LOCAL_FILE, reg_id, 0} + * to identify the object. + * + * \retval object object(dt_object) found or created + * \retval ERR_PTR(errno) ERR_PTR(errno) if not get the object. + */ +static struct dt_object +*osp_find_or_create_local_file(const struct lu_env *env, struct osp_device *osp, + struct lu_attr *attr, __u32 reg_id) { - /* - * we might have lost precreated objects due to VBR and precreate - * orphans, the gap in objid can be calculated properly only here - */ - if (objid > le64_to_cpu(d->opd_last_used_id)) { - if (objid - le64_to_cpu(d->opd_last_used_id) > 1) { - d->opd_gap_start = le64_to_cpu(d->opd_last_used_id) + 1; - d->opd_gap_count = objid - d->opd_gap_start; - CDEBUG(D_HA, "Gap in objids: %d, start = %llu\n", - d->opd_gap_count, d->opd_gap_start); - } - d->opd_last_used_id = cpu_to_le64(objid); + struct osp_thread_info *osi = osp_env_info(env); + struct dt_object_format dof = { 0 }; + struct dt_object *dto; + int rc; + ENTRY; + + lu_local_obj_fid(&osi->osi_fid, reg_id); + attr->la_valid = LA_MODE; + attr->la_mode = S_IFREG | 0644; + dof.dof_type = DFT_REGULAR; + /* Find or create the local object by osi_fid. */ + dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid, + &dof, attr); + if (IS_ERR(dto)) + RETURN(dto); + + /* Get attributes of the local object. */ + rc = dt_attr_get(env, dto, attr); + if (rc) { + CERROR("%s: can't be initialized: rc = %d\n", + osp->opd_obd->obd_name, rc); + lu_object_put(env, &dto->do_lu); + RETURN(ERR_PTR(rc)); } + RETURN(dto); } -static int osp_last_used_init(const struct lu_env *env, struct osp_device *m) +/** + * Write data buffer to a local file object. + * + * \param[in] env execution environment + * \param[in] osp OSP device + * \param[in] dt_obj object written to + * \param[in] buf buffer containing byte array and length + * \param[in] offset write offset in the object in bytes + * + * \retval 0 0 if write succeed + * \retval -EFAULT -EFAULT if only part of buffer is written. + * \retval negative other negative errno if write failed. + */ +static int osp_write_local_file(const struct lu_env *env, + struct osp_device *osp, + struct dt_object *dt_obj, + struct lu_buf *buf, + loff_t offset) { - struct osp_thread_info *osi = osp_env_info(env); - struct dt_object_format dof = { 0 }; - struct dt_object *o; - int rc; + struct thandle *th; + int rc; - ENTRY; + th = dt_trans_create(env, osp->opd_storage); + if (IS_ERR(th)) + RETURN(PTR_ERR(th)); - osi->osi_attr.la_valid = LA_MODE; - osi->osi_attr.la_mode = S_IFREG | 0644; - lu_local_obj_fid(&osi->osi_fid, MDD_LOV_OBJ_OID); - dof.dof_type = DFT_REGULAR; - o = dt_find_or_create(env, m->opd_storage, &osi->osi_fid, &dof, - &osi->osi_attr); - if (IS_ERR(o)) - RETURN(PTR_ERR(o)); - - rc = dt_attr_get(env, o, &osi->osi_attr, NULL); + rc = dt_declare_record_write(env, dt_obj, buf, offset, th); + if (rc) + GOTO(out, rc); + rc = dt_trans_start_local(env, osp->opd_storage, th); if (rc) GOTO(out, rc); + rc = dt_record_write(env, dt_obj, buf, &offset, th); +out: + dt_trans_stop(env, osp->opd_storage, th); + RETURN(rc); +} + +/** + * Initialize last ID object. + * + * This function initializes the LAST_ID file, which stores the current last + * used id of data objects. The MDT will use the last used id and the last_seq + * (\see osp_init_last_seq()) to synchronize the precreate object cache with + * OSTs. + * + * \param[in] env execution environment + * \param[in] osp OSP device + * + * \retval 0 0 if initialization succeed + * \retval negative negative errno if initialization failed + */ +static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp) +{ + struct osp_thread_info *osi = osp_env_info(env); + struct lu_fid *fid = &osp->opd_last_used_fid; + struct dt_object *dto; + int rc = -EFAULT; + ENTRY; + + dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr, + MDD_LOV_OBJ_OID); + if (IS_ERR(dto)) + RETURN(PTR_ERR(dto)); + /* object will be released in device cleanup path */ - m->opd_last_used_file = o; + if (osi->osi_attr.la_size >= + sizeof(osi->osi_id) * (osp->opd_index + 1)) { + osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid, + osp->opd_index); + rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off); + if (rc != 0 && rc != -EFAULT) + GOTO(out, rc); + } - if (osi->osi_attr.la_size >= sizeof(osi->osi_id) * - (m->opd_index + 1)) { - osp_objid_buf_prep(osi, m, m->opd_index); - rc = dt_record_read(env, o, &osi->osi_lb, &osi->osi_off); + if (rc == -EFAULT) { /* fresh LAST_ID */ + fid->f_oid = 0; + osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid, + osp->opd_index); + rc = osp_write_local_file(env, osp, dto, &osi->osi_lb, + osi->osi_off); if (rc != 0) GOTO(out, rc); - } else { - /* reset value to 0, just to make sure and change file's size */ - struct thandle *th; + } + osp->opd_last_used_oid_file = dto; + RETURN(0); +out: + /* object will be released in device cleanup path */ + CERROR("%s: can't initialize lov_objid: rc = %d\n", + osp->opd_obd->obd_name, rc); + lu_object_put(env, &dto->do_lu); + osp->opd_last_used_oid_file = NULL; + RETURN(rc); +} - m->opd_last_used_id = 0; - osp_objid_buf_prep(osi, m, m->opd_index); +/** + * Initialize last sequence object. + * + * This function initializes the LAST_SEQ file in the local OSD, which stores + * the current last used sequence of data objects. The MDT will use the last + * sequence and last id (\see osp_init_last_objid()) to synchronize the + * precreate object cache with OSTs. + * + * \param[in] env execution environment + * \param[in] osp OSP device + * + * \retval 0 0 if initialization succeed + * \retval negative negative errno if initialization failed + */ +static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp) +{ + struct osp_thread_info *osi = osp_env_info(env); + struct lu_fid *fid = &osp->opd_last_used_fid; + struct dt_object *dto; + int rc = -EFAULT; + ENTRY; - th = dt_trans_create(env, m->opd_storage); - if (IS_ERR(th)) - GOTO(out, rc = PTR_ERR(th)); + dto = osp_find_or_create_local_file(env, osp, &osi->osi_attr, + MDD_LOV_OBJ_OSEQ); + if (IS_ERR(dto)) + RETURN(PTR_ERR(dto)); - rc = dt_declare_record_write(env, m->opd_last_used_file, - osi->osi_lb.lb_len, osi->osi_off, - th); - if (rc) { - dt_trans_stop(env, m->opd_storage, th); + /* object will be released in device cleanup path */ + if (osi->osi_attr.la_size >= + sizeof(osi->osi_id) * (osp->opd_index + 1)) { + osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq, + osp->opd_index); + rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off); + if (rc != 0 && rc != -EFAULT) GOTO(out, rc); - } + } + + if (rc == -EFAULT) { /* fresh OSP */ + fid->f_seq = 0; + osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq, + osp->opd_index); + rc = osp_write_local_file(env, osp, dto, &osi->osi_lb, + osi->osi_off); + if (rc != 0) + GOTO(out, rc); + } + osp->opd_last_used_seq_file = dto; + RETURN(0); +out: + /* object will be released in device cleanup path */ + CERROR("%s: can't initialize lov_seq: rc = %d\n", + osp->opd_obd->obd_name, rc); + lu_object_put(env, &dto->do_lu); + osp->opd_last_used_seq_file = NULL; + RETURN(rc); +} + +/** + * Initialize last OID and sequence object. + * + * If the MDT is just upgraded to 2.4 from the lower version, where the + * LAST_SEQ file does not exist, the file will be created and IDIF sequence + * will be written into the file. + * + * \param[in] env execution environment + * \param[in] osp OSP device + * + * \retval 0 0 if initialization succeed + * \retval negative negative error if initialization failed + */ +static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp) +{ + struct osp_thread_info *osi = osp_env_info(env); + int rc; + ENTRY; - rc = dt_trans_start_local(env, m->opd_storage, th); + fid_zero(&osp->opd_last_used_fid); + rc = osp_init_last_objid(env, osp); + if (rc < 0) { + CERROR("%s: Can not get ids %d from old objid!\n", + osp->opd_obd->obd_name, rc); + RETURN(rc); + } + + rc = osp_init_last_seq(env, osp); + if (rc < 0) { + CERROR("%s: Can not get ids %d from old objid!\n", + osp->opd_obd->obd_name, rc); + GOTO(out, rc); + } + + if (fid_oid(&osp->opd_last_used_fid) != 0 && + fid_seq(&osp->opd_last_used_fid) == 0) { + /* Just upgrade from the old version, + * set the seq to be IDIF */ + osp->opd_last_used_fid.f_seq = + fid_idif_seq(fid_oid(&osp->opd_last_used_fid), + osp->opd_index); + osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, + &osp->opd_last_used_fid.f_seq, + osp->opd_index); + rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file, + &osi->osi_lb, osi->osi_off); if (rc) { - dt_trans_stop(env, m->opd_storage, th); + CERROR("%s : Can not write seq file: rc = %d\n", + osp->opd_obd->obd_name, rc); GOTO(out, rc); } + } - rc = dt_record_write(env, m->opd_last_used_file, &osi->osi_lb, - &osi->osi_off, th); - dt_trans_stop(env, m->opd_storage, th); - if (rc) - GOTO(out, rc); + if (!fid_is_zero(&osp->opd_last_used_fid) && + !fid_is_sane(&osp->opd_last_used_fid)) { + CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name, + PFID(&osp->opd_last_used_fid)); + GOTO(out, rc = -EINVAL); } - CDEBUG(D_HA, "%s: Read last used ID: "LPU64"\n", m->opd_obd->obd_name, - le64_to_cpu(m->opd_last_used_id)); - RETURN(0); + + CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n", + osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid)); out: - CERROR("%s: can't initialize lov_objid: %d\n", - m->opd_obd->obd_name, rc); - lu_object_put(env, &o->do_lu); - m->opd_last_used_file = NULL; - return rc; + if (rc != 0) { + if (osp->opd_last_used_oid_file != NULL) { + lu_object_put(env, &osp->opd_last_used_oid_file->do_lu); + osp->opd_last_used_oid_file = NULL; + } + if (osp->opd_last_used_seq_file != NULL) { + lu_object_put(env, &osp->opd_last_used_seq_file->do_lu); + osp->opd_last_used_seq_file = NULL; + } + } + + RETURN(rc); } -static void osp_last_used_fini(const struct lu_env *env, struct osp_device *d) +/** + * Release the last sequence and OID file objects in OSP device. + * + * \param[in] env execution environment + * \param[in] osp OSP device + */ +static void osp_last_used_fini(const struct lu_env *env, struct osp_device *osp) { - if (d->opd_last_used_file != NULL) { - lu_object_put(env, &d->opd_last_used_file->do_lu); - d->opd_last_used_file = NULL; + /* release last_used file */ + if (osp->opd_last_used_oid_file != NULL) { + lu_object_put(env, &osp->opd_last_used_oid_file->do_lu); + osp->opd_last_used_oid_file = NULL; + } + + if (osp->opd_last_used_seq_file != NULL) { + lu_object_put(env, &osp->opd_last_used_seq_file->do_lu); + osp->opd_last_used_seq_file = NULL; } } -int osp_disconnect(struct osp_device *d) +/** + * Disconnects the connection between OSP and its correspondent MDT or OST, and + * the import will be marked as inactive. It will only be called during OSP + * cleanup process. + * + * \param[in] d OSP device being disconnected + * + * \retval 0 0 if disconnection succeed + * \retval negative negative errno if disconnection failed + */ +static int osp_disconnect(struct osp_device *d) { struct obd_import *imp; int rc = 0; @@ -202,9 +464,9 @@ int osp_disconnect(struct osp_device *d) * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't * fully deactivate the import, or that would drop all requests. */ LASSERT(imp != NULL); - cfs_spin_lock(&imp->imp_lock); + spin_lock(&imp->imp_lock); imp->imp_deactive = 1; - cfs_spin_unlock(&imp->imp_lock); + spin_unlock(&imp->imp_lock); ptlrpc_deactivate_import(imp); @@ -214,9 +476,7 @@ int osp_disconnect(struct osp_device *d) (void)ptlrpc_pinger_del_import(imp); rc = ptlrpc_disconnect_import(imp, 0); - if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN) - rc = 0; - if (rc) + if (rc != 0) CERROR("%s: can't disconnect: rc = %d\n", d->opd_obd->obd_name, rc); @@ -225,52 +485,167 @@ int osp_disconnect(struct osp_device *d) RETURN(rc); } -static int osp_shutdown(const struct lu_env *env, struct osp_device *d) +/** + * Initialize the osp_update structure in OSP device + * + * Allocate osp update structure and start update thread. + * + * \param[in] osp OSP device + * + * \retval 0 if initialization succeeds. + * \retval negative errno if initialization fails. + */ +static int osp_update_init(struct osp_device *osp) { - int rc = 0; + struct l_wait_info lwi = { 0 }; + struct task_struct *task; + ENTRY; - if (is_osp_on_ost(d->opd_obd->obd_name)) { - rc = osp_disconnect(d); + LASSERT(osp->opd_connect_mdt); + + OBD_ALLOC_PTR(osp->opd_update); + if (osp->opd_update == NULL) + RETURN(-ENOMEM); + + init_waitqueue_head(&osp->opd_update_thread.t_ctl_waitq); + init_waitqueue_head(&osp->opd_update->ou_waitq); + spin_lock_init(&osp->opd_update->ou_lock); + INIT_LIST_HEAD(&osp->opd_update->ou_list); + osp->opd_update->ou_rpc_version = 1; + osp->opd_update->ou_version = 1; + + /* start thread handling sending updates to the remote MDT */ + task = kthread_run(osp_send_update_thread, osp, + "osp_up%u-%u", osp->opd_index, osp->opd_group); + if (IS_ERR(task)) { + int rc = PTR_ERR(task); + + OBD_FREE_PTR(osp->opd_update); + osp->opd_update = NULL; + CERROR("%s: can't start precreate thread: rc = %d\n", + osp->opd_obd->obd_name, rc); RETURN(rc); } + l_wait_event(osp->opd_update_thread.t_ctl_waitq, + osp_send_update_thread_running(osp) || + osp_send_update_thread_stopped(osp), &lwi); + + RETURN(0); +} + +/** + * Finialize osp_update structure in OSP device + * + * Stop the OSP update sending thread, then delete the left + * osp thandle in the sending list. + * + * \param [in] osp OSP device. + */ +static void osp_update_fini(const struct lu_env *env, struct osp_device *osp) +{ + struct osp_update_request *our; + struct osp_update_request *tmp; + struct osp_updates *ou = osp->opd_update; + + if (ou == NULL) + return; + + osp->opd_update_thread.t_flags = SVC_STOPPING; + wake_up(&ou->ou_waitq); + + wait_event(osp->opd_update_thread.t_ctl_waitq, + osp->opd_update_thread.t_flags & SVC_STOPPED); + + /* Remove the left osp thandle from the list */ + spin_lock(&ou->ou_lock); + list_for_each_entry_safe(our, tmp, &ou->ou_list, + our_list) { + list_del_init(&our->our_list); + LASSERT(our->our_th != NULL); + osp_trans_callback(env, our->our_th, -EIO); + /* our will be destroyed in osp_thandle_put() */ + osp_thandle_put(our->our_th); + } + spin_unlock(&ou->ou_lock); + + OBD_FREE_PTR(ou); + osp->opd_update = NULL; +} + +/** + * Cleanup OSP, which includes disconnect import, cleanup unlink log, stop + * precreate threads etc. + * + * \param[in] env execution environment. + * \param[in] d OSP device being disconnected. + * + * \retval 0 0 if cleanup succeed + * \retval negative negative errno if cleanup failed + */ +static int osp_shutdown(const struct lu_env *env, struct osp_device *d) +{ + int rc = 0; + ENTRY; + LASSERT(env); - /* release last_used file */ - osp_last_used_fini(env, d); rc = osp_disconnect(d); - /* stop precreate thread */ - osp_precreate_fini(d); + if (!d->opd_connect_mdt) { + /* stop sync thread */ + osp_sync_fini(d); + + /* stop precreate thread */ + osp_precreate_fini(d); + + /* release last_used file */ + osp_last_used_fini(env, d); + } - /* stop sync thread */ - osp_sync_fini(d); + obd_fid_fini(d->opd_obd); RETURN(rc); } +/** + * Implementation of osp_lu_ops::ldo_process_config + * + * This function processes config log records in OSP layer. It is usually + * called from the top layer of MDT stack, and goes through the stack by calling + * ldo_process_config of next layer. + * + * \param[in] env execution environment + * \param[in] dev lu_device of OSP + * \param[in] lcfg config log + * + * \retval 0 0 if the config log record is executed correctly. + * \retval negative negative errno if the record execution fails. + */ static int osp_process_config(const struct lu_env *env, struct lu_device *dev, struct lustre_cfg *lcfg) { struct osp_device *d = lu2osp_dev(dev); - struct lprocfs_static_vars lvars = { 0 }; + struct obd_device *obd = d->opd_obd; int rc; ENTRY; switch (lcfg->lcfg_command) { + case LCFG_PRE_CLEANUP: + rc = osp_disconnect(d); + osp_update_fini(env, d); + break; case LCFG_CLEANUP: - if (!is_osp_on_ost(d->opd_obd->obd_name)) - lu_dev_del_linkage(dev->ld_site, dev); + lu_dev_del_linkage(dev->ld_site, dev); rc = osp_shutdown(env, d); break; case LCFG_PARAM: - lprocfs_osp_init_vars(&lvars); - - LASSERT(d->opd_obd); - rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars, - lcfg, d->opd_obd); + LASSERT(obd); + rc = class_process_proc_param(d->opd_connect_mdt ? + PARAM_OSP : PARAM_OSC, + obd->obd_vars, lcfg, obd); if (rc > 0) rc = 0; if (rc == -ENOSYS) { @@ -294,16 +669,29 @@ static int osp_process_config(const struct lu_env *env, RETURN(rc); } +/** + * Implementation of osp_lu_ops::ldo_recovery_complete + * + * This function is called after recovery is finished, and OSP layer + * will wake up precreate thread here. + * + * \param[in] env execution environment + * \param[in] dev lu_device of OSP + * + * \retval 0 0 unconditionally + */ static int osp_recovery_complete(const struct lu_env *env, struct lu_device *dev) { struct osp_device *osp = lu2osp_dev(dev); - int rc = 0; ENTRY; osp->opd_recovery_completed = 1; - cfs_waitq_signal(&osp->opd_pre_waitq); - RETURN(rc); + + if (!osp->opd_connect_mdt && osp->opd_pre != NULL) + wake_up(&osp->opd_pre_waitq); + + RETURN(0); } const struct lu_device_operations osp_lu_ops = { @@ -313,19 +701,39 @@ const struct lu_device_operations osp_lu_ops = { }; /** - * provides with statfs from corresponded OST + * Implementation of dt_device_operations::dt_statfs + * + * This function provides statfs status (for precreation) from + * corresponding OST. Note: this function only retrieves the status + * from the OSP device, and the real statfs RPC happens inside + * precreate thread (\see osp_statfs_update). Note: OSP for MDT does + * not need to retrieve statfs data for now. + * + * \param[in] env execution environment. + * \param[in] dev dt_device of OSP. + * \param[out] sfs holds the retrieved statfs data. * + * \retval 0 0 statfs data was retrieved successfully or + * retrieval was not needed + * \retval negative negative errno if get statfs failed. */ static int osp_statfs(const struct lu_env *env, struct dt_device *dev, struct obd_statfs *sfs) { struct osp_device *d = dt2osp_dev(dev); + struct obd_import *imp = d->opd_obd->u.cli.cl_import; ENTRY; + if (imp->imp_state == LUSTRE_IMP_CLOSED) + RETURN(-ESHUTDOWN); + if (unlikely(d->opd_imp_active == 0)) RETURN(-ENOTCONN); + if (d->opd_pre == NULL) + RETURN(0); + /* return recently updated data */ *sfs = d->opd_statfs; @@ -333,10 +741,16 @@ static int osp_statfs(const struct lu_env *env, struct dt_device *dev, * layer above osp (usually lod) can use ffree to estimate * how many objects are available for immediate creation */ - cfs_spin_lock(&d->opd_pre_lock); - sfs->os_fprecreated = d->opd_pre_last_created - d->opd_pre_used_id; + spin_lock(&d->opd_pre_lock); + LASSERTF(fid_seq(&d->opd_pre_last_created_fid) == + fid_seq(&d->opd_pre_used_fid), + "last_created "DFID", next_fid "DFID"\n", + PFID(&d->opd_pre_last_created_fid), + PFID(&d->opd_pre_used_fid)); + sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) - + fid_oid(&d->opd_pre_used_fid); sfs->os_fprecreated -= d->opd_pre_reserved; - cfs_spin_unlock(&d->opd_pre_lock); + spin_unlock(&d->opd_pre_lock); LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2); @@ -347,23 +761,153 @@ static int osp_statfs(const struct lu_env *env, struct dt_device *dev, RETURN(0); } +static int osp_sync_timeout(void *data) +{ + return 1; +} + +/** + * Implementation of dt_device_operations::dt_sync + * + * This function synchronizes the OSP cache to the remote target. It wakes + * up unlink log threads and sends out unlink records to the remote OST. + * + * \param[in] env execution environment + * \param[in] dev dt_device of OSP + * + * \retval 0 0 if synchronization succeeds + * \retval negative negative errno if synchronization fails + */ static int osp_sync(const struct lu_env *env, struct dt_device *dev) { + struct osp_device *d = dt2osp_dev(dev); + cfs_time_t expire; + struct l_wait_info lwi = { 0 }; + unsigned long id, old; + int rc = 0; + unsigned long start = cfs_time_current(); ENTRY; - /* - * XXX: wake up sync thread, command it to start flushing asap? - */ + /* No Sync between MDTs yet. */ + if (d->opd_connect_mdt) + RETURN(0); - RETURN(0); + if (unlikely(d->opd_imp_active == 0)) + RETURN(-ENOTCONN); + + id = d->opd_syn_last_used_id; + down_write(&d->opd_async_updates_rwsem); + + CDEBUG(D_OTHER, "%s: async updates %d\n", d->opd_obd->obd_name, + atomic_read(&d->opd_async_updates_count)); + + /* make sure the connection is fine */ + expire = cfs_time_shift(obd_timeout); + lwi = LWI_TIMEOUT(expire - cfs_time_current(), osp_sync_timeout, d); + rc = l_wait_event(d->opd_syn_barrier_waitq, + atomic_read(&d->opd_async_updates_count) == 0, + &lwi); + up_write(&d->opd_async_updates_rwsem); + if (rc != 0) + GOTO(out, rc); + + CDEBUG(D_CACHE, "%s: id: used %lu, processed %lu\n", + d->opd_obd->obd_name, id, d->opd_syn_last_processed_id); + + /* wait till all-in-line are processed */ + while (d->opd_syn_last_processed_id < id) { + + old = d->opd_syn_last_processed_id; + + /* make sure the connection is fine */ + expire = cfs_time_shift(obd_timeout); + lwi = LWI_TIMEOUT(expire - cfs_time_current(), + osp_sync_timeout, d); + l_wait_event(d->opd_syn_barrier_waitq, + d->opd_syn_last_processed_id >= id, + &lwi); + + if (d->opd_syn_last_processed_id >= id) + break; + + if (d->opd_syn_last_processed_id != old) { + /* some progress have been made, + * keep trying... */ + continue; + } + + /* no changes and expired, something is wrong */ + GOTO(out, rc = -ETIMEDOUT); + } + + /* block new processing (barrier>0 - few callers are possible */ + atomic_inc(&d->opd_syn_barrier); + + CDEBUG(D_CACHE, "%s: %u in flight\n", d->opd_obd->obd_name, + d->opd_syn_rpc_in_flight); + + /* wait till all-in-flight are replied, so executed by the target */ + /* XXX: this is used by LFSCK at the moment, which doesn't require + * all the changes to be committed, but in general it'd be + * better to wait till commit */ + while (d->opd_syn_rpc_in_flight > 0) { + + old = d->opd_syn_rpc_in_flight; + + expire = cfs_time_shift(obd_timeout); + lwi = LWI_TIMEOUT(expire - cfs_time_current(), + osp_sync_timeout, d); + l_wait_event(d->opd_syn_barrier_waitq, + d->opd_syn_rpc_in_flight == 0, &lwi); + + if (d->opd_syn_rpc_in_flight == 0) + break; + + if (d->opd_syn_rpc_in_flight != old) { + /* some progress have been made */ + continue; + } + + /* no changes and expired, something is wrong */ + GOTO(out, rc = -ETIMEDOUT); + } + +out: + /* resume normal processing (barrier=0) */ + atomic_dec(&d->opd_syn_barrier); + __osp_sync_check_for_work(d); + + CDEBUG(D_CACHE, "%s: done in %lu: rc = %d\n", d->opd_obd->obd_name, + cfs_time_current() - start, rc); + + RETURN(rc); } const struct dt_device_operations osp_dt_ops = { - .dt_statfs = osp_statfs, - .dt_sync = osp_sync, + .dt_statfs = osp_statfs, + .dt_sync = osp_sync, + .dt_trans_create = osp_trans_create, + .dt_trans_start = osp_trans_start, + .dt_trans_stop = osp_trans_stop, + .dt_trans_cb_add = osp_trans_cb_add, }; -static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m, +/** + * Connect OSP to local OSD. + * + * Locate the local OSD referenced by \a nextdev and connect to it. Sometimes, + * OSP needs to access the local OSD to store some information. For example, + * during precreate, it needs to update last used OID and sequence file + * (LAST_SEQ) in local OSD. + * + * \param[in] env execution environment + * \param[in] osp OSP device + * \param[in] nextdev the name of local OSD + * + * \retval 0 0 connection succeeded + * \retval negative negative errno connection failed + */ +static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *osp, const char *nextdev) { struct obd_connect_data *data = NULL; @@ -372,7 +916,7 @@ static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m, ENTRY; - LASSERT(m->opd_storage_exp == NULL); + LASSERT(osp->opd_storage_exp == NULL); OBD_ALLOC_PTR(data); if (data == NULL) @@ -381,82 +925,188 @@ static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m, obd = class_name2obd(nextdev); if (obd == NULL) { CERROR("%s: can't locate next device: %s\n", - m->opd_obd->obd_name, nextdev); + osp->opd_obd->obd_name, nextdev); GOTO(out, rc = -ENOTCONN); } - rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data, + rc = obd_connect(env, &osp->opd_storage_exp, obd, &obd->obd_uuid, data, NULL); if (rc) { CERROR("%s: cannot connect to next dev %s: rc = %d\n", - m->opd_obd->obd_name, nextdev, rc); + osp->opd_obd->obd_name, nextdev, rc); GOTO(out, rc); } - m->opd_dt_dev.dd_lu_dev.ld_site = - m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site; - LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site); - m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev); + osp->opd_dt_dev.dd_lu_dev.ld_site = + osp->opd_storage_exp->exp_obd->obd_lu_dev->ld_site; + LASSERT(osp->opd_dt_dev.dd_lu_dev.ld_site); + osp->opd_storage = lu2dt_dev(osp->opd_storage_exp->exp_obd->obd_lu_dev); out: OBD_FREE_PTR(data); RETURN(rc); } -static int osp_init0(const struct lu_env *env, struct osp_device *m, +/** + * Determine if the lock needs to be cancelled + * + * Determine if the unused lock should be cancelled before replay, see + * (ldlm_cancel_no_wait_policy()). Currently, only inode bits lock exists + * between MDTs. + * + * \param[in] lock lock to be checked. + * + * \retval 1 if the lock needs to be cancelled before replay. + * \retval 0 if the lock does not need to be cancelled before + * replay. + */ +static int osp_cancel_weight(struct ldlm_lock *lock) +{ + if (lock->l_resource->lr_type != LDLM_IBITS) + RETURN(0); + + RETURN(1); +} + +/** + * Initialize OSP device according to the parameters in the configuration + * log \a cfg. + * + * Reconstruct the local device name from the configuration profile, and + * initialize necessary threads and structures according to the OSP type + * (MDT or OST). + * + * Since there is no record in the MDT configuration for the local disk + * device, we have to extract this from elsewhere in the profile. + * The only information we get at setup is from the OSC records: + * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID + * + * Note: configs generated by Lustre 1.8 are missing the -MDTxxxx part, + * so, we need to reconstruct the name of the underlying OSD from this: + * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd". + * + * \param[in] env execution environment + * \param[in] osp OSP device + * \param[in] ldt lu device type of OSP + * \param[in] cfg configuration log + * + * \retval 0 0 if OSP initialization succeeded. + * \retval negative negative errno if OSP initialization failed. + */ +static int osp_init0(const struct lu_env *env, struct osp_device *osp, struct lu_device_type *ldt, struct lustre_cfg *cfg) { - struct obd_device *obd; - struct obd_import *imp; - class_uuid_t uuid; - char *src, *ost, *mdt, *osdname = NULL; - int rc, idx; + struct obd_device *obd; + struct obd_import *imp; + class_uuid_t uuid; + char *src, *tgt, *mdt, *osdname = NULL; + int rc; + long idx; ENTRY; + mutex_init(&osp->opd_async_requests_mutex); + INIT_LIST_HEAD(&osp->opd_async_updates); + init_rwsem(&osp->opd_async_updates_rwsem); + atomic_set(&osp->opd_async_updates_count, 0); + obd = class_name2obd(lustre_cfg_string(cfg, 0)); if (obd == NULL) { CERROR("Cannot find obd with name %s\n", lustre_cfg_string(cfg, 0)); RETURN(-ENODEV); } - m->opd_obd = obd; - - /* There is no record in the MDT configuration for the local disk - * device, so we have to extract this from elsewhere in the profile. - * The only information we get at setup is from the OSC records: - * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID - * Note that 1.8 generated configs are missing the -MDTxxxx part. - * We need to reconstruct the name of the underlying OSD from this: - * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd". We - * also need to determine the OST index from this - will be used - * to calculate the offset in shared lov_objids file later */ + osp->opd_obd = obd; src = lustre_cfg_string(cfg, 0); if (src == NULL) RETURN(-EINVAL); - ost = strstr(src, "-OST"); - if (ost == NULL) + tgt = strrchr(src, '-'); + if (tgt == NULL) { + CERROR("%s: invalid target name %s: rc = %d\n", + osp->opd_obd->obd_name, lustre_cfg_string(cfg, 0), + -EINVAL); RETURN(-EINVAL); - - idx = simple_strtol(ost + 4, &mdt, 16); - if (mdt[0] != '-' || idx > INT_MAX || idx < 0) { - CERROR("%s: invalid OST index in '%s'\n", obd->obd_name, src); - GOTO(out_fini, rc = -EINVAL); } - m->opd_index = idx; - idx = ost - src; + if (strncmp(tgt, "-osc", 4) == 0) { + /* Old OSC name fsname-OSTXXXX-osc */ + for (tgt--; tgt > src && *tgt != '-'; tgt--) + ; + if (tgt == src) { + CERROR("%s: invalid target name %s: rc = %d\n", + osp->opd_obd->obd_name, + lustre_cfg_string(cfg, 0), -EINVAL); + RETURN(-EINVAL); + } + + if (strncmp(tgt, "-OST", 4) != 0) { + CERROR("%s: invalid target name %s: rc = %d\n", + osp->opd_obd->obd_name, + lustre_cfg_string(cfg, 0), -EINVAL); + RETURN(-EINVAL); + } + + idx = simple_strtol(tgt + 4, &mdt, 16); + if (mdt[0] != '-' || idx > INT_MAX || idx < 0) { + CERROR("%s: invalid OST index in '%s': rc = %d\n", + osp->opd_obd->obd_name, src, -EINVAL); + RETURN(-EINVAL); + } + osp->opd_index = idx; + osp->opd_group = 0; + idx = tgt - src; + } else { + /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */ + if (strncmp(tgt, "-MDT", 4) != 0 && + strncmp(tgt, "-OST", 4) != 0) { + CERROR("%s: invalid target name %s: rc = %d\n", + osp->opd_obd->obd_name, + lustre_cfg_string(cfg, 0), -EINVAL); + RETURN(-EINVAL); + } + + idx = simple_strtol(tgt + 4, &mdt, 16); + if (*mdt != '\0' || idx > INT_MAX || idx < 0) { + CERROR("%s: invalid OST index in '%s': rc = %d\n", + osp->opd_obd->obd_name, src, -EINVAL); + RETURN(-EINVAL); + } + + /* Get MDT index from the name and set it to opd_group, + * which will be used by OSP to connect with OST */ + osp->opd_group = idx; + if (tgt - src <= 12) { + CERROR("%s: invalid mdt index from %s: rc =%d\n", + osp->opd_obd->obd_name, + lustre_cfg_string(cfg, 0), -EINVAL); + RETURN(-EINVAL); + } + + if (strncmp(tgt - 12, "-MDT", 4) == 0) + osp->opd_connect_mdt = 1; + + idx = simple_strtol(tgt - 8, &mdt, 16); + if (mdt[0] != '-' || idx > INT_MAX || idx < 0) { + CERROR("%s: invalid OST index in '%s': rc =%d\n", + osp->opd_obd->obd_name, src, -EINVAL); + RETURN(-EINVAL); + } + + osp->opd_index = idx; + idx = tgt - src - 12; + } /* check the fsname length, and after this everything else will fit */ if (idx > MTI_NAME_MAXLEN) { - CERROR("%s: fsname too long in '%s'\n", obd->obd_name, src); - GOTO(out_fini, rc = -EINVAL); + CERROR("%s: fsname too long in '%s': rc = %d\n", + osp->opd_obd->obd_name, src, -EINVAL); + RETURN(-EINVAL); } OBD_ALLOC(osdname, MAX_OBD_NAME); if (osdname == NULL) - GOTO(out_fini, rc = -ENOMEM); + RETURN(-ENOMEM); memcpy(osdname, src, idx); /* copy just the fsname part */ osdname[idx] = '\0'; @@ -469,11 +1119,21 @@ static int osp_init0(const struct lu_env *env, struct osp_device *m, strcat(osdname, "-osd"); CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src); - m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops; - m->opd_dt_dev.dd_ops = &osp_dt_ops; - obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev; + if (osp->opd_connect_mdt) { + struct client_obd *cli = &osp->opd_obd->u.cli; + + OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock)); + if (!cli->cl_rpc_lock) + GOTO(out_fini, rc = -ENOMEM); + osp_init_rpc_lock(cli->cl_rpc_lock); + } - rc = osp_connect_to_osd(env, m, osdname); + osp->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops; + osp->opd_dt_dev.dd_ops = &osp_dt_ops; + + obd->obd_lu_dev = &osp->opd_dt_dev.dd_lu_dev; + + rc = osp_connect_to_osd(env, osp, osdname); if (rc) GOTO(out_fini, rc); @@ -483,39 +1143,56 @@ static int osp_init0(const struct lu_env *env, struct osp_device *m, rc = client_obd_setup(obd, cfg); if (rc) { - CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc); + CERROR("%s: can't setup obd: rc = %d\n", osp->opd_obd->obd_name, + rc); GOTO(out_ref, rc); } - osp_lprocfs_init(m); + osp_lprocfs_init(osp); - /* - * Initialize last id from the storage - will be used in orphan cleanup - */ - rc = osp_last_used_init(env, m); - if (rc) + rc = obd_fid_init(osp->opd_obd, NULL, osp->opd_connect_mdt ? + LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA); + if (rc) { + CERROR("%s: fid init error: rc = %d\n", + osp->opd_obd->obd_name, rc); GOTO(out_proc, rc); + } - /* - * Initialize precreation thread, it handles new connections as well - */ - rc = osp_init_precreate(m); - if (rc) - GOTO(out_last_used, rc); + if (!osp->opd_connect_mdt) { + /* Initialize last id from the storage - will be + * used in orphan cleanup. */ + rc = osp_last_used_init(env, osp); + if (rc) + GOTO(out_fid, rc); - /* - * Initialize synhronization mechanism taking care of propogating - * changes to OST in near transactional manner - */ - rc = osp_sync_init(env, m); - if (rc) - GOTO(out_precreat, rc); + + /* Initialize precreation thread, it handles new + * connections as well. */ + rc = osp_init_precreate(osp); + if (rc) + GOTO(out_last_used, rc); + + /* + * Initialize synhronization mechanism taking + * care of propogating changes to OST in near + * transactional manner. + */ + rc = osp_sync_init(env, osp); + if (rc < 0) + GOTO(out_precreat, rc); + } else { + rc = osp_update_init(osp); + if (rc != 0) + GOTO(out_fid, rc); + } + + ns_register_cancel(obd->obd_namespace, osp_cancel_weight); /* * Initiate connect to OST */ ll_generate_random_uuid(uuid); - class_uuid_unparse(uuid, &m->opd_cluuid); + class_uuid_unparse(uuid, &osp->opd_cluuid); imp = obd->u.cli.cl_import; @@ -527,110 +1204,168 @@ static int osp_init0(const struct lu_env *env, struct osp_device *m, RETURN(0); out: - /* stop sync thread */ - osp_sync_fini(m); + if (!osp->opd_connect_mdt) + /* stop sync thread */ + osp_sync_fini(osp); out_precreat: /* stop precreate thread */ - osp_precreate_fini(m); + if (!osp->opd_connect_mdt) + osp_precreate_fini(osp); + else + osp_update_fini(env, osp); out_last_used: - osp_last_used_fini(env, m); + if (!osp->opd_connect_mdt) + osp_last_used_fini(env, osp); +out_fid: + obd_fid_fini(osp->opd_obd); out_proc: ptlrpc_lprocfs_unregister_obd(obd); lprocfs_obd_cleanup(obd); - class_destroy_import(obd->u.cli.cl_import); + if (osp->opd_symlink) + lprocfs_remove(&osp->opd_symlink); client_obd_cleanup(obd); out_ref: ptlrpcd_decref(); out_disconnect: - obd_disconnect(m->opd_storage_exp); + if (osp->opd_connect_mdt) { + struct client_obd *cli = &osp->opd_obd->u.cli; + if (cli->cl_rpc_lock != NULL) { + OBD_FREE_PTR(cli->cl_rpc_lock); + cli->cl_rpc_lock = NULL; + } + } + obd_disconnect(osp->opd_storage_exp); out_fini: if (osdname) OBD_FREE(osdname, MAX_OBD_NAME); RETURN(rc); } +/** + * Implementation of lu_device_type_operations::ldto_device_free + * + * Free the OSP device in memory. No return value is needed for now, + * so always return NULL to comply with the interface. + * + * \param[in] env execution environment + * \param[in] lu lu_device of OSP + * + * \retval NULL NULL unconditionally + */ static struct lu_device *osp_device_free(const struct lu_env *env, struct lu_device *lu) { - struct osp_device *m = lu2osp_dev(lu); - - ENTRY; + struct osp_device *osp = lu2osp_dev(lu); - if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) { + if (atomic_read(&lu->ld_ref) && lu->ld_site) { LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL); lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer); } - dt_device_fini(&m->opd_dt_dev); - OBD_FREE_PTR(m); - RETURN(NULL); + dt_device_fini(&osp->opd_dt_dev); + OBD_FREE_PTR(osp); + + return NULL; } +/** + * Implementation of lu_device_type_operations::ldto_device_alloc + * + * This function allocates and initializes OSP device in memory according to + * the config log. + * + * \param[in] env execution environment + * \param[in] type device type of OSP + * \param[in] lcfg config log + * + * \retval pointer the pointer of allocated OSP if succeed. + * \retval ERR_PTR(errno) ERR_PTR(errno) if failed. + */ static struct lu_device *osp_device_alloc(const struct lu_env *env, - struct lu_device_type *t, + struct lu_device_type *type, struct lustre_cfg *lcfg) { - struct osp_device *m; - struct lu_device *l; + struct osp_device *osp; + struct lu_device *ld; - OBD_ALLOC_PTR(m); - if (m == NULL) { - l = ERR_PTR(-ENOMEM); + OBD_ALLOC_PTR(osp); + if (osp == NULL) { + ld = ERR_PTR(-ENOMEM); } else { int rc; - l = osp2lu_dev(m); - dt_device_init(&m->opd_dt_dev, t); - if (is_osp_on_ost(lustre_cfg_string(lcfg, 0))) - rc = osp_init_for_ost(env, m, t, lcfg); - else - rc = osp_init0(env, m, t, lcfg); + ld = osp2lu_dev(osp); + dt_device_init(&osp->opd_dt_dev, type); + rc = osp_init0(env, osp, type, lcfg); if (rc != 0) { - osp_device_free(env, l); - l = ERR_PTR(rc); + osp_device_free(env, ld); + ld = ERR_PTR(rc); } } - return l; + return ld; } +/** + * Implementation of lu_device_type_operations::ldto_device_fini + * + * This function cleans up the OSP device, i.e. release and free those + * attached items in osp_device. + * + * \param[in] env execution environment + * \param[in] ld lu_device of OSP + * + * \retval NULL NULL if cleanup succeeded. + * \retval ERR_PTR(errno) ERR_PTR(errno) if cleanup failed. + */ static struct lu_device *osp_device_fini(const struct lu_env *env, - struct lu_device *d) + struct lu_device *ld) { - struct osp_device *m = lu2osp_dev(d); + struct osp_device *osp = lu2osp_dev(ld); struct obd_import *imp; int rc; ENTRY; - if (m->opd_storage_exp) - obd_disconnect(m->opd_storage_exp); - - if (is_osp_on_ost(m->opd_obd->obd_name)) - osp_fini_for_ost(m); + if (osp->opd_async_requests != NULL) { + osp_update_request_destroy(osp->opd_async_requests); + osp->opd_async_requests = NULL; + } - imp = m->opd_obd->u.cli.cl_import; + if (osp->opd_storage_exp) + obd_disconnect(osp->opd_storage_exp); - if (imp->imp_rq_pool) { - ptlrpc_free_rq_pool(imp->imp_rq_pool); - imp->imp_rq_pool = NULL; - } + imp = osp->opd_obd->u.cli.cl_import; - obd_cleanup_client_import(m->opd_obd); + if (osp->opd_symlink) + lprocfs_remove(&osp->opd_symlink); - if (m->opd_symlink) - lprocfs_remove(&m->opd_symlink); + LASSERT(osp->opd_obd); + ptlrpc_lprocfs_unregister_obd(osp->opd_obd); + lprocfs_obd_cleanup(osp->opd_obd); - LASSERT(m->opd_obd); - ptlrpc_lprocfs_unregister_obd(m->opd_obd); - lprocfs_obd_cleanup(m->opd_obd); + if (osp->opd_connect_mdt) { + struct client_obd *cli = &osp->opd_obd->u.cli; + if (cli->cl_rpc_lock != NULL) { + OBD_FREE_PTR(cli->cl_rpc_lock); + cli->cl_rpc_lock = NULL; + } + } - rc = client_obd_cleanup(m->opd_obd); - LASSERTF(rc == 0, "error %d\n", rc); + rc = client_obd_cleanup(osp->opd_obd); + if (rc != 0) { + ptlrpcd_decref(); + RETURN(ERR_PTR(rc)); + } ptlrpcd_decref(); RETURN(NULL); } +/** + * Implementation of obd_ops::o_reconnect + * + * This function is empty and does not need to do anything for now. + */ static int osp_reconnect(const struct lu_env *env, struct obd_export *exp, struct obd_device *obd, struct obd_uuid *cluuid, @@ -641,7 +1376,24 @@ static int osp_reconnect(const struct lu_env *env, } /* - * we use exports to track all LOD users + * Implementation of obd_ops::o_connect + * + * Connect OSP to the remote target (MDT or OST). Allocate the + * export and return it to the LOD, which calls this function + * for each OSP to connect it to the remote target. This function + * is currently only called once per OSP. + * + * \param[in] env execution environment + * \param[out] exp export connected to OSP + * \param[in] obd OSP device + * \param[in] cluuid OSP device client uuid + * \param[in] data connect_data to be used to connect to the remote + * target + * \param[in] localdata necessary for the API interface, but not used in + * this function + * + * \retval 0 0 if the connection succeeded. + * \retval negative negative errno if the connection failed. */ static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp, struct obd_device *obd, struct obd_uuid *cluuid, @@ -662,35 +1414,23 @@ static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp, RETURN(rc); *exp = class_conn2export(&conn); - if (is_osp_on_ost(obd->obd_name)) - osp->opd_exp = *exp; - /* Why should there ever be more than 1 connect? */ osp->opd_connects++; LASSERT(osp->opd_connects == 1); + osp->opd_exp = *exp; + imp = osp->opd_obd->u.cli.cl_import; imp->imp_dlm_handle = conn; + LASSERT(data != NULL); + LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX); ocd = &imp->imp_connect_data; - ocd->ocd_connect_flags = OBD_CONNECT_AT | - OBD_CONNECT_FULL20 | - OBD_CONNECT_INDEX | -#ifdef HAVE_LRU_RESIZE_SUPPORT - OBD_CONNECT_LRU_RESIZE | -#endif - OBD_CONNECT_MDS | - OBD_CONNECT_OSS_CAPA | - OBD_CONNECT_REQPORTAL | - OBD_CONNECT_SKIP_ORPHAN | - OBD_CONNECT_VERSION | - OBD_CONNECT_FID; - - if (is_osp_on_ost(osp->opd_obd->obd_name)) - ocd->ocd_connect_flags |= OBD_CONNECT_LIGHTWEIGHT; + *ocd = *data; + + imp->imp_connect_flags_orig = ocd->ocd_connect_flags; ocd->ocd_version = LUSTRE_VERSION_CODE; - LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX); ocd->ocd_index = data->ocd_index; imp->imp_connect_flags_orig = ocd->ocd_connect_flags; @@ -698,17 +1438,27 @@ static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp, if (rc) { CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc); GOTO(out, rc); + } else { + osp->opd_obd->u.cli.cl_seq->lcs_exp = + class_export_get(osp->opd_exp); } ptlrpc_pinger_add_import(imp); - out: RETURN(rc); } -/* - * once last export (we don't count self-export) disappeared - * osp can be released +/** + * Implementation of obd_ops::o_disconnect + * + * Disconnect the export for the OSP. This is called by LOD to release the + * OSP during cleanup (\see lod_del_device()). The OSP will be released after + * the export is released. + * + * \param[in] exp export to be disconnected. + * + * \retval 0 0 if disconnection succeed + * \retval negative negative errno if disconnection failed */ static int osp_obd_disconnect(struct obd_export *exp) { @@ -729,17 +1479,29 @@ static int osp_obd_disconnect(struct obd_export *exp) } /* destroy the device */ - if (!is_osp_on_ost(obd->obd_name)) - class_manual_cleanup(obd); + class_manual_cleanup(obd); RETURN(rc); } -/* - * lprocfs helpers still use OBD API, let's keep obd_statfs() support +/** + * Implementation of obd_ops::o_statfs + * + * Send a RPC to the remote target to get statfs status. This is only used + * in lprocfs helpers by obd_statfs. + * + * \param[in] env execution environment + * \param[in] exp connection state from this OSP to the parent (LOD) + * device + * \param[out] osfs hold the statfs result + * \param[in] unused Not used in this function for now + * \param[in] flags flags to indicate how OSP will issue the RPC + * + * \retval 0 0 if statfs succeeded. + * \retval negative negative errno if statfs failed. */ static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp, - struct obd_statfs *osfs, __u64 max_age, __u32 flags) + struct obd_statfs *osfs, __u64 unused, __u32 flags) { struct obd_statfs *msfs; struct ptlrpc_request *req; @@ -750,19 +1512,13 @@ static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp, /* Since the request might also come from lprocfs, so we need * sync this with client_disconnect_export Bug15684 */ - cfs_down_read(&exp->exp_obd->u.cli.cl_sem); + down_read(&exp->exp_obd->u.cli.cl_sem); if (exp->exp_obd->u.cli.cl_import) imp = class_import_get(exp->exp_obd->u.cli.cl_import); - cfs_up_read(&exp->exp_obd->u.cli.cl_sem); + up_read(&exp->exp_obd->u.cli.cl_sem); if (!imp) RETURN(-ENODEV); - /* We could possibly pass max_age in the request (as an absolute - * timestamp or a "seconds.usec ago") so the target can avoid doing - * extra calls into the filesystem if that isn't necessary (e.g. - * during mount that would help a bit). Having relative timestamps - * is not so great if request processing is slow, while absolute - * timestamps are not ideal because they need time synchronization. */ req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS); class_import_put(imp); @@ -801,6 +1557,20 @@ out: return rc; } +/** + * Implementation of obd_ops::o_import_event + * + * This function is called when some related import event happens. It will + * mark the necessary flags according to the event and notify the necessary + * threads (mainly precreate thread). + * + * \param[in] obd OSP OBD device + * \param[in] imp import attached from OSP to remote (OST/MDT) service + * \param[in] event event related to remote service (IMP_EVENT_*) + * + * \retval 0 0 if the event handling succeeded. + * \retval negative negative errno if the event handling failed. + */ static int osp_import_event(struct obd_device *obd, struct obd_import *imp, enum obd_import_event event) { @@ -810,33 +1580,49 @@ static int osp_import_event(struct obd_device *obd, struct obd_import *imp, case IMP_EVENT_DISCON: d->opd_got_disconnected = 1; d->opd_imp_connected = 0; - if (is_osp_on_ost(d->opd_obd->obd_name)) + if (d->opd_connect_mdt) break; - osp_pre_update_status(d, -ENODEV); - cfs_waitq_signal(&d->opd_pre_waitq); + + if (d->opd_pre != NULL) { + osp_pre_update_status(d, -ENODEV); + wake_up(&d->opd_pre_waitq); + } + CDEBUG(D_HA, "got disconnected\n"); break; case IMP_EVENT_INACTIVE: d->opd_imp_active = 0; - if (is_osp_on_ost(d->opd_obd->obd_name)) + d->opd_imp_connected = 0; + if (d->opd_connect_mdt) break; - osp_pre_update_status(d, -ENODEV); - cfs_waitq_signal(&d->opd_pre_waitq); + + if (d->opd_pre != NULL) { + osp_pre_update_status(d, -ENODEV); + wake_up(&d->opd_pre_waitq); + } + CDEBUG(D_HA, "got inactive\n"); break; case IMP_EVENT_ACTIVE: d->opd_imp_active = 1; + if (d->opd_got_disconnected) d->opd_new_connection = 1; d->opd_imp_connected = 1; d->opd_imp_seen_connected = 1; - if (is_osp_on_ost(d->opd_obd->obd_name)) + if (d->opd_connect_mdt) break; - cfs_waitq_signal(&d->opd_pre_waitq); + + if (d->opd_pre != NULL) + wake_up(&d->opd_pre_waitq); + __osp_sync_check_for_work(d); CDEBUG(D_HA, "got connected\n"); break; case IMP_EVENT_INVALIDATE: + if (d->opd_connect_mdt) + osp_invalidate_request(d); + if (obd->obd_namespace == NULL) break; ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY); @@ -852,8 +1638,25 @@ static int osp_import_event(struct obd_device *obd, struct obd_import *imp, return 0; } +/** + * Implementation of obd_ops: o_iocontrol + * + * This function is the ioctl handler for OSP. Note: lctl will access the OSP + * directly by ioctl, instead of through the MDS stack. + * + * param[in] cmd ioctl command. + * param[in] exp export of this OSP. + * param[in] len data length of \a karg. + * param[in] karg input argument which is packed as + * obd_ioctl_data + * param[out] uarg pointer to userspace buffer (must access by + * copy_to_user()). + * + * \retval 0 0 if the ioctl handling succeeded. + * \retval negative negative errno if the ioctl handling failed. + */ static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len, - void *karg, void *uarg) + void *karg, void __user *uarg) { struct obd_device *obd = exp->exp_obd; struct osp_device *d; @@ -866,8 +1669,9 @@ static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len, d = lu2osp_dev(obd->obd_lu_dev); LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops); - if (!cfs_try_module_get(THIS_MODULE)) { - CERROR("%s: can't get module. Is it alive?", obd->obd_name); + if (!try_module_get(THIS_MODULE)) { + CERROR("%s: cannot get module '%s'\n", obd->obd_name, + module_name(THIS_MODULE)); return -EINVAL; } @@ -887,27 +1691,144 @@ static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len, break; default: CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name, - cmd, cfs_curproc_comm()); + cmd, current_comm()); rc = -ENOTTY; } - cfs_module_put(THIS_MODULE); + module_put(THIS_MODULE); return rc; } -static int osp_obd_health_check(const struct lu_env *env, - struct obd_device *obd) + +/** + * Implementation of obd_ops::o_get_info + * + * Retrieve information by key. Retrieval starts from the top layer + * (MDT) of the MDS stack and traverses the stack by calling the + * obd_get_info() method of the next sub-layer. + * + * \param[in] env execution environment + * \param[in] exp export of this OSP + * \param[in] keylen length of \a key + * \param[in] key the key + * \param[out] vallen length of \a val + * \param[out] val holds the value returned by the key + * + * \retval 0 0 if getting information succeeded. + * \retval negative negative errno if getting information failed. + */ +static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp, + __u32 keylen, void *key, __u32 *vallen, void *val) { - struct osp_device *d = lu2osp_dev(obd->obd_lu_dev); + int rc = -EINVAL; + + if (KEY_IS(KEY_OSP_CONNECTED)) { + struct obd_device *obd = exp->exp_obd; + struct osp_device *osp; + + if (!obd->obd_set_up || obd->obd_stopping) + RETURN(-EAGAIN); + + osp = lu2osp_dev(obd->obd_lu_dev); + LASSERT(osp); + /* + * 1.8/2.0 behaviour is that OST being connected once at least + * is considered "healthy". and one "healthy" OST is enough to + * allow lustre clients to connect to MDS + */ + RETURN(!osp->opd_imp_seen_connected); + } + + RETURN(rc); +} + +static int osp_obd_set_info_async(const struct lu_env *env, + struct obd_export *exp, + u32 keylen, void *key, + u32 vallen, void *val, + struct ptlrpc_request_set *set) +{ + struct obd_device *obd = exp->exp_obd; + struct obd_import *imp = obd->u.cli.cl_import; + struct osp_device *osp; + struct ptlrpc_request *req; + char *tmp; + int rc; + if (KEY_IS(KEY_SPTLRPC_CONF)) { + sptlrpc_conf_client_adapt(exp->exp_obd); + RETURN(0); + } + + LASSERT(set != NULL); + if (!obd->obd_set_up || obd->obd_stopping) + RETURN(-EAGAIN); + osp = lu2osp_dev(obd->obd_lu_dev); + + req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO); + if (req == NULL) + RETURN(-ENOMEM); + + req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY, + RCL_CLIENT, keylen); + req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL, + RCL_CLIENT, vallen); + if (osp->opd_connect_mdt) + rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SET_INFO); + else + rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO); + if (rc) { + ptlrpc_request_free(req); + RETURN(rc); + } + + tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY); + memcpy(tmp, key, keylen); + tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL); + memcpy(tmp, val, vallen); + + ptlrpc_request_set_replen(req); + ptlrpc_set_add_req(set, req); + ptlrpc_check_set(NULL, set); + + RETURN(0); +} + +/** + * Implementation of obd_ops: o_fid_alloc + * + * Allocate a FID. There are two cases in which OSP performs + * FID allocation. + * + * 1. FID precreation for data objects, which is done in + * osp_precreate_fids() instead of this function. + * 2. FID allocation for each sub-stripe of a striped directory. + * Similar to other FID clients, OSP requests the sequence + * from its corresponding remote MDT, which in turn requests + * sequences from the sequence controller (MDT0). + * + * \param[in] env execution environment + * \param[in] exp export of the OSP + * \param[out] fid FID being allocated + * \param[in] unused necessary for the interface but unused. + * + * \retval 0 0 FID allocated successfully. + * \retval 1 1 FID allocated successfully and new sequence + * requested from seq meta server + * \retval negative negative errno if FID allocation failed. + */ +static int osp_fid_alloc(const struct lu_env *env, struct obd_export *exp, + struct lu_fid *fid, struct md_op_data *unused) +{ + struct client_obd *cli = &exp->exp_obd->u.cli; + struct osp_device *osp = lu2osp_dev(exp->exp_obd->obd_lu_dev); + struct lu_client_seq *seq = cli->cl_seq; ENTRY; - /* - * 1.8/2.0 behaviour is that OST being connected once at least - * is considired "healthy". and one "healty" OST is enough to - * allow lustre clients to connect to MDS - */ - LASSERT(d); - RETURN(!d->opd_imp_seen_connected); + LASSERT(osp->opd_obd->u.cli.cl_seq != NULL); + /* Sigh, fid client is not ready yet */ + LASSERT(osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL); + + RETURN(seq_client_alloc_fid(env, seq, fid)); } /* context key constructor/destructor: mdt_key_init, mdt_key_fini */ @@ -931,7 +1852,7 @@ struct lu_context_key osp_thread_key = { LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info); struct lu_context_key osp_txn_key = { - .lct_tags = LCT_OSP_THREAD, + .lct_tags = LCT_OSP_THREAD | LCT_TX_HANDLE, .lct_init = osp_txn_key_init, .lct_fini = osp_txn_key_fini }; @@ -954,7 +1875,7 @@ static struct lu_device_type osp_device_type = { .ldt_tags = LU_DEVICE_DT, .ldt_name = LUSTRE_OSP_NAME, .ldt_ops = &osp_device_type_ops, - .ldt_ctx_tags = LCT_MD_THREAD + .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD, }; static struct obd_ops osp_obd_device_ops = { @@ -964,54 +1885,93 @@ static struct obd_ops osp_obd_device_ops = { .o_reconnect = osp_reconnect, .o_connect = osp_obd_connect, .o_disconnect = osp_obd_disconnect, - .o_health_check = osp_obd_health_check, + .o_get_info = osp_obd_get_info, + .o_set_info_async = osp_obd_set_info_async, .o_import_event = osp_import_event, .o_iocontrol = osp_iocontrol, .o_statfs = osp_obd_statfs, + .o_fid_init = client_fid_init, + .o_fid_fini = client_fid_fini, + .o_fid_alloc = osp_fid_alloc, }; -static int __init osp_mod_init(void) +struct llog_operations osp_mds_ost_orig_logops; + +/** + * Initialize OSP module. + * + * Register device types OSP and Light Weight Proxy (LWP) (\see lwp_dev.c) + * in obd_types (\see class_obd.c). Initialize procfs for the + * the OSP device. Note: OSP was called OSC before Lustre 2.4, + * so for compatibility it still uses the name "osc" in procfs. + * This is called at module load time. + * + * \retval 0 0 if initialization succeeds. + * \retval negative negative errno if initialization failed. + */ +static int __init osp_init(void) { - struct lprocfs_static_vars lvars; - cfs_proc_dir_entry_t *osc_proc_dir; - int rc; + struct obd_type *type; + int rc; rc = lu_kmem_init(osp_caches); if (rc) return rc; - lprocfs_osp_init_vars(&lvars); - rc = class_register_type(&osp_obd_device_ops, NULL, lvars.module_vars, + rc = class_register_type(&osp_obd_device_ops, NULL, true, NULL, LUSTRE_OSP_NAME, &osp_device_type); + if (rc != 0) { + lu_kmem_fini(osp_caches); + return rc; + } - /* create "osc" entry in procfs for compatibility purposes */ + rc = class_register_type(&lwp_obd_device_ops, NULL, true, NULL, + LUSTRE_LWP_NAME, &lwp_device_type); if (rc != 0) { + class_unregister_type(LUSTRE_OSP_NAME); lu_kmem_fini(osp_caches); return rc; } - osc_proc_dir = lprocfs_srch(proc_lustre_root, "osc"); - if (osc_proc_dir == NULL) { - osc_proc_dir = lprocfs_register("osc", proc_lustre_root, NULL, - NULL); - if (IS_ERR(osc_proc_dir)) - CERROR("osp: can't create compat entry \"osc\": %d\n", - (int) PTR_ERR(osc_proc_dir)); + /* Note: add_rec/delcare_add_rec will be only used by catalogs */ + osp_mds_ost_orig_logops = llog_osd_ops; + osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec; + osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec; + + /* create "osc" entry in procfs for compatibility purposes */ + type = class_search_type(LUSTRE_OSC_NAME); + if (type != NULL && type->typ_procroot != NULL) + return rc; + + type = class_search_type(LUSTRE_OSP_NAME); + type->typ_procsym = lprocfs_register("osc", proc_lustre_root, + NULL, NULL); + if (IS_ERR(type->typ_procsym)) { + CERROR("osp: can't create compat entry \"osc\": %d\n", + (int) PTR_ERR(type->typ_procsym)); + type->typ_procsym = NULL; } return rc; } -static void __exit osp_mod_exit(void) +/** + * Finalize OSP module. + * + * This callback is called when kernel unloads OSP module from memory, and + * it will deregister OSP and LWP device type from obd_types (\see class_obd.c). + */ +static void __exit osp_exit(void) { - lprocfs_try_remove_proc_entry("osc", proc_lustre_root); - + class_unregister_type(LUSTRE_LWP_NAME); class_unregister_type(LUSTRE_OSP_NAME); lu_kmem_fini(osp_caches); } -MODULE_AUTHOR("Intel, Inc. "); -MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")"); +MODULE_AUTHOR("OpenSFS, Inc. "); +MODULE_DESCRIPTION("Lustre OSD Storage Proxy ("LUSTRE_OSP_NAME")"); +MODULE_VERSION(LUSTRE_VERSION_STRING); MODULE_LICENSE("GPL"); -cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit); +module_init(osp_init); +module_exit(osp_exit);