X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fliblustre%2Frw.c;h=2fefc3113fd035d83f3ad9fbc0535c85e542b3e6;hp=70fd26e43a69205d98cf92b7e9b670865dcde8b5;hb=93fe562c5dd3829939e3bc0533918f66b19776a4;hpb=c5050e412572b00cbe93d8517d2d1f767bebfa92 diff --git a/lustre/liblustre/rw.c b/lustre/liblustre/rw.c index 70fd26e..2fefc31 100644 --- a/lustre/liblustre/rw.c +++ b/lustre/liblustre/rw.c @@ -1,24 +1,41 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Lustre Light Super operations + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * Copyright (c) 2002, 2003 Cluster File Systems, Inc. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * This file is part of Lustre, http://www.lustre.org. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * 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 * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * 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. * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * GPL HEADER END + */ +/* + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2011, 2013, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lustre/liblustre/rw.c + * + * Lustre Light block IO */ #define DEBUG_SUBSYSTEM S_LLITE @@ -28,314 +45,258 @@ #include #include #include +#include #include - -#include -#include -#include -#include -#include +#include +#include #include "llite_lib.h" -#if 0 -void llu_pgcache_remove_extent(struct inode *inode, struct lov_stripe_md *lsm, - struct ldlm_lock *lock) -{ - clear_bit(LLI_F_HAVE_SIZE_LOCK, &(llu_i2info(inode)->lli_flags)); -#if 0 - struct ldlm_extent *extent = &lock->l_extent; - unsigned long start, end, count, skip, i, j; - struct page *page; - int ret; - ENTRY; - - CDEBUG(D_INODE, "obdo %lu inode %p ["LPU64"->"LPU64"] size: %llu\n", - inode->i_ino, inode, extent->start, extent->end, inode->i_size); - - start = extent->start >> PAGE_CACHE_SHIFT; - count = ~0; - skip = 0; - end = (extent->end >> PAGE_CACHE_SHIFT) + 1; - if ((end << PAGE_CACHE_SHIFT) < extent->end) - end = ~0; - if (lsm->lsm_stripe_count > 1) { - struct { - char name[16]; - struct ldlm_lock *lock; - struct lov_stripe_md *lsm; - } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm }; - __u32 stripe; - __u32 vallen = sizeof(stripe); - int rc; - - /* get our offset in the lov */ - rc = obd_get_info(ll_i2obdconn(inode), sizeof(key), - &key, &vallen, &stripe); - if (rc != 0) { - CERROR("obd_get_info: rc = %d\n", rc); - LBUG(); - } - LASSERT(stripe < lsm->lsm_stripe_count); - - count = lsm->lsm_stripe_size >> PAGE_CACHE_SHIFT; - skip = (lsm->lsm_stripe_count - 1) * count; - start += (start/count * skip) + (stripe * count); - if (end != ~0) - end += (end/count * skip) + (stripe * count); - } - - i = (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT; - if (end >= i) - clear_bit(LLI_F_HAVE_SIZE_LOCK, &(ll_i2info(inode)->lli_flags)); - if (i < end) - end = i; - - CDEBUG(D_INODE, "start: %lu j: %lu count: %lu skip: %lu end: %lu\n", - start, start % count, count, skip, end); - - /* start writeback on dirty pages in the extent when its PW */ - for (i = start, j = start % count; - lock->l_granted_mode == LCK_PW && i < end; j++, i++) { - if (j == count) { - i += skip; - j = 0; - } - /* its unlikely, but give us a chance to bail when we're out */ - PGCACHE_WRLOCK(inode->i_mapping); - if (list_empty(&inode->i_mapping->dirty_pages)) { - CDEBUG(D_INODE, "dirty list empty\n"); - PGCACHE_WRUNLOCK(inode->i_mapping); - break; - } - PGCACHE_WRUNLOCK(inode->i_mapping); - - if (need_resched()) - schedule(); - - /* always do a getattr for the first person to pop out of lock - * acquisition.. the DID_GETATTR flag and semaphore serialize - * this initial race. we used to make a decision based on whether - * the lock was matched or acquired, but the matcher could win the - * waking race with the first issuer so that was no good.. - */ - if (test_bit(LLI_F_DID_GETATTR, &lli->lli_flags)) - RETURN(ELDLM_OK); - - down(&lli->lli_getattr_sem); +typedef ssize_t llu_file_piov_t(const struct iovec *iovec, int iovlen, + _SYSIO_OFF_T pos, ssize_t len, + void *private); - if (!test_bit(LLI_F_DID_GETATTR, &lli->lli_flags)) { - rc = ll_inode_getattr(inode, lsm); - if (rc == 0) { - set_bit(LLI_F_DID_GETATTR, &lli->lli_flags); - } else { - unlock_page(page); - } - page_cache_release(page); +size_t llap_cookie_size; - } - - /* our locks are page granular thanks to osc_enqueue, we invalidate the - * whole page. */ - LASSERT((extent->start & ~PAGE_CACHE_MASK) == 0); - LASSERT(((extent->end+1) & ~PAGE_CACHE_MASK) == 0); - for (i = start, j = start % count ; i < end ; j++, i++) { - if ( j == count ) { - i += skip; - j = 0; - } - PGCACHE_WRLOCK(inode->i_mapping); - if (list_empty(&inode->i_mapping->dirty_pages) && - list_empty(&inode->i_mapping->clean_pages) && - list_empty(&inode->i_mapping->locked_pages)) { - CDEBUG(D_INODE, "nothing left\n"); - PGCACHE_WRUNLOCK(inode->i_mapping); - break; - } - PGCACHE_WRUNLOCK(inode->i_mapping); - if (need_resched()) - schedule(); - page = find_get_page(inode->i_mapping, i); - if (page == NULL) - continue; - CDEBUG(D_INODE, "dropping page %p at %lu\n", page, page->index); - lock_page(page); - if (page->mapping) /* might have raced */ -#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)) - truncate_complete_page(page); -#else - truncate_complete_page(page->mapping, page); -#endif - unlock_page(page); - page_cache_release(page); - } - EXIT; -#endif -} - -int llu_lock_callback(struct ldlm_lock *lock, struct ldlm_lock_desc *new, - void *data, int flag) +static int llu_lock_to_stripe_offset(struct obd_export *exp, + struct lov_stripe_md *lsm, + struct ldlm_lock *lock) { - struct inode *inode = data; - struct llu_inode_info *lli = llu_i2info(inode); - struct lustre_handle lockh = {0}; + struct { + char name[16]; + struct ldlm_lock *lock; + } key = { .name = KEY_LOCK_TO_STRIPE, .lock = lock }; + __u32 stripe, vallen = sizeof(stripe); int rc; ENTRY; - if (inode == NULL) - LBUG(); + if (lsm == NULL || lsm->lsm_stripe_count == 1) + RETURN(0); - switch (flag) { - case LDLM_CB_BLOCKING: - ldlm_lock2handle(lock, &lockh); - rc = ldlm_cli_cancel(&lockh); - if (rc != ELDLM_OK) - CERROR("ldlm_cli_cancel failed: %d\n", rc); - break; - case LDLM_CB_CANCELING: { - /* FIXME: we could be given 'canceling intents' so that we - * could know to write-back or simply throw away the pages - * based on if the cancel comes from a desire to, say, - * read or truncate.. */ - llu_pgcache_remove_extent(inode, lli->lli_smd, lock); - break; - } - default: + /* get our offset in the lov */ + rc = obd_get_info(NULL, exp, sizeof(key), &key, &vallen, &stripe, lsm); + if (rc != 0) { + CERROR("obd_get_info: rc = %d\n", rc); LBUG(); } - - RETURN(0); + LASSERT(stripe < lsm->lsm_stripe_count); + RETURN(stripe); } -#endif -static int llu_extent_lock_callback(struct ldlm_lock *lock, - struct ldlm_lock_desc *new, void *data, - int flag) +int llu_extent_lock_cancel_cb(struct ldlm_lock *lock, + struct ldlm_lock_desc *new, void *data, + int flag) { struct lustre_handle lockh = { 0 }; int rc; ENTRY; - if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) { LDLM_ERROR(lock, "cancelling lock with bad data %p", data); LBUG(); } - + switch (flag) { case LDLM_CB_BLOCKING: ldlm_lock2handle(lock, &lockh); - rc = ldlm_cli_cancel(&lockh); + rc = ldlm_cli_cancel(&lockh, 0); if (rc != ELDLM_OK) CERROR("ldlm_cli_cancel failed: %d\n", rc); break; case LDLM_CB_CANCELING: { - struct inode *inode = llu_inode_from_lock(lock); + struct inode *inode; struct llu_inode_info *lli; - + struct lov_stripe_md *lsm; + __u32 stripe; + __u64 kms; + + /* This lock wasn't granted, don't try to evict pages */ + if (lock->l_req_mode != lock->l_granted_mode) + RETURN(0); + + inode = llu_inode_from_lock(lock); if (!inode) RETURN(0); lli= llu_i2info(inode); - if (!lli) { - I_RELE(inode); - RETURN(0); - } - if (!lli->lli_smd) { - I_RELE(inode); - RETURN(0); - } - -/* - ll_pgcache_remove_extent(inode, lli->lli_smd, lock); - iput(inode); -*/ + if (!lli) + goto iput; + if (!lli->lli_has_smd) + goto iput; + + lsm = ccc_inode_lsm_get(inode); + if (lsm == NULL) + goto iput; + + stripe = llu_lock_to_stripe_offset(llu_i2obdexp(inode), + lsm, lock); + lock_res_and_lock(lock); + kms = ldlm_extent_shift_kms(lock, + lsm->lsm_oinfo[stripe]->loi_kms); + unlock_res_and_lock(lock); + if (lsm->lsm_oinfo[stripe]->loi_kms != kms) + LDLM_DEBUG(lock, "updating kms from "LPU64" to "LPU64, + lsm->lsm_oinfo[stripe]->loi_kms, kms); + loi_kms_set(lsm->lsm_oinfo[stripe], kms); + ccc_inode_lsm_put(inode, lsm); +iput: I_RELE(inode); break; } default: LBUG(); } - + RETURN(0); } -int llu_extent_lock_no_validate(struct ll_file_data *fd, - struct inode *inode, - struct lov_stripe_md *lsm, - int mode, - struct ldlm_extent *extent, - struct lustre_handle *lockh, - int ast_flags) +static int llu_glimpse_callback(struct ldlm_lock *lock, void *reqp) { - struct llu_sb_info *sbi = llu_i2sbi(inode); - struct llu_inode_info *lli = llu_i2info(inode); - int rc; + struct ptlrpc_request *req = reqp; + struct inode *inode = llu_inode_from_lock(lock); + struct llu_inode_info *lli; + struct ost_lvb *lvb; + struct lov_stripe_md *lsm; + int rc, stripe = 0; ENTRY; - LASSERT(lockh->cookie == 0); - -#if 0 - /* XXX phil: can we do this? won't it screw the file size up? */ - if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) || - (sbi->ll_flags & LL_SBI_NOLCK)) - RETURN(0); -#endif - - CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n", - lli->lli_st_ino, extent->start, extent->end); + if (inode == NULL) + GOTO(out, rc = -ELDLM_NO_LOCK_DATA); + lli = llu_i2info(inode); + if (lli == NULL) + GOTO(iput, rc = -ELDLM_NO_LOCK_DATA); + + req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK); + if (exp_connect_lvb_type(req->rq_export)) + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, + sizeof(*lvb)); + else + req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, + sizeof(struct ost_lvb_v1)); + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) { + CERROR("failed pack reply: %d\n", rc); + GOTO(iput, rc); + } - rc = obd_enqueue(sbi->ll_osc_exp, lsm, NULL, LDLM_EXTENT, extent, - sizeof(extent), mode, &ast_flags, - llu_extent_lock_callback, inode, lockh); + lsm = ccc_inode_lsm_get(inode); + if (lsm == NULL) + GOTO(iput, rc = -ELDLM_NO_LOCK_DATA); + + /* First, find out which stripe index this lock corresponds to. */ + stripe = llu_lock_to_stripe_offset(llu_i2obdexp(inode), lsm, lock); + + lvb = req_capsule_server_get(&req->rq_pill, &RMF_DLM_LVB); + lvb->lvb_size = lsm->lsm_oinfo[stripe]->loi_kms; + ccc_inode_lsm_put(inode, lsm); + + LDLM_DEBUG(lock, "i_size: "LPU64" -> stripe number %u -> kms "LPU64, + (__u64)llu_i2stat(inode)->st_size, stripe,lvb->lvb_size); + iput: + I_RELE(inode); + out: + /* These errors are normal races, so we don't want to fill the console + * with messages by calling ptlrpc_error() */ + if (rc == -ELDLM_NO_LOCK_DATA) + lustre_pack_reply(req, 1, NULL, NULL); + + req->rq_status = rc; + return rc; +} - RETURN(rc); +int llu_merge_lvb(const struct lu_env *env, struct inode *inode) +{ + struct llu_inode_info *lli = llu_i2info(inode); + struct cl_object *obj = lli->lli_clob; + struct intnl_stat *st = llu_i2stat(inode); + struct cl_attr *attr = ccc_env_thread_attr(env); + struct ost_lvb lvb; + int rc; + ENTRY; + + /* merge timestamps the most recently obtained from mds with + timestamps obtained from osts */ + LTIME_S(inode->i_atime) = lli->lli_lvb.lvb_atime; + LTIME_S(inode->i_mtime) = lli->lli_lvb.lvb_mtime; + LTIME_S(inode->i_ctime) = lli->lli_lvb.lvb_ctime; + + inode_init_lvb(inode, &lvb); + + cl_object_attr_lock(obj); + rc = cl_object_attr_get(env, obj, attr); + cl_object_attr_unlock(obj); + if (rc == 0) { + if (lvb.lvb_atime < attr->cat_atime) + lvb.lvb_atime = attr->cat_atime; + if (lvb.lvb_ctime < attr->cat_ctime) + lvb.lvb_ctime = attr->cat_ctime; + if (lvb.lvb_mtime < attr->cat_mtime) + lvb.lvb_mtime = attr->cat_mtime; + + st->st_size = lvb.lvb_size; + st->st_blocks = lvb.lvb_blocks; + st->st_mtime = lvb.lvb_mtime; + st->st_atime = lvb.lvb_atime; + st->st_ctime = lvb.lvb_ctime; + } + + RETURN(rc); } -/* - * this grabs a lock and manually implements behaviour that makes it look like - * the OST is returning the file size with each lock acquisition. - */ int llu_extent_lock(struct ll_file_data *fd, struct inode *inode, struct lov_stripe_md *lsm, int mode, - struct ldlm_extent *extent, struct lustre_handle *lockh) + ldlm_policy_data_t *policy, struct lustre_handle *lockh, + int ast_flags) { - struct llu_inode_info *lli = llu_i2info(inode); - struct obd_export *exp = llu_i2obdexp(inode); - struct ldlm_extent size_lock; - struct lustre_handle match_lockh = {0}; - int flags, rc, matched; + struct llu_sb_info *sbi = llu_i2sbi(inode); + struct intnl_stat *st = llu_i2stat(inode); + struct ldlm_enqueue_info einfo = { 0 }; + struct obd_info oinfo = { { { 0 } } }; + struct ost_lvb lvb; + int rc; ENTRY; - rc = llu_extent_lock_no_validate(fd, inode, lsm, mode, extent, lockh, 0); - if (rc != ELDLM_OK) - RETURN(rc); + LASSERT(!lustre_handle_is_used(lockh)); + CLASSERT(ELDLM_OK == 0); - if (test_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags)) + /* XXX phil: can we do this? won't it screw the file size up? */ + if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) || + (sbi->ll_flags & LL_SBI_NOLCK) || mode == LCK_NL) RETURN(0); - rc = llu_inode_getattr(inode, lsm); - if (rc) { - llu_extent_unlock(fd, inode, lsm, mode, lockh); - RETURN(rc); + CDEBUG(D_DLMTRACE, "Locking inode "LPU64", start "LPU64" end "LPU64"\n", + (__u64)st->st_ino, policy->l_extent.start, + policy->l_extent.end); + + einfo.ei_type = LDLM_EXTENT; + einfo.ei_mode = mode; + einfo.ei_cb_bl = llu_extent_lock_cancel_cb; + einfo.ei_cb_cp = ldlm_completion_ast; + einfo.ei_cb_gl = llu_glimpse_callback; + einfo.ei_cbdata = inode; + + oinfo.oi_policy = *policy; + oinfo.oi_lockh = lockh; + oinfo.oi_md = lsm; + oinfo.oi_flags = ast_flags; + + rc = obd_enqueue(sbi->ll_dt_exp, &oinfo, &einfo, NULL); + *policy = oinfo.oi_policy; + if (rc > 0) + rc = -EIO; + + inode_init_lvb(inode, &lvb); + obd_merge_lvb(sbi->ll_dt_exp, lsm, &lvb, 1); + if (policy->l_extent.start == 0 && + policy->l_extent.end == OBD_OBJECT_EOF) + st->st_size = lvb.lvb_size; + + if (rc == 0) { + st->st_mtime = lvb.lvb_mtime; + st->st_atime = lvb.lvb_atime; + st->st_ctime = lvb.lvb_ctime; } - size_lock.start = lli->lli_st_size; - size_lock.end = OBD_OBJECT_EOF; - - /* XXX I bet we should be checking the lock ignore flags.. */ - flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED; - matched = obd_match(exp, lsm, LDLM_EXTENT, &size_lock, - sizeof(size_lock), LCK_PR, &flags, inode, - &match_lockh); - - /* hey, alright, we hold a size lock that covers the size we - * just found, its not going to change for a while.. */ - if (matched == 1) { - set_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags); - obd_cancel(exp, lsm, LCK_PR, &match_lockh); - } - - RETURN(0); + RETURN(rc); } int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode, @@ -345,458 +306,236 @@ int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode, struct llu_sb_info *sbi = llu_i2sbi(inode); int rc; ENTRY; -#if 0 + + CLASSERT(ELDLM_OK == 0); + /* XXX phil: can we do this? won't it screw the file size up? */ if ((fd && (fd->fd_flags & LL_FILE_IGNORE_LOCK)) || - (sbi->ll_flags & LL_SBI_NOLCK)) + (sbi->ll_flags & LL_SBI_NOLCK) || mode == LCK_NL) RETURN(0); -#endif - rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh); - - RETURN(rc); -} - -#define LLAP_MAGIC 12346789 - -struct ll_async_page { - int llap_magic; - void *llap_cookie; - int llap_queued; - struct page *llap_page; - struct inode *llap_inode; -}; - -static struct ll_async_page *llap_from_cookie(void *cookie) -{ - struct ll_async_page *llap = cookie; - if (llap->llap_magic != LLAP_MAGIC) - return ERR_PTR(-EINVAL); - return llap; -}; - -static void llu_ap_fill_obdo(void *data, int cmd, struct obdo *oa) -{ - struct ll_async_page *llap; - struct inode *inode; - struct lov_stripe_md *lsm; - obd_flag valid_flags; - ENTRY; - - llap = llap_from_cookie(data); - if (IS_ERR(llap)) { - EXIT; - return; - } - inode = llap->llap_inode; - lsm = llu_i2info(inode)->lli_smd; + rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh); - oa->o_id = lsm->lsm_object_id; - oa->o_valid = OBD_MD_FLID; - valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME; - if (cmd == OBD_BRW_WRITE) - valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME; - - obdo_from_inode(oa, inode, valid_flags); - EXIT; -} - -/* called for each page in a completed rpc.*/ -static void llu_ap_completion(void *data, int cmd, int rc) -{ - struct ll_async_page *llap; - struct page *page; - - llap = llap_from_cookie(data); - if (IS_ERR(llap)) { - EXIT; - return; - } - - llap->llap_queued = 0; - page = llap->llap_page; - - if (rc != 0) { - if (cmd == OBD_BRW_WRITE) - CERROR("writeback error on page %p index %ld: %d\n", - page, page->index, rc); - } - EXIT; + RETURN(rc); } -static struct obd_async_page_ops llu_async_page_ops = { - .ap_make_ready = NULL, - .ap_refresh_count = NULL, - .ap_fill_obdo = llu_ap_fill_obdo, - .ap_completion = llu_ap_completion, -}; - static -struct llu_sysio_cookie* get_sysio_cookie(struct inode *inode, int npages) +ssize_t llu_file_prwv(const struct iovec *iovec, int iovlen, + _SYSIO_OFF_T pos, ssize_t len, + void *private) { - struct llu_sysio_cookie *cookie; - - OBD_ALLOC(cookie, LLU_SYSIO_COOKIE_SIZE(npages)); - if (cookie) { - I_REF(inode); - cookie->lsc_inode = inode; - cookie->lsc_npages = npages; - cookie->lsc_llap = (struct ll_async_page *)(cookie + 1); - cookie->lsc_pages = (struct page *) (cookie->lsc_llap + npages); + struct llu_io_session *session = (struct llu_io_session *) private; + struct inode *inode = session->lis_inode; + struct llu_inode_info *lli = llu_i2info(inode); + int err; + struct lu_env *env; + struct cl_io *io; + struct slp_io *sio; + int refcheck; + ENTRY; - osic_init(&cookie->lsc_osic); - } + /* in a large iov read/write we'll be repeatedly called. + * so give a chance to answer cancel ast here + */ + liblustre_wait_event(0); - return cookie; -} + if (len == 0 || iovlen == 0) + RETURN(0); -static -void put_sysio_cookie(struct llu_sysio_cookie *cookie) -{ - struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd; - struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode); - struct ll_async_page *llap = cookie->lsc_llap; - int i; - - for (i = 0; i< cookie->lsc_npages; i++) { - if (llap[i].llap_cookie) - obd_teardown_async_page(exp, lsm, NULL, - llap[i].llap_cookie); + if (pos + len > lli->lli_maxbytes) + RETURN(-ERANGE); + + env = cl_env_get(&refcheck); + if (IS_ERR(env)) + RETURN(PTR_ERR(env)); + + /* Do NOT call "ccc_env_thread_io()" again to prevent reinitializing */ + io = &ccc_env_info(env)->cti_io; + if (cl_io_rw_init(env, io, session->lis_cmd == OBD_BRW_WRITE?CIT_WRITE: + CIT_READ, + pos, len) == 0) { + struct ccc_io *cio; + sio = slp_env_io(env); + cio = ccc_env_io(env); + /* XXX this is not right: cio->cui_iov can be modified. */ + cio->cui_iov = (struct iovec *)iovec; + cio->cui_nrsegs = iovlen; + cio->cui_tot_nrsegs = iovlen; + sio->sio_session = session; + err = cl_io_loop(env, io); + } else { + /* XXX WTF? */ + LBUG(); } + cl_io_fini(env, io); + cl_env_put(env, &refcheck); - I_RELE(cookie->lsc_inode); + if (err < 0) + RETURN(err); - OBD_FREE(cookie, LLU_SYSIO_COOKIE_SIZE(cookie->lsc_npages)); + RETURN(len); } static -int llu_prep_async_io(struct llu_sysio_cookie *cookie, int cmd, - char *buf, loff_t pos, size_t count) +struct llu_io_session *get_io_session(struct inode *ino, int ngroups, int cmd) { - struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd; - struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode); - struct page *pages = cookie->lsc_pages; - struct ll_async_page *llap = cookie->lsc_llap; - int i, rc, npages = 0; - ENTRY; + struct llu_io_session *session; - if (!exp) - RETURN(-EINVAL); - - cookie->lsc_rwcount = count; - - /* prepare the pages array */ - do { - unsigned long index, offset, bytes; - - offset = (pos & ~PAGE_CACHE_MASK); - index = pos >> PAGE_CACHE_SHIFT; - bytes = PAGE_CACHE_SIZE - offset; - if (bytes > count) - bytes = count; - - /* prepare page for this index */ - pages[npages].index = index; - pages[npages].addr = buf - offset; - - pages[npages]._offset = offset; - pages[npages]._count = bytes; - - npages++; - count -= bytes; - pos += bytes; - buf += bytes; - } while (count); - - for (i = 0; i < npages; i++) { - llap[i].llap_magic = LLAP_MAGIC; - rc = obd_prep_async_page(exp, lsm, NULL, &pages[i], - (obd_off)pages[i].index << PAGE_SHIFT, - &llu_async_page_ops, - &llap[i], &llap[i].llap_cookie); - if (rc) { - llap[i].llap_cookie = NULL; - RETURN(rc); - } - CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", - &llap[i], &pages[i], llap[i].llap_cookie, - (obd_off)pages[i].index << PAGE_SHIFT); - pages[i].private = (unsigned long)&llap[i]; - llap[i].llap_page = &pages[i]; - llap[i].llap_inode = cookie->lsc_inode; - - rc = obd_queue_sync_io(exp, lsm, NULL, &cookie->lsc_osic, - llap[i].llap_cookie, cmd, - pages[i]._offset, pages[i]._count, 0); - if (rc) - RETURN(rc); - - llap[i].llap_queued = 1; - } + OBD_ALLOC_PTR(session); + if (!session) + return NULL; - RETURN(0); + I_REF(ino); + session->lis_inode = ino; + session->lis_max_groups = ngroups; + session->lis_cmd = cmd; + return session; } -static -int llu_start_async_io(struct llu_sysio_cookie *cookie) +static void put_io_session(struct llu_io_session *session) { - struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd; - struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode); - - return obd_trigger_sync_io(exp, lsm, NULL, &cookie->lsc_osic); + I_RELE(session->lis_inode); + OBD_FREE_PTR(session); } -/* - * read/write a continuous buffer for an inode (zero-copy) - */ -struct llu_sysio_cookie* -llu_rw(int cmd, struct inode *inode, char *buf, size_t count, loff_t pos) +static int llu_file_rwx(struct inode *ino, + struct ioctx *ioctx, + int read) { - struct llu_sysio_cookie *cookie; - int max_pages, rc; + struct llu_io_session *session; + ssize_t cc; + int cmd = read ? OBD_BRW_READ : OBD_BRW_WRITE; ENTRY; - max_pages = (count >> PAGE_SHIFT) + 2; - - cookie = get_sysio_cookie(inode, max_pages); - if (!cookie) - RETURN(ERR_PTR(-ENOMEM)); + LASSERT(ioctx->ioctx_xtvlen >= 0); + LASSERT(ioctx->ioctx_iovlen >= 0); - rc = llu_prep_async_io(cookie, cmd, buf, pos, count); - if (rc) - GOTO(out_cleanup, rc); + liblustre_wait_event(0); - rc = llu_start_async_io(cookie); - if (rc) - GOTO(out_cleanup, rc); + if (!ioctx->ioctx_xtvlen) + RETURN(0); -/* - rc = osic_wait(&osic); - if (rc) { - CERROR("file i/o error!\n"); - rw_count = rc; + /* XXX consider other types later */ + if (S_ISDIR(llu_i2stat(ino)->st_mode)) + RETURN(-EISDIR); + if (!S_ISREG(llu_i2stat(ino)->st_mode)) + RETURN(-EOPNOTSUPP); + + session = get_io_session(ino, ioctx->ioctx_xtvlen * 2, cmd); + if (!session) + RETURN(-ENOMEM); + + cc = _sysio_enumerate_extents(ioctx->ioctx_xtv, ioctx->ioctx_xtvlen, + ioctx->ioctx_iov, ioctx->ioctx_iovlen, + llu_file_prwv, session); + + if (cc >= 0) { + LASSERT(!ioctx->ioctx_cc); + ioctx->ioctx_private = session; + cc = 0; + } else { + put_io_session(session); } -*/ - RETURN(cookie); -out_cleanup: - put_sysio_cookie(cookie); - RETURN(ERR_PTR(rc)); + liblustre_wait_event(0); + RETURN(cc); } -struct llu_sysio_callback_args* -llu_file_write(struct inode *inode, const struct iovec *iovec, - size_t iovlen, loff_t pos) +void llu_io_init(struct cl_io *io, struct inode *inode, int write) { struct llu_inode_info *lli = llu_i2info(inode); - struct ll_file_data *fd = lli->lli_file_data; - struct lustre_handle lockh = {0}; - struct lov_stripe_md *lsm = lli->lli_smd; - struct llu_sysio_callback_args *lsca; - struct llu_sysio_cookie *cookie; - struct ldlm_extent extent; - ldlm_error_t err; - int iovidx; - ENTRY; - /* XXX consider other types later */ - if (!S_ISREG(lli->lli_st_mode)) - LBUG(); - - LASSERT(iovlen <= MAX_IOVEC); - - OBD_ALLOC(lsca, sizeof(*lsca)); - if (!lsca) - RETURN(ERR_PTR(-ENOMEM)); - - /* FIXME optimize the following extent locking */ - for (iovidx = 0; iovidx < iovlen; iovidx++) { - char *buf = iovec[iovidx].iov_base; - size_t count = iovec[iovidx].iov_len; - - if (count == 0) - continue; - - /* FIXME libsysio haven't consider the open flags - * such as O_APPEND */ -#if 0 - if (!S_ISBLK(lli->lli_st_mode) && file->f_flags & O_APPEND) { - extent.start = 0; - extent.end = OBD_OBJECT_EOF; - } else { - extent.start = *ppos; - extent.end = *ppos + count - 1; - } -#else - extent.start = pos; - extent.end = pos + count - 1; -#endif - - err = llu_extent_lock(fd, inode, lsm, LCK_PW, &extent, &lockh); - if (err != ELDLM_OK) - GOTO(err_out, err = -ENOLCK); - - CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n", - lli->lli_st_ino, count, pos); - - cookie = llu_rw(OBD_BRW_WRITE, inode, buf, count, pos); - if (!IS_ERR(cookie)) { - /* save cookie */ - lsca->cookies[lsca->ncookies++] = cookie; - pos += count; - /* file size grow. XXX should be done here? */ - if (pos > lli->lli_st_size) { - lli->lli_st_size = pos; - set_bit(LLI_F_PREFER_EXTENDED_SIZE, - &lli->lli_flags); - } - } else { - llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh); - GOTO(err_out, err = PTR_ERR(cookie)); - } - - /* XXX errors? */ - err = llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh); - if (err) - CERROR("extent unlock error %d\n", err); - } + io->u.ci_rw.crw_nonblock = lli->lli_open_flags & O_NONBLOCK; + if (write) + io->u.ci_wr.wr_append = lli->lli_open_flags & O_APPEND; + io->ci_obj = llu_i2info(inode)->lli_clob; - RETURN(lsca); - -err_out: - /* teardown all async stuff */ - while (lsca->ncookies--) { - put_sysio_cookie(lsca->cookies[lsca->ncookies]); - } - OBD_FREE(lsca, sizeof(*lsca)); - - RETURN(ERR_PTR(err)); + if ((lli->lli_open_flags & O_APPEND) && write) + io->ci_lockreq = CILR_MANDATORY; + else + io->ci_lockreq = CILR_NEVER; } -#if 0 -static void llu_update_atime(struct inode *inode) +int llu_iop_read(struct inode *ino, + struct ioctx *ioctx) { - struct llu_inode_info *lli = llu_i2info(inode); + struct intnl_stat *st = llu_i2stat(ino); + struct lu_env *env; + struct cl_io *io; + int refcheck; + int ret; + + /* BUG: 5972 */ + st->st_atime = CFS_CURRENT_TIME; -#ifdef USE_ATIME - struct iattr attr; + env = cl_env_get(&refcheck); + if (IS_ERR(env)) + RETURN(PTR_ERR(env)); - attr.ia_atime = LTIME_S(CURRENT_TIME); - attr.ia_valid = ATTR_ATIME; + io = ccc_env_thread_io(env); + llu_io_init(io, ino, 0); - if (lli->lli_st_atime == attr.ia_atime) return; - if (IS_RDONLY(inode)) return; - if (IS_NOATIME(inode)) return; + ret = llu_file_rwx(ino, ioctx, 1); - /* ll_inode_setattr() sets inode->i_atime from attr.ia_atime */ - llu_inode_setattr(inode, &attr, 0); -#else - /* update atime, but don't explicitly write it out just this change */ - inode->i_atime = CURRENT_TIME; -#endif + cl_env_put(env, &refcheck); + return ret; } -#endif -struct llu_sysio_callback_args* -llu_file_read(struct inode *inode, const struct iovec *iovec, - size_t iovlen, loff_t pos) +int llu_iop_write(struct inode *ino, + struct ioctx *ioctx) { - struct llu_inode_info *lli = llu_i2info(inode); - struct ll_file_data *fd = lli->lli_file_data; - struct lov_stripe_md *lsm = lli->lli_smd; - struct lustre_handle lockh = { 0 }; - struct ldlm_extent extent; - struct llu_sysio_callback_args *lsca; - struct llu_sysio_cookie *cookie; - int iovidx; + struct intnl_stat *st = llu_i2stat(ino); + struct lu_env *env; + struct cl_io *io; + int refcheck; + int ret; - ldlm_error_t err; - ENTRY; + st->st_mtime = st->st_ctime = CFS_CURRENT_TIME; - OBD_ALLOC(lsca, sizeof(*lsca)); - if (!lsca) - RETURN(ERR_PTR(-ENOMEM)); - - for (iovidx = 0; iovidx < iovlen; iovidx++) { - char *buf = iovec[iovidx].iov_base; - size_t count = iovec[iovidx].iov_len; - - /* "If nbyte is 0, read() will return 0 and have no other results." - * -- Single Unix Spec */ - if (count == 0) - continue; - - extent.start = pos; - extent.end = pos + count - 1; - - err = llu_extent_lock(fd, inode, lsm, LCK_PR, &extent, &lockh); - if (err != ELDLM_OK) - GOTO(err_out, err = -ENOLCK); - - CDEBUG(D_INFO, "Reading inode %lu, "LPSZ" bytes, offset %Ld\n", - lli->lli_st_ino, count, pos); - - cookie = llu_rw(OBD_BRW_READ, inode, buf, count, pos); - if (!IS_ERR(cookie)) { - /* save cookie */ - lsca->cookies[lsca->ncookies++] = cookie; - pos += count; - } else { - llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh); - GOTO(err_out, err = PTR_ERR(cookie)); - } - - /* XXX errors? */ - err = llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh); - if (err) - CERROR("extent_unlock fail: %d\n", err); - } -#if 0 - if (readed > 0) - llu_update_atime(inode); -#endif - RETURN(lsca); - -err_out: - /* teardown all async stuff */ - while (lsca->ncookies--) { - put_sysio_cookie(lsca->cookies[lsca->ncookies]); - } - OBD_FREE(lsca, sizeof(*lsca)); + env = cl_env_get(&refcheck); + if (IS_ERR(env)) + RETURN(PTR_ERR(env)); - RETURN(ERR_PTR(err)); + io = ccc_env_thread_io(env); + llu_io_init(io, ino, 1); + + ret = llu_file_rwx(ino, ioctx, 0); + cl_env_put(env, &refcheck); + return ret; } -int llu_iop_iodone(struct ioctx *ioctxp) +int llu_iop_iodone(struct ioctx *ioctx) { - struct llu_sysio_callback_args *lsca = ioctxp->ioctx_private; - struct llu_sysio_cookie *cookie; - int i, err = 0, rc = 0; + struct llu_io_session *session; + struct lu_env *env; + struct cl_io *io; + int refcheck; ENTRY; - /* write/read(fd, buf, 0) */ - if (!lsca) - return 1; - - LASSERT(!IS_ERR(lsca)); - - for (i = 0; i < lsca->ncookies; i++) { - cookie = lsca->cookies[i]; - if (cookie) { - err = osic_wait(&cookie->lsc_osic); - if (err && !rc) - rc = err; - if (!rc) - ioctxp->ioctx_cc += cookie->lsc_rwcount; - put_sysio_cookie(cookie); - } + liblustre_wait_event(0); + + env = cl_env_get(&refcheck); + if (IS_ERR(env)) + RETURN(PTR_ERR(env)); + + io = &ccc_env_info(env)->cti_io; + cl_io_fini(env, io); + cl_env_put(env, &refcheck); + session = (struct llu_io_session *) ioctx->ioctx_private; + LASSERT(session); + LASSERT(!IS_ERR(session)); + + if (session->lis_rc == 0) { + ioctx->ioctx_cc = session->lis_rwcount; + } else { + LASSERT(session->lis_rc < 0); + ioctx->ioctx_cc = -1; + ioctx->ioctx_errno = -session->lis_rc; } - if (rc) - ioctxp->ioctx_cc = rc; - - OBD_FREE(lsca, sizeof(*lsca)); - ioctxp->ioctx_private = NULL; + put_io_session(session); + ioctx->ioctx_private = NULL; + liblustre_wait_event(0); RETURN(1); }