X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;ds=sidebyside;f=lustre%2Fliblustre%2Frw.c;h=b716544be8e97c4e090ed32abc463c118d094409;hb=d2d56f38da01001c92a09afc6b52b5acbd9bc13c;hp=6fe7431ccf16c5561998ebb48581e5bdc334f84c;hpb=ff17cc7e0282d9b1522810e0c5d12171c4d46a2d;p=fs%2Flustre-release.git diff --git a/lustre/liblustre/rw.c b/lustre/liblustre/rw.c index 6fe7431..b716544 100644 --- a/lustre/liblustre/rw.c +++ b/lustre/liblustre/rw.c @@ -1,9 +1,9 @@ /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- * vim:expandtab:shiftwidth=8:tabstop=8: * - * Lustre Light Super operations + * Lustre Light block IO * - * Copyright (c) 2002, 2003 Cluster File Systems, Inc. + * Copyright (c) 2002-2004 Cluster File Systems, Inc. * * This file is part of Lustre, http://www.lustre.org. * @@ -28,18 +28,63 @@ #include #include #include +#include #include +#include +#include +#ifdef HAVE_XTIO_H +#include +#endif #include #include #include #include +#ifdef HAVE_FILE_H #include +#endif #undef LIST_HEAD #include "llite_lib.h" +struct llu_io_group +{ + struct obd_io_group *lig_oig; + struct inode *lig_inode; + struct lustre_rw_params *lig_params; + int lig_maxpages; + int lig_npages; + __u64 lig_rwcount; + struct ll_async_page *lig_llaps; + struct page *lig_pages; + void *lig_llap_cookies; +}; + +#define LLU_IO_GROUP_SIZE(x) \ + (sizeof(struct llu_io_group) + \ + (sizeof(struct ll_async_page) + \ + sizeof(struct page) + \ + llap_cookie_size) * (x)) + +struct llu_io_session +{ + struct inode *lis_inode; + int lis_cmd; + int lis_max_groups; + int lis_ngroups; + struct llu_io_group *lis_groups[0]; +}; +#define LLU_IO_SESSION_SIZE(x) \ + (sizeof(struct llu_io_session) + (x) * 2 * sizeof(void *)) + + +typedef ssize_t llu_file_piov_t(const struct iovec *iovec, int iovlen, + _SYSIO_OFF_T pos, ssize_t len, + void *private); + +size_t llap_cookie_size; + static int llu_lock_to_stripe_offset(struct inode *inode, struct ldlm_lock *lock) { struct llu_inode_info *lli = llu_i2info(inode); @@ -74,13 +119,12 @@ static int llu_extent_lock_callback(struct ldlm_lock *lock, 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); @@ -89,12 +133,17 @@ static int llu_extent_lock_callback(struct ldlm_lock *lock, 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); @@ -105,12 +154,14 @@ static int llu_extent_lock_callback(struct ldlm_lock *lock, lsm = lli->lli_smd; stripe = llu_lock_to_stripe_offset(inode, lock); + lock_res_and_lock(lock); kms = ldlm_extent_shift_kms(lock, - lsm->lsm_oinfo[stripe].loi_kms); - if (lsm->lsm_oinfo[stripe].loi_kms != kms) + 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); - lsm->lsm_oinfo[stripe].loi_kms = kms; + lsm->lsm_oinfo[stripe]->loi_kms, kms); + lsm->lsm_oinfo[stripe]->loi_kms = kms; iput: I_RELE(inode); break; @@ -118,7 +169,7 @@ iput: default: LBUG(); } - + RETURN(0); } @@ -126,209 +177,196 @@ static int llu_glimpse_callback(struct ldlm_lock *lock, void *reqp) { struct ptlrpc_request *req = reqp; struct inode *inode = llu_inode_from_lock(lock); - struct obd_export *exp; struct llu_inode_info *lli; struct ost_lvb *lvb; - struct { - int stripe_number; - __u64 size; - struct lov_stripe_md *lsm; - } data; - __u32 vallen = sizeof(data); - int rc, size = sizeof(*lvb); + int size[2] = { sizeof(struct ptlrpc_body), sizeof(*lvb) }; + int rc, stripe = 0; ENTRY; if (inode == NULL) - RETURN(0); + GOTO(out, rc = -ELDLM_NO_LOCK_DATA); lli = llu_i2info(inode); if (lli == NULL) - goto iput; + GOTO(iput, rc = -ELDLM_NO_LOCK_DATA); if (lli->lli_smd == NULL) - goto iput; - exp = llu_i2obdexp(inode); + GOTO(iput, rc = -ELDLM_NO_LOCK_DATA); /* First, find out which stripe index this lock corresponds to. */ if (lli->lli_smd->lsm_stripe_count > 1) - data.stripe_number = llu_lock_to_stripe_offset(inode, lock); - else - data.stripe_number = 0; - - data.size = lli->lli_st_size; - data.lsm = lli->lli_smd; - - rc = obd_get_info(exp, strlen("size_to_stripe"), "size_to_stripe", - &vallen, &data); - if (rc != 0) { - CERROR("obd_get_info: rc = %d\n", rc); - LBUG(); - } - - LDLM_DEBUG(lock, "i_size: %Lu -> stripe number %d -> size %Lu", - lli->lli_st_size, data.stripe_number, data.size); + stripe = llu_lock_to_stripe_offset(inode, lock); - rc = lustre_pack_reply(req, 1, &size, NULL); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) { CERROR("lustre_pack_reply: %d\n", rc); - goto iput; + GOTO(iput, rc); } - lvb = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*lvb)); - lvb->lvb_size = data.size; - ptlrpc_reply(req); + lvb = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*lvb)); + lvb->lvb_size = lli->lli_smd->lsm_oinfo[stripe]->loi_kms; + LDLM_DEBUG(lock, "i_size: %llu -> stripe number %u -> kms "LPU64, + (long long)llu_i2stat(inode)->st_size, stripe,lvb->lvb_size); iput: I_RELE(inode); - RETURN(0); + 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; } -__u64 lov_merge_size(struct lov_stripe_md *lsm, int kms); -__u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time); - -/* NB: lov_merge_size will prefer locally cached writes if they extend the - * file (because it prefers KMS over RSS when larger) */ -int llu_glimpse_size(struct inode *inode, struct ost_lvb *lvb) +static void llu_merge_lvb(struct inode *inode) { struct llu_inode_info *lli = llu_i2info(inode); struct llu_sb_info *sbi = llu_i2sbi(inode); - ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } }; - struct lustre_handle lockh; - int rc, flags = LDLM_FL_HAS_INTENT; + struct intnl_stat *st = llu_i2stat(inode); + struct ost_lvb lvb; ENTRY; - CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", lli->lli_st_ino); - - rc = obd_enqueue(sbi->ll_osc_exp, lli->lli_smd, LDLM_EXTENT, &policy, - LCK_PR, &flags, llu_extent_lock_callback, - ldlm_completion_ast, llu_glimpse_callback, inode, - sizeof(*lvb), lustre_swab_ost_lvb, &lockh); - if (rc > 0) - RETURN(-EIO); - - lvb->lvb_size = lov_merge_size(lli->lli_smd, 0); - //inode->i_mtime = lov_merge_mtime(lli->lli_smd, inode->i_mtime); - - CDEBUG(D_DLMTRACE, "glimpse: size: "LPU64"\n", lvb->lvb_size); + inode_init_lvb(inode, &lvb); + obd_merge_lvb(sbi->ll_dt_exp, lli->lli_smd, &lvb, 0); + 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; + EXIT; +} - obd_cancel(sbi->ll_osc_exp, lli->lli_smd, LCK_PR, &lockh); +int llu_local_size(struct inode *inode) +{ + ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } }; + struct llu_inode_info *lli = llu_i2info(inode); + struct llu_sb_info *sbi = llu_i2sbi(inode); + struct lustre_handle lockh = { 0 }; + int flags = 0; + int rc; + ENTRY; - RETURN(rc); + if (lli->lli_smd->lsm_stripe_count == 0) + RETURN(0); + + rc = obd_match(sbi->ll_dt_exp, lli->lli_smd, LDLM_EXTENT, + &policy, LCK_PR | LCK_PW, &flags, inode, &lockh); + if (rc < 0) + RETURN(rc); + else if (rc == 0) + RETURN(-ENODATA); + + llu_merge_lvb(inode); + obd_cancel(sbi->ll_dt_exp, lli->lli_smd, LCK_PR, &lockh); + RETURN(0); } -int llu_extent_lock(struct ll_file_data *fd, struct inode *inode, - struct lov_stripe_md *lsm, int mode, - ldlm_policy_data_t *policy, struct lustre_handle *lockh, - int ast_flags) +/* NB: lov_merge_size will prefer locally cached writes if they extend the + * file (because it prefers KMS over RSS when larger) */ +int llu_glimpse_size(struct inode *inode) { - struct llu_sb_info *sbi = llu_i2sbi(inode); struct llu_inode_info *lli = llu_i2info(inode); + struct intnl_stat *st = llu_i2stat(inode); + struct llu_sb_info *sbi = llu_i2sbi(inode); + struct lustre_handle lockh = { 0 }; + struct obd_enqueue_info einfo = { 0 }; + struct obd_info oinfo = { { { 0 } } }; int rc; ENTRY; - LASSERT(lockh->cookie == 0); + /* If size is cached on the mds, skip glimpse. */ + if (lli->lli_flags & LLIF_MDS_SIZE_LOCK) + RETURN(0); + + CDEBUG(D_DLMTRACE, "Glimpsing inode %llu\n", (long long)st->st_ino); - /* 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)) + if (!lli->lli_smd) { + CDEBUG(D_DLMTRACE, "No objects for inode %llu\n", + (long long)st->st_ino); RETURN(0); + } - CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n", - lli->lli_st_ino, policy->l_extent.start, policy->l_extent.end); + einfo.ei_type = LDLM_EXTENT; + einfo.ei_mode = LCK_PR; + einfo.ei_flags = LDLM_FL_HAS_INTENT; + einfo.ei_cb_bl = llu_extent_lock_callback; + einfo.ei_cb_cp = ldlm_completion_ast; + einfo.ei_cb_gl = llu_glimpse_callback; + einfo.ei_cbdata = inode; - rc = obd_enqueue(sbi->ll_osc_exp, lsm, LDLM_EXTENT, policy, mode, - &ast_flags, llu_extent_lock_callback, - ldlm_completion_ast, llu_glimpse_callback, inode, - sizeof(struct ost_lvb), lustre_swab_ost_lvb, lockh); - if (rc > 0) - rc = -EIO; + oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF; + oinfo.oi_lockh = &lockh; + oinfo.oi_md = lli->lli_smd; - if (policy->l_extent.start == 0 && - policy->l_extent.end == OBD_OBJECT_EOF) - lli->lli_st_size = lov_merge_size(lsm, 1); + rc = obd_enqueue_rqset(sbi->ll_dt_exp, &oinfo, &einfo); + if (rc) { + CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc); + RETURN(rc > 0 ? -EIO : rc); + } - //inode->i_mtime = lov_merge_mtime(lsm, inode->i_mtime); + llu_merge_lvb(inode); + CDEBUG(D_DLMTRACE, "glimpse: size: %llu, blocks: %llu\n", + (long long)st->st_size, (long long)st->st_blocks); RETURN(rc); } -#if 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) +int llu_extent_lock(struct ll_file_data *fd, struct inode *inode, + struct lov_stripe_md *lsm, int mode, + ldlm_policy_data_t *policy, struct lustre_handle *lockh, + int ast_flags) { struct llu_sb_info *sbi = llu_i2sbi(inode); - struct llu_inode_info *lli = llu_i2info(inode); + struct intnl_stat *st = llu_i2stat(inode); + struct obd_enqueue_info einfo = { 0 }; + struct obd_info oinfo = { { { 0 } } }; + struct ost_lvb lvb; int rc; ENTRY; - LASSERT(lockh->cookie == 0); + LASSERT(!lustre_handle_is_used(lockh)); + 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); - CDEBUG(D_DLMTRACE, "Locking inode %lu, start "LPU64" end "LPU64"\n", - lli->lli_st_ino, extent->start, extent->end); + CDEBUG(D_DLMTRACE, "Locking inode %llu, start "LPU64" end "LPU64"\n", + (long long)st->st_ino, policy->l_extent.start, + policy->l_extent.end); - rc = obd_enqueue(sbi->ll_osc_exp, lsm, NULL, LDLM_EXTENT, extent, - sizeof(extent), mode, &ast_flags, - llu_extent_lock_callback, inode, lockh); + einfo.ei_type = LDLM_EXTENT; + einfo.ei_mode = mode; + einfo.ei_flags = ast_flags; + einfo.ei_cb_bl = llu_extent_lock_callback; + einfo.ei_cb_cp = ldlm_completion_ast; + einfo.ei_cb_gl = llu_glimpse_callback; + einfo.ei_cbdata = inode; - RETURN(rc); -} + oinfo.oi_policy = *policy; + oinfo.oi_lockh = lockh; + oinfo.oi_md = lsm; -/* - * 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) -{ - 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; - ENTRY; - - rc = llu_extent_lock_no_validate(fd, inode, lsm, mode, extent, lockh, 0); - if (rc != ELDLM_OK) - RETURN(rc); - - if (test_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags)) - RETURN(0); + rc = obd_enqueue(sbi->ll_dt_exp, &oinfo, &einfo); + *policy = oinfo.oi_policy; + if (rc > 0) + rc = -EIO; - rc = llu_inode_getattr(inode, lsm); - if (rc) { - llu_extent_unlock(fd, inode, lsm, mode, lockh); - RETURN(rc); - } + 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; - size_lock.start = lli->lli_st_size; - size_lock.end = OBD_OBJECT_EOF; - - /* XXX I bet we should be checking the lock ignore flags.. */ - /* FIXME use LDLM_FL_TEST_LOCK instead */ - 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); + if (rc == 0) { + st->st_mtime = lvb.lvb_mtime; + st->st_atime = lvb.lvb_atime; + st->st_ctime = lvb.lvb_ctime; } - RETURN(0); + RETURN(rc); } -#endif int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode, struct lov_stripe_md *lsm, int mode, @@ -338,12 +376,14 @@ int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode, int rc; ENTRY; + 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); - rc = obd_cancel(sbi->ll_osc_exp, lsm, mode, lockh); + rc = obd_cancel(sbi->ll_dt_exp, lsm, mode, lockh); RETURN(rc); } @@ -358,14 +398,6 @@ struct ll_async_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; @@ -374,564 +406,522 @@ static void llu_ap_fill_obdo(void *data, int cmd, struct obdo *oa) obd_flag valid_flags; ENTRY; - llap = llap_from_cookie(data); - if (IS_ERR(llap)) { - EXIT; - return; - } - + llap = LLAP_FROM_COOKIE(data); inode = llap->llap_inode; lsm = llu_i2info(inode)->lli_smd; 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; + if (cmd & OBD_BRW_WRITE) + valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME | + OBD_MD_FLUID | OBD_MD_FLGID | + OBD_MD_FLFID | OBD_MD_FLGENER; obdo_from_inode(oa, inode, valid_flags); EXIT; } +static void llu_ap_update_obdo(void *data, int cmd, struct obdo *oa, + obd_valid valid) +{ + struct ll_async_page *llap; + ENTRY; + + llap = LLAP_FROM_COOKIE(data); + obdo_from_inode(oa, llap->llap_inode, valid); + + EXIT; +} + /* called for each page in a completed rpc.*/ -static void llu_ap_completion(void *data, int cmd, int rc) +static int llu_ap_completion(void *data, int cmd, struct obdo *oa, int rc) { struct ll_async_page *llap; struct page *page; + ENTRY; - llap = llap_from_cookie(data); - if (IS_ERR(llap)) { - EXIT; - return; - } - + llap = LLAP_FROM_COOKIE(data); 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", + if (cmd & OBD_BRW_WRITE) + CERROR("writeback error on page %p index %ld: %d\n", page, page->index, rc); } - EXIT; + RETURN(0); } 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_update_obdo = llu_ap_update_obdo, .ap_completion = llu_ap_completion, }; -static -struct llu_sysio_cookie* get_sysio_cookie(struct inode *inode, int maxpages) -{ - struct llu_sysio_cookie *cookie; - int rc; - - OBD_ALLOC(cookie, LLU_SYSIO_COOKIE_SIZE(maxpages)); - if (cookie == NULL) - goto out; - - I_REF(inode); - cookie->lsc_inode = inode; - cookie->lsc_maxpages = maxpages; - cookie->lsc_llap = (struct ll_async_page *)(cookie + 1); - cookie->lsc_pages = (struct page *) (cookie->lsc_llap + maxpages); - - rc = oig_init(&cookie->lsc_oig); - if (rc) { - OBD_FREE(cookie, LLU_SYSIO_COOKIE_SIZE(maxpages)); - cookie = NULL; - } - -out: - return cookie; -} - -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; -#ifdef LIBLUSTRE_HANDLE_UNALIGNED_PAGE - struct page *pages = cookie->lsc_pages; -#endif - int i; - - for (i = 0; i< cookie->lsc_maxpages; i++) { - if (llap[i].llap_cookie) - obd_teardown_async_page(exp, lsm, NULL, - llap[i].llap_cookie); -#ifdef LIBLUSTRE_HANDLE_UNALIGNED_PAGE - if (pages[i]._managed) { - free(pages[i].addr); - pages[i]._managed = 0; - } -#endif - } - - I_RELE(cookie->lsc_inode); - - oig_release(cookie->lsc_oig); - OBD_FREE(cookie, LLU_SYSIO_COOKIE_SIZE(cookie->lsc_maxpages)); -} - -#ifdef LIBLUSTRE_HANDLE_UNALIGNED_PAGE -/* Note: these code should be removed finally, don't need - * more cleanup - */ -static -int prepare_unaligned_write(struct llu_sysio_cookie *cookie) -{ - struct inode *inode = cookie->lsc_inode; - struct llu_inode_info *lli = llu_i2info(inode); - struct lov_stripe_md *lsm = lli->lli_smd; - struct obdo oa; - struct page *pages = cookie->lsc_pages; - int i, pgidx[2] = {0, cookie->lsc_npages-1}; - int rc; - ENTRY; - - for (i = 0; i < 2; i++) { - struct page *oldpage = &pages[pgidx[i]]; - struct page newpage; - struct brw_page pg; - char *newbuf; - - if (i == 0 && pgidx[0] == pgidx[1]) - continue; - - LASSERT(oldpage->_offset + oldpage->_count <= PAGE_CACHE_SIZE); - - if (oldpage->_count == PAGE_CACHE_SIZE) - continue; - - if (oldpage->index << PAGE_CACHE_SHIFT >= - lli->lli_st_size) - continue; - - newbuf = malloc(PAGE_CACHE_SIZE); - if (!newbuf) - return -ENOMEM; - - newpage.index = oldpage->index; - newpage.addr = newbuf; - - pg.pg = &newpage; - pg.off = ((obd_off)newpage.index << PAGE_CACHE_SHIFT); - if (pg.off + PAGE_CACHE_SIZE > lli->lli_st_size) - pg.count = lli->lli_st_size % PAGE_CACHE_SIZE; - else - pg.count = PAGE_CACHE_SIZE; - pg.flag = 0; - - oa.o_id = lsm->lsm_object_id; - oa.o_mode = lli->lli_st_mode; - oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE; - - /* issue read */ - rc = obd_brw(OBD_BRW_READ, llu_i2obdexp(inode), &oa, lsm, 1, &pg, NULL); - if (rc) { - free(newbuf); - RETURN(rc); - } - - /* copy page content, and reset page params */ - memcpy(newbuf + oldpage->_offset, - (char*)oldpage->addr + oldpage->_offset, - oldpage->_count); - - oldpage->addr = newbuf; - if ((((obd_off)oldpage->index << PAGE_CACHE_SHIFT) + - oldpage->_offset + oldpage->_count) > lli->lli_st_size) - oldpage->_count += oldpage->_offset; - else - oldpage->_count = PAGE_CACHE_SIZE; - oldpage->_offset = 0; - oldpage->_managed = 1; - } - - RETURN(0); -} -#endif - -static -int llu_prep_async_io(struct llu_sysio_cookie *cookie, int cmd, - char *buf, loff_t pos, size_t count) +static int llu_queue_pio(int cmd, struct llu_io_group *group, + char *buf, size_t count, loff_t pos) { - struct llu_inode_info *lli = llu_i2info(cookie->lsc_inode); + struct llu_inode_info *lli = llu_i2info(group->lig_inode); + struct intnl_stat *st = llu_i2stat(group->lig_inode); struct lov_stripe_md *lsm = lli->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; + struct obd_export *exp = llu_i2obdexp(group->lig_inode); + struct page *pages = &group->lig_pages[group->lig_npages],*page = pages; + struct ll_async_page *llap = &group->lig_llaps[group->lig_npages]; + void *llap_cookie = group->lig_llap_cookies + + llap_cookie_size * group->lig_npages; + int i, rc, npages = 0, ret_bytes = 0; + int local_lock; ENTRY; if (!exp) RETURN(-EINVAL); + local_lock = group->lig_params->lrp_lock_mode != LCK_NL; /* 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; + offset = (pos & ~CFS_PAGE_MASK); + index = pos >> CFS_PAGE_SHIFT; + bytes = CFS_PAGE_SIZE - offset; if (bytes > count) bytes = count; /* prevent read beyond file range */ - if ((cmd == OBD_BRW_READ) && - (pos + bytes) >= lli->lli_st_size) { - if (pos >= lli->lli_st_size) + if (/* local_lock && */ + cmd == OBD_BRW_READ && pos + bytes >= st->st_size) { + if (pos >= st->st_size) break; - bytes = lli->lli_st_size - pos; + bytes = st->st_size - pos; } /* prepare page for this index */ - pages[npages].index = index; - pages[npages].addr = buf - offset; + page->index = index; + page->addr = buf - offset; - pages[npages]._offset = offset; - pages[npages]._count = bytes; + page->_offset = offset; + page->_count = bytes; + page++; npages++; count -= bytes; pos += bytes; buf += bytes; - cookie->lsc_rwcount += bytes; + group->lig_rwcount += bytes; + ret_bytes += bytes; } while (count); - cookie->lsc_npages = npages; + group->lig_npages += npages; -#ifdef LIBLUSTRE_HANDLE_UNALIGNED_PAGE - if (cmd == OBD_BRW_WRITE) { - rc = prepare_unaligned_write(cookie); - if (rc) - RETURN(rc); - } -#endif - - 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, + for (i = 0, page = pages; i < npages; + i++, page++, llap++, llap_cookie += llap_cookie_size){ + llap->llap_magic = LLAP_MAGIC; + llap->llap_cookie = llap_cookie; + rc = obd_prep_async_page(exp, lsm, NULL, page, + (obd_off)page->index << CFS_PAGE_SHIFT, &llu_async_page_ops, - &llap[i], &llap[i].llap_cookie); + llap, &llap->llap_cookie); if (rc) { - llap[i].llap_cookie = NULL; + LASSERT(rc < 0); + llap->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_group_io(exp, lsm, NULL, cookie->lsc_oig, - llap[i].llap_cookie, cmd, - pages[i]._offset, pages[i]._count, 0, + CDEBUG(D_CACHE, "llap %p page %p group %p obj off "LPU64"\n", + llap, page, llap->llap_cookie, + (obd_off)pages->index << CFS_PAGE_SHIFT); + page->private = (unsigned long)llap; + llap->llap_page = page; + llap->llap_inode = group->lig_inode; + + rc = obd_queue_group_io(exp, lsm, NULL, group->lig_oig, + llap->llap_cookie, cmd, + page->_offset, page->_count, + group->lig_params->lrp_brw_flags, ASYNC_READY | ASYNC_URGENT | ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC); - if (rc) + if (!local_lock && cmd == OBD_BRW_READ) { + /* + * In OST-side locking case short reads cannot be + * detected properly. + * + * The root of the problem is that + * + * kms = lov_merge_size(lsm, 1); + * if (end >= kms) + * glimpse_size(inode); + * else + * st->st_size = kms; + * + * logic in the read code (both llite and liblustre) + * only works correctly when client holds DLM lock on + * [start, end]. Without DLM lock KMS can be + * completely out of date, and client can either make + * spurious short-read (missing concurrent write), or + * return stale data (missing concurrent + * truncate). For llite client this is fatal, because + * incorrect data are cached and can be later sent + * back to the server (vide bug 5047). This is hard to + * fix by handling short-reads on the server, as there + * is no easy way to communicate file size (or amount + * of bytes read/written) back to the client, + * _especially_ because OSC pages can be sliced and + * dices into multiple RPCs arbitrary. Fortunately, + * liblustre doesn't cache data and the worst case is + * that we get race with concurrent write or truncate. + */ + } + if (rc) { + LASSERT(rc < 0); RETURN(rc); + } - llap[i].llap_queued = 1; + llap->llap_queued = 1; } - RETURN(0); + RETURN(ret_bytes); } static -int llu_start_async_io(struct llu_sysio_cookie *cookie) +struct llu_io_group * get_io_group(struct inode *inode, int maxpages, + struct lustre_rw_params *params) { - struct lov_stripe_md *lsm = llu_i2info(cookie->lsc_inode)->lli_smd; - struct obd_export *exp = llu_i2obdexp(cookie->lsc_inode); + struct llu_io_group *group; + int rc; - return obd_trigger_group_io(exp, lsm, NULL, cookie->lsc_oig); -} + if (!llap_cookie_size) + llap_cookie_size = obd_prep_async_page(llu_i2obdexp(inode), + NULL, NULL, NULL, 0, + NULL, NULL, NULL); -/* - * 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) -{ - struct llu_sysio_cookie *cookie; - int max_pages, rc; - ENTRY; + OBD_ALLOC(group, LLU_IO_GROUP_SIZE(maxpages)); + if (!group) + return ERR_PTR(-ENOMEM); - max_pages = (count >> PAGE_SHIFT) + 2; + I_REF(inode); + group->lig_inode = inode; + group->lig_maxpages = maxpages; + group->lig_params = params; + group->lig_llaps = (struct ll_async_page *)(group + 1); + group->lig_pages = (struct page *)(&group->lig_llaps[maxpages]); + group->lig_llap_cookies = (void *)(&group->lig_pages[maxpages]); + + rc = oig_init(&group->lig_oig); + if (rc) { + OBD_FREE(group, LLU_IO_GROUP_SIZE(maxpages)); + return ERR_PTR(rc); + } - cookie = get_sysio_cookie(inode, max_pages); - if (!cookie) - RETURN(ERR_PTR(-ENOMEM)); + return group; +} - rc = llu_prep_async_io(cookie, cmd, buf, pos, count); - if (rc) - GOTO(out_cleanup, rc); +static int max_io_pages(ssize_t len, int iovlen) +{ + return (((len + CFS_PAGE_SIZE -1) / CFS_PAGE_SIZE) + 2 + iovlen - 1); +} - rc = llu_start_async_io(cookie); - if (rc) - GOTO(out_cleanup, rc); +static +void put_io_group(struct llu_io_group *group) +{ + struct lov_stripe_md *lsm = llu_i2info(group->lig_inode)->lli_smd; + struct obd_export *exp = llu_i2obdexp(group->lig_inode); + struct ll_async_page *llap = group->lig_llaps; + int i; -/* - rc = oig_wait(&oig); - if (rc) { - CERROR("file i/o error!\n"); - rw_count = rc; + for (i = 0; i < group->lig_npages; i++, llap++) { + if (llap->llap_cookie) + obd_teardown_async_page(exp, lsm, NULL, + llap->llap_cookie); } -*/ - RETURN(cookie); -out_cleanup: - put_sysio_cookie(cookie); - RETURN(ERR_PTR(rc)); -} + I_RELE(group->lig_inode); -void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm, - obd_off size); + oig_release(group->lig_oig); + OBD_FREE(group, LLU_IO_GROUP_SIZE(group->lig_maxpages)); +} -struct llu_sysio_callback_args* -llu_file_write(struct inode *inode, const struct iovec *iovec, - size_t iovlen, loff_t pos) +static +ssize_t llu_file_prwv(const struct iovec *iovec, int iovlen, + _SYSIO_OFF_T pos, ssize_t len, + void *private) { + struct llu_io_session *session = (struct llu_io_session *) private; + struct inode *inode = session->lis_inode; struct llu_inode_info *lli = llu_i2info(inode); + struct intnl_stat *st = llu_i2stat(inode); struct ll_file_data *fd = lli->lli_file_data; struct lustre_handle lockh = {0}; struct lov_stripe_md *lsm = lli->lli_smd; struct obd_export *exp = NULL; - ldlm_policy_data_t policy; - struct llu_sysio_callback_args *lsca; - struct llu_sysio_cookie *cookie; - ldlm_error_t err; - int iovidx; + struct llu_io_group *iogroup; + struct lustre_rw_params p; + struct ost_lvb lvb; + __u64 kms; + int err, is_read, iovidx, ret; + int local_lock; + ssize_t ret_len = len; ENTRY; - /* XXX consider other types later */ - if (!S_ISREG(lli->lli_st_mode)) - LBUG(); - - LASSERT(iovlen <= MAX_IOVEC); + /* in a large iov read/write we'll be repeatedly called. + * so give a chance to answer cancel ast here + */ + liblustre_wait_event(0); exp = llu_i2obdexp(inode); if (exp == NULL) - RETURN(ERR_PTR(-EINVAL)); + RETURN(-EINVAL); - OBD_ALLOC(lsca, sizeof(*lsca)); - if (!lsca) - RETURN(ERR_PTR(-ENOMEM)); + if (len == 0 || iovlen == 0) + RETURN(0); + + if (pos + len > lli->lli_maxbytes) + RETURN(-ERANGE); + + lustre_build_lock_params(session->lis_cmd, lli->lli_open_flags, + lli->lli_sbi->ll_lco.lco_flags, + pos, len, &p); + + iogroup = get_io_group(inode, max_io_pages(len, iovlen), &p); + if (IS_ERR(iogroup)) + RETURN(PTR_ERR(iogroup)); + + local_lock = p.lrp_lock_mode != LCK_NL; + + err = llu_extent_lock(fd, inode, lsm, p.lrp_lock_mode, &p.lrp_policy, + &lockh, p.lrp_ast_flags); + if (err != ELDLM_OK) + GOTO(err_put, err); + + is_read = (session->lis_cmd == OBD_BRW_READ); + if (is_read) { + /* + * If OST-side locking is used, KMS can be completely out of + * date, and, hence, cannot be used for short-read + * detection. Rely in OST to handle short reads in that case. + */ + inode_init_lvb(inode, &lvb); + obd_merge_lvb(exp, lsm, &lvb, 1); + kms = lvb.lvb_size; + /* extent.end is last byte of the range */ + if (p.lrp_policy.l_extent.end >= kms) { + /* A glimpse is necessary to determine whether + * we return a short read or some zeroes at + * the end of the buffer + * + * In the case of OST-side locking KMS can be + * completely out of date and short-reads maybe + * mishandled. See llu_queue_pio() for more detailed + * comment. + */ + if ((err = llu_glimpse_size(inode))) { + GOTO(err_unlock, err); + } + } else { + st->st_size = kms; + } + } else if (lli->lli_open_flags & O_APPEND) { + pos = st->st_size; + } - /* FIXME optimize the following extent locking */ for (iovidx = 0; iovidx < iovlen; iovidx++) { - char *buf = (char*)iovec[iovidx].iov_base; + char *buf = (char *) iovec[iovidx].iov_base; size_t count = iovec[iovidx].iov_len; - if (count == 0) + if (!count) continue; + if (len < count) + count = len; + if (IS_BAD_PTR(buf) || IS_BAD_PTR(buf + count)) { + GOTO(err_unlock, err = -EFAULT); + } - if (pos + count > lli->lli_maxbytes) - GOTO(err_out, err = -ERANGE); - - /* FIXME libsysio haven't handle O_APPEND?? */ - policy.l_extent.start = pos; - policy.l_extent.end = pos + count - 1; - - err = llu_extent_lock(fd, inode, lsm, LCK_PW, &policy, - &lockh, 0); - 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; - lov_increase_kms(exp, lsm, pos); - /* file size grow */ - if (pos > lli->lli_st_size) - lli->lli_st_size = pos; + if (is_read) { + if (/* local_lock && */ pos >= st->st_size) + break; } else { - llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh); - GOTO(err_out, err = PTR_ERR(cookie)); + if (pos >= lli->lli_maxbytes) { + GOTO(err_unlock, err = -EFBIG); + } + if (pos + count >= lli->lli_maxbytes) + count = lli->lli_maxbytes - pos; } - /* XXX errors? */ - err = llu_extent_unlock(fd, inode, lsm, LCK_PW, &lockh); - if (err) - CERROR("extent unlock error %d\n", err); + ret = llu_queue_pio(session->lis_cmd, iogroup, buf, count, pos); + if (ret < 0) { + GOTO(err_unlock, err = ret); + } else { + pos += ret; + if (!is_read) { + LASSERT(ret == count); + obd_adjust_kms(exp, lsm, pos, 0); + /* file size grow immediately */ + if (pos > st->st_size) + st->st_size = pos; + } + len -= ret; + if (!len) + break; + } } + LASSERT(len == 0 || is_read); /* libsysio should guarantee this */ - RETURN(lsca); + err = obd_trigger_group_io(exp, lsm, NULL, iogroup->lig_oig); + if (err) + GOTO(err_unlock, err); -err_out: - /* teardown all async stuff */ - while (lsca->ncookies--) { - put_sysio_cookie(lsca->cookies[lsca->ncookies]); + err = oig_wait(iogroup->lig_oig); + if (err) { + CERROR("sync error %d, data corruption possible\n", err); + GOTO(err_unlock, err); } - OBD_FREE(lsca, sizeof(*lsca)); - RETURN(ERR_PTR(err)); + ret = llu_extent_unlock(fd, inode, lsm, p.lrp_lock_mode, &lockh); + if (ret) + CERROR("extent unlock error %d\n", ret); + + session->lis_groups[session->lis_ngroups++] = iogroup; + RETURN(ret_len); + +err_unlock: + llu_extent_unlock(fd, inode, lsm, p.lrp_lock_mode, &lockh); +err_put: + put_io_group(iogroup); + RETURN((ssize_t)err); } -#if 0 -static void llu_update_atime(struct inode *inode) +static +struct llu_io_session *get_io_session(struct inode *ino, int ngroups, int cmd) { - struct llu_inode_info *lli = llu_i2info(inode); + struct llu_io_session *session; + + OBD_ALLOC(session, LLU_IO_SESSION_SIZE(ngroups)); + if (!session) + return NULL; -#ifdef USE_ATIME - struct iattr attr; + I_REF(ino); + session->lis_inode = ino; + session->lis_max_groups = ngroups; + session->lis_cmd = cmd; + return session; +} - attr.ia_atime = LTIME_S(CURRENT_TIME); - attr.ia_valid = ATTR_ATIME; +static void put_io_session(struct llu_io_session *session) +{ + int i; - if (lli->lli_st_atime == attr.ia_atime) return; - if (IS_RDONLY(inode)) return; - if (IS_NOATIME(inode)) return; + for (i = 0; i < session->lis_ngroups; i++) { + if (session->lis_groups[i]) { + put_io_group(session->lis_groups[i]); + session->lis_groups[i] = NULL; + } + } - /* 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 + I_RELE(session->lis_inode); + OBD_FREE(session, LLU_IO_SESSION_SIZE(session->lis_max_groups)); } -#endif -struct llu_sysio_callback_args* -llu_file_read(struct inode *inode, const struct iovec *iovec, - size_t iovlen, loff_t pos) +static int llu_file_rwx(struct inode *ino, + struct ioctx *ioctx, + int read) { - 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 }; - ldlm_policy_data_t policy; - struct llu_sysio_callback_args *lsca; - struct llu_sysio_cookie *cookie; - __u64 kms; - int iovidx; - - ldlm_error_t err; + struct llu_io_session *session; + ssize_t cc; + int cmd = read ? OBD_BRW_READ : OBD_BRW_WRITE; ENTRY; - 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; + LASSERT(ioctx->ioctx_xtvlen >= 0); + LASSERT(ioctx->ioctx_iovlen >= 0); - /* "If nbyte is 0, read() will return 0 and have no other results." - * -- Single Unix Spec */ - if (count == 0) - continue; + liblustre_wait_event(0); - policy.l_extent.start = pos; - policy.l_extent.end = pos + count - 1; - - err = llu_extent_lock(fd, inode, lsm, LCK_PR, &policy, &lockh, 0); - if (err != ELDLM_OK) - GOTO(err_out, err = -ENOLCK); - - kms = lov_merge_size(lsm, 1); - if (policy.l_extent.end > kms) { - /* A glimpse is necessary to determine whether we - * return a short read or some zeroes at the end of - * the buffer */ - struct ost_lvb lvb; - if (llu_glimpse_size(inode, &lvb)) { - llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh); - GOTO(err_out, err = -ENOLCK); - } - lli->lli_st_size = lvb.lvb_size; - } else { - lli->lli_st_size = kms; - } + if (!ioctx->ioctx_xtvlen) + RETURN(0); - CDEBUG(D_INFO, "Reading inode %lu, "LPSZ" bytes, offset %Ld, " - "i_size "LPU64"\n", lli->lli_st_ino, count, pos, - lli->lli_st_size); + /* 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); + } - if (pos >= lli->lli_st_size) { - llu_extent_unlock(fd, inode, lsm, LCK_PR, &lockh); - break; - } + liblustre_wait_event(0); + RETURN(cc); +} - 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)); - } +int llu_iop_read(struct inode *ino, + struct ioctx *ioctx) +{ + /* BUG: 5972 */ + struct intnl_stat *st = llu_i2stat(ino); + st->st_atime = CURRENT_TIME; - /* 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); + return llu_file_rwx(ino, ioctx, 1); +} -err_out: - /* teardown all async stuff */ - while (lsca->ncookies--) { - put_sysio_cookie(lsca->cookies[lsca->ncookies]); - } - OBD_FREE(lsca, sizeof(*lsca)); +int llu_iop_write(struct inode *ino, + struct ioctx *ioctx) +{ + struct intnl_stat *st = llu_i2stat(ino); + st->st_mtime = st->st_ctime = CURRENT_TIME; - RETURN(ERR_PTR(err)); + return llu_file_rwx(ino, ioctx, 0); } -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; + struct llu_io_session *session; + struct llu_io_group *group; int i, err = 0, rc = 0; ENTRY; - /* write/read(fd, buf, 0) */ - if (!lsca) { - ioctxp->ioctx_cc = 0; - RETURN(1); - } + liblustre_wait_event(0); - LASSERT(!IS_ERR(lsca)); + session = (struct llu_io_session *) ioctx->ioctx_private; + LASSERT(session); + LASSERT(!IS_ERR(session)); - for (i = 0; i < lsca->ncookies; i++) { - cookie = lsca->cookies[i]; - if (cookie) { - err = oig_wait(cookie->lsc_oig); - if (err && !rc) - rc = err; + for (i = 0; i < session->lis_ngroups; i++) { + group = session->lis_groups[i]; + if (group) { + if (!rc) { + err = oig_wait(group->lig_oig); + if (err) + rc = err; + } if (!rc) - ioctxp->ioctx_cc += cookie->lsc_rwcount; - put_sysio_cookie(cookie); + ioctx->ioctx_cc += group->lig_rwcount; + put_io_group(group); + session->lis_groups[i] = NULL; } } if (rc) { LASSERT(rc < 0); - ioctxp->ioctx_cc = -1; - ioctxp->ioctx_errno = -rc; + ioctx->ioctx_cc = -1; + ioctx->ioctx_errno = -rc; } - OBD_FREE(lsca, sizeof(*lsca)); - ioctxp->ioctx_private = NULL; + put_io_session(session); + ioctx->ioctx_private = NULL; + liblustre_wait_event(0); RETURN(1); }