X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fliblustre%2Frw.c;h=eb22812c15ab6f910494087633afd2092b4114f6;hb=78f134e96678afef1c8c426aaed9f75f5a70fccf;hp=847b1d02660a5d0eba3dbae199bff92a33c85056;hpb=96ec6856f91f7f9031cfce4273c714d72cfe59ae;p=fs%2Flustre-release.git diff --git a/lustre/liblustre/rw.c b/lustre/liblustre/rw.c index 847b1d0..eb22812 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. * @@ -25,495 +25,864 @@ #include #include -#include #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" -int llu_iop_iodone(struct ioctx *ioctxp __IS_UNUSED) +struct llu_io_group { - return 1; -} + 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 *)) -/* - * 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) + +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) { -#if 0 - struct ll_inode_info *lli = ll_i2info(inode); + struct llu_inode_info *lli = llu_i2info(inode); + struct lov_stripe_md *lsm = lli->lli_smd; + struct obd_export *exp = llu_i2obdexp(inode); + struct { + char name[16]; + struct ldlm_lock *lock; + struct lov_stripe_md *lsm; + } key = { .name = "lock_to_stripe", .lock = lock, .lsm = lsm }; + __u32 stripe, vallen = sizeof(stripe); int rc; ENTRY; - rc = ll_extent_lock_no_validate(fd, inode, lsm, mode, extent, lockh); - if (rc != ELDLM_OK) - RETURN(rc); + if (lsm->lsm_stripe_count == 1) + RETURN(0); - /* 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); + /* get our offset in the lov */ + rc = obd_get_info(exp, sizeof(key), &key, &vallen, &stripe); + if (rc != 0) { + CERROR("obd_get_info: rc = %d\n", rc); + LBUG(); + } + LASSERT(stripe < lsm->lsm_stripe_count); + RETURN(stripe); +} - down(&lli->lli_getattr_sem); +static int llu_extent_lock_callback(struct ldlm_lock *lock, + struct ldlm_lock_desc *new, void *data, + int flag) +{ + struct lustre_handle lockh = { 0 }; + int rc; + ENTRY; - if (!test_bit(LLI_F_DID_GETATTR, &lli->lli_flags)) { - rc = ll_inode_getattr(inode, lsm, fd ? &fd->fd_ost_och : NULL); - if (rc == 0) { - set_bit(LLI_F_DID_GETATTR, &lli->lli_flags); - } else { - /* XXX can this fail? */ - ll_extent_unlock(fd, inode, lsm, mode, lockh); - } + if ((unsigned long)data > 0 && (unsigned long)data < 0x1000) { + LDLM_ERROR(lock, "cancelling lock with bad data %p", data); + LBUG(); } - up(&lli->lli_getattr_sem); - RETURN(rc); -#else - return ELDLM_OK; -#endif + 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: { + 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) + goto iput; + if (!lli->lli_smd) + goto iput; + 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); + 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; +iput: + I_RELE(inode); + break; + } + default: + LBUG(); + } + + RETURN(0); } -int ll_extent_unlock(struct ll_file_data *fd, struct inode *inode, - struct lov_stripe_md *lsm, int mode, - struct lustre_handle *lockh) +static int llu_glimpse_callback(struct ldlm_lock *lock, void *reqp) { -#if 0 - struct ll_sb_info *sbi = ll_i2sbi(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; + int size[2] = { sizeof(struct ptlrpc_body), sizeof(*lvb) }; + int rc, stripe = 0; ENTRY; - /* 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); + if (inode == NULL) + GOTO(out, rc = -ELDLM_NO_LOCK_DATA); + lli = llu_i2info(inode); + if (lli == NULL) + GOTO(iput, rc = -ELDLM_NO_LOCK_DATA); + if (lli->lli_smd == NULL) + GOTO(iput, rc = -ELDLM_NO_LOCK_DATA); - rc = obd_cancel(&sbi->ll_osc_conn, lsm, mode, lockh); + /* First, find out which stripe index this lock corresponds to. */ + if (lli->lli_smd->lsm_stripe_count > 1) + stripe = llu_lock_to_stripe_offset(inode, lock); - RETURN(rc); -#else - return 0; -#endif + rc = lustre_pack_reply(req, 2, size, NULL); + if (rc) { + CERROR("lustre_pack_reply: %d\n", rc); + GOTO(iput, rc); + } + + 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); + 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; } -static int llu_brw(int cmd, struct inode *inode, struct page *page, int 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_inode_info *lli = llu_i2info(inode); - struct lov_stripe_md *lsm = lli->lli_smd; - struct brw_page pg; + 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 } } }; + struct ost_lvb lvb; int rc; ENTRY; - pg.pg = page; - pg.off = ((obd_off)page->index) << PAGE_SHIFT; - - /* FIXME FIXME FIXME FIXME FIXME FIXME FIXME FIXME FIXME */ -#if 0 - if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > lli->lli_st_size)) - pg.count = lli->lli_st_size % PAGE_SIZE; - else -#endif - pg.count = PAGE_SIZE; + CDEBUG(D_DLMTRACE, "Glimpsing inode %llu\n", (long long)st->st_ino); - CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n", - cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, lli->lli_st_ino, - pg.off, pg.off); - if (pg.count == 0) { - LBUG(); + if (!lli->lli_smd) { + CDEBUG(D_DLMTRACE, "No objects for inode %llu\n", + (long long)st->st_ino); + RETURN(0); } - pg.flag = flags; + 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; + + oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF; + oinfo.oi_lockh = &lockh; + oinfo.oi_md = lli->lli_smd; - rc = obd_brw(cmd, llu_i2obdconn(inode), lsm, 1, &pg, set, NULL); + rc = obd_enqueue_rqset(sbi->ll_osc_exp, &oinfo, &einfo); if (rc) { - CERROR("error from obd_brw: rc = %d\n", rc); + CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc); + RETURN(rc > 0 ? -EIO : rc); } + inode_init_lvb(inode, &lvb); + obd_merge_lvb(sbi->ll_osc_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; + + CDEBUG(D_DLMTRACE, "glimpse: size: %llu, blocks: %llu\n", + (long long)st->st_size, (long long)st->st_blocks); + RETURN(rc); } -static int llu_prepare_write(struct inode *inode, struct page *page, - unsigned from, unsigned to) +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_inode_info *lli = llu_i2info(inode); - obd_off offset = ((obd_off)page->index) << PAGE_SHIFT; - int rc = 0; + struct llu_sb_info *sbi = llu_i2sbi(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; -#if 0 - if (!PageLocked(page)) - LBUG(); + LASSERT(!lustre_handle_is_used(lockh)); + CLASSERT(ELDLM_OK == 0); - if (PageUptodate(page)) + /* 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); - //POISON(addr + from, 0xca, to - from); -#endif - /* We're completely overwriting an existing page, so _don't_ set it up - * to date until commit_write */ - if (from == 0 && to == PAGE_SIZE) + CDEBUG(D_DLMTRACE, "Locking inode %llu, start "LPU64" end "LPU64"\n", + (long long)st->st_ino, policy->l_extent.start, + policy->l_extent.end); + + 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; + + oinfo.oi_policy = *policy; + oinfo.oi_lockh = lockh; + oinfo.oi_md = lsm; + + rc = obd_enqueue(sbi->ll_osc_exp, &oinfo, &einfo); + *policy = oinfo.oi_policy; + if (rc > 0) + rc = -EIO; + + inode_init_lvb(inode, &lvb); + obd_merge_lvb(sbi->ll_osc_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; + } + + RETURN(rc); +} + +int llu_extent_unlock(struct ll_file_data *fd, struct inode *inode, + struct lov_stripe_md *lsm, int mode, + struct lustre_handle *lockh) +{ + struct llu_sb_info *sbi = llu_i2sbi(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) || mode == LCK_NL) RETURN(0); - /* If are writing to a new page, no need to read old data. - * the extent locking and getattr procedures in ll_file_write have - * guaranteed that i_size is stable enough for our zeroing needs */ - if (lli->lli_st_size <= offset) { - memset(kmap(page), 0, PAGE_SIZE); - kunmap(page); - GOTO(prepare_done, rc = 0); - } + 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 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); + inode = llap->llap_inode; + lsm = llu_i2info(inode)->lli_smd; - rc = llu_brw(OBD_BRW_READ, inode, page, 0); + 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 | + OBD_MD_FLUID | OBD_MD_FLGID | + OBD_MD_FLFID | OBD_MD_FLGENER; + obdo_from_inode(oa, inode, valid_flags); EXIT; +} - prepare_done: - return rc; +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; } -static int llu_commit_write(struct inode *inode, struct page *page, - unsigned from, unsigned to) +/* called for each page in a completed rpc.*/ +static int llu_ap_completion(void *data, int cmd, struct obdo *oa, int rc) { - struct llu_inode_info *lli = llu_i2info(inode); - loff_t size; - int rc; + struct ll_async_page *llap; + struct page *page; ENTRY; -#if 0 - LASSERT(inode == file->f_dentry->d_inode); - LASSERT(PageLocked(page)); - - CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n", - inode, page, from, to, page->index); - CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu,from=%d,to=%d\n", - inode->i_ino, from, to); - /* to match full page case in prepare_write */ - SetPageUptodate(page); - /* mark the page dirty, put it on mapping->dirty, - * mark the inode PAGES_DIRTY, put it on sb->dirty */ - set_page_dirty(page); -#endif - rc = llu_brw(OBD_BRW_WRITE, inode, page, 0); - if (rc) - return rc; - /* this is matched by a hack in obdo_to_inode at the moment */ - size = (((obd_off)page->index) << PAGE_SHIFT) + to; - if (size > lli->lli_st_size) - lli->lli_st_size = size; + 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", + page, page->index, rc); + } RETURN(0); -} /* ll_commit_write */ +} -ssize_t -llu_generic_file_write(struct inode *inode, const char *buf, - size_t count, loff_t pos) +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 int llu_queue_pio(int cmd, struct llu_io_group *group, + char *buf, size_t count, loff_t pos) { - struct page *page; - ssize_t written; - long status = 0; - int err; - unsigned bytes; - - if ((ssize_t) count < 0) - return -EINVAL; -#if 0 - down(&inode->i_sem); -#endif - if (pos < 0) - return -EINVAL; + 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(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; - written = 0; + if (!exp) + RETURN(-EINVAL); -#if 0 - remove_suid(inode); - update_inode_times(inode); -#endif + local_lock = group->lig_params->lrp_lock_mode != LCK_NL; + /* prepare the pages array */ do { - unsigned long index, offset; - char *kaddr; - - /* - * Try to find the page in the cache. If it isn't there, - * allocate a free page. - */ - offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */ - index = pos >> PAGE_CACHE_SHIFT; - bytes = PAGE_CACHE_SIZE - offset; - if (bytes > count) { - bytes = count; - } - - status = -ENOMEM; /* we'll assign it later anyway */ - page = __grab_cache_page(index); - if (!page) - break; - - kaddr = kmap(page); - status = llu_prepare_write(inode, page, offset, offset+bytes); - if (status) - goto sync_failure; - - memcpy(kaddr+offset, buf, bytes); - - status = llu_commit_write(inode, page, offset, offset+bytes); - if (!status) - status = bytes; - - if (status >= 0) { - written += status; - count -= status; - pos += status; - buf += status; - } -unlock: - kunmap(page); - page_cache_release(page); - - if (status < 0) - break; - } while (count); -done: - err = written ? written : status; - -#if 0 - up(&inode->i_sem); -#endif - return err; - - status = -EFAULT; - goto unlock; - -sync_failure: - /* - * If blocksize < pagesize, prepare_write() may have instantiated a - * few blocks outside i_size. Trim these off again. - */ - kunmap(page); - page_cache_release(page); - goto done; + unsigned long index, offset, bytes; + + 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 (/* local_lock && */ + cmd == OBD_BRW_READ && pos + bytes >= st->st_size) { + if (pos >= st->st_size) + break; + bytes = st->st_size - pos; + } + + /* prepare page for this index */ + page->index = index; + page->addr = buf - offset; + + page->_offset = offset; + page->_count = bytes; + + page++; + npages++; + count -= bytes; + pos += bytes; + buf += bytes; + + group->lig_rwcount += bytes; + ret_bytes += bytes; + } while (count); + + group->lig_npages += npages; + + 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, &llap->llap_cookie); + if (rc) { + LASSERT(rc < 0); + llap->llap_cookie = NULL; + RETURN(rc); + } + 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 (!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->llap_queued = 1; + } + + RETURN(ret_bytes); } -ssize_t llu_file_write(struct inode *inode, const struct iovec *iovec, - size_t iovlen, loff_t pos) +static +struct llu_io_group * get_io_group(struct inode *inode, int maxpages, + struct lustre_rw_params *params) { + struct llu_io_group *group; + int rc; + + if (!llap_cookie_size) + llap_cookie_size = obd_prep_async_page(llu_i2obdexp(inode), + NULL, NULL, NULL, 0, + NULL, NULL, NULL); + + OBD_ALLOC(group, LLU_IO_GROUP_SIZE(maxpages)); + if (!group) + return ERR_PTR(-ENOMEM); + + 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); + } + + return group; +} + +static int max_io_pages(ssize_t len, int iovlen) +{ + return (((len + CFS_PAGE_SIZE -1) >> CFS_PAGE_SHIFT) + 2 + iovlen - 1); +} + +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; + + for (i = 0; i < group->lig_npages; i++, llap++) { + if (llap->llap_cookie) + obd_teardown_async_page(exp, lsm, NULL, + llap->llap_cookie); + } + + I_RELE(group->lig_inode); + + oig_release(group->lig_oig); + OBD_FREE(group, LLU_IO_GROUP_SIZE(group->lig_maxpages)); +} + +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 ll_file_data *fd = lli->lli_file_data; /* XXX not ready don't use it now */ - struct lustre_handle lockh = { 0 }; + 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 ldlm_extent extent; - ldlm_error_t err; - ssize_t retval = 0; + struct obd_export *exp = NULL; + 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(); -#if 0 - CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu,size="LPSZ",offset=%Ld\n", - inode->i_ino, count, *ppos); - - /* - * sleep doing some writeback work of this mount's dirty data - * if the VM thinks we're low on memory.. other dirtying code - * paths should think about doing this, too, but they should be - * careful not to hold locked pages while they do so. like - * ll_prepare_write. *cough* + /* in a large iov read/write we'll be repeatedly called. + * so give a chance to answer cancel ast here */ - ll_check_dirty(inode->i_sb); -#endif - while (iovlen--) { - const char *buf = iovec[iovlen].iov_base; - size_t count = iovec[iovlen].iov_len; + liblustre_wait_event(0); + + exp = llu_i2obdexp(inode); + if (exp == NULL) + RETURN(-EINVAL); + + 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; + } + + for (iovidx = 0; iovidx < iovlen; iovidx++) { + char *buf = (char *) iovec[iovidx].iov_base; + size_t count = iovec[iovidx].iov_len; - /* POSIX, but surprised the VFS doesn't check this already */ - 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 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; + if (is_read) { + if (/* local_lock && */ pos >= st->st_size) + break; + } else { + if (pos >= lli->lli_maxbytes) { + GOTO(err_unlock, err = -EFBIG); + } + if (pos + count >= lli->lli_maxbytes) + count = lli->lli_maxbytes - pos; } -#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) - RETURN(-ENOLCK); + 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 */ -#if 0 - if (!S_ISBLK(inode->i_mode) && file->f_flags & O_APPEND) - *ppos = inode->i_size; + err = obd_trigger_group_io(exp, lsm, NULL, iogroup->lig_oig); + if (err) + GOTO(err_unlock, err); - CDEBUG(D_INFO, "Writing inode %lu, "LPSZ" bytes, offset %Lu\n", - inode->i_ino, count, *ppos); -#endif - retval += llu_generic_file_write(inode, buf, count, pos); + err = oig_wait(iogroup->lig_oig); + if (err) { + CERROR("sync error %d, data corruption possible\n", err); + GOTO(err_unlock, err); } - /* XXX errors? */ - ll_extent_unlock(fd, inode, lsm, LCK_PW, &lockh); - return(retval); -} + ret = llu_extent_unlock(fd, inode, lsm, p.lrp_lock_mode, &lockh); + if (ret) + CERROR("extent unlock error %d\n", ret); -static void llu_update_atime(struct inode *inode) -{ -#if 0 - struct llu_inode_info *lli = llu_i2info(inode); + session->lis_groups[session->lis_ngroups++] = iogroup; + RETURN(ret_len); -#ifdef USE_ATIME - struct iattr attr; +err_unlock: + llu_extent_unlock(fd, inode, lsm, p.lrp_lock_mode, &lockh); +err_put: + put_io_group(iogroup); + RETURN((ssize_t)err); +} - attr.ia_atime = LTIME_S(CURRENT_TIME); - attr.ia_valid = ATTR_ATIME; +static +struct llu_io_session *get_io_session(struct inode *ino, int ngroups, int cmd) +{ + struct llu_io_session *session; - if (lli->lli_st_atime == attr.ia_atime) return; - if (IS_RDONLY(inode)) return; - if (IS_NOATIME(inode)) return; + OBD_ALLOC(session, LLU_IO_SESSION_SIZE(ngroups)); + if (!session) + return 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 -#endif + I_REF(ino); + session->lis_inode = ino; + session->lis_max_groups = ngroups; + session->lis_cmd = cmd; + return session; } -static size_t llu_generic_file_read(struct inode *inode, char *buf, - size_t count, loff_t pos) +static void put_io_session(struct llu_io_session *session) { - struct llu_inode_info *lli = llu_i2info(inode); - unsigned long index, offset; - int error = 0; - size_t readed = 0; + int i; - index = pos >> PAGE_CACHE_SHIFT; - offset = pos & ~PAGE_CACHE_MASK; - - do { - struct page *page; - unsigned long end_index, nr; - - end_index = lli->lli_st_size >> PAGE_CACHE_SHIFT; - - if (index > end_index) - break; - nr = PAGE_CACHE_SIZE; - if (index == end_index) { - nr = lli->lli_st_size & ~PAGE_CACHE_MASK; - if (nr <= offset) - break; - } - - nr = nr - offset; - if (nr > count) - nr = count; - - page = grab_cache_page(index); - if (!page) { - error = -ENOMEM; - break; + 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; } + } - error = llu_brw(OBD_BRW_READ, inode, page, 0); - if (error) { - page_cache_release(page); - break; - } - - memcpy(buf, kmap(page)+offset, nr); - offset += nr; - index += offset >> PAGE_CACHE_SHIFT; - offset &= ~PAGE_CACHE_MASK; - readed += nr; - count -= nr; - - page_cache_release(page); - } while (count); - - if (error) - return error; - return readed; + I_RELE(session->lis_inode); + OBD_FREE(session, LLU_IO_SESSION_SIZE(session->lis_max_groups)); } -ssize_t 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 }; -#if 0 - struct ll_read_extent rextent; -#else - struct ldlm_extent extent; -#endif - ldlm_error_t err; - ssize_t retval = 0; + struct llu_io_session *session; + ssize_t cc; + int cmd = read ? OBD_BRW_READ : OBD_BRW_WRITE; ENTRY; - while (iovlen--) { - char *buf = iovec[iovlen].iov_base; - size_t count = iovec[iovlen].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) - RETURN(0); + liblustre_wait_event(0); -#if 0 - rextent.re_extent.start = pos; - rextent.re_extent.end = pos + count - 1; -#else - extent.start = pos; - extent.end = pos + count - 1; -#endif - err = llu_extent_lock(fd, inode, lsm, LCK_PR, &extent, &lockh); - if (err != ELDLM_OK) - RETURN(-ENOLCK); -#if 0 - rextent.re_task = current; - spin_lock(&lli->lli_read_extent_lock); - list_add(&rextent.re_lli_item, &lli->lli_read_extents); - spin_unlock(&lli->lli_read_extent_lock); -#endif - CDEBUG(D_INFO, "Reading inode %lu, "LPSZ" bytes, offset %Ld\n", - lli->lli_st_ino, count, pos); - retval = llu_generic_file_read(inode, buf, count, pos); -#if 0 - spin_lock(&lli->lli_read_extent_lock); - list_del(&rextent.re_lli_item); - spin_unlock(&lli->lli_read_extent_lock); -#endif + if (!ioctx->ioctx_xtvlen) + RETURN(0); + + /* 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 (retval > 0) - llu_update_atime(inode); + liblustre_wait_event(0); + RETURN(cc); +} + +int llu_iop_read(struct inode *ino, + struct ioctx *ioctx) +{ + /* BUG: 5972 */ + struct intnl_stat *st = llu_i2stat(ino); + st->st_atime = CURRENT_TIME; + + return llu_file_rwx(ino, ioctx, 1); +} + +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; - /* XXX errors? */ - ll_extent_unlock(fd, inode, lsm, LCK_PR, &lockh); - RETURN(retval); + return llu_file_rwx(ino, ioctx, 0); } +int llu_iop_iodone(struct ioctx *ioctx) +{ + struct llu_io_session *session; + struct llu_io_group *group; + int i, err = 0, rc = 0; + ENTRY; + + liblustre_wait_event(0); + + session = (struct llu_io_session *) ioctx->ioctx_private; + LASSERT(session); + LASSERT(!IS_ERR(session)); + + 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) + ioctx->ioctx_cc += group->lig_rwcount; + put_io_group(group); + session->lis_groups[i] = NULL; + } + } + + if (rc) { + LASSERT(rc < 0); + ioctx->ioctx_cc = -1; + ioctx->ioctx_errno = -rc; + } + + put_io_session(session); + ioctx->ioctx_private = NULL; + liblustre_wait_event(0); + + RETURN(1); +}