X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fvvp_io.c;h=e7ca14e889d36597accaf1e4806e16952e107b3d;hp=cbf02cd293b536e7deff1ef19b6a1f60625e8c55;hb=a6476823f1c9697762403e33808f8598cde2bbcc;hpb=b801d865eb36890b98f3186727d0bf77313273f1 diff --git a/lustre/llite/vvp_io.c b/lustre/llite/vvp_io.c index cbf02cd..e7ca14e 100644 --- a/lustre/llite/vvp_io.c +++ b/lustre/llite/vvp_io.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -29,8 +27,7 @@ * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011 Whamcloud, Inc. - * + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -39,6 +36,7 @@ * Implementation of cl_io for VVP layer. * * Author: Nikita Danilov + * Author: Jinshan Xiong */ #define DEBUG_SUBSYSTEM S_LLITE @@ -67,6 +65,39 @@ int cl_is_normalio(const struct lu_env *env, const struct cl_io *io) return vio->cui_io_subtype == IO_NORMAL; } +/** + * For swapping layout. The file's layout may have changed. + * To avoid populating pages to a wrong stripe, we have to verify the + * correctness of layout. It works because swapping layout processes + * have to acquire group lock. + */ +static bool can_populate_pages(const struct lu_env *env, struct cl_io *io, + struct inode *inode) +{ + struct ll_inode_info *lli = ll_i2info(inode); + struct ccc_io *cio = ccc_env_io(env); + bool rc = true; + + switch (io->ci_type) { + case CIT_READ: + case CIT_WRITE: + /* don't need lock here to check lli_layout_gen as we have held + * extent lock and GROUP lock has to hold to swap layout */ + if (lli->lli_layout_gen != cio->cui_layout_gen) { + io->ci_need_restart = 1; + /* this will return application a short read/write */ + io->ci_continue = 0; + rc = false; + } + case CIT_FAULT: + /* fault is okay because we've already had a page. */ + default: + break; + } + + return rc; +} + /***************************************************************************** * * io operations. @@ -87,18 +118,25 @@ static int vvp_io_fault_iter_init(const struct lu_env *env, static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios) { - struct cl_io *io = ios->cis_io; - struct cl_object *obj = io->ci_obj; + struct cl_io *io = ios->cis_io; + struct cl_object *obj = io->ci_obj; + struct ccc_io *cio = cl2ccc_io(env, ios); CLOBINVRNT(env, obj, ccc_object_invariant(obj)); - if (io->ci_type == CIT_READ) { - struct vvp_io *vio = cl2vvp_io(env, ios); - struct ccc_io *cio = cl2ccc_io(env, ios); - if (vio->cui_ra_window_set) - ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead); - } + CDEBUG(D_VFSTRACE, "ignore/verify layout %d/%d, layout version %d.\n", + io->ci_ignore_layout, io->ci_verify_layout, cio->cui_layout_gen); + if (!io->ci_ignore_layout && io->ci_verify_layout) { + __u32 gen = 0; + + /* check layout version */ + ll_layout_refresh(ccc_object_inode(obj), &gen); + io->ci_need_restart = cio->cui_layout_gen != gen; + if (io->ci_need_restart) + CDEBUG(D_VFSTRACE, "layout changed from %d to %d.\n", + cio->cui_layout_gen, gen); + } } static void vvp_io_fault_fini(const struct lu_env *env, @@ -133,6 +171,7 @@ static int vvp_mmap_locks(const struct lu_env *env, struct ccc_io *vio, struct cl_io *io) { struct ccc_thread_info *cti = ccc_env_info(env); + struct mm_struct *mm = current->mm; struct vm_area_struct *vma; struct cl_lock_descr *descr = &cti->cti_descr; ldlm_policy_data_t policy; @@ -147,6 +186,13 @@ static int vvp_mmap_locks(const struct lu_env *env, if (!cl_is_normalio(env, io)) RETURN(0); + if (vio->cui_iov == NULL) /* nfs or loop back device write */ + RETURN(0); + + /* No MM (e.g. NFS)? No vmas too. */ + if (mm == NULL) + RETURN(0); + for (seg = 0; seg < vio->cui_nrsegs; seg++) { const struct iovec *iv = &vio->cui_iov[seg]; @@ -157,12 +203,14 @@ static int vvp_mmap_locks(const struct lu_env *env, count += addr & (~CFS_PAGE_MASK); addr &= CFS_PAGE_MASK; - while((vma = our_vma(addr, count)) != NULL) { + + down_read(&mm->mmap_sem); + while((vma = our_vma(mm, addr, count)) != NULL) { struct inode *inode = vma->vm_file->f_dentry->d_inode; int flags = CEF_MUST; if (ll_file_nolock(vma->vm_file)) { - /* + /* * For no lock case, a lockless lock will be * generated. */ @@ -197,6 +245,7 @@ static int vvp_mmap_locks(const struct lu_env *env, count -= vma->vm_end - addr; addr = vma->vm_end; } + up_read(&mm->mmap_sem); } RETURN(0); } @@ -230,7 +279,7 @@ static int vvp_io_read_lock(const struct lu_env *env, ENTRY; /* XXX: Layer violation, we shouldn't see lsm at llite level. */ - if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */ + if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */ result = vvp_io_rw_lock(env, io, CLM_READ, io->u.ci_rd.rd.crw_pos, io->u.ci_rd.rd.crw_pos + @@ -271,24 +320,9 @@ static int vvp_io_write_lock(const struct lu_env *env, } static int vvp_io_setattr_iter_init(const struct lu_env *env, - const struct cl_io_slice *ios) + const struct cl_io_slice *ios) { - struct ccc_io *cio = ccc_env_io(env); - struct inode *inode = ccc_object_inode(ios->cis_obj); - - /* - * We really need to get our PW lock before we change inode->i_size. - * If we don't we can race with other i_size updaters on our node, - * like ll_file_read. We can also race with i_size propogation to - * other nodes through dirtying and writeback of final cached pages. - * This last one is especially bad for racing o_append users on other - * nodes. - */ - UNLOCK_INODE_MUTEX(inode); - if (cl_io_is_trunc(ios->cis_io)) - UP_WRITE_I_ALLOC_SEM(inode); - cio->u.setattr.cui_locks_released = 1; - return 0; + return 0; } /** @@ -299,10 +333,10 @@ static int vvp_io_setattr_iter_init(const struct lu_env *env, static int vvp_io_setattr_lock(const struct lu_env *env, const struct cl_io_slice *ios) { - struct ccc_io *cio = ccc_env_io(env); - struct cl_io *io = ios->cis_io; - size_t new_size; - __u32 enqflags = 0; + struct ccc_io *cio = ccc_env_io(env); + struct cl_io *io = ios->cis_io; + __u64 new_size; + __u32 enqflags = 0; if (cl_io_is_trunc(io)) { new_size = io->u.ci_setattr.sa_attr.lvb_size; @@ -323,62 +357,23 @@ static int vvp_io_setattr_lock(const struct lu_env *env, static int vvp_do_vmtruncate(struct inode *inode, size_t size) { - int result; - /* - * Only ll_inode_size_lock is taken at this level. lov_stripe_lock() - * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If - * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock() - */ - ll_inode_size_lock(inode, 0); - result = vmtruncate(inode, size); - if (result != 0) - ll_inode_size_unlock(inode, 0); - - return result; + int result; + /* + * Only ll_inode_size_lock is taken at this level. + */ + ll_inode_size_lock(inode); + result = vmtruncate(inode, size); + ll_inode_size_unlock(inode); + + return result; } static int vvp_io_setattr_trunc(const struct lu_env *env, const struct cl_io_slice *ios, struct inode *inode, loff_t size) { - struct vvp_io *vio = cl2vvp_io(env, ios); - struct cl_io *io = ios->cis_io; - struct cl_object *obj = ios->cis_obj; - pgoff_t start = cl_index(obj, size); - int result; - - DOWN_WRITE_I_ALLOC_SEM(inode); - - result = vvp_do_vmtruncate(inode, size); - - /* - * If a page is partially truncated, keep it owned across truncate to - * prevent... races. - * - * XXX this properly belongs to osc, because races in question are OST - * specific. - */ - if (cl_offset(obj, start) != size) { - struct cl_object_header *hdr; - - hdr = cl_object_header(obj); - cfs_spin_lock(&hdr->coh_page_guard); - vio->cui_partpage = cl_page_lookup(hdr, start); - cfs_spin_unlock(&hdr->coh_page_guard); - - if (vio->cui_partpage != NULL) - /* - * Wait for the transfer completion for a partially - * truncated page to avoid dead-locking an OST with - * the concurrent page-wise overlapping WRITE and - * PUNCH requests. BUG:17397. - * - * Partial page is disowned in vvp_io_trunc_end(). - */ - cl_page_own(env, io, vio->cui_partpage); - } else - vio->cui_partpage = NULL; - return result; + inode_dio_wait(inode); + return 0; } static int vvp_io_setattr_time(const struct lu_env *env, @@ -407,60 +402,38 @@ static int vvp_io_setattr_time(const struct lu_env *env, } static int vvp_io_setattr_start(const struct lu_env *env, - const struct cl_io_slice *ios) + const struct cl_io_slice *ios) { - struct ccc_io *cio = cl2ccc_io(env, ios); - struct cl_io *io = ios->cis_io; - struct inode *inode = ccc_object_inode(io->ci_obj); - - LASSERT(cio->u.setattr.cui_locks_released); - - LOCK_INODE_MUTEX(inode); - cio->u.setattr.cui_locks_released = 0; - - if (cl_io_is_trunc(io)) - return vvp_io_setattr_trunc(env, ios, inode, - io->u.ci_setattr.sa_attr.lvb_size); - else - return vvp_io_setattr_time(env, ios); + struct cl_io *io = ios->cis_io; + struct inode *inode = ccc_object_inode(io->ci_obj); + + mutex_lock(&inode->i_mutex); + if (cl_io_is_trunc(io)) + return vvp_io_setattr_trunc(env, ios, inode, + io->u.ci_setattr.sa_attr.lvb_size); + else + return vvp_io_setattr_time(env, ios); } static void vvp_io_setattr_end(const struct lu_env *env, const struct cl_io_slice *ios) { - struct vvp_io *vio = cl2vvp_io(env, ios); - struct cl_io *io = ios->cis_io; - struct inode *inode = ccc_object_inode(io->ci_obj); - - if (!cl_io_is_trunc(io)) - return; - if (vio->cui_partpage != NULL) { - cl_page_disown(env, ios->cis_io, vio->cui_partpage); - cl_page_put(env, vio->cui_partpage); - vio->cui_partpage = NULL; - } - - /* - * Do vmtruncate again, to remove possible stale pages populated by - * competing read threads. bz20645. - */ - vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size); + struct cl_io *io = ios->cis_io; + struct inode *inode = ccc_object_inode(io->ci_obj); + + if (cl_io_is_trunc(io)) { + /* Truncate in memory pages - they must be clean pages + * because osc has already notified to destroy osc_extents. */ + vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size); + inode_dio_write_done(inode); + } + mutex_unlock(&inode->i_mutex); } static void vvp_io_setattr_fini(const struct lu_env *env, - const struct cl_io_slice *ios) + const struct cl_io_slice *ios) { - struct ccc_io *cio = ccc_env_io(env); - struct cl_io *io = ios->cis_io; - struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj); - - if (cio->u.setattr.cui_locks_released) { - LOCK_INODE_MUTEX(inode); - if (cl_io_is_trunc(io)) - DOWN_WRITE_I_ALLOC_SEM(inode); - cio->u.setattr.cui_locks_released = 0; - } - vvp_io_fini(env, ios); + vvp_io_fini(env, ios); } #ifdef HAVE_FILE_READV @@ -512,7 +485,10 @@ static int vvp_io_read_start(const struct lu_env *env, CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt); - result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed); + if (!can_populate_pages(env, io, inode)) + return 0; + + result = ccc_prep_size(env, obj, io, pos, tot, &exceed); if (result != 0) return result; else if (exceed != 0) @@ -554,6 +530,10 @@ static int vvp_io_read_start(const struct lu_env *env, result = generic_file_splice_read(file, &pos, vio->u.splice.cui_pipe, cnt, vio->u.splice.cui_flags); + /* LU-1109: do splice read stripe by stripe otherwise if it + * may make nfsd stuck if this read occupied all internal pipe + * buffers. */ + io->ci_continue = 0; break; #endif default: @@ -562,15 +542,27 @@ static int vvp_io_read_start(const struct lu_env *env, } out: - if (result >= 0) { - if (result < cnt) - io->ci_continue = 0; - io->ci_nob += result; - ll_rw_stats_tally(ll_i2sbi(inode), current->pid, - cio->cui_fd, pos, result, 0); - result = 0; - } - return result; + if (result >= 0) { + if (result < cnt) + io->ci_continue = 0; + io->ci_nob += result; + ll_rw_stats_tally(ll_i2sbi(inode), current->pid, cio->cui_fd, + pos, result, READ); + result = 0; + } + + return result; +} + +static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios) +{ + struct vvp_io *vio = cl2vvp_io(env, ios); + struct ccc_io *cio = cl2ccc_io(env, ios); + + if (vio->cui_ra_window_set) + ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead); + + vvp_io_fini(env, ios); } static int vvp_io_write_start(const struct lu_env *env, @@ -587,6 +579,9 @@ static int vvp_io_write_start(const struct lu_env *env, ENTRY; + if (!can_populate_pages(env, io, inode)) + return 0; + if (cl_io_is_append(io)) { /* * PARALLEL IO This has to be changed for parallel IO doing @@ -605,15 +600,16 @@ static int vvp_io_write_start(const struct lu_env *env, else result = lustre_generic_file_write(file, cio, &pos); - if (result > 0) { - if (result < cnt) - io->ci_continue = 0; - io->ci_nob += result; - ll_rw_stats_tally(ll_i2sbi(inode), current->pid, - cio->cui_fd, pos, result, 0); - result = 0; - } - RETURN(result); + if (result > 0) { + if (result < cnt) + io->ci_continue = 0; + io->ci_nob += result; + ll_rw_stats_tally(ll_i2sbi(inode), current->pid, + cio->cui_fd, pos, result, WRITE); + result = 0; + } + + RETURN(result); } #ifndef HAVE_VM_OP_FAULT @@ -638,45 +634,43 @@ static int vvp_io_kernel_fault(struct vvp_fault_io *cfio) cfio->nopage.ft_address, (long)cfio->nopage.ft_type); cfio->ft_vmpage = vmpage; + lock_page(vmpage); return 0; } #else static int vvp_io_kernel_fault(struct vvp_fault_io *cfio) { - cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, cfio->fault.ft_vmf); - - if (cfio->fault.ft_vmf->page) { - LL_CDEBUG_PAGE(D_PAGE, cfio->fault.ft_vmf->page, - "got addr %p type NOPAGE\n", - cfio->fault.ft_vmf->virtual_address); - /*XXX workaround to bug in CLIO - he deadlocked with - lock cancel if page locked */ - if (likely(cfio->fault.ft_flags & VM_FAULT_LOCKED)) { - unlock_page(cfio->fault.ft_vmf->page); - cfio->fault.ft_flags &= ~VM_FAULT_LOCKED; + struct vm_fault *vmf = cfio->fault.ft_vmf; + + cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf); + + if (vmf->page) { + LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n", + vmf->virtual_address); + if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) { + lock_page(vmf->page); + cfio->fault.ft_flags &= VM_FAULT_LOCKED; } - cfio->ft_vmpage = cfio->fault.ft_vmf->page; + cfio->ft_vmpage = vmf->page; return 0; } - if (unlikely (cfio->fault.ft_flags & VM_FAULT_ERROR)) { - CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", - cfio->fault.ft_vmf->virtual_address); + if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) { + CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address); return -EFAULT; } - if (unlikely (cfio->fault.ft_flags & VM_FAULT_NOPAGE)) { - CDEBUG(D_PAGE, "got addr %p - OOM\n", - cfio->fault.ft_vmf->virtual_address); + if (cfio->fault.ft_flags & VM_FAULT_OOM) { + CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address); return -ENOMEM; } - if (unlikely(cfio->fault.ft_flags & VM_FAULT_RETRY)) + if (cfio->fault.ft_flags & VM_FAULT_RETRY) return -EAGAIN; - CERROR("unknow error in page fault!\n"); + CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags); return -EINVAL; } @@ -685,18 +679,18 @@ static int vvp_io_kernel_fault(struct vvp_fault_io *cfio) static int vvp_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios) { - struct vvp_io *vio = cl2vvp_io(env, ios); - struct cl_io *io = ios->cis_io; - struct cl_object *obj = io->ci_obj; - struct inode *inode = ccc_object_inode(obj); - struct cl_fault_io *fio = &io->u.ci_fault; - struct vvp_fault_io *cfio = &vio->u.fault; - loff_t offset; - int kernel_result = 0; - int result = 0; - struct cl_page *page; - loff_t size; - pgoff_t last; /* last page in a file data region */ + struct vvp_io *vio = cl2vvp_io(env, ios); + struct cl_io *io = ios->cis_io; + struct cl_object *obj = io->ci_obj; + struct inode *inode = ccc_object_inode(obj); + struct cl_fault_io *fio = &io->u.ci_fault; + struct vvp_fault_io *cfio = &vio->u.fault; + loff_t offset; + int result = 0; + cfs_page_t *vmpage = NULL; + struct cl_page *page; + loff_t size; + pgoff_t last; /* last page in a file data region */ if (fio->ft_executable && LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime) @@ -707,68 +701,145 @@ static int vvp_io_fault_start(const struct lu_env *env, /* offset of the last byte on the page */ offset = cl_offset(obj, fio->ft_index + 1) - 1; LASSERT(cl_index(obj, offset) == fio->ft_index); - result = ccc_prep_size(env, obj, io, 0, offset + 1, 0, NULL); + result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL); if (result != 0) return result; - /* must return unlocked page */ - kernel_result = vvp_io_kernel_fault(cfio); - if (kernel_result != 0) - return kernel_result; + /* must return locked page */ + if (fio->ft_mkwrite) { + LASSERT(cfio->ft_vmpage != NULL); + lock_page(cfio->ft_vmpage); + } else { + result = vvp_io_kernel_fault(cfio); + if (result != 0) + return result; + } - if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE)) { - truncate_inode_pages_range(inode->i_mapping, - cl_offset(obj, fio->ft_index), offset); - } + vmpage = cfio->ft_vmpage; + LASSERT(PageLocked(vmpage)); - /* Temporarily lock vmpage to keep cl_page_find() happy. */ - lock_page(cfio->ft_vmpage); + if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE)) + ll_invalidate_page(vmpage); + size = i_size_read(inode); /* Though we have already held a cl_lock upon this page, but * it still can be truncated locally. */ - if (unlikely(cfio->ft_vmpage->mapping == NULL)) { - unlock_page(cfio->ft_vmpage); - + if (unlikely((vmpage->mapping != inode->i_mapping) || + (page_offset(vmpage) > size))) { CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n"); /* return +1 to stop cl_io_loop() and ll_fault() will catch * and retry. */ - return +1; + GOTO(out, result = +1); } - page = cl_page_find(env, obj, fio->ft_index, cfio->ft_vmpage, - CPT_CACHEABLE); - unlock_page(cfio->ft_vmpage); - if (IS_ERR(page)) { - page_cache_release(cfio->ft_vmpage); - cfio->ft_vmpage = NULL; - return PTR_ERR(page); - } - size = i_size_read(inode); - last = cl_index(obj, size - 1); + if (fio->ft_mkwrite ) { + pgoff_t last_index; + /* + * Capture the size while holding the lli_trunc_sem from above + * we want to make sure that we complete the mkwrite action + * while holding this lock. We need to make sure that we are + * not past the end of the file. + */ + last_index = cl_index(obj, size - 1); + if (last_index < fio->ft_index) { + CDEBUG(D_PAGE, + "llite: mkwrite and truncate race happened: " + "%p: 0x%lx 0x%lx\n", + vmpage->mapping,fio->ft_index,last_index); + /* + * We need to return if we are + * passed the end of the file. This will propagate + * up the call stack to ll_page_mkwrite where + * we will return VM_FAULT_NOPAGE. Any non-negative + * value returned here will be silently + * converted to 0. If the vmpage->mapping is null + * the error code would be converted back to ENODATA + * in ll_page_mkwrite0. Thus we return -ENODATA + * to handle both cases + */ + GOTO(out, result = -ENODATA); + } + } + + page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE); + if (IS_ERR(page)) + GOTO(out, result = PTR_ERR(page)); + + /* if page is going to be written, we should add this page into cache + * earlier. */ + if (fio->ft_mkwrite) { + wait_on_page_writeback(vmpage); + if (set_page_dirty(vmpage)) { + struct ccc_page *cp; + + /* vvp_page_assume() calls wait_on_page_writeback(). */ + cl_page_assume(env, io, page); + + cp = cl2ccc_page(cl_page_at(page, &vvp_device_type)); + vvp_write_pending(cl2ccc(obj), cp); + + /* Do not set Dirty bit here so that in case IO is + * started before the page is really made dirty, we + * still have chance to detect it. */ + result = cl_page_cache_add(env, io, page, CRT_WRITE); + LASSERT(cl_page_is_owned(page, io)); + + vmpage = NULL; + if (result < 0) { + cl_page_unmap(env, io, page); + cl_page_discard(env, io, page); + cl_page_disown(env, io, page); + + cl_page_put(env, page); + + /* we're in big trouble, what can we do now? */ + if (result == -EDQUOT) + result = -ENOSPC; + GOTO(out, result); + } else + cl_page_disown(env, io, page); + } + } + + last = cl_index(obj, size - 1); + /* + * The ft_index is only used in the case of + * a mkwrite action. We need to check + * our assertions are correct, since + * we should have caught this above + */ + LASSERT(!fio->ft_mkwrite || fio->ft_index <= last); if (fio->ft_index == last) /* * Last page is mapped partially. */ fio->ft_nob = size - cl_offset(obj, fio->ft_index); - else + else fio->ft_nob = cl_page_size(obj); - lu_ref_add(&page->cp_reference, "fault", io); - fio->ft_page = page; - /* - * Certain 2.6 kernels return not-NULL from - * filemap_nopage() when page is beyond the file size, - * on the grounds that "An external ptracer can access - * pages that normally aren't accessible.." Don't - * propagate such page fault to the lower layers to - * avoid side-effects like KMS updates. - */ - if (fio->ft_index > last) - result = +1; + lu_ref_add(&page->cp_reference, "fault", io); + fio->ft_page = page; + EXIT; - return result; +out: + /* return unlocked vmpage to avoid deadlocking */ + if (vmpage != NULL) + unlock_page(vmpage); +#ifdef HAVE_VM_OP_FAULT + cfio->fault.ft_flags &= ~VM_FAULT_LOCKED; +#endif + return result; +} + +static int vvp_io_fsync_start(const struct lu_env *env, + const struct cl_io_slice *ios) +{ + /* we should mark TOWRITE bit to each dirty page in radix tree to + * verify pages have been written, but this is difficult because of + * race. */ + return 0; } static int vvp_io_read_page(const struct lu_env *env, @@ -792,7 +863,8 @@ static int vvp_io_read_page(const struct lu_env *env, ENTRY; - if (sbi->ll_ra_info.ra_max_pages_per_file) + if (sbi->ll_ra_info.ra_max_pages_per_file && + sbi->ll_ra_info.ra_max_pages) ras_update(sbi, inode, ras, page->cp_index, cp->cpg_defer_uptodate); @@ -815,7 +887,8 @@ static int vvp_io_read_page(const struct lu_env *env, * this will unlock it automatically as part of cl_page_list_disown(). */ cl_2queue_add(queue, page); - if (sbi->ll_ra_info.ra_max_pages_per_file) + if (sbi->ll_ra_info.ra_max_pages_per_file && + sbi->ll_ra_info.ra_max_pages) ll_readahead(env, io, ras, vmpage->mapping, &queue->c2_qin, fd->fd_flags); @@ -834,7 +907,7 @@ static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io, queue = &io->ci_queue; cl_2queue_init_page(queue, page); - result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0); + result = cl_io_submit_sync(env, io, crt, queue, 0); LASSERT(cl_page_is_owned(page, io)); if (crt == CRT_READ) @@ -870,10 +943,10 @@ static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io, * purposes here we can treat it like i_size. */ if (attr->cat_kms <= offset) { - char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0); + char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0); memset(kaddr, 0, cl_page_size(obj)); - kunmap_atomic(kaddr, KM_USER0); + ll_kunmap_atomic(kaddr, KM_USER0); } else if (cp->cpg_defer_uptodate) cp->cpg_ra_used = 1; else @@ -939,6 +1012,7 @@ static int vvp_io_commit_write(const struct lu_env *env, struct cl_page *pg = slice->cpl_page; struct inode *inode = ccc_object_inode(obj); struct ll_sb_info *sbi = ll_i2sbi(inode); + struct ll_inode_info *lli = ll_i2info(inode); cfs_page_t *vmpage = cp->cpg_page; int result; @@ -975,13 +1049,12 @@ static int vvp_io_commit_write(const struct lu_env *env, */ if (!PageDirty(vmpage)) { tallyop = LPROC_LL_DIRTY_MISSES; - vvp_write_pending(cl2ccc(obj), cp); - set_page_dirty(vmpage); - /* ll_set_page_dirty() does the same for now, but - * it will not soon. */ - vvp_write_pending(cl2ccc(obj), cp); result = cl_page_cache_add(env, io, pg, CRT_WRITE); - if (result == -EDQUOT) { + if (result == 0) { + /* page was added into cache successfully. */ + set_page_dirty(vmpage); + vvp_write_pending(cl2ccc(obj), cp); + } else if (result == -EDQUOT) { pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT; bool need_clip = true; @@ -1020,9 +1093,23 @@ static int vvp_io_commit_write(const struct lu_env *env, } ll_stats_ops_tally(sbi, tallyop, 1); + /* Inode should be marked DIRTY even if no new page was marked DIRTY + * because page could have been not flushed between 2 modifications. + * It is important the file is marked DIRTY as soon as the I/O is done + * Indeed, when cache is flushed, file could be already closed and it + * is too late to warn the MDT. + * It is acceptable that file is marked DIRTY even if I/O is dropped + * for some reasons before being flushed to OST. + */ + if (result == 0) { + spin_lock(&lli->lli_lock); + lli->lli_flags |= LLIF_DATA_MODIFIED; + spin_unlock(&lli->lli_lock); + } + size = cl_offset(obj, pg->cp_index) + to; - ll_inode_size_lock(inode, 0); + ll_inode_size_lock(inode); if (result == 0) { if (size > i_size_read(inode)) { cl_isize_write_nolock(inode, size); @@ -1035,14 +1122,14 @@ static int vvp_io_commit_write(const struct lu_env *env, if (size > i_size_read(inode)) cl_page_discard(env, io, pg); } - ll_inode_size_unlock(inode, 0); - RETURN(result); + ll_inode_size_unlock(inode); + RETURN(result); } static const struct cl_io_operations vvp_io_ops = { .op = { [CIT_READ] = { - .cio_fini = vvp_io_fini, + .cio_fini = vvp_io_read_fini, .cio_lock = vvp_io_read_lock, .cio_start = vvp_io_read_start, .cio_advance = ccc_io_advance @@ -1067,6 +1154,10 @@ static const struct cl_io_operations vvp_io_ops = { .cio_start = vvp_io_fault_start, .cio_end = ccc_io_end }, + [CIT_FSYNC] = { + .cio_start = vvp_io_fsync_start, + .cio_fini = vvp_io_fini + }, [CIT_MISC] = { .cio_fini = vvp_io_fini } @@ -1079,10 +1170,9 @@ static const struct cl_io_operations vvp_io_ops = { int vvp_io_init(const struct lu_env *env, struct cl_object *obj, struct cl_io *io) { - struct vvp_io *vio = vvp_env_io(env); - struct ccc_io *cio = ccc_env_io(env); - struct inode *inode = ccc_object_inode(obj); - struct ll_sb_info *sbi = ll_i2sbi(inode); + struct vvp_io *vio = vvp_env_io(env); + struct ccc_io *cio = ccc_env_io(env); + struct inode *inode = ccc_object_inode(obj); int result; CLOBINVRNT(env, obj, ccc_object_invariant(obj)); @@ -1091,9 +1181,10 @@ int vvp_io_init(const struct lu_env *env, struct cl_object *obj, CL_IO_SLICE_CLEAN(cio, cui_cl); cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops); vio->cui_ra_window_set = 0; - result = 0; - if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) { - size_t count; + result = 0; + if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) { + size_t count; + struct ll_inode_info *lli = ll_i2info(inode); count = io->u.ci_rw.crw_count; /* "If nbyte is 0, read() will return 0 and have no other @@ -1104,14 +1195,42 @@ int vvp_io_init(const struct lu_env *env, struct cl_object *obj, cio->cui_tot_count = count; cio->cui_tot_nrsegs = 0; } - } else if (io->ci_type == CIT_SETATTR) { - if (cl_io_is_trunc(io)) - /* lockless truncate? */ - ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1); - else - io->ci_lockreq = CILR_MANDATORY; - } - RETURN(result); + /* for read/write, we store the jobid in the inode, and + * it'll be fetched by osc when building RPC. + * + * it's not accurate if the file is shared by different + * jobs. + */ + lustre_get_jobid(lli->lli_jobid); + } else if (io->ci_type == CIT_SETATTR) { + if (!cl_io_is_trunc(io)) + io->ci_lockreq = CILR_MANDATORY; + } + + /* ignore layout change for generic CIT_MISC but not for glimpse. + * io context for glimpse must set ci_verify_layout to true, + * see cl_glimpse_size0() for details. */ + if (io->ci_type == CIT_MISC && !io->ci_verify_layout) + io->ci_ignore_layout = 1; + + /* Enqueue layout lock and get layout version. We need to do this + * even for operations requiring to open file, such as read and write, + * because it might not grant layout lock in IT_OPEN. */ + if (result == 0 && !io->ci_ignore_layout) { + result = ll_layout_refresh(inode, &cio->cui_layout_gen); + if (result == -ENOENT) + /* If the inode on MDS has been removed, but the objects + * on OSTs haven't been destroyed (async unlink), layout + * fetch will return -ENOENT, we'd ingore this error + * and continue with dirty flush. LU-3230. */ + result = 0; + if (result < 0) + CERROR("%s: refresh file layout " DFID " error %d.\n", + ll_get_fsname(inode->i_sb, NULL, 0), + PFID(lu_object_fid(&obj->co_lu)), result); + } + + RETURN(result); } static struct vvp_io *cl2vvp_io(const struct lu_env *env,