X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fvvp_io.c;h=b393042dbcf5eab195d7161423dd7a19d166fe3b;hp=f2958365250d861ff8fa8312b10da7d2e5f75079;hb=f384d789b21689f72dc1ad03bfbf16b114748ea2;hpb=e3a7c58aebafce40323db54bf6056029e5af4a70 diff --git a/lustre/llite/vvp_io.c b/lustre/llite/vvp_io.c index f295836..b393042 100644 --- a/lustre/llite/vvp_io.c +++ b/lustre/llite/vvp_io.c @@ -27,7 +27,7 @@ * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Whamcloud, Inc. + * Copyright (c) 2011, 2014, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -36,32 +36,181 @@ * Implementation of cl_io for VVP layer. * * Author: Nikita Danilov + * Author: Jinshan Xiong */ #define DEBUG_SUBSYSTEM S_LLITE -#ifndef __KERNEL__ -# error This file is kernel only. -#endif #include -#include - +#include "llite_internal.h" #include "vvp_internal.h" static struct vvp_io *cl2vvp_io(const struct lu_env *env, - const struct cl_io_slice *slice); + const struct cl_io_slice *slice) +{ + struct vvp_io *vio; + + vio = container_of(slice, struct vvp_io, vui_cl); + LASSERT(vio == vvp_env_io(env)); + + return vio; +} /** - * True, if \a io is a normal io, False for sendfile() / splice_{read|write} + * True, if \a io is a normal io, False for splice_{read,write} */ -int cl_is_normalio(const struct lu_env *env, const struct cl_io *io) +static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io) { - struct vvp_io *vio = vvp_env_io(env); + struct vvp_io *vio = vvp_env_io(env); - LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); + LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); + + return vio->vui_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 vvp_io *vio = vvp_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 (ll_layout_version_get(lli) != vio->vui_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; +} + +static void vvp_object_size_lock(struct cl_object *obj) +{ + struct inode *inode = vvp_object_inode(obj); + + ll_inode_size_lock(inode); + cl_object_attr_lock(obj); +} + +static void vvp_object_size_unlock(struct cl_object *obj) +{ + struct inode *inode = vvp_object_inode(obj); + + cl_object_attr_unlock(obj); + ll_inode_size_unlock(inode); +} - return vio->cui_io_subtype == IO_NORMAL; +/** + * Helper function that if necessary adjusts file size (inode->i_size), when + * position at the offset \a pos is accessed. File size can be arbitrary stale + * on a Lustre client, but client at least knows KMS. If accessed area is + * inside [0, KMS], set file size to KMS, otherwise glimpse file size. + * + * Locking: i_size_lock is used to serialize changes to inode size and to + * protect consistency between inode size and cl_object + * attributes. cl_object_size_lock() protects consistency between cl_attr's of + * top-object and sub-objects. + */ +static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj, + struct cl_io *io, loff_t start, size_t count, + int *exceed) +{ + struct cl_attr *attr = ccc_env_thread_attr(env); + struct inode *inode = vvp_object_inode(obj); + loff_t pos = start + count - 1; + loff_t kms; + int result; + + /* + * Consistency guarantees: following possibilities exist for the + * relation between region being accessed and real file size at this + * moment: + * + * (A): the region is completely inside of the file; + * + * (B-x): x bytes of region are inside of the file, the rest is + * outside; + * + * (C): the region is completely outside of the file. + * + * This classification is stable under DLM lock already acquired by + * the caller, because to change the class, other client has to take + * DLM lock conflicting with our lock. Also, any updates to ->i_size + * by other threads on this client are serialized by + * ll_inode_size_lock(). This guarantees that short reads are handled + * correctly in the face of concurrent writes and truncates. + */ + vvp_object_size_lock(obj); + result = cl_object_attr_get(env, obj, attr); + if (result == 0) { + kms = attr->cat_kms; + if (pos > kms) { + /* + * A glimpse is necessary to determine whether we + * return a short read (B) or some zeroes at the end + * of the buffer (C) + */ + vvp_object_size_unlock(obj); + result = cl_glimpse_lock(env, io, inode, obj, 0); + if (result == 0 && exceed != NULL) { + /* If objective page index exceed end-of-file + * page index, return directly. Do not expect + * kernel will check such case correctly. + * linux-2.6.18-128.1.1 miss to do that. + * --bug 17336 */ + loff_t size = i_size_read(inode); + unsigned long cur_index = start >> + PAGE_CACHE_SHIFT; + + if ((size == 0 && cur_index != 0) || + (((size - 1) >> PAGE_CACHE_SHIFT) < + cur_index)) + *exceed = 1; + } + + return result; + } else { + /* + * region is within kms and, hence, within real file + * size (A). We need to increase i_size to cover the + * read region so that generic_file_read() will do its + * job, but that doesn't mean the kms size is + * _correct_, it is only the _minimum_ size. If + * someone does a stat they will get the correct size + * which will always be >= the kms value here. + * b=11081 + */ + if (i_size_read(inode) < kms) { + i_size_write(inode, kms); + CDEBUG(D_VFSTRACE, + DFID" updating i_size "LPU64"\n", + PFID(lu_object_fid(&obj->co_lu)), + (__u64)i_size_read(inode)); + } + } + } + + vvp_object_size_unlock(obj); + + return result; } /***************************************************************************** @@ -70,32 +219,141 @@ int cl_is_normalio(const struct lu_env *env, const struct cl_io *io) * */ -static int vvp_io_fault_iter_init(const struct lu_env *env, - const struct cl_io_slice *ios) +static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io, + __u32 enqflags, enum cl_lock_mode mode, + pgoff_t start, pgoff_t end) { - struct vvp_io *vio = cl2vvp_io(env, ios); - struct inode *inode = ccc_object_inode(ios->cis_obj); + struct vvp_io *vio = vvp_env_io(env); + struct cl_lock_descr *descr = &vio->vui_link.cill_descr; + struct cl_object *obj = io->ci_obj; + + CLOBINVRNT(env, obj, vvp_object_invariant(obj)); + ENTRY; + + CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end); + + memset(&vio->vui_link, 0, sizeof vio->vui_link); + + if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) { + descr->cld_mode = CLM_GROUP; + descr->cld_gid = vio->vui_fd->fd_grouplock.cg_gid; + } else { + descr->cld_mode = mode; + } + + descr->cld_obj = obj; + descr->cld_start = start; + descr->cld_end = end; + descr->cld_enq_flags = enqflags; - LASSERT(inode == - cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode); - vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime); - return 0; + cl_io_lock_add(env, io, &vio->vui_link); + + RETURN(0); } -static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios) +static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io, + __u32 enqflags, enum cl_lock_mode mode, + loff_t start, loff_t end) { - struct cl_io *io = ios->cis_io; - struct cl_object *obj = io->ci_obj; + struct cl_object *obj = io->ci_obj; - 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); + return vvp_io_one_lock_index(env, io, enqflags, mode, + cl_index(obj, start), cl_index(obj, end)); +} - if (vio->cui_ra_window_set) - ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead); - } +static int vvp_io_write_iter_init(const struct lu_env *env, + const struct cl_io_slice *ios) +{ + struct vvp_io *vio = cl2vvp_io(env, ios); + + cl_page_list_init(&vio->u.write.vui_queue); + vio->u.write.vui_written = 0; + vio->u.write.vui_from = 0; + vio->u.write.vui_to = PAGE_SIZE; + + return 0; +} +static void vvp_io_write_iter_fini(const struct lu_env *env, + const struct cl_io_slice *ios) +{ + struct vvp_io *vio = cl2vvp_io(env, ios); + + LASSERT(vio->u.write.vui_queue.pl_nr == 0); +} + +static int vvp_io_fault_iter_init(const struct lu_env *env, + const struct cl_io_slice *ios) +{ + struct vvp_io *vio = cl2vvp_io(env, ios); + struct inode *inode = vvp_object_inode(ios->cis_obj); + + LASSERT(inode == vio->vui_fd->fd_file->f_dentry->d_inode); + vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime); + + return 0; +} + +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 vvp_io *vio = cl2vvp_io(env, ios); + struct inode *inode = vvp_object_inode(obj); + + CLOBINVRNT(env, obj, vvp_object_invariant(obj)); + + CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d " + "restore needed %d\n", + PFID(lu_object_fid(&obj->co_lu)), + io->ci_ignore_layout, io->ci_verify_layout, + vio->vui_layout_gen, io->ci_restore_needed); + + if (io->ci_restore_needed == 1) { + int rc; + + /* file was detected release, we need to restore it + * before finishing the io + */ + rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF); + /* if restore registration failed, no restart, + * we will return -ENODATA */ + /* The layout will change after restore, so we need to + * block on layout lock hold by the MDT + * as MDT will not send new layout in lvb (see LU-3124) + * we have to explicitly fetch it, all this will be done + * by ll_layout_refresh() + */ + if (rc == 0) { + io->ci_restore_needed = 0; + io->ci_need_restart = 1; + io->ci_verify_layout = 1; + } else { + io->ci_restore_needed = 1; + io->ci_need_restart = 0; + io->ci_verify_layout = 0; + io->ci_result = rc; + } + } + + if (!io->ci_ignore_layout && io->ci_verify_layout) { + __u32 gen = 0; + + /* check layout version */ + ll_layout_refresh(inode, &gen); + io->ci_need_restart = vio->vui_layout_gen != gen; + if (io->ci_need_restart) { + CDEBUG(D_VFSTRACE, + DFID" layout changed from %d to %d.\n", + PFID(lu_object_fid(&obj->co_lu)), + vio->vui_layout_gen, gen); + /* today successful restore is the only possible + * case */ + /* restore was done, clear restoring state */ + ll_i2info(vvp_object_inode(obj))->lli_flags &= + ~LLIF_FILE_RESTORING; + } + } } static void vvp_io_fault_fini(const struct lu_env *env, @@ -104,7 +362,7 @@ static void vvp_io_fault_fini(const struct lu_env *env, struct cl_io *io = ios->cis_io; struct cl_page *page = io->u.ci_fault.ft_page; - CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj)); + CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj)); if (page != NULL) { lu_ref_del(&page->cp_reference, "fault", io); @@ -114,7 +372,7 @@ static void vvp_io_fault_fini(const struct lu_env *env, vvp_io_fini(env, ios); } -enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma) +static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma) { /* * we only want to hold PW locks if the mmap() can generate @@ -127,16 +385,17 @@ enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma) } static int vvp_mmap_locks(const struct lu_env *env, - struct ccc_io *vio, struct cl_io *io) + struct vvp_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; unsigned long addr; unsigned long seg; ssize_t count; - int result; + int result = 0; ENTRY; LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); @@ -144,11 +403,15 @@ 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); + if (vio->vui_iov == NULL) /* nfs or loop back device write */ + RETURN(0); - for (seg = 0; seg < vio->cui_nrsegs; seg++) { - const struct iovec *iv = &vio->cui_iov[seg]; + /* No MM (e.g. NFS)? No vmas too. */ + if (mm == NULL) + RETURN(0); + + for (seg = 0; seg < vio->vui_nrsegs; seg++) { + const struct iovec *iv = &vio->vui_iov[seg]; addr = (unsigned long)iv->iov_base; count = iv->iov_len; @@ -157,17 +420,19 @@ 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. - */ - flags = CEF_NEVER; - } + if (ll_file_nolock(vma->vm_file)) { + /* + * For no lock case is not allowed for mmap + */ + result = -EINVAL; + break; + } /* * XXX: Required lock mode can be weakened: CIT_WRITE @@ -188,56 +453,124 @@ static int vvp_mmap_locks(const struct lu_env *env, descr->cld_mode, descr->cld_start, descr->cld_end); - if (result < 0) - RETURN(result); + if (result < 0) + break; - if (vma->vm_end - addr >= count) - break; + if (vma->vm_end - addr >= count) + break; - count -= vma->vm_end - addr; - addr = vma->vm_end; - } - } - RETURN(0); + count -= vma->vm_end - addr; + addr = vma->vm_end; + } + up_read(&mm->mmap_sem); + if (result < 0) + break; + } + RETURN(result); +} + +static void vvp_io_advance(const struct lu_env *env, + const struct cl_io_slice *ios, + size_t nob) +{ + struct vvp_io *vio = cl2vvp_io(env, ios); + struct cl_io *io = ios->cis_io; + struct cl_object *obj = ios->cis_io->ci_obj; + + CLOBINVRNT(env, obj, vvp_object_invariant(obj)); + + if (!cl_is_normalio(env, io)) + return; + + LASSERT(vio->vui_tot_nrsegs >= vio->vui_nrsegs); + LASSERT(vio->vui_tot_count >= nob); + + vio->vui_iov += vio->vui_nrsegs; + vio->vui_tot_nrsegs -= vio->vui_nrsegs; + vio->vui_tot_count -= nob; + + /* update the iov */ + if (vio->vui_iov_olen > 0) { + struct iovec *iv; + + vio->vui_iov--; + vio->vui_tot_nrsegs++; + iv = &vio->vui_iov[0]; + if (io->ci_continue) { + iv->iov_base += iv->iov_len; + LASSERT(vio->vui_iov_olen > iv->iov_len); + iv->iov_len = vio->vui_iov_olen - iv->iov_len; + } else { + /* restore the iov_len, in case of restart io. */ + iv->iov_len = vio->vui_iov_olen; + } + vio->vui_iov_olen = 0; + } +} + +static void vvp_io_update_iov(const struct lu_env *env, + struct vvp_io *vio, struct cl_io *io) +{ + int i; + size_t size = io->u.ci_rw.crw_count; + + vio->vui_iov_olen = 0; + if (!cl_is_normalio(env, io) || vio->vui_tot_nrsegs == 0) + return; + + for (i = 0; i < vio->vui_tot_nrsegs; i++) { + struct iovec *iv = &vio->vui_iov[i]; + + if (iv->iov_len < size) { + size -= iv->iov_len; + } else { + if (iv->iov_len > size) { + vio->vui_iov_olen = iv->iov_len; + iv->iov_len = size; + } + break; + } + } + + vio->vui_nrsegs = i + 1; + LASSERTF(vio->vui_tot_nrsegs >= vio->vui_nrsegs, + "tot_nrsegs: %lu, nrsegs: %lu\n", + vio->vui_tot_nrsegs, vio->vui_nrsegs); } static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io, enum cl_lock_mode mode, loff_t start, loff_t end) { - struct ccc_io *cio = ccc_env_io(env); - int result; - int ast_flags = 0; + struct vvp_io *vio = vvp_env_io(env); + int result; + int ast_flags = 0; - LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); - ENTRY; + LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); + ENTRY; - ccc_io_update_iov(env, cio, io); + vvp_io_update_iov(env, vio, io); + + if (io->u.ci_rw.crw_nonblock) + ast_flags |= CEF_NONBLOCK; + + result = vvp_mmap_locks(env, vio, io); + if (result == 0) + result = vvp_io_one_lock(env, io, ast_flags, mode, start, end); - if (io->u.ci_rw.crw_nonblock) - ast_flags |= CEF_NONBLOCK; - result = vvp_mmap_locks(env, cio, io); - if (result == 0) - result = ccc_io_one_lock(env, io, ast_flags, mode, start, end); RETURN(result); } static int vvp_io_read_lock(const struct lu_env *env, const struct cl_io_slice *ios) { - struct cl_io *io = ios->cis_io; - struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj)); - int result; - - 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 */ - result = vvp_io_rw_lock(env, io, CLM_READ, - io->u.ci_rd.rd.crw_pos, - io->u.ci_rd.rd.crw_pos + - io->u.ci_rd.rd.crw_count - 1); - else - result = 0; - RETURN(result); + struct cl_io *io = ios->cis_io; + struct cl_io_rw_common *rd = &io->u.ci_rd.rd; + int result; + + ENTRY; + result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos, + rd->crw_pos + rd->crw_count - 1); + RETURN(result); } static int vvp_io_fault_lock(const struct lu_env *env, @@ -248,9 +581,11 @@ static int vvp_io_fault_lock(const struct lu_env *env, /* * XXX LDLM_FL_CBPENDING */ - return ccc_io_one_lock_index - (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma), - io->u.ci_fault.ft_index, io->u.ci_fault.ft_index); + return vvp_io_one_lock_index(env, + io, 0, + vvp_mode_from_vma(vio->u.fault.ft_vma), + io->u.ci_fault.ft_index, + io->u.ci_fault.ft_index); } static int vvp_io_write_lock(const struct lu_env *env, @@ -271,24 +606,9 @@ 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 +619,10 @@ 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 vvp_io *vio = vvp_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; @@ -316,67 +636,39 @@ static int vvp_io_setattr_lock(const struct lu_env *env, return 0; new_size = 0; } - cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK; - return ccc_io_one_lock(env, io, enqflags, CLM_WRITE, - new_size, OBD_OBJECT_EOF); + + vio->u.setattr.vui_local_lock = SETATTR_EXTENT_LOCK; + + return vvp_io_one_lock(env, io, enqflags, CLM_WRITE, + new_size, OBD_OBJECT_EOF); } 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(). - */ - ll_inode_size_lock(inode, 0); - result = vmtruncate(inode, size); - 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 = inode_newsize_ok(inode, size); + if (result < 0) { + ll_inode_size_unlock(inode); + return result; + } + i_size_write(inode, size); + + ll_truncate_pagecache(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, @@ -398,119 +690,72 @@ static int vvp_io_setattr_time(const struct lu_env *env, attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime; valid |= CAT_MTIME; } - result = cl_object_attr_set(env, obj, attr, valid); - cl_object_attr_unlock(obj); + result = cl_object_attr_update(env, obj, attr, valid); + cl_object_attr_unlock(obj); - return result; + return result; } 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 = vvp_object_inode(io->ci_obj); + int result = 0; + + mutex_lock(&inode->i_mutex); + if (cl_io_is_trunc(io)) + result = vvp_io_setattr_trunc(env, ios, inode, + io->u.ci_setattr.sa_attr.lvb_size); + if (result == 0) + result = vvp_io_setattr_time(env, ios); + return result; } 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 = vvp_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) -{ - 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); -} - -#ifdef HAVE_FILE_READV -static ssize_t lustre_generic_file_read(struct file *file, - struct ccc_io *vio, loff_t *ppos) -{ - return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos); -} - -static ssize_t lustre_generic_file_write(struct file *file, - struct ccc_io *vio, loff_t *ppos) -{ - return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos); -} -#else -static ssize_t lustre_generic_file_read(struct file *file, - struct ccc_io *vio, loff_t *ppos) + const struct cl_io_slice *ios) { - return generic_file_aio_read(vio->cui_iocb, vio->cui_iov, - vio->cui_nrsegs, *ppos); + vvp_io_fini(env, ios); } -static ssize_t lustre_generic_file_write(struct file *file, - struct ccc_io *vio, loff_t *ppos) -{ - return generic_file_aio_write(vio->cui_iocb, vio->cui_iov, - vio->cui_nrsegs, *ppos); -} -#endif - static int vvp_io_read_start(const struct lu_env *env, - const struct cl_io_slice *ios) + const struct cl_io_slice *ios) { - struct vvp_io *vio = cl2vvp_io(env, ios); - struct ccc_io *cio = cl2ccc_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 ll_ra_read *bead = &vio->cui_bead; - struct file *file = cio->cui_fd->fd_file; - - int result; - loff_t pos = io->u.ci_rd.rd.crw_pos; - long cnt = io->u.ci_rd.rd.crw_count; - long tot = cio->cui_tot_count; + 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 = vvp_object_inode(obj); + struct file *file = vio->vui_fd->fd_file; + + int result; + loff_t pos = io->u.ci_rd.rd.crw_pos; + long cnt = io->u.ci_rd.rd.crw_count; + long tot = vio->vui_tot_count; int exceed = 0; - CLOBINVRNT(env, obj, ccc_object_invariant(obj)); + CLOBINVRNT(env, obj, vvp_object_invariant(obj)); CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt); - result = ccc_prep_size(env, obj, io, pos, tot, &exceed); + if (!can_populate_pages(env, io, inode)) + return 0; + + result = vvp_prep_size(env, obj, io, pos, tot, &exceed); if (result != 0) return result; else if (exceed != 0) @@ -520,183 +765,384 @@ static int vvp_io_read_start(const struct lu_env *env, "Read ino %lu, %lu bytes, offset %lld, size %llu\n", inode->i_ino, cnt, pos, i_size_read(inode)); - /* turn off the kernel's read-ahead */ - cio->cui_fd->fd_file->f_ra.ra_pages = 0; + /* turn off the kernel's read-ahead */ + vio->vui_fd->fd_file->f_ra.ra_pages = 0; + + /* initialize read-ahead window once per syscall */ + if (!vio->vui_ra_valid) { + vio->vui_ra_valid = true; + vio->vui_ra_start = cl_index(obj, pos); + vio->vui_ra_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1); + ll_ras_enter(file); + } + + /* BUG: 5972 */ + file_accessed(file); + switch (vio->vui_io_subtype) { + case IO_NORMAL: + LASSERT(vio->vui_iocb->ki_pos == pos); + result = generic_file_aio_read(vio->vui_iocb, + vio->vui_iov, vio->vui_nrsegs, + vio->vui_iocb->ki_pos); + break; + case IO_SPLICE: + result = generic_file_splice_read(file, &pos, + vio->u.splice.vui_pipe, cnt, + vio->u.splice.vui_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; + default: + CERROR("Wrong IO type %u\n", vio->vui_io_subtype); + LBUG(); + } - /* initialize read-ahead window once per syscall */ - if (!vio->cui_ra_window_set) { - vio->cui_ra_window_set = 1; - bead->lrr_start = cl_index(obj, pos); - /* - * XXX: explicit CFS_PAGE_SIZE - */ - bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1); - ll_ra_read_in(file, bead); - } +out: + if (result >= 0) { + if (result < cnt) + io->ci_continue = 0; + io->ci_nob += result; + ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd, + pos, result, READ); + result = 0; + } + + return result; +} - /* BUG: 5972 */ - file_accessed(file); - switch (vio->cui_io_subtype) { - case IO_NORMAL: - result = lustre_generic_file_read(file, cio, &pos); - break; -#ifdef HAVE_KERNEL_SENDFILE - case IO_SENDFILE: - result = generic_file_sendfile(file, &pos, cnt, - vio->u.sendfile.cui_actor, - vio->u.sendfile.cui_target); - break; -#endif -#ifdef HAVE_KERNEL_SPLICE_READ - case IO_SPLICE: - 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: - CERROR("Wrong IO type %u\n", vio->cui_io_subtype); - LBUG(); - } +static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io, + struct cl_page_list *plist, int from, int to) +{ + struct cl_2queue *queue = &io->ci_queue; + struct cl_page *page; + unsigned int bytes = 0; + int rc = 0; + ENTRY; + + if (plist->pl_nr == 0) + RETURN(0); + + if (from > 0 || to != PAGE_SIZE) { + page = cl_page_list_first(plist); + if (plist->pl_nr == 1) { + cl_page_clip(env, page, from, to); + } else { + if (from > 0) + cl_page_clip(env, page, from, PAGE_SIZE); + if (to != PAGE_SIZE) { + page = cl_page_list_last(plist); + cl_page_clip(env, page, 0, to); + } + } + } + + cl_2queue_init(queue); + cl_page_list_splice(plist, &queue->c2_qin); + rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0); + + /* plist is not sorted any more */ + cl_page_list_splice(&queue->c2_qin, plist); + cl_page_list_splice(&queue->c2_qout, plist); + cl_2queue_fini(env, queue); + + if (rc == 0) { + /* calculate bytes */ + bytes = plist->pl_nr << PAGE_SHIFT; + bytes -= from + PAGE_SIZE - to; + + while (plist->pl_nr > 0) { + page = cl_page_list_first(plist); + cl_page_list_del(env, plist, page); + + cl_page_clip(env, page, 0, PAGE_SIZE); + + SetPageUptodate(cl_page_vmpage(page)); + cl_page_disown(env, io, page); + + /* held in ll_cl_init() */ + lu_ref_del(&page->cp_reference, "cl_io", io); + cl_page_put(env, page); + } + } + + RETURN(bytes > 0 ? bytes : rc); +} -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; +static void write_commit_callback(const struct lu_env *env, struct cl_io *io, + struct cl_page *page) +{ + struct vvp_page *vpg; + struct page *vmpage = page->cp_vmpage; + struct cl_object *clob = cl_io_top(io)->ci_obj; + + SetPageUptodate(vmpage); + set_page_dirty(vmpage); + + vpg = cl2vvp_page(cl_object_page_slice(clob, page)); + vvp_write_pending(cl2vvp(clob), vpg); + + cl_page_disown(env, io, page); + + /* held in ll_cl_init() */ + lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io)); + cl_page_put(env, page); +} + +/* make sure the page list is contiguous */ +static bool page_list_sanity_check(struct cl_object *obj, + struct cl_page_list *plist) +{ + struct cl_page *page; + pgoff_t index = CL_PAGE_EOF; + + cl_page_list_for_each(page, plist) { + struct vvp_page *vpg = cl_object_page_slice(obj, page); + + if (index == CL_PAGE_EOF) { + index = vvp_index(vpg); + continue; + } + + ++index; + if (index == vvp_index(vpg)) + continue; + + return false; + } + return true; +} + +/* Return how many bytes have queued or written */ +int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io) +{ + struct cl_object *obj = io->ci_obj; + struct inode *inode = vvp_object_inode(obj); + struct vvp_io *vio = vvp_env_io(env); + struct cl_page_list *queue = &vio->u.write.vui_queue; + struct cl_page *page; + int rc = 0; + int bytes = 0; + unsigned int npages = vio->u.write.vui_queue.pl_nr; + ENTRY; + + if (npages == 0) + RETURN(0); + + CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n", + npages, vio->u.write.vui_from, vio->u.write.vui_to); + + LASSERT(page_list_sanity_check(obj, queue)); + + /* submit IO with async write */ + rc = cl_io_commit_async(env, io, queue, + vio->u.write.vui_from, vio->u.write.vui_to, + write_commit_callback); + npages -= queue->pl_nr; /* already committed pages */ + if (npages > 0) { + /* calculate how many bytes were written */ + bytes = npages << PAGE_SHIFT; + + /* first page */ + bytes -= vio->u.write.vui_from; + if (queue->pl_nr == 0) /* last page */ + bytes -= PAGE_SIZE - vio->u.write.vui_to; + LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages); + + vio->u.write.vui_written += bytes; + + CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n", + npages, bytes, vio->u.write.vui_written); + + /* the first page must have been written. */ + vio->u.write.vui_from = 0; + } + LASSERT(page_list_sanity_check(obj, queue)); + LASSERT(ergo(rc == 0, queue->pl_nr == 0)); + + /* out of quota, try sync write */ + if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) { + rc = vvp_io_commit_sync(env, io, queue, + vio->u.write.vui_from, + vio->u.write.vui_to); + if (rc > 0) { + vio->u.write.vui_written += rc; + rc = 0; + } + } + + /* update inode size */ + ll_merge_attr(env, inode); + + /* Now the pages in queue were failed to commit, discard them + * unless they were dirtied before. */ + while (queue->pl_nr > 0) { + page = cl_page_list_first(queue); + cl_page_list_del(env, queue, page); + + if (!PageDirty(cl_page_vmpage(page))) + cl_page_discard(env, io, page); + + cl_page_disown(env, io, page); + + /* held in ll_cl_init() */ + lu_ref_del(&page->cp_reference, "cl_io", io); + cl_page_put(env, page); + } + cl_page_list_fini(env, queue); + + RETURN(rc); } static int vvp_io_write_start(const struct lu_env *env, const struct cl_io_slice *ios) { - struct ccc_io *cio = cl2ccc_io(env, 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 file *file = cio->cui_fd->fd_file; + struct inode *inode = vvp_object_inode(obj); ssize_t result = 0; loff_t pos = io->u.ci_wr.wr.crw_pos; size_t cnt = io->u.ci_wr.wr.crw_count; ENTRY; - if (cl_io_is_append(io)) { - /* - * PARALLEL IO This has to be changed for parallel IO doing - * out-of-order writes. - */ - pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode); -#ifndef HAVE_FILE_WRITEV - cio->cui_iocb->ki_pos = pos; -#endif - } - - CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt); - - if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */ - result = 0; - 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 (!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 + * out-of-order writes. + */ + ll_merge_attr(env, inode); + pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode); + vio->vui_iocb->ki_pos = pos; + } else { + LASSERT(vio->vui_iocb->ki_pos == pos); + } + + CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt); + + if (vio->vui_iov == NULL) { + /* from a temp io in ll_cl_init(). */ + result = 0; + } else { + /* + * When using the locked AIO function (generic_file_aio_write()) + * testing has shown the inode mutex to be a limiting factor + * with multi-threaded single shared file performance. To get + * around this, we now use the lockless version. To maintain + * consistency, proper locking to protect against writes, + * trucates, etc. is handled in the higher layers of lustre. + */ + result = __generic_file_aio_write(vio->vui_iocb, + vio->vui_iov, vio->vui_nrsegs, + &vio->vui_iocb->ki_pos); + if (result > 0 || result == -EIOCBQUEUED) { + ssize_t err; + + err = generic_write_sync(vio->vui_iocb->ki_filp, + pos, result); + if (err < 0 && result > 0) + result = err; + } + + } + if (result > 0) { + result = vvp_io_write_commit(env, io); + if (vio->u.write.vui_written > 0) { + result = vio->u.write.vui_written; + io->ci_nob += result; + + CDEBUG(D_VFSTRACE, "write: nob %zd, result: %zd\n", + io->ci_nob, result); + } + } + if (result > 0) { + struct ll_inode_info *lli = ll_i2info(inode); + + spin_lock(&lli->lli_lock); + lli->lli_flags |= LLIF_DATA_MODIFIED; + spin_unlock(&lli->lli_lock); + + if (result < cnt) + io->ci_continue = 0; + ll_rw_stats_tally(ll_i2sbi(inode), current->pid, + vio->vui_fd, pos, result, WRITE); + result = 0; + } + + RETURN(result); } -#ifndef HAVE_VM_OP_FAULT static int vvp_io_kernel_fault(struct vvp_fault_io *cfio) { - cfs_page_t *vmpage; - - vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address, - cfio->nopage.ft_type); - - if (vmpage == NOPAGE_SIGBUS) { - CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n", - cfio->nopage.ft_address,(long)cfio->nopage.ft_type); - return -EFAULT; - } else if (vmpage == NOPAGE_OOM) { - CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n", - cfio->nopage.ft_address, (long)cfio->nopage.ft_type); - return -ENOMEM; - } - - LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n", - cfio->nopage.ft_address, (long)cfio->nopage.ft_type); + struct vm_fault *vmf = cfio->ft_vmf; - cfio->ft_vmpage = vmpage; - lock_page(vmpage); + cfio->ft_flags = filemap_fault(cfio->ft_vma, vmf); + cfio->ft_flags_valid = 1; - return 0; -} -#else -static int vvp_io_kernel_fault(struct vvp_fault_io *cfio) -{ - struct vm_fault *vmf = cfio->fault.ft_vmf; + if (vmf->page) { + LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n", + vmf->virtual_address); + if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) { + lock_page(vmf->page); + cfio->ft_flags |= VM_FAULT_LOCKED; + } - cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf); + cfio->ft_vmpage = vmf->page; - 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; - } + return 0; + } - cfio->ft_vmpage = vmf->page; - return 0; - } + if (cfio->ft_flags & VM_FAULT_SIGBUS) { + CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address); + return -EFAULT; + } - if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) { - CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address); - return -EFAULT; - } + if (cfio->ft_flags & VM_FAULT_OOM) { + CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address); + return -ENOMEM; + } - if (cfio->fault.ft_flags & VM_FAULT_OOM) { - CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address); - return -ENOMEM; - } + if (cfio->ft_flags & VM_FAULT_RETRY) + return -EAGAIN; - if (cfio->fault.ft_flags & VM_FAULT_RETRY) - return -EAGAIN; + CERROR("unknown error in page fault %d\n", cfio->ft_flags); - CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags); - return -EINVAL; + return -EINVAL; } -#endif +static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io, + struct cl_page *page) +{ + struct vvp_page *vpg; + struct cl_object *clob = cl_io_top(io)->ci_obj; + + set_page_dirty(page->cp_vmpage); + + vpg = cl2vvp_page(cl_object_page_slice(clob, page)); + vvp_write_pending(cl2vvp(clob), vpg); +} 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 result = 0; - cfs_page_t *vmpage = NULL; - 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 = vvp_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; + struct page *vmpage = NULL; + struct cl_page *page; + loff_t size; + pgoff_t last_index; + ENTRY; if (fio->ft_executable && LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime) @@ -707,29 +1153,31 @@ 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, NULL); - if (result != 0) - return 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; - } - - vmpage = cfio->ft_vmpage; - LASSERT(PageLocked(vmpage)); - - if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE)) - ll_invalidate_page(vmpage); - + result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL); + if (result != 0) + RETURN(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); + } + + vmpage = cfio->ft_vmpage; + LASSERT(PageLocked(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(vmpage->mapping == NULL)) { + 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 @@ -737,43 +1185,90 @@ static int vvp_io_fault_start(const struct lu_env *env, GOTO(out, result = +1); } - 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); - if (result < 0) { - cl_page_unassume(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); - } - } - } - - size = i_size_read(inode); - last = cl_index(obj, size - 1); - LASSERT(fio->ft_index <= last); - if (fio->ft_index == last) + last_index = cl_index(obj, size - 1); + + if (fio->ft_mkwrite ) { + /* + * 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. + */ + 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 (!PageDirty(vmpage)) { + struct cl_page_list *plist = &io->ci_queue.c2_qin; + struct vvp_page *vpg = cl_object_page_slice(obj, page); + int to = PAGE_SIZE; + + /* vvp_page_assume() calls wait_on_page_writeback(). */ + cl_page_assume(env, io, page); + + cl_page_list_init(plist); + cl_page_list_add(plist, page); + + /* size fixup */ + if (last_index == vvp_index(vpg)) + to = size & ~CFS_PAGE_MASK; + + /* 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_io_commit_async(env, io, plist, 0, to, + mkwrite_commit_callback); + LASSERT(cl_page_is_owned(page, io)); + cl_page_list_fini(env, plist); + + vmpage = NULL; + if (result < 0) { + 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); + } + } + + /* + * 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_index); + if (fio->ft_index == last_index) /* * Last page is mapped partially. */ @@ -786,297 +1281,83 @@ static int vvp_io_fault_start(const struct lu_env *env, EXIT; out: - /* return unlocked vmpage to avoid deadlocking */ - unlock_page(vmpage); -#ifdef HAVE_VM_OP_FAULT - cfio->fault.ft_flags &= ~VM_FAULT_LOCKED; -#endif - return result; -} + /* return unlocked vmpage to avoid deadlocking */ + if (vmpage != NULL) + unlock_page(vmpage); -static int vvp_io_read_page(const struct lu_env *env, - const struct cl_io_slice *ios, - const struct cl_page_slice *slice) -{ - struct cl_io *io = ios->cis_io; - struct cl_object *obj = slice->cpl_obj; - struct ccc_page *cp = cl2ccc_page(slice); - struct cl_page *page = slice->cpl_page; - struct inode *inode = ccc_object_inode(obj); - struct ll_sb_info *sbi = ll_i2sbi(inode); - struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd; - struct ll_readahead_state *ras = &fd->fd_ras; - cfs_page_t *vmpage = cp->cpg_page; - struct cl_2queue *queue = &io->ci_queue; - int rc; - - CLOBINVRNT(env, obj, ccc_object_invariant(obj)); - LASSERT(slice->cpl_obj == obj); - - ENTRY; - - 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); - - /* Sanity check whether the page is protected by a lock. */ - rc = cl_page_is_under_lock(env, io, page); - if (rc != -EBUSY) { - CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n", - rc == -ENODATA ? "without a lock" : - "match failed", rc); - if (rc != -ENODATA) - RETURN(rc); - } + cfio->ft_flags &= ~VM_FAULT_LOCKED; - if (cp->cpg_defer_uptodate) { - cp->cpg_ra_used = 1; - cl_page_export(env, page, 1); - } - /* - * Add page into the queue even when it is marked uptodate above. - * 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 && - sbi->ll_ra_info.ra_max_pages) - ll_readahead(env, io, ras, - vmpage->mapping, &queue->c2_qin, fd->fd_flags); - - RETURN(0); + return result; } -static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io, - struct cl_page *page, struct ccc_page *cp, - enum cl_req_type crt) +static int vvp_io_fsync_start(const struct lu_env *env, + const struct cl_io_slice *ios) { - struct cl_2queue *queue; - int result; - - LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); - - queue = &io->ci_queue; - cl_2queue_init_page(queue, page); - - result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0); - LASSERT(cl_page_is_owned(page, io)); - - if (crt == CRT_READ) - /* - * in CRT_WRITE case page is left locked even in case of - * error. - */ - cl_page_list_disown(env, io, &queue->c2_qin); - cl_2queue_fini(env, queue); - - return result; + /* 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; } -/** - * Prepare partially written-to page for a write. - */ -static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io, - struct cl_object *obj, struct cl_page *pg, - struct ccc_page *cp, - unsigned from, unsigned to) -{ - struct cl_attr *attr = ccc_env_thread_attr(env); - loff_t offset = cl_offset(obj, pg->cp_index); - int result; - - cl_object_attr_lock(obj); - result = cl_object_attr_get(env, obj, attr); - cl_object_attr_unlock(obj); - if (result == 0) { - /* - * If are writing to a new page, no need to read old data. - * The extent locking will have updated the KMS, and for our - * purposes here we can treat it like i_size. - */ - if (attr->cat_kms <= offset) { - char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0); - - memset(kaddr, 0, cl_page_size(obj)); - kunmap_atomic(kaddr, KM_USER0); - } else if (cp->cpg_defer_uptodate) - cp->cpg_ra_used = 1; - else - result = vvp_page_sync_io(env, io, pg, cp, CRT_READ); - /* - * In older implementations, obdo_refresh_inode is called here - * to update the inode because the write might modify the - * object info at OST. However, this has been proven useless, - * since LVB functions will be called when user space program - * tries to retrieve inode attribute. Also, see bug 15909 for - * details. -jay - */ - if (result == 0) - cl_page_export(env, pg, 1); - } - return result; -} - -static int vvp_io_prepare_write(const struct lu_env *env, - const struct cl_io_slice *ios, - const struct cl_page_slice *slice, - unsigned from, unsigned to) +static int vvp_io_read_page(const struct lu_env *env, + const struct cl_io_slice *ios, + const struct cl_page_slice *slice) { - struct cl_object *obj = slice->cpl_obj; - struct ccc_page *cp = cl2ccc_page(slice); - struct cl_page *pg = slice->cpl_page; - cfs_page_t *vmpage = cp->cpg_page; - - int result; - - ENTRY; - - LINVRNT(cl_page_is_vmlocked(env, pg)); - LASSERT(vmpage->mapping->host == ccc_object_inode(obj)); - - result = 0; - - CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to); - if (!PageUptodate(vmpage)) { - /* - * We're completely overwriting an existing page, so _don't_ - * set it up to date until commit_write - */ - if (from == 0 && to == CFS_PAGE_SIZE) { - CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n"); - POISON_PAGE(page, 0x11); - } else - result = vvp_io_prepare_partial(env, ios->cis_io, obj, - pg, cp, from, to); - } else - CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n"); - RETURN(result); + struct cl_io *io = ios->cis_io; + struct vvp_page *vpg = cl2vvp_page(slice); + struct cl_page *page = slice->cpl_page; + struct inode *inode = vvp_object_inode(slice->cpl_obj); + struct ll_sb_info *sbi = ll_i2sbi(inode); + struct ll_file_data *fd = cl2vvp_io(env, ios)->vui_fd; + struct ll_readahead_state *ras = &fd->fd_ras; + struct cl_2queue *queue = &io->ci_queue; + + ENTRY; + + if (sbi->ll_ra_info.ra_max_pages_per_file > 0 && + sbi->ll_ra_info.ra_max_pages > 0) + ras_update(sbi, inode, ras, vvp_index(vpg), + vpg->vpg_defer_uptodate); + + if (vpg->vpg_defer_uptodate) { + vpg->vpg_ra_used = 1; + cl_page_export(env, page, 1); + } + + /* + * Add page into the queue even when it is marked uptodate above. + * 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 > 0 && + sbi->ll_ra_info.ra_max_pages > 0) + ll_readahead(env, io, &queue->c2_qin, ras, + vpg->vpg_defer_uptodate); + + RETURN(0); } -static int vvp_io_commit_write(const struct lu_env *env, - const struct cl_io_slice *ios, - const struct cl_page_slice *slice, - unsigned from, unsigned to) +static void vvp_io_end(const struct lu_env *env, const struct cl_io_slice *ios) { - struct cl_object *obj = slice->cpl_obj; - struct cl_io *io = ios->cis_io; - struct ccc_page *cp = cl2ccc_page(slice); - struct cl_page *pg = slice->cpl_page; - struct inode *inode = ccc_object_inode(obj); - struct ll_sb_info *sbi = ll_i2sbi(inode); - cfs_page_t *vmpage = cp->cpg_page; - - int result; - int tallyop; - loff_t size; - - ENTRY; - - LINVRNT(cl_page_is_vmlocked(env, pg)); - LASSERT(vmpage->mapping->host == inode); - - LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n"); - CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to); - - /* - * queue a write for some time in the future the first time we - * dirty the page. - * - * This is different from what other file systems do: they usually - * just mark page (and some of its buffers) dirty and rely on - * balance_dirty_pages() to start a write-back. Lustre wants write-back - * to be started earlier for the following reasons: - * - * (1) with a large number of clients we need to limit the amount - * of cached data on the clients a lot; - * - * (2) large compute jobs generally want compute-only then io-only - * and the IO should complete as quickly as possible; - * - * (3) IO is batched up to the RPC size and is async until the - * client max cache is hit - * (/proc/fs/lustre/osc/OSC.../max_dirty_mb) - * - */ - 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) { - pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT; - bool need_clip = true; - - /* - * Client ran out of disk space grant. Possible - * strategies are: - * - * (a) do a sync write, renewing grant; - * - * (b) stop writing on this stripe, switch to the - * next one. - * - * (b) is a part of "parallel io" design that is the - * ultimate goal. (a) is what "old" client did, and - * what the new code continues to do for the time - * being. - */ - if (last_index > pg->cp_index) { - to = CFS_PAGE_SIZE; - need_clip = false; - } else if (last_index == pg->cp_index) { - int size_to = i_size_read(inode) & ~CFS_PAGE_MASK; - if (to < size_to) - to = size_to; - } - if (need_clip) - cl_page_clip(env, pg, 0, to); - result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE); - if (result) - CERROR("Write page %lu of inode %p failed %d\n", - pg->cp_index, inode, result); - } - } else { - tallyop = LPROC_LL_DIRTY_HITS; - result = 0; - } - ll_stats_ops_tally(sbi, tallyop, 1); - - size = cl_offset(obj, pg->cp_index) + to; - - ll_inode_size_lock(inode, 0); - if (result == 0) { - if (size > i_size_read(inode)) { - cl_isize_write_nolock(inode, size); - CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n", - PFID(lu_object_fid(&obj->co_lu)), - (unsigned long)size); - } - cl_page_export(env, pg, 1); - } else { - if (size > i_size_read(inode)) - cl_page_discard(env, io, pg); - } - ll_inode_size_unlock(inode, 0); - RETURN(result); + CLOBINVRNT(env, ios->cis_io->ci_obj, + vvp_object_invariant(ios->cis_io->ci_obj)); } static const struct cl_io_operations vvp_io_ops = { - .op = { - [CIT_READ] = { - .cio_fini = vvp_io_fini, - .cio_lock = vvp_io_read_lock, - .cio_start = vvp_io_read_start, - .cio_advance = ccc_io_advance - }, + .op = { + [CIT_READ] = { + .cio_fini = vvp_io_fini, + .cio_lock = vvp_io_read_lock, + .cio_start = vvp_io_read_start, + .cio_advance = vvp_io_advance, + }, [CIT_WRITE] = { - .cio_fini = vvp_io_fini, - .cio_lock = vvp_io_write_lock, - .cio_start = vvp_io_write_start, - .cio_advance = ccc_io_advance + .cio_fini = vvp_io_fini, + .cio_iter_init = vvp_io_write_iter_init, + .cio_iter_fini = vvp_io_write_iter_fini, + .cio_lock = vvp_io_write_lock, + .cio_start = vvp_io_write_start, + .cio_advance = vvp_io_advance, }, [CIT_SETATTR] = { .cio_fini = vvp_io_setattr_fini, @@ -1090,55 +1371,87 @@ static const struct cl_io_operations vvp_io_ops = { .cio_iter_init = vvp_io_fault_iter_init, .cio_lock = vvp_io_fault_lock, .cio_start = vvp_io_fault_start, - .cio_end = ccc_io_end + .cio_end = vvp_io_end, }, + [CIT_FSYNC] = { + .cio_start = vvp_io_fsync_start, + .cio_fini = vvp_io_fini + }, [CIT_MISC] = { .cio_fini = vvp_io_fini } }, .cio_read_page = vvp_io_read_page, - .cio_prepare_write = vvp_io_prepare_write, - .cio_commit_write = vvp_io_commit_write }; 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); - int result; - - CLOBINVRNT(env, obj, ccc_object_invariant(obj)); - ENTRY; - - 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; + struct vvp_io *vio = vvp_env_io(env); + struct inode *inode = vvp_object_inode(obj); + int result; + + CLOBINVRNT(env, obj, vvp_object_invariant(obj)); + ENTRY; + + CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d " + "restore needed %d\n", + PFID(lu_object_fid(&obj->co_lu)), + io->ci_ignore_layout, io->ci_verify_layout, + vio->vui_layout_gen, io->ci_restore_needed); + + CL_IO_SLICE_CLEAN(vio, vui_cl); + cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops); + vio->vui_ra_valid = false; + 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 * results." -- Single Unix Spec */ if (count == 0) result = 1; - else { - cio->cui_tot_count = count; - cio->cui_tot_nrsegs = 0; - } - } else if (io->ci_type == CIT_SETATTR) { - if (!cl_io_is_trunc(io)) - io->ci_lockreq = CILR_MANDATORY; - } - RETURN(result); + else { + vio->vui_tot_count = count; + vio->vui_tot_nrsegs = 0; + } + + /* 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, &vio->vui_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, - const struct cl_io_slice *slice) -{ - /* Caling just for assertion */ - cl2ccc_io(env, slice); - return vvp_env_io(env); -} -