X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fvvp_io.c;h=9f92f9eace37f6500fd2f2ece553c15ea8234683;hp=dc0d2340ec4824631ec5d565542483dfaac142a7;hb=1515e409cc57af5eaef809eee6d8f8d6725d092b;hpb=488f93f92541e07d269c7e3d94905f96e9cd0609 diff --git a/lustre/llite/vvp_io.c b/lustre/llite/vvp_io.c index dc0d234..9f92f9e 100644 --- a/lustre/llite/vvp_io.c +++ b/lustre/llite/vvp_io.c @@ -26,8 +26,11 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved. + * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011 Whamcloud, Inc. + * */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -52,6 +55,18 @@ static struct vvp_io *cl2vvp_io(const struct lu_env *env, const struct cl_io_slice *slice); +/** + * True, if \a io is a normal io, False for sendfile() / splice_{read|write} + */ +int cl_is_normalio(const struct lu_env *env, const struct cl_io *io) +{ + struct vvp_io *vio = vvp_env_io(env); + + LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); + + return vio->cui_io_subtype == IO_NORMAL; +} + /***************************************************************************** * * io operations. @@ -76,9 +91,7 @@ static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios) struct cl_object *obj = io->ci_obj; CLOBINVRNT(env, obj, ccc_object_invariant(obj)); - if (io->ci_type == CIT_WRITE) - up(&ll_i2info(ccc_object_inode(obj))->lli_write_sem); - else { + if (io->ci_type == CIT_READ) { struct vvp_io *vio = cl2vvp_io(env, ios); struct ccc_io *cio = cl2ccc_io(env, ios); @@ -123,7 +136,6 @@ static int vvp_mmap_locks(const struct lu_env *env, struct vm_area_struct *vma; struct cl_lock_descr *descr = &cti->cti_descr; ldlm_policy_data_t policy; - struct inode *inode; unsigned long addr; unsigned long seg; ssize_t count; @@ -132,7 +144,7 @@ static int vvp_mmap_locks(const struct lu_env *env, LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); - if (cl_io_is_sendfile(io)) + if (!cl_is_normalio(env, io)) RETURN(0); for (seg = 0; seg < vio->cui_nrsegs; seg++) { @@ -146,16 +158,16 @@ 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) { - struct file *file = vma->vm_file; - struct ll_file_data *fd; - - LASSERT(file); - fd = LUSTRE_FPRIVATE(file); + struct inode *inode = vma->vm_file->f_dentry->d_inode; + int flags = CEF_MUST; - inode = file->f_dentry->d_inode; - if (!(fd->fd_flags & LL_FILE_IGNORE_LOCK || - ll_i2sbi(inode)->ll_flags & LL_SBI_NOLCK)) - goto cont; + if (ll_file_nolock(vma->vm_file)) { + /* + * For no lock case, a lockless lock will be + * generated. + */ + flags = CEF_NEVER; + } /* * XXX: Required lock mode can be weakened: CIT_WRITE @@ -169,13 +181,19 @@ static int vvp_mmap_locks(const struct lu_env *env, policy.l_extent.start); descr->cld_end = cl_index(descr->cld_obj, policy.l_extent.end); - result = cl_io_lock_alloc_add(env, io, descr, CEF_MUST); + descr->cld_enq_flags = flags; + result = cl_io_lock_alloc_add(env, io, descr); + + CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n", + descr->cld_mode, descr->cld_start, + descr->cld_end); + if (result < 0) RETURN(result); - cont: if (vma->vm_end - addr >= count) break; + count -= vma->vm_end - addr; addr = vma->vm_end; } @@ -183,36 +201,6 @@ static int vvp_mmap_locks(const struct lu_env *env, RETURN(0); } -static void vvp_io_update_iov(const struct lu_env *env, - struct ccc_io *vio, struct cl_io *io) -{ - int i; - size_t size = io->u.ci_rw.crw_count; - - vio->cui_iov_olen = 0; - if (cl_io_is_sendfile(io) || size == vio->cui_tot_count) - return; - - if (vio->cui_tot_nrsegs == 0) - vio->cui_tot_nrsegs = vio->cui_nrsegs; - - for (i = 0; i < vio->cui_tot_nrsegs; i++) { - struct iovec *iv = &vio->cui_iov[i]; - - if (iv->iov_len < size) - size -= iv->iov_len; - else { - if (iv->iov_len > size) { - vio->cui_iov_olen = iv->iov_len; - iv->iov_len = size; - } - break; - } - } - - vio->cui_nrsegs = i + 1; -} - 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) { @@ -221,10 +209,9 @@ static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io, int ast_flags = 0; LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); - LASSERT(vvp_env_io(env)->cui_oneshot == 0); ENTRY; - vvp_io_update_iov(env, cio, io); + ccc_io_update_iov(env, cio, io); if (io->u.ci_rw.crw_nonblock) ast_flags |= CEF_NONBLOCK; @@ -272,26 +259,21 @@ static int vvp_io_write_lock(const struct lu_env *env, struct cl_io *io = ios->cis_io; loff_t start; loff_t end; - int result; - if (cl2vvp_io(env, ios)->cui_oneshot == 0) { - if (io->u.ci_wr.wr_append) { - start = 0; - end = OBD_OBJECT_EOF; - } else { - start = io->u.ci_wr.wr.crw_pos; - end = start + io->u.ci_wr.wr.crw_count - 1; - } - result = vvp_io_rw_lock(env, io, CLM_WRITE, start, end); - } else - result = 0; - return result; + if (io->u.ci_wr.wr_append) { + start = 0; + end = OBD_OBJECT_EOF; + } else { + start = io->u.ci_wr.wr.crw_pos; + end = start + io->u.ci_wr.wr.crw_count - 1; + } + return vvp_io_rw_lock(env, io, CLM_WRITE, start, end); } -static int vvp_io_trunc_iter_init(const struct lu_env *env, - const struct cl_io_slice *ios) +static int vvp_io_setattr_iter_init(const struct lu_env *env, + const struct cl_io_slice *ios) { - struct ccc_io *vio = cl2ccc_io(env, ios); + struct ccc_io *cio = ccc_env_io(env); struct inode *inode = ccc_object_inode(ios->cis_obj); /* @@ -302,61 +284,73 @@ static int vvp_io_trunc_iter_init(const struct lu_env *env, * This last one is especially bad for racing o_append users on other * nodes. */ - UNLOCK_INODE_MUTEX(inode); - UP_WRITE_I_ALLOC_SEM(inode); - vio->u.trunc.cui_locks_released = 1; + if (cl_io_is_trunc(ios->cis_io)) + UP_WRITE_I_ALLOC_SEM(inode); + cio->u.setattr.cui_locks_released = 1; return 0; } /** - * Implementation of cl_io_operations::vio_lock() method for CIT_TRUNC io. + * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io. * * Handles "lockless io" mode when extent locking is done by server. */ -static int vvp_io_trunc_lock(const struct lu_env *env, - const struct cl_io_slice *ios) +static int vvp_io_setattr_lock(const struct lu_env *env, + const struct cl_io_slice *ios) { - struct ccc_io *vio = cl2ccc_io(env, ios); - struct cl_io *io = ios->cis_io; - size_t new_size = io->u.ci_truncate.tr_size; - __u32 enqflags = new_size == 0 ? CEF_DISCARD_DATA : 0; - int result; + struct ccc_io *cio = ccc_env_io(env); + struct cl_io *io = ios->cis_io; + size_t new_size; + __u32 enqflags = 0; + + if (cl_io_is_trunc(io)) { + new_size = io->u.ci_setattr.sa_attr.lvb_size; + if (new_size == 0) + enqflags = CEF_DISCARD_DATA; + } else { + if ((io->u.ci_setattr.sa_attr.lvb_mtime >= + io->u.ci_setattr.sa_attr.lvb_ctime) || + (io->u.ci_setattr.sa_attr.lvb_atime >= + io->u.ci_setattr.sa_attr.lvb_ctime)) + 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); +} + +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); - vio->u.trunc.cui_local_lock = TRUNC_EXTENT; - result = ccc_io_one_lock(env, io, enqflags, CLM_WRITE, - new_size, OBD_OBJECT_EOF); return result; } -static int vvp_io_trunc_start(const struct lu_env *env, - const struct cl_io_slice *ios) +static int vvp_io_setattr_trunc(const struct lu_env *env, + const struct cl_io_slice *ios, + struct inode *inode, loff_t size) { - struct ccc_io *cio = cl2ccc_io(env, 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); struct cl_object *obj = ios->cis_obj; - size_t size = io->u.ci_truncate.tr_size; pgoff_t start = cl_index(obj, size); int result; - LASSERT(cio->u.trunc.cui_locks_released); - LASSERT(vio->cui_oneshot == 0); - - LOCK_INODE_MUTEX(inode); DOWN_WRITE_I_ALLOC_SEM(inode); - cio->u.trunc.cui_locks_released = 0; - /* - * 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); + result = vvp_do_vmtruncate(inode, size); + /* * If a page is partially truncated, keep it owned across truncate to * prevent... races. @@ -368,9 +362,9 @@ static int vvp_io_trunc_start(const struct lu_env *env, struct cl_object_header *hdr; hdr = cl_object_header(obj); - spin_lock(&hdr->coh_page_guard); + cfs_spin_lock(&hdr->coh_page_guard); vio->cui_partpage = cl_page_lookup(hdr, start); - spin_unlock(&hdr->coh_page_guard); + cfs_spin_unlock(&hdr->coh_page_guard); if (vio->cui_partpage != NULL) /* @@ -387,28 +381,84 @@ static int vvp_io_trunc_start(const struct lu_env *env, return result; } -static void vvp_io_trunc_end(const struct lu_env *env, - const struct cl_io_slice *ios) +static int vvp_io_setattr_time(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 cl_attr *attr = ccc_env_thread_attr(env); + int result; + unsigned valid = CAT_CTIME; + + cl_object_attr_lock(obj); + attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime; + if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) { + attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime; + valid |= CAT_ATIME; + } + if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) { + 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); + + return result; +} +static int vvp_io_setattr_start(const struct lu_env *env, + 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); +} + +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); } -static void vvp_io_trunc_fini(const struct lu_env *env, - const struct cl_io_slice *ios) +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.trunc.cui_locks_released) { + if (cio->u.setattr.cui_locks_released) { LOCK_INODE_MUTEX(inode); - DOWN_WRITE_I_ALLOC_SEM(inode); - cio->u.trunc.cui_locks_released = 0; + if (cl_io_is_trunc(io)) + DOWN_WRITE_I_ALLOC_SEM(inode); + cio->u.setattr.cui_locks_released = 0; } vvp_io_fini(env, ios); } @@ -454,14 +504,13 @@ static int vvp_io_read_start(const struct lu_env *env, int result; loff_t pos = io->u.ci_rd.rd.crw_pos; - size_t cnt = io->u.ci_rd.rd.crw_count; - size_t tot = cio->cui_tot_count; + long cnt = io->u.ci_rd.rd.crw_count; + long tot = cio->cui_tot_count; int exceed = 0; CLOBINVRNT(env, obj, ccc_object_invariant(obj)); - LASSERT(vio->cui_oneshot == 0); - CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + (long long)cnt); + CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt); result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed); if (result != 0) @@ -470,7 +519,7 @@ static int vvp_io_read_start(const struct lu_env *env, goto out; LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, - "Read ino %lu, "LPSZ" bytes, offset %lld, size %llu\n", + "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 */ @@ -489,11 +538,27 @@ static int vvp_io_read_start(const struct lu_env *env, /* BUG: 5972 */ file_accessed(file); - if (cl_io_is_sendfile(io)) { + 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.read.cui_actor, vio->u.read.cui_target); - } else { - result = lustre_generic_file_read(file, cio, &pos); + 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); + break; +#endif + default: + CERROR("Wrong IO type %u\n", vio->cui_io_subtype); + LBUG(); } out: @@ -535,7 +600,7 @@ static int vvp_io_write_start(const struct lu_env *env, CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt); - if (cl2vvp_io(env, ios)->cui_oneshot > 0) + if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */ result = 0; else result = lustre_generic_file_write(file, cio, &pos); @@ -551,6 +616,72 @@ static int vvp_io_write_start(const struct lu_env *env, 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); + + cfio->ft_vmpage = 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; + } + + cfio->ft_vmpage = cfio->fault.ft_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); + 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); + return -ENOMEM; + } + + if (unlikely(cfio->fault.ft_flags & VM_FAULT_RETRY)) + return -EAGAIN; + + CERROR("unknow error in page fault!\n"); + return -EINVAL; +} + +#endif + static int vvp_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios) { @@ -560,11 +691,12 @@ static int vvp_io_fault_start(const struct lu_env *env, struct inode *inode = ccc_object_inode(obj); struct cl_fault_io *fio = &io->u.ci_fault; struct vvp_fault_io *cfio = &vio->u.fault; - cfs_page_t *vmpage; loff_t offset; + int kernel_result = 0; int result = 0; - - LASSERT(vio->cui_oneshot == 0); + 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) @@ -579,87 +711,64 @@ static int vvp_io_fault_start(const struct lu_env *env, if (result != 0) return result; - vmpage = filemap_nopage(cfio->ft_vma, cfio->ft_address, cfio->ft_type); - if (vmpage != NOPAGE_SIGBUS && vmpage != NOPAGE_OOM) - LL_CDEBUG_PAGE(D_PAGE, vmpage, - "got addr %lu type %lx\n", - cfio->ft_address, (long)cfio->ft_type); - else - CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n", - cfio->ft_address, (long)cfio->ft_type); - - if (vmpage == NOPAGE_SIGBUS) - result = -EFAULT; - else if (vmpage == NOPAGE_OOM) - result = -ENOMEM; - else { - struct cl_page *page; - loff_t size; - pgoff_t last; /* last page in a file data region */ - - /* Temporarily lock vmpage to keep cl_page_find() happy. */ - lock_page(vmpage); - page = cl_page_find(env, obj, fio->ft_index, vmpage, - CPT_CACHEABLE); - unlock_page(vmpage); - if (!IS_ERR(page)) { - size = i_size_read(inode); - last = cl_index(obj, size - 1); - if (fio->ft_index == last) - /* - * Last page is mapped partially. - */ - fio->ft_nob = size - cl_offset(obj, - fio->ft_index); - 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; - } else - result = PTR_ERR(page); - } - return result; -} + /* must return unlocked page */ + kernel_result = vvp_io_kernel_fault(cfio); + if (kernel_result != 0) + return kernel_result; -static void vvp_io_advance(const struct lu_env *env, - const struct cl_io_slice *ios, size_t nob) -{ - struct ccc_io *vio = cl2ccc_io(env, ios); - struct cl_io *io = ios->cis_io; - struct cl_object *obj = ios->cis_io->ci_obj; + if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE)) { + truncate_inode_pages_range(inode->i_mapping, + cl_offset(obj, fio->ft_index), offset); + } - CLOBINVRNT(env, obj, ccc_object_invariant(obj)); + /* Temporarily lock vmpage to keep cl_page_find() happy. */ + lock_page(cfio->ft_vmpage); - if (!cl_io_is_sendfile(io) && io->ci_continue) { - /* update the iov */ - LASSERT(vio->cui_tot_nrsegs >= vio->cui_nrsegs); - LASSERT(vio->cui_tot_count >= nob); + /* 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); - vio->cui_iov += vio->cui_nrsegs; - vio->cui_tot_nrsegs -= vio->cui_nrsegs; - vio->cui_tot_count -= nob; + CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n"); - if (vio->cui_iov_olen) { - struct iovec *iv; + /* return +1 to stop cl_io_loop() and ll_fault() will catch + * and retry. */ + return +1; + } - vio->cui_iov--; - vio->cui_tot_nrsegs++; - iv = &vio->cui_iov[0]; - iv->iov_base += iv->iov_len; - LASSERT(vio->cui_iov_olen > iv->iov_len); - iv->iov_len = vio->cui_iov_olen - iv->iov_len; - } + 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_index == last) + /* + * Last page is mapped partially. + */ + fio->ft_nob = size - cl_offset(obj, fio->ft_index); + 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; + + return result; } static int vvp_io_read_page(const struct lu_env *env, @@ -676,41 +785,39 @@ static int vvp_io_read_page(const struct lu_env *env, 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(cl2vvp_io(env, ios)->cui_oneshot == 0); LASSERT(slice->cpl_obj == obj); ENTRY; - if (sbi->ll_ra_info.ra_max_pages) + 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. */ - if (likely(!(fd->fd_flags & LL_FILE_IGNORE_LOCK))) { - int rc; - - rc = cl_page_is_under_lock(env, io, page); - if (rc != -EBUSY) { - CL_PAGE_HEADER(D_WARNING, env, page, "%s: %i\n", - rc == -ENODATA ? "without a lock" : - "match failed", rc); - if (rc != -ENODATA) - RETURN(rc); - } + 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); } if (cp->cpg_defer_uptodate) { cp->cpg_ra_used = 1; - cl_page_export(env, page); + 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) + 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); @@ -719,35 +826,18 @@ static int vvp_io_read_page(const struct lu_env *env, static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io, struct cl_page *page, struct ccc_page *cp, - int to, enum cl_req_type crt) + enum cl_req_type crt) { struct cl_2queue *queue; - struct ccc_object *cobo = cl2ccc(page->cp_obj); - struct cl_sync_io *anchor = &ccc_env_info(env)->cti_sync_io; - - int writing = io->ci_type == CIT_WRITE; int result; LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); queue = &io->ci_queue; - cl_2queue_init_page(queue, page); - if (writing) - /* Do not pass llap here as it is sync write. */ - vvp_write_pending(cobo, cp); - - cl_sync_io_init(anchor, 1); - cp->cpg_sync_io = anchor; - cl_page_clip(env, page, 0, to); - result = cl_io_submit_rw(env, io, crt, queue, CRP_NORMAL); - if (result == 0) - result = cl_sync_io_wait(env, io, &queue->c2_qout, anchor); - else - cp->cpg_sync_io = NULL; + result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0); LASSERT(cl_page_is_owned(page, io)); - cl_page_clip(env, page, 0, CFS_PAGE_SIZE); if (crt == CRT_READ) /* @@ -768,7 +858,7 @@ static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io, struct ccc_page *cp, unsigned from, unsigned to) { - struct cl_attr *attr = &ccc_env_info(env)->cti_attr; + struct cl_attr *attr = ccc_env_thread_attr(env); loff_t offset = cl_offset(obj, pg->cp_index); int result; @@ -789,8 +879,7 @@ static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io, } else if (cp->cpg_defer_uptodate) cp->cpg_ra_used = 1; else - result = vvp_page_sync_io(env, io, pg, cp, - CFS_PAGE_SIZE, CRT_READ); + 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 @@ -800,7 +889,7 @@ static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io, * details. -jay */ if (result == 0) - cl_page_export(env, pg); + cl_page_export(env, pg, 1); } return result; } @@ -890,8 +979,14 @@ static int vvp_io_commit_write(const struct lu_env *env, 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 == -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: @@ -906,8 +1001,21 @@ static int vvp_io_commit_write(const struct lu_env *env, * what the new code continues to do for the time * being. */ - result = vvp_page_sync_io(env, io, pg, cp, - to, CRT_WRITE); + 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; @@ -916,12 +1024,20 @@ static int vvp_io_commit_write(const struct lu_env *env, 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)) - i_size_write(inode, size); - cl_page_export(env, pg); - } else if (size > i_size_read(inode)) - cl_page_discard(env, io, pg); + cl_page_discard(env, io, pg); + } + ll_inode_size_unlock(inode, 0); RETURN(result); } @@ -931,20 +1047,20 @@ static const struct cl_io_operations vvp_io_ops = { .cio_fini = vvp_io_fini, .cio_lock = vvp_io_read_lock, .cio_start = vvp_io_read_start, - .cio_advance = vvp_io_advance + .cio_advance = ccc_io_advance }, [CIT_WRITE] = { .cio_fini = vvp_io_fini, .cio_lock = vvp_io_write_lock, .cio_start = vvp_io_write_start, - .cio_advance = vvp_io_advance + .cio_advance = ccc_io_advance }, - [CIT_TRUNC] = { - .cio_fini = vvp_io_trunc_fini, - .cio_iter_init = vvp_io_trunc_iter_init, - .cio_lock = vvp_io_trunc_lock, - .cio_start = vvp_io_trunc_start, - .cio_end = vvp_io_trunc_end + [CIT_SETATTR] = { + .cio_fini = vvp_io_setattr_fini, + .cio_iter_init = vvp_io_setattr_iter_init, + .cio_lock = vvp_io_setattr_lock, + .cio_start = vvp_io_setattr_start, + .cio_end = vvp_io_setattr_end }, [CIT_FAULT] = { .cio_fini = vvp_io_fault_fini, @@ -976,18 +1092,12 @@ 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_oneshot = 0; vio->cui_ra_window_set = 0; result = 0; if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) { - int op; size_t count; count = io->u.ci_rw.crw_count; - op = io->ci_type == CIT_READ ? - LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES; - if (io->ci_type == CIT_WRITE) - down(&ll_i2info(inode)->lli_write_sem); /* "If nbyte is 0, read() will return 0 and have no other * results." -- Single Unix Spec */ if (count == 0) @@ -995,11 +1105,13 @@ int vvp_io_init(const struct lu_env *env, struct cl_object *obj, else { cio->cui_tot_count = count; cio->cui_tot_nrsegs = 0; - ll_stats_ops_tally(sbi, op, count); } - } else if (io->ci_type == CIT_TRUNC) { - /* lockless truncate? */ - ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1); + } 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); }