X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fvvp_io.c;h=253d13a04fd64c1fd498a847da366ae26db4f582;hp=89bb0486fed4d678eb7c8e06533d08d80794737b;hb=56e56ed49fc5392048872e4b9502a4ea66be2d90;hpb=0b72da3f303132ec09d80c65156c7a20fa03c721 diff --git a/lustre/llite/vvp_io.c b/lustre/llite/vvp_io.c index 89bb048..253d13a 100644 --- a/lustre/llite/vvp_io.c +++ b/lustre/llite/vvp_io.c @@ -26,8 +26,10 @@ * 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, 2012, Whamcloud, Inc. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -144,6 +146,9 @@ 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); + for (seg = 0; seg < vio->cui_nrsegs; seg++) { const struct iovec *iv = &vio->cui_iov[seg]; @@ -180,6 +185,11 @@ static int vvp_mmap_locks(const struct lu_env *env, policy.l_extent.end); 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); @@ -262,10 +272,10 @@ static int vvp_io_write_lock(const struct lu_env *env, 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); /* @@ -276,31 +286,41 @@ 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::cio_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; - loff_t new_size = io->u.ci_truncate.tr_size; - __u32 enqflags = new_size == 0 ? CEF_DISCARD_DATA : 0; - int result; - - 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; + 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) @@ -308,34 +328,26 @@ 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() + * is grabbed by ll_truncate() only over call to obd_adjust_kms(). */ ll_inode_size_lock(inode, 0); result = vmtruncate(inode, size); - if (result != 0) - ll_inode_size_unlock(inode, 0); + ll_inode_size_unlock(inode, 0); 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; - loff_t size = io->u.ci_truncate.tr_size; pgoff_t start = cl_index(obj, size); int result; - LASSERT(cio->u.trunc.cui_locks_released); - - LOCK_INODE_MUTEX(inode); DOWN_WRITE_I_ALLOC_SEM(inode); - cio->u.trunc.cui_locks_released = 0; result = vvp_do_vmtruncate(inode, size); @@ -369,14 +381,59 @@ 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 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 vvp_io *vio = cl2vvp_io(env, 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); - loff_t size = io->u.ci_truncate.tr_size; + 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); @@ -387,19 +444,21 @@ static void vvp_io_trunc_end(const struct lu_env *env, * Do vmtruncate again, to remove possible stale pages populated by * competing read threads. bz20645. */ - vvp_do_vmtruncate(inode, size); + 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); } @@ -453,7 +512,7 @@ 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); + result = ccc_prep_size(env, obj, io, pos, tot, &exceed); if (result != 0) return result; else if (exceed != 0) @@ -495,6 +554,10 @@ 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: @@ -579,42 +642,43 @@ 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; } - CERROR("unknow error in page fault!\n"); + if (cfio->fault.ft_flags & VM_FAULT_RETRY) + return -EAGAIN; + + CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags); return -EINVAL; } @@ -630,8 +694,8 @@ static int vvp_io_fault_start(const struct lu_env *env, 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; + cfs_page_t *vmpage = NULL; struct cl_page *page; loff_t size; pgoff_t last; /* last page in a file data region */ @@ -645,48 +709,90 @@ 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 locked page */ - kernel_result = vvp_io_kernel_fault(cfio); - if (kernel_result != 0) - return kernel_result; - /* Temporarily lock vmpage to keep cl_page_find() happy. */ - lock_page(cfio->ft_vmpage); - 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); + 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); + + /* Though we have already held a cl_lock upon this page, but + * it still can be truncated locally. */ + if (unlikely(vmpage->mapping == NULL)) { + CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n"); + + /* return +1 to stop cl_io_loop() and ll_fault() will catch + * and retry. */ + 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 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; +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; } @@ -711,14 +817,15 @@ 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); /* 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: %i\n", + CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n", rc == -ENODATA ? "without a lock" : "match failed", rc); if (rc != -ENODATA) @@ -734,7 +841,8 @@ 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); @@ -743,7 +851,7 @@ 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; int result; @@ -751,13 +859,10 @@ static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io, LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE); queue = &io->ci_queue; - cl_2queue_init_page(queue, page); - cl_page_clip(env, page, 0, to); - + 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) /* @@ -799,8 +904,7 @@ 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 @@ -904,7 +1008,10 @@ static int vvp_io_commit_write(const struct lu_env *env, * 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: @@ -919,11 +1026,21 @@ 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; @@ -934,15 +1051,18 @@ static int vvp_io_commit_write(const struct lu_env *env, ll_inode_size_lock(inode, 0); if (result == 0) { - if (size > i_size_read(inode)) + 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); } @@ -960,12 +1080,12 @@ static const struct cl_io_operations vvp_io_ops = { .cio_start = vvp_io_write_start, .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, @@ -1000,12 +1120,9 @@ int vvp_io_init(const struct lu_env *env, struct cl_object *obj, 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 nbyte is 0, read() will return 0 and have no other * results." -- Single Unix Spec */ if (count == 0) @@ -1013,11 +1130,13 @@ 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); }