X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fvvp_io.c;h=4b031ac8fc17468af3239d3cf5d7446d32c1a7be;hp=7ab77609c5d8217648042cd8308f437d8eeb8d3a;hb=7af2428511c286788d9c73c436cb43d6f791821f;hpb=b4e0446ad189788529e4d69d4dfa9c822f2e3bae diff --git a/lustre/llite/vvp_io.c b/lustre/llite/vvp_io.c index 7ab7760..4b031ac 100644 --- a/lustre/llite/vvp_io.c +++ b/lustre/llite/vvp_io.c @@ -28,6 +28,9 @@ /* * 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/ @@ -323,13 +326,11 @@ 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; } @@ -509,7 +510,7 @@ 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) @@ -635,42 +636,43 @@ 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; } @@ -686,8 +688,8 @@ 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 */ @@ -701,48 +703,90 @@ 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; } @@ -767,7 +811,8 @@ static int vvp_io_read_page(const struct lu_env *env, ENTRY; - if (sbi->ll_ra_info.ra_max_pages_per_file) + if (sbi->ll_ra_info.ra_max_pages_per_file && + sbi->ll_ra_info.ra_max_pages) ras_update(sbi, inode, ras, page->cp_index, cp->cpg_defer_uptodate); @@ -790,7 +835,8 @@ 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); @@ -799,7 +845,7 @@ 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; @@ -807,13 +853,10 @@ 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) /* @@ -855,8 +898,7 @@ 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 @@ -960,7 +1002,10 @@ 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: @@ -975,11 +1020,21 @@ 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; @@ -1002,7 +1057,6 @@ static int vvp_io_commit_write(const struct lu_env *env, cl_page_discard(env, io, pg); } ll_inode_size_unlock(inode, 0); - RETURN(result); } @@ -1060,12 +1114,9 @@ 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) @@ -1073,7 +1124,6 @@ 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_SETATTR) { if (cl_io_is_trunc(io))