-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
/*
* 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/
struct ccc_io *vio, struct cl_io *io)
{
struct ccc_thread_info *cti = ccc_env_info(env);
+ struct mm_struct *mm = current->mm;
struct vm_area_struct *vma;
struct cl_lock_descr *descr = &cti->cti_descr;
ldlm_policy_data_t policy;
if (!cl_is_normalio(env, io))
RETURN(0);
+ if (vio->cui_iov == NULL) /* nfs or loop back device write */
+ RETURN(0);
+
+ /* No MM (e.g. NFS)? No vmas too. */
+ if (mm == NULL)
+ RETURN(0);
+
for (seg = 0; seg < vio->cui_nrsegs; seg++) {
const struct iovec *iv = &vio->cui_iov[seg];
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.
*/
count -= vma->vm_end - addr;
addr = vma->vm_end;
}
+ up_read(&mm->mmap_sem);
}
RETURN(0);
}
if (new_size == 0)
enqflags = CEF_DISCARD_DATA;
} else {
- LASSERT((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));
+ 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;
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;
}
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)
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:
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;
}
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 */
/* 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;
}
+static void vvp_io_fsync_end(const struct lu_env *env,
+ const struct cl_io_slice *ios)
+{
+ /* never try to verify there is no dirty pages in sync range
+ * because page_mkwrite() can generate new dirty pages any time. */
+}
+
static int vvp_io_read_page(const struct lu_env *env,
const struct cl_io_slice *ios,
const struct cl_page_slice *slice)
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);
* 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);
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;
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)
/*
} 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
* 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:
* 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;
cl_page_discard(env, io, pg);
}
ll_inode_size_unlock(inode, 0);
-
RETURN(result);
}
.cio_start = vvp_io_fault_start,
.cio_end = ccc_io_end
},
+ [CIT_FSYNC] = {
+ .cio_end = vvp_io_fsync_end,
+ .cio_fini = vvp_io_fini
+ },
[CIT_MISC] = {
.cio_fini = vvp_io_fini
}
{
struct vvp_io *vio = vvp_env_io(env);
struct ccc_io *cio = ccc_env_io(env);
- struct inode *inode = ccc_object_inode(obj);
- struct ll_sb_info *sbi = ll_i2sbi(inode);
int result;
CLOBINVRNT(env, obj, ccc_object_invariant(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;
+ struct ll_inode_info *lli = ll_i2info(ccc_object_inode(obj));
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)
else {
cio->cui_tot_count = count;
cio->cui_tot_nrsegs = 0;
- ll_stats_ops_tally(sbi, op, count);
}
+ /* 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))
- /* lockless truncate? */
- ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
- else
+ if (!cl_io_is_trunc(io))
io->ci_lockreq = CILR_MANDATORY;
}
RETURN(result);