-/* -*- 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.
* 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, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* Implementation of cl_io for VVP layer.
*
* Author: Nikita Danilov <nikita.danilov@sun.com>
+ * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
*/
#define DEBUG_SUBSYSTEM S_LLITE
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;
+}
+
+/**
+ * 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 ccc_io *cio = ccc_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 (lli->lli_layout_gen != cio->cui_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;
+}
+
/*****************************************************************************
*
* io operations.
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 cl_io *io = ios->cis_io;
+ struct cl_object *obj = io->ci_obj;
+ struct ccc_io *cio = cl2ccc_io(env, ios);
CLOBINVRNT(env, obj, ccc_object_invariant(obj));
- if (io->ci_type == CIT_WRITE)
- up(&ll_i2info(ccc_object_inode(obj))->lli_write_sem);
- else {
- struct vvp_io *vio = cl2vvp_io(env, ios);
- struct ccc_io *cio = cl2ccc_io(env, ios);
-
- if (vio->cui_ra_window_set)
- ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
- }
+ CDEBUG(D_VFSTRACE, "ignore/verify layout %d/%d, layout version %d.\n",
+ io->ci_ignore_layout, io->ci_verify_layout, cio->cui_layout_gen);
+
+ if (!io->ci_ignore_layout && io->ci_verify_layout) {
+ __u32 gen = 0;
+
+ /* check layout version */
+ ll_layout_refresh(ccc_object_inode(obj), &gen);
+ io->ci_need_restart = cio->cui_layout_gen != gen;
+ if (io->ci_need_restart)
+ CDEBUG(D_VFSTRACE, "layout changed from %d to %d.\n",
+ cio->cui_layout_gen, gen);
+ }
}
static void vvp_io_fault_fini(const struct lu_env *env,
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;
- struct inode *inode;
unsigned long addr;
unsigned long seg;
ssize_t count;
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);
+
+ 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++) {
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);
+ 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;
- 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
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;
}
+ up_read(&mm->mmap_sem);
}
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)
{
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;
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 */
+ if (lli->lli_has_smd) /* lsm-less file doesn'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 +
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 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);
- UP_WRITE_I_ALLOC_SEM(inode);
- vio->u.trunc.cui_locks_released = 1;
- return 0;
+ 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;
-
- 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;
+ __u64 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_io_trunc_start(const struct lu_env *env,
- const struct cl_io_slice *ios)
+static int vvp_do_vmtruncate(struct inode *inode, size_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);
+ int result;
+ /*
+ * Only ll_inode_size_lock is taken at this level.
+ */
+ ll_inode_size_lock(inode);
+ result = vmtruncate(inode, size);
+ ll_inode_size_unlock(inode);
+
+ return result;
+}
- LOCK_INODE_MUTEX(inode);
- DOWN_WRITE_I_ALLOC_SEM(inode);
- cio->u.trunc.cui_locks_released = 0;
+static int vvp_io_setattr_trunc(const struct lu_env *env,
+ const struct cl_io_slice *ios,
+ struct inode *inode, loff_t size)
+{
+ inode_dio_wait(inode);
+ return 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);
- /*
- * 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;
+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;
- hdr = cl_object_header(obj);
- spin_lock(&hdr->coh_page_guard);
- vio->cui_partpage = cl_page_lookup(hdr, start);
- spin_unlock(&hdr->coh_page_guard);
+ 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);
- 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;
}
-static void vvp_io_trunc_end(const struct lu_env *env,
- const struct cl_io_slice *ios)
+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);
-
- 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;
- }
+ struct cl_io *io = ios->cis_io;
+ struct inode *inode = ccc_object_inode(io->ci_obj);
+
+ mutex_lock(&inode->i_mutex);
+ 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_trunc_fini(const struct lu_env *env,
- const struct cl_io_slice *ios)
+static void vvp_io_setattr_end(const struct lu_env *env,
+ const struct cl_io_slice *ios)
{
- struct ccc_io *cio = ccc_env_io(env);
- struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
+ struct cl_io *io = ios->cis_io;
+ struct inode *inode = ccc_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);
+}
- if (cio->u.trunc.cui_locks_released) {
- LOCK_INODE_MUTEX(inode);
- DOWN_WRITE_I_ALLOC_SEM(inode);
- cio->u.trunc.cui_locks_released = 0;
- }
- vvp_io_fini(env, ios);
+static void vvp_io_setattr_fini(const struct lu_env *env,
+ const struct cl_io_slice *ios)
+{
+ vvp_io_fini(env, ios);
}
#ifdef HAVE_FILE_READV
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);
+
+ if (!can_populate_pages(env, io, inode))
+ return 0;
- 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)
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 */
/* 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);
+ /* 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();
}
out:
return result;
}
+static void vvp_io_read_fini(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);
+
+ if (vio->cui_ra_window_set)
+ ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
+
+ vvp_io_fini(env, ios);
+}
+
static int vvp_io_write_start(const struct lu_env *env,
const struct cl_io_slice *ios)
{
ENTRY;
- if (cl_io_is_append(io))
+ 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.
*/
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 (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);
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;
+ lock_page(vmpage);
+
+ return 0;
+}
+#else
+static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
+{
+ 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 = vmf->page;
+ return 0;
+ }
+
+ if (cfio->fault.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_OOM) {
+ CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
+ return -ENOMEM;
+ }
+
+ if (cfio->fault.ft_flags & VM_FAULT_RETRY)
+ return -EAGAIN;
+
+ CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
+ return -EINVAL;
+}
+
+#endif
+
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;
- cfs_page_t *vmpage;
- loff_t offset;
- int result = 0;
-
- LASSERT(vio->cui_oneshot == 0);
+ 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 */
if (fio->ft_executable &&
LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
/* 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;
- 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);
+ /* 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 != 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
+ * and retry. */
+ GOTO(out, result = +1);
}
- return 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;
- CLOBINVRNT(env, obj, ccc_object_invariant(obj));
- 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);
+ if (fio->ft_mkwrite ) {
+ pgoff_t last_index;
+ /*
+ * 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.
+ */
+ last_index = cl_index(obj, size - 1);
+ 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 (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);
+ LASSERT(cl_page_is_owned(page, io));
+
+ vmpage = NULL;
+ if (result < 0) {
+ cl_page_unmap(env, io, page);
+ 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);
+ }
+ }
+
+ last = cl_index(obj, size - 1);
+ /*
+ * 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);
+ 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);
- vio->cui_iov += vio->cui_nrsegs;
- vio->cui_tot_nrsegs -= vio->cui_nrsegs;
- vio->cui_tot_count -= nob;
+ lu_ref_add(&page->cp_reference, "fault", io);
+ fio->ft_page = page;
+ EXIT;
- if (vio->cui_iov_olen) {
- struct iovec *iv;
+out:
+ /* return unlocked vmpage to avoid deadlocking */
+ if (vmpage != NULL)
+ unlock_page(vmpage);
+#ifdef HAVE_VM_OP_FAULT
+ cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
+#endif
+ return result;
+}
- 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;
- }
- }
+static int vvp_io_fsync_start(const struct lu_env *env,
+ const struct cl_io_slice *ios)
+{
+ /* 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;
}
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);
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, 0);
LASSERT(cl_page_is_owned(page, io));
- cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
if (crt == CRT_READ)
/*
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;
* purposes here we can treat it like i_size.
*/
if (attr->cat_kms <= offset) {
- char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
+ char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
memset(kaddr, 0, cl_page_size(obj));
- kunmap_atomic(kaddr, KM_USER0);
+ ll_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,
- 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
* details. -jay
*/
if (result == 0)
- cl_page_export(env, pg);
+ cl_page_export(env, pg, 1);
}
return result;
}
struct cl_page *pg = slice->cpl_page;
struct inode *inode = ccc_object_inode(obj);
struct ll_sb_info *sbi = ll_i2sbi(inode);
+ struct ll_inode_info *lli = ll_i2info(inode);
cfs_page_t *vmpage = cp->cpg_page;
int result;
*/
if (!PageDirty(vmpage)) {
tallyop = LPROC_LL_DIRTY_MISSES;
- vvp_write_pending(cl2ccc(obj), cp);
- set_page_dirty(vmpage);
result = cl_page_cache_add(env, io, pg, CRT_WRITE);
- if (result == -EDQUOT)
+ if (result == 0) {
+ /* page was added into cache successfully. */
+ set_page_dirty(vmpage);
+ vvp_write_pending(cl2ccc(obj), cp);
+ } else 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;
}
ll_stats_ops_tally(sbi, tallyop, 1);
+ /* Inode should be marked DIRTY even if no new page was marked DIRTY
+ * because page could have been not flushed between 2 modifications.
+ * It is important the file is marked DIRTY as soon as the I/O is done
+ * Indeed, when cache is flushed, file could be already closed and it
+ * is too late to warn the MDT.
+ * It is acceptable that file is marked DIRTY even if I/O is dropped
+ * for some reasons before being flushed to OST.
+ */
+ if (result == 0) {
+ spin_lock(&lli->lli_lock);
+ lli->lli_flags |= LLIF_DATA_MODIFIED;
+ spin_unlock(&lli->lli_lock);
+ }
+
size = cl_offset(obj, pg->cp_index) + to;
+ ll_inode_size_lock(inode);
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);
- RETURN(result);
+ cl_page_discard(env, io, pg);
+ }
+ ll_inode_size_unlock(inode);
+ RETURN(result);
}
static const struct cl_io_operations vvp_io_ops = {
.op = {
[CIT_READ] = {
- .cio_fini = vvp_io_fini,
+ .cio_fini = vvp_io_read_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,
.cio_start = vvp_io_fault_start,
.cio_end = ccc_io_end
},
+ [CIT_FSYNC] = {
+ .cio_start = vvp_io_fsync_start,
+ .cio_fini = vvp_io_fini
+ },
[CIT_MISC] = {
.cio_fini = vvp_io_fini
}
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);
- struct inode *inode = ccc_object_inode(obj);
- struct ll_sb_info *sbi = ll_i2sbi(inode);
+ struct vvp_io *vio = vvp_env_io(env);
+ struct ccc_io *cio = ccc_env_io(env);
+ struct inode *inode = ccc_object_inode(obj);
int result;
CLOBINVRNT(env, obj, ccc_object_invariant(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;
+ 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;
- 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)
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);
- }
- RETURN(result);
+ /* 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, &cio->cui_layout_gen);
+
+ RETURN(result);
}
static struct vvp_io *cl2vvp_io(const struct lu_env *env,