-/* -*- 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, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* Implementation of cl_io for OSC layer.
*
* Author: Nikita Danilov <nikita.danilov@sun.com>
+ * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
*/
#define DEBUG_SUBSYSTEM S_OSC
return container_of(oap, struct osc_page, ops_oap)->ops_cl.cpl_page;
}
-static void osc_io_unplug(const struct lu_env *env, struct osc_object *osc,
- struct client_obd *cli)
-{
- loi_list_maint(cli, osc->oo_oinfo);
- osc_check_rpcs(env, cli);
- client_obd_list_unlock(&cli->cl_loi_list_lock);
-}
-
-/**
- * How many pages osc_io_submit() queues before checking whether an RPC is
- * ready.
- */
-#define OSC_QUEUE_GRAIN (32)
-
/**
* An implementation of cl_io_operations::cio_io_submit() method for osc
* layer. Iterates over pages in the in-queue, prepares each for io by calling
* cl_page_prep() and then either submits them through osc_io_submit_page()
* or, if page is already submitted, changes osc flags through
- * osc_set_async_flags_base().
+ * osc_set_async_flags().
*/
static int osc_io_submit(const struct lu_env *env,
const struct cl_io_slice *ios,
LASSERT(qin->pl_nr > 0);
- CDEBUG(D_INFO, "%i %i\n", qin->pl_nr, crt);
+ CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, crt);
/*
* NOTE: here @page is a top-level page. This is done to avoid
* creation of sub-page-list.
exp = osc_export(osc);
if (priority > CRP_NORMAL) {
- spin_lock(&oap->oap_lock);
+ cfs_spin_lock(&oap->oap_lock);
oap->oap_async_flags |= ASYNC_HP;
- spin_unlock(&oap->oap_lock);
- }
- /*
- * This can be checked without cli->cl_loi_list_lock, because
- * ->oap_*_item are always manipulated when the page is owned.
- */
- if (!list_empty(&oap->oap_urgent_item) ||
- !list_empty(&oap->oap_rpc_item)) {
- result = -EBUSY;
- break;
+ cfs_spin_unlock(&oap->oap_lock);
}
if (osc0 == NULL) { /* first iteration */
cli = &exp->exp_obd->u.cli;
osc0 = osc;
+ client_obd_list_lock(&cli->cl_loi_list_lock);
} else /* check that all pages are against the same object
* (for now) */
LASSERT(osc == osc0);
- if (queued++ == 0)
- client_obd_list_lock(&cli->cl_loi_list_lock);
+
+ if (!cfs_list_empty(&oap->oap_urgent_item) ||
+ !cfs_list_empty(&oap->oap_rpc_item)) {
+ result = -EBUSY;
+ break;
+ }
+
result = cl_page_prep(env, io, page, crt);
if (result == 0) {
+ ++queued;
cl_page_list_move(qout, qin, page);
- if (list_empty(&oap->oap_pending_item)) {
+ if (cfs_list_empty(&oap->oap_pending_item)) {
osc_io_submit_page(env, cl2osc_io(env, ios),
opg, crt);
} else {
- result = osc_set_async_flags_base(cli,
- osc->oo_oinfo,
- oap,
- OSC_FLAGS);
- /*
- * bug 18881: we can't just break out here when
- * error occurrs after cl_page_prep has been
- * called against the page. The correct
- * way is to call page's completion routine,
- * as in osc_oap_interrupted. For simplicity,
- * we just force osc_set_async_flags_base() to
- * not return error.
- */
+ result = osc_set_async_flags(osc, opg,
+ OSC_FLAGS);
+ /*
+ * bug 18881: we can't just break out here when
+ * error occurs after cl_page_prep has been
+ * called against the page. The correct
+ * way is to call page's completion routine,
+ * as in osc_oap_interrupted. For simplicity,
+ * we just force osc_set_async_flags() to
+ * not return error.
+ */
LASSERT(result == 0);
}
opg->ops_submit_time = cfs_time_current();
*/
result = 0;
}
+
/*
- * Don't keep client_obd_list_lock() for too long.
- *
- * XXX client_obd_list lock has to be unlocked periodically to
- * avoid soft-lockups that tend to happen otherwise (see bug
- * 16651). On the other hand, osc_io_submit_page() queues a
- * page with ASYNC_URGENT flag and so all pages queued up
- * until this point are sent out immediately by
- * osc_io_unplug() resulting in sub-optimal RPCs (sub-optimal
- * RPCs only happen during `warm up' phase when less than
- * cl_max_rpcs_in_flight RPCs are in flight). To balance these
- * conflicting requirements, one might unplug once enough
- * pages to form a large RPC were queued (i.e., use
- * cli->cl_max_pages_per_rpc as OSC_QUEUE_GRAIN, see
- * lop_makes_rpc()), or ignore soft-lockup issue altogether.
+ * We might hold client_obd_list_lock() for too long and cause
+ * soft-lockups (see bug 16651). But on the other hand, pages
+ * are queued here with ASYNC_URGENT flag, thus will be sent
+ * out immediately once osc_io_unplug() be called, possibly
+ * resulting sub-optimal RPCs.
*
- * XXX lock_need_resched() should be used here, but it is not
- * available in the older of supported kernels.
+ * We think creating optimal-sized RPCs is more important than
+ * avoiding the transient soft-lockups, plus I believe the
+ * soft-locks only happen in full debug testing.
*/
- if (queued > OSC_QUEUE_GRAIN || cfs_need_resched()) {
- queued = 0;
- osc_io_unplug(env, osc, cli);
- cfs_cond_resched();
- }
}
LASSERT(ergo(result == 0, cli != NULL));
LASSERT(ergo(result == 0, osc == osc0));
if (queued > 0)
- osc_io_unplug(env, osc, cli);
- CDEBUG(D_INFO, "%i/%i %i\n", qin->pl_nr, qout->pl_nr, result);
+ osc_io_unplug(env, cli, osc, PDL_POLICY_ROUND);
+ if (osc0)
+ client_obd_list_unlock(&cli->cl_loi_list_lock);
+ CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
return qout->pl_nr > 0 ? 0 : result;
}
{
struct osc_device *dev = lu2osc_dev(slice->cpl_obj->co_lu.lo_dev);
struct obd_import *imp = class_exp2cliimp(dev->od_exp);
-
+ struct osc_io *oio = cl2osc_io(env, ios);
+ int result = 0;
ENTRY;
/*
* This implements OBD_BRW_CHECK logic from old client.
*/
- RETURN(imp == NULL || imp->imp_invalid ? -EIO : 0);
+ if (imp == NULL || imp->imp_invalid)
+ result = -EIO;
+ if (result == 0 && oio->oi_lockless)
+ /* this page contains `invalid' data, but who cares?
+ * nobody can access the invalid data.
+ * in osc_io_commit_write(), we're going to write exact
+ * [from, to) bytes of this page to OST. -jay */
+ cl_page_export(env, slice->cpl_page, 1);
+
+ RETURN(result);
}
static int osc_io_commit_write(const struct lu_env *env,
const struct cl_page_slice *slice,
unsigned from, unsigned to)
{
+ struct osc_io *oio = cl2osc_io(env, ios);
struct osc_page *opg = cl2osc_page(slice);
struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
struct osc_async_page *oap = &opg->ops_oap;
cfs_capable(CFS_CAP_SYS_RESOURCE))
oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
+ if (oio->oi_lockless)
+ /* see osc_io_prepare_write() for lockless io handling. */
+ cl_page_clip(env, slice->cpl_page, from, to);
+
RETURN(0);
}
io = ios->cis_io;
fio = &io->u.ci_fault;
- CDEBUG(D_INFO, "%lu %i %i\n",
+ CDEBUG(D_INFO, "%lu %d %d\n",
fio->ft_index, fio->ft_writable, fio->ft_nob);
/*
* If mapping is writeable, adjust kms to cover this page,
RETURN(0);
}
-static int osc_punch_upcall(void *a, int rc)
+static int osc_async_upcall(void *a, int rc)
{
- struct osc_punch_cbargs *args = a;
+ struct osc_async_cbargs *args = a;
args->opc_rc = rc;
- complete(&args->opc_sync);
+ cfs_complete(&args->opc_sync);
return 0;
}
-#ifdef __KERNEL__
+/* Disable osc_trunc_check() because it is naturally race between read and
+ * truncate. See bug 20645 for details.
+ */
+#if 0 && defined(__KERNEL__)
/**
* Checks that there are no pages being written in the extent being truncated.
*/
* XXX this is quite expensive check.
*/
cl_page_list_init(list);
- cl_page_gang_lookup(env, clob, io, start + partial, CL_PAGE_EOF, list, 0);
+ cl_page_gang_lookup(env, clob, io, start + partial, CL_PAGE_EOF, list);
cl_page_list_for_each(page, list)
CL_PAGE_DEBUG(D_ERROR, env, page, "exists %lu\n", start);
cl_page_list_disown(env, io, list);
cl_page_list_fini(env, list);
- spin_lock(&obj->oo_seatbelt);
- list_for_each_entry(cp, &obj->oo_inflight[CRT_WRITE], ops_inflight) {
+ cfs_spin_lock(&obj->oo_seatbelt);
+ cfs_list_for_each_entry(cp, &obj->oo_inflight[CRT_WRITE],
+ ops_inflight) {
page = cp->ops_cl.cpl_page;
if (page->cp_index >= start + partial) {
cfs_task_t *submitter;
/*
* XXX Linux specific debugging stuff.
*/
- CL_PAGE_DEBUG(D_ERROR, env, page, "%s/%i %lu\n",
+ CL_PAGE_DEBUG(D_ERROR, env, page, "%s/%d %lu\n",
submitter->comm, submitter->pid, start);
libcfs_debug_dumpstack(submitter);
}
}
- spin_unlock(&obj->oo_seatbelt);
+ cfs_spin_unlock(&obj->oo_seatbelt);
}
#else /* __KERNEL__ */
# define osc_trunc_check(env, io, oio, size) do {;} while (0)
#endif
-static int osc_io_trunc_start(const struct lu_env *env,
- const struct cl_io_slice *slice)
+static int osc_io_setattr_start(const struct lu_env *env,
+ const struct cl_io_slice *slice)
{
struct cl_io *io = slice->cis_io;
struct osc_io *oio = cl2osc_io(env, slice);
struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
struct cl_attr *attr = &osc_env_info(env)->oti_attr;
struct obdo *oa = &oio->oi_oa;
- struct osc_punch_cbargs *cbargs = &oio->oi_punch_cbarg;
- struct obd_capa *capa;
- loff_t size = io->u.ci_truncate.tr_size;
+ struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
+ loff_t size = io->u.ci_setattr.sa_attr.lvb_size;
+ unsigned int ia_valid = io->u.ci_setattr.sa_valid;
int result = 0;
+ struct obd_info oinfo = { { { 0 } } };
-
- memset(oa, 0, sizeof(*oa));
-
- osc_trunc_check(env, io, oio, size);
+ if (ia_valid & ATTR_SIZE)
+ osc_trunc_check(env, io, oio, size);
if (oio->oi_lockless == 0) {
cl_object_attr_lock(obj);
result = cl_object_attr_get(env, obj, attr);
if (result == 0) {
- attr->cat_size = attr->cat_kms = size;
- result = cl_object_attr_set(env, obj, attr,
- CAT_SIZE|CAT_KMS);
+ unsigned int cl_valid = 0;
+
+ if (ia_valid & ATTR_SIZE) {
+ attr->cat_size = attr->cat_kms = size;
+ cl_valid = (CAT_SIZE | CAT_KMS);
+ }
+ if (ia_valid & ATTR_MTIME_SET) {
+ attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
+ cl_valid |= CAT_MTIME;
+ }
+ if (ia_valid & ATTR_ATIME_SET) {
+ attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
+ cl_valid |= CAT_ATIME;
+ }
+ if (ia_valid & ATTR_CTIME_SET) {
+ attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
+ cl_valid |= CAT_CTIME;
+ }
+ result = cl_object_attr_set(env, obj, attr, cl_valid);
}
cl_object_attr_unlock(obj);
}
-
+ memset(oa, 0, sizeof(*oa));
if (result == 0) {
oa->o_id = loi->loi_id;
- oa->o_gr = loi->loi_gr;
+ oa->o_seq = loi->loi_seq;
oa->o_mtime = attr->cat_mtime;
oa->o_atime = attr->cat_atime;
oa->o_ctime = attr->cat_ctime;
oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLATIME |
OBD_MD_FLCTIME | OBD_MD_FLMTIME;
- if (oio->oi_lockless) {
- oa->o_flags = OBD_FL_TRUNCLOCK;
- oa->o_valid |= OBD_MD_FLFLAGS;
+ if (ia_valid & ATTR_SIZE) {
+ oa->o_size = size;
+ oa->o_blocks = OBD_OBJECT_EOF;
+ oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
+
+ if (oio->oi_lockless) {
+ oa->o_flags = OBD_FL_SRVLOCK;
+ oa->o_valid |= OBD_MD_FLFLAGS;
+ }
+ } else {
+ LASSERT(oio->oi_lockless == 0);
}
- oa->o_size = size;
- oa->o_blocks = OBD_OBJECT_EOF;
- oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
-
- capa = io->u.ci_truncate.tr_capa;
- init_completion(&cbargs->opc_sync);
- result = osc_punch_base(osc_export(cl2osc(obj)), oa, capa,
- osc_punch_upcall, cbargs, PTLRPCD_SET);
+
+ oinfo.oi_oa = oa;
+ oinfo.oi_capa = io->u.ci_setattr.sa_capa;
+ cfs_init_completion(&cbargs->opc_sync);
+
+ if (ia_valid & ATTR_SIZE)
+ result = osc_punch_base(osc_export(cl2osc(obj)),
+ &oinfo, osc_async_upcall,
+ cbargs, PTLRPCD_SET);
+ else
+ result = osc_setattr_async_base(osc_export(cl2osc(obj)),
+ &oinfo, NULL,
+ osc_async_upcall,
+ cbargs, PTLRPCD_SET);
}
return result;
}
-static void osc_io_trunc_end(const struct lu_env *env,
- const struct cl_io_slice *slice)
+static void osc_io_setattr_end(const struct lu_env *env,
+ const struct cl_io_slice *slice)
{
struct cl_io *io = slice->cis_io;
struct osc_io *oio = cl2osc_io(env, slice);
- struct osc_punch_cbargs *cbargs = &oio->oi_punch_cbarg;
- struct obdo *oa = &oio->oi_oa;
+ struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
int result;
- wait_for_completion(&cbargs->opc_sync);
+ cfs_wait_for_completion(&cbargs->opc_sync);
result = io->ci_result = cbargs->opc_rc;
if (result == 0) {
struct cl_object *obj = slice->cis_obj;
- if (oio->oi_lockless == 0) {
- struct cl_attr *attr = &osc_env_info(env)->oti_attr;
- int valid = 0;
-
- /* Update kms & size */
- if (oa->o_valid & OBD_MD_FLSIZE) {
- attr->cat_size = oa->o_size;
- attr->cat_kms = oa->o_size;
- valid |= CAT_KMS|CAT_SIZE;
- }
- if (oa->o_valid & OBD_MD_FLBLOCKS) {
- attr->cat_blocks = oa->o_blocks;
- valid |= CAT_BLOCKS;
- }
- if (oa->o_valid & OBD_MD_FLMTIME) {
- attr->cat_mtime = oa->o_mtime;
- valid |= CAT_MTIME;
- }
- if (oa->o_valid & OBD_MD_FLCTIME) {
- attr->cat_ctime = oa->o_ctime;
- valid |= CAT_CTIME;
- }
- if (oa->o_valid & OBD_MD_FLATIME) {
- attr->cat_atime = oa->o_atime;
- valid |= CAT_ATIME;
- }
- cl_object_attr_lock(obj);
- result = cl_object_attr_set(env, obj, attr, valid);
- cl_object_attr_unlock(obj);
- } else { /* lockless truncate */
+ if (oio->oi_lockless) {
+ /* lockless truncate */
struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
+
+ LASSERT(cl_io_is_trunc(io));
/* XXX: Need a lock. */
osd->od_stats.os_lockless_truncates++;
}
}
+}
- /* return result; */
+static int osc_io_read_start(const struct lu_env *env,
+ const struct cl_io_slice *slice)
+{
+ struct osc_io *oio = cl2osc_io(env, slice);
+ struct cl_object *obj = slice->cis_obj;
+ struct cl_attr *attr = &osc_env_info(env)->oti_attr;
+ int result = 0;
+ ENTRY;
+
+ if (oio->oi_lockless == 0) {
+ cl_object_attr_lock(obj);
+ result = cl_object_attr_get(env, obj, attr);
+ if (result == 0) {
+ attr->cat_atime = LTIME_S(CFS_CURRENT_TIME);
+ result = cl_object_attr_set(env, obj, attr,
+ CAT_ATIME);
+ }
+ cl_object_attr_unlock(obj);
+ }
+ RETURN(result);
+}
+
+static int osc_io_write_start(const struct lu_env *env,
+ const struct cl_io_slice *slice)
+{
+ struct osc_io *oio = cl2osc_io(env, slice);
+ struct cl_object *obj = slice->cis_obj;
+ struct cl_attr *attr = &osc_env_info(env)->oti_attr;
+ int result = 0;
+ ENTRY;
+
+ if (oio->oi_lockless == 0) {
+ cl_object_attr_lock(obj);
+ result = cl_object_attr_get(env, obj, attr);
+ if (result == 0) {
+ attr->cat_mtime = attr->cat_ctime =
+ LTIME_S(CFS_CURRENT_TIME);
+ result = cl_object_attr_set(env, obj, attr,
+ CAT_MTIME | CAT_CTIME);
+ }
+ cl_object_attr_unlock(obj);
+ }
+ RETURN(result);
+}
+
+static int osc_io_fsync_start(const struct lu_env *env,
+ const struct cl_io_slice *slice)
+{
+ struct cl_io *io = slice->cis_io;
+ struct osc_io *oio = cl2osc_io(env, slice);
+ struct obdo *oa = &oio->oi_oa;
+ struct obd_info *oinfo = &oio->oi_info;
+ struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
+ struct cl_object *obj = slice->cis_obj;
+ struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
+ int result = 0;
+ ENTRY;
+
+ memset(oa, 0, sizeof(*oa));
+ oa->o_id = loi->loi_id;
+ oa->o_seq = loi->loi_seq;
+ oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
+
+ /* reload size and blocks for start and end of sync range */
+ oa->o_size = io->u.ci_fsync.fi_start;
+ oa->o_blocks = io->u.ci_fsync.fi_end;
+ oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
+
+ obdo_set_parent_fid(oa, io->u.ci_fsync.fi_fid);
+
+ memset(oinfo, 0, sizeof(*oinfo));
+ oinfo->oi_oa = oa;
+ oinfo->oi_capa = io->u.ci_fsync.fi_capa;
+ cfs_init_completion(&cbargs->opc_sync);
+
+ result = osc_sync_base(osc_export(cl2osc(obj)), oinfo,
+ osc_async_upcall, cbargs, PTLRPCD_SET);
+ RETURN(result);
+}
+
+static void osc_io_fsync_end(const struct lu_env *env,
+ const struct cl_io_slice *slice)
+{
+ struct cl_io *io = slice->cis_io;
+ struct osc_io *oio = cl2osc_io(env, slice);
+ struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
+
+ cfs_wait_for_completion(&cbargs->opc_sync);
+ io->ci_result = cbargs->opc_rc;
}
static const struct cl_io_operations osc_io_ops = {
.op = {
[CIT_READ] = {
+ .cio_start = osc_io_read_start,
.cio_fini = osc_io_fini
},
[CIT_WRITE] = {
+ .cio_start = osc_io_write_start,
.cio_fini = osc_io_fini
},
- [CIT_TRUNC] = {
- .cio_start = osc_io_trunc_start,
- .cio_end = osc_io_trunc_end
+ [CIT_SETATTR] = {
+ .cio_start = osc_io_setattr_start,
+ .cio_end = osc_io_setattr_end
},
[CIT_FAULT] = {
.cio_fini = osc_io_fini,
.cio_start = osc_io_fault_start
},
+ [CIT_FSYNC] = {
+ .cio_start = osc_io_fsync_start,
+ .cio_end = osc_io_fsync_end,
+ .cio_fini = osc_io_fini
+ },
[CIT_MISC] = {
.cio_fini = osc_io_fini
}
/**
* Implementation of struct cl_req_operations::cro_attr_set() for osc
- * layer. osc is responsible for struct obdo::o_id and struct obdo::o_gr
+ * layer. osc is responsible for struct obdo::o_id and struct obdo::o_seq
* fields.
*/
static void osc_req_attr_set(const struct lu_env *env,
oa->o_valid |= OBD_MD_FLID;
}
if (flags & OBD_MD_FLGROUP) {
- oa->o_gr = oinfo->loi_gr;
+ oa->o_seq = oinfo->loi_seq;
oa->o_valid |= OBD_MD_FLGROUP;
}
if (flags & OBD_MD_FLHANDLE) {
clerq = slice->crs_req;
- LASSERT(!list_empty(&clerq->crq_pages));
+ LASSERT(!cfs_list_empty(&clerq->crq_pages));
apage = container_of(clerq->crq_pages.next,
struct cl_page, cp_flight);
opg = osc_cl_page_osc(apage);
apage = opg->ops_cl.cpl_page; /* now apage is a sub-page */
lock = cl_lock_at_page(env, apage->cp_obj, apage, NULL, 1, 1);
- LASSERT(lock != NULL);
+ if (lock == NULL) {
+ struct cl_object_header *head;
+ struct cl_lock *scan;
+
+ head = cl_object_header(apage->cp_obj);
+ cfs_list_for_each_entry(scan, &head->coh_locks,
+ cll_linkage)
+ CL_LOCK_DEBUG(D_ERROR, env, scan,
+ "no cover page!\n");
+ CL_PAGE_DEBUG(D_ERROR, env, apage,
+ "dump uncover page!\n");
+ libcfs_debug_dumpstack(NULL);
+ LBUG();
+ }
+
olck = osc_lock_at(lock);
LASSERT(olck != NULL);
+ LASSERT(ergo(opg->ops_srvlock, olck->ols_lock == NULL));
/* check for lockless io. */
if (olck->ols_lock != NULL) {
oa->o_handle = olck->ols_lock->l_remote_handle;