#include "osc_cl_internal.h"
+/*
+ * Comment out osc_page_protected because it may sleep inside the
+ * the client_obd_list_lock.
+ * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
+ * -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
+ * -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
+ */
+#if 0
static int osc_page_is_dlocked(const struct lu_env *env,
const struct osc_page *opg,
enum cl_lock_mode mode, int pending, int unref)
ldlm_mode_t dlmmode;
int flags;
+ might_sleep();
+
info = osc_env_info(env);
resname = &info->oti_resname;
policy = &info->oti_policy;
dlmmode, &flags, NULL, lockh, unref);
}
+/**
+ * Checks an invariant that a page in the cache is covered by a lock, as
+ * needed.
+ */
static int osc_page_protected(const struct lu_env *env,
const struct osc_page *opg,
enum cl_lock_mode mode, int unref)
LINVRNT(!opg->ops_temp);
+ page = opg->ops_cl.cpl_page;
+ if (page->cp_owner != NULL &&
+ cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
+ /*
+ * If IO is done without locks (liblustre, or lloop), lock is
+ * not required.
+ */
+ result = 1;
+ else
+ /* otherwise check for a DLM lock */
result = osc_page_is_dlocked(env, opg, mode, 1, unref);
if (result == 0) {
/* maybe this page is a part of a lockless io? */
hdr = cl_object_header(opg->ops_cl.cpl_obj);
- page = opg->ops_cl.cpl_page;
descr = &osc_env_info(env)->oti_descr;
descr->cld_mode = mode;
descr->cld_start = page->cp_index;
}
return result;
}
+#else
+static int osc_page_protected(const struct lu_env *env,
+ const struct osc_page *opg,
+ enum cl_lock_mode mode, int unref)
+{
+ return 1;
+}
+#endif
/*****************************************************************************
*
{
struct osc_object *obj;
+ LINVRNT(cl_page_is_vmlocked(env, opg->ops_cl.cpl_page));
+
obj = cl2osc(opg->ops_cl.cpl_obj);
spin_lock(&obj->oo_seatbelt);
list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
static int osc_page_cache_add(const struct lu_env *env,
const struct cl_page_slice *slice,
- struct cl_io *_)
+ struct cl_io *unused)
{
struct osc_page *opg = cl2osc_page(slice);
struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
ENTRY;
/* Set the OBD_BRW_SRVLOCK before the page is queued. */
- brw_flags = oio->oi_lockless ? OBD_BRW_SRVLOCK : 0;
+ brw_flags = osc_io_srvlock(oio) ? OBD_BRW_SRVLOCK : 0;
if (!client_is_remote(osc_export(obj)) &&
cfs_capable(CFS_CAP_SYS_RESOURCE)) {
brw_flags |= OBD_BRW_NOQUOTA;
static int osc_page_is_under_lock(const struct lu_env *env,
const struct cl_page_slice *slice,
- struct cl_io *_)
+ struct cl_io *unused)
{
struct cl_lock *lock;
int result;
}
static int osc_page_fail(const struct lu_env *env,
- const struct cl_page_slice *slice, struct cl_io *_)
+ const struct cl_page_slice *slice,
+ struct cl_io *unused)
{
/*
* Cached read?
return list_empty(head) ? "-" : "+";
}
+static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
+{
+ if (opg->ops_submit_time == 0)
+ return 0;
+
+ return (cfs_time_current() - opg->ops_submit_time);
+}
+
static int osc_page_print(const struct lu_env *env,
const struct cl_page_slice *slice,
void *cookie, lu_printer_t printer)
struct osc_async_page *oap = &opg->ops_oap;
return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
- "%#x %d %u %s %s %s %llu %u %#x %p %p %p %p %p\n",
- opg, oap->oap_magic, oap->oap_cmd,
+ "< %#x %d %u %s %s %s >"
+ "< %llu %u %#x %#x %p %p %p %p %p >"
+ "< %s %p %d %lu >\n",
+ opg,
+ /* 1 */
+ oap->oap_magic, oap->oap_cmd,
oap->oap_interrupted,
osc_list(&oap->oap_pending_item),
osc_list(&oap->oap_urgent_item),
osc_list(&oap->oap_rpc_item),
+ /* 2 */
oap->oap_obj_off, oap->oap_page_off,
- oap->oap_async_flags, oap->oap_request,
+ oap->oap_async_flags, oap->oap_brw_flags,
+ oap->oap_request,
oap->oap_cli, oap->oap_loi, oap->oap_caller_ops,
- oap->oap_caller_data);
+ oap->oap_caller_data,
+ /* 3 */
+ osc_list(&opg->ops_inflight),
+ opg->ops_submitter, opg->ops_transfer_pinned,
+ osc_submit_duration(opg));
}
static void osc_page_delete(const struct lu_env *env,
CDEBUG(D_TRACE, "%p\n", opg);
osc_page_transfer_put(env, opg);
rc = osc_teardown_async_page(osc_export(obj), NULL, obj->oo_oinfo, oap);
- LASSERTF(rc == 0, "%i\n", rc);
+ if (rc) {
+ CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
+ "Trying to teardown failed: %d\n", rc);
+ LASSERT(0);
+ }
spin_lock(&obj->oo_seatbelt);
list_del_init(&opg->ops_inflight);
spin_unlock(&obj->oo_seatbelt);
opg->ops_from = from;
opg->ops_to = to;
+ spin_lock(&oap->oap_lock);
oap->oap_async_flags |= ASYNC_COUNT_STABLE;
+ spin_unlock(&oap->oap_lock);
}
static int osc_page_cancel(const struct lu_env *env,
ENTRY;
result = cl_page_make_ready(env, page, CRT_WRITE);
+ if (result == 0)
+ opg->ops_submit_time = cfs_time_current();
RETURN(result);
}
enum cl_req_type crt;
LINVRNT(osc_page_protected(env, opg, CLM_READ, 1));
+ LINVRNT(cl_page_is_vmlocked(env, page));
ENTRY;
LASSERT(page->cp_req == NULL);
/* As the transfer for this page is being done, clear the flags */
+ spin_lock(&oap->oap_lock);
oap->oap_async_flags = 0;
+ spin_unlock(&oap->oap_lock);
crt = cmd == OBD_BRW_READ ? CRT_READ : CRT_WRITE;
/* Clear opg->ops_transfer_pinned before VM lock is released. */
list_del_init(&opg->ops_inflight);
spin_unlock(&obj->oo_seatbelt);
+ opg->ops_submit_time = 0;
+
cl_page_completion(env, page, crt, rc);
/* statistic */
struct osc_page *opg;
int result;
- OBD_SLAB_ALLOC_PTR(opg, osc_page_kmem);
+ OBD_SLAB_ALLOC_PTR_GFP(opg, osc_page_kmem, CFS_ALLOC_IO);
if (opg != NULL) {
void *oap = &opg->ops_oap;
return ERR_PTR(result);
}
+/**
+ * Helper function called by osc_io_submit() for every page in an immediate
+ * transfer (i.e., transferred synchronously).
+ */
void osc_io_submit_page(const struct lu_env *env,
struct osc_io *oio, struct osc_page *opg,
enum cl_req_type crt)
{
struct osc_async_page *oap = &opg->ops_oap;
struct client_obd *cli = oap->oap_cli;
+ int flags = 0;
LINVRNT(osc_page_protected(env, opg,
crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
oap->oap_page_off = opg->ops_from;
oap->oap_count = opg->ops_to - opg->ops_from;
+ /* Give a hint to OST that requests are coming from kswapd - bug19529 */
+ if (libcfs_memory_pressure_get())
+ oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
oap->oap_brw_flags |= OBD_BRW_SYNC;
- if (oio->oi_lockless)
+ if (osc_io_srvlock(oio))
oap->oap_brw_flags |= OBD_BRW_SRVLOCK;
oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
oap->oap_cmd |= OBD_BRW_NOQUOTA;
}
- oap->oap_async_flags |= OSC_FLAGS;
if (oap->oap_cmd & OBD_BRW_READ)
- oap->oap_async_flags |= ASYNC_COUNT_STABLE;
+ flags = ASYNC_COUNT_STABLE;
else if (!(oap->oap_brw_page.flag & OBD_BRW_FROM_GRANT))
osc_enter_cache_try(env, cli, oap->oap_loi, oap, 1);
+ spin_lock(&oap->oap_lock);
+ oap->oap_async_flags |= OSC_FLAGS | flags;
+ spin_unlock(&oap->oap_lock);
+
osc_oap_to_pending(oap);
osc_page_transfer_get(opg, "transfer\0imm");
osc_page_transfer_add(env, opg, crt);