staging/lustre: Limit reply buffer size
When allocating a reply buffer for the striping information don't
assume the unlikely worst case. Instead, assume the common case
and size the buffer based on the observed default ea/cookie size.
The default size is initialized to a single stripe and allowed to
grow up to an entire page if needed. This means that for smallish
filesystems (less than ~21 OSTs) where the worst case striping
information can fit in a single page there is effectively no
change. Only for larger filesystem will the default be less than
the maximum. This has a number of advantages.
* By limiting the default reply buffer size we avoid always
vmalloc()'ing the buffer because it exceeds four pages in size
and instead kmalloc() it. This prevents the client from
thrashing on the global vmalloc() spin lock.
* A reply buffer of exactly the right size (no larger) is allocated
in the overflow case. These larger reply buffers are still
unlikely to exceed the 16k limit where a vmalloc() will occur.
* Saves memory in the common case. Wide striped files exceeded
the default are expected to be the exception.
The reason this patch works is because the ptlrpc layer is smart
enough to reallocate the reply buffer when an overflow occurs.
Therefore the client doesn't have to drop the incoming reply and
send a new request with a larger reply buffer.
It's also worth mentioning that the reply buffer always contains
a significant amount of extra padding because they are rounded up
to the nearest power of two. This means that even files striped
wider than the default have a good chance of fitting in the
allocated reply buffer.
Also remove client eadatasize check in mdt xattr packing because
as said above client can handle -EOVERFLOW.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Lai Siyao <lai.siyao@intel.com>
Reviewed-on: http://review.whamcloud.com/6339
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-3338
Reviewed-by: James Simmons <uja.ornl@gmail.com>
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Reviewed-by: Bob Glossman <bob.glossman@intel.com>
Signed-off-by: Oleg Drokin <oleg.drokin@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
diff --git a/drivers/staging/lustre/lustre/mdc/mdc_request.c b/drivers/staging/lustre/lustre/mdc/mdc_request.c
index bde9f93..fca43cf 100644
--- a/drivers/staging/lustre/lustre/mdc/mdc_request.c
+++ b/drivers/staging/lustre/lustre/mdc/mdc_request.c
@@ -903,9 +903,9 @@
mdc_close_pack(req, op_data);
req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
- obd->u.cli.cl_max_mds_easize);
+ obd->u.cli.cl_default_mds_easize);
req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES, RCL_SERVER,
- obd->u.cli.cl_max_mds_cookiesize);
+ obd->u.cli.cl_default_mds_cookiesize);
ptlrpc_request_set_replen(req);
@@ -2153,12 +2153,40 @@
if (*vallen != sizeof(int))
return -EINVAL;
- mdsize = *(int*)val;
+ mdsize = *(int *)val;
if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
max_easize = val;
*max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
return 0;
+ } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
+ int *default_easize;
+
+ if (*vallen != sizeof(int))
+ return -EINVAL;
+ default_easize = val;
+ *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
+ return 0;
+ } else if (KEY_IS(KEY_MAX_COOKIESIZE)) {
+ int mdsize, *max_cookiesize;
+
+ if (*vallen != sizeof(int))
+ return -EINVAL;
+ mdsize = *(int *)val;
+ if (mdsize > exp->exp_obd->u.cli.cl_max_mds_cookiesize)
+ exp->exp_obd->u.cli.cl_max_mds_cookiesize = mdsize;
+ max_cookiesize = val;
+ *max_cookiesize = exp->exp_obd->u.cli.cl_max_mds_cookiesize;
+ return 0;
+ } else if (KEY_IS(KEY_DEFAULT_COOKIESIZE)) {
+ int *default_cookiesize;
+
+ if (*vallen != sizeof(int))
+ return -EINVAL;
+ default_cookiesize = val;
+ *default_cookiesize =
+ exp->exp_obd->u.cli.cl_default_mds_cookiesize;
+ return 0;
} else if (KEY_IS(KEY_CONN_DATA)) {
struct obd_import *imp = class_exp2cliimp(exp);
struct obd_connect_data *data = val;
@@ -2439,11 +2467,15 @@
}
/* Initialize the default and maximum LOV EA and cookie sizes. This allows
- * us to make MDS RPCs with large enough reply buffers to hold the
- * maximum-sized (= maximum striped) EA and cookie without having to
- * calculate this (via a call into the LOV + OSCs) each time we make an RPC. */
+ * us to make MDS RPCs with large enough reply buffers to hold a default
+ * sized EA and cookie without having to calculate this (via a call into the
+ * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
+ * but not used to avoid wastefully vmalloc()'ing large reply buffers when
+ * a large number of stripes is possible. If a larger reply buffer is
+ * required it will be reallocated in the ptlrpc layer due to overflow.
+ */
static int mdc_init_ea_size(struct obd_export *exp, int easize,
- int def_easize, int cookiesize)
+ int def_easize, int cookiesize, int def_cookiesize)
{
struct obd_device *obd = exp->exp_obd;
struct client_obd *cli = &obd->u.cli;
@@ -2457,6 +2489,9 @@
if (cli->cl_max_mds_cookiesize < cookiesize)
cli->cl_max_mds_cookiesize = cookiesize;
+ if (cli->cl_default_mds_cookiesize < def_cookiesize)
+ cli->cl_default_mds_cookiesize = def_cookiesize;
+
return 0;
}