Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
Merge patch series "scsi: target: Add support for completing commands from backend context"
The following patches made over Linus's current tree allow users to
tell the target layer to perform direct completions instead of always
deferring to LIO's completion workqueue. When the frontend driver
already has multiple worker threads (or you are doing a LUN per target
with a single thread per target) then bypassing the LIO workqueue can
increase performance 20-30%.
Link: https://patch.msgid.link/20260222232946.7637-1-michael.christie@oracle.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
+129
-14
+1
drivers/infiniband/ulp/srpt/ib_srpt.c
+1
drivers/infiniband/ulp/srpt/ib_srpt.c
+2
drivers/scsi/elx/efct/efct_lio.c
+2
drivers/scsi/elx/efct/efct_lio.c
···
1612
1612
.sess_get_initiator_sid = NULL,
1613
1613
.tfc_tpg_base_attrs = efct_lio_tpg_attrs,
1614
1614
.tfc_tpg_attrib_attrs = efct_lio_tpg_attrib_attrs,
1615
+
.default_compl_type = TARGET_QUEUE_COMPL,
1615
1616
.default_submit_type = TARGET_DIRECT_SUBMIT,
1616
1617
.direct_submit_supp = 1,
1617
1618
};
···
1651
1650
.tfc_tpg_base_attrs = efct_lio_npiv_tpg_attrs,
1652
1651
.tfc_tpg_attrib_attrs = efct_lio_npiv_tpg_attrib_attrs,
1653
1652
1653
+
.default_compl_type = TARGET_QUEUE_COMPL,
1654
1654
.default_submit_type = TARGET_DIRECT_SUBMIT,
1655
1655
.direct_submit_supp = 1,
1656
1656
};
+1
drivers/scsi/ibmvscsi_tgt/ibmvscsi_tgt.c
+1
drivers/scsi/ibmvscsi_tgt/ibmvscsi_tgt.c
+2
drivers/scsi/qla2xxx/tcm_qla2xxx.c
+2
drivers/scsi/qla2xxx/tcm_qla2xxx.c
···
1841
1841
.tfc_tpg_base_attrs = tcm_qla2xxx_tpg_attrs,
1842
1842
.tfc_tpg_attrib_attrs = tcm_qla2xxx_tpg_attrib_attrs,
1843
1843
1844
+
.default_compl_type = TARGET_QUEUE_COMPL,
1844
1845
.default_submit_type = TARGET_DIRECT_SUBMIT,
1845
1846
.direct_submit_supp = 1,
1846
1847
};
···
1882
1881
1883
1882
.tfc_wwn_attrs = tcm_qla2xxx_wwn_attrs,
1884
1883
1884
+
.default_compl_type = TARGET_QUEUE_COMPL,
1885
1885
.default_submit_type = TARGET_DIRECT_SUBMIT,
1886
1886
.direct_submit_supp = 1,
1887
1887
};
+1
drivers/target/iscsi/iscsi_target_configfs.c
+1
drivers/target/iscsi/iscsi_target_configfs.c
+1
drivers/target/loopback/tcm_loop.c
+1
drivers/target/loopback/tcm_loop.c
···
1107
1107
.tfc_wwn_attrs = tcm_loop_wwn_attrs,
1108
1108
.tfc_tpg_base_attrs = tcm_loop_tpg_attrs,
1109
1109
.tfc_tpg_attrib_attrs = tcm_loop_tpg_attrib_attrs,
1110
+
.default_compl_type = TARGET_QUEUE_COMPL,
1110
1111
.default_submit_type = TARGET_QUEUE_SUBMIT,
1111
1112
.direct_submit_supp = 0,
1112
1113
};
+1
drivers/target/sbp/sbp_target.c
+1
drivers/target/sbp/sbp_target.c
+22
drivers/target/target_core_configfs.c
+22
drivers/target/target_core_configfs.c
···
578
578
DEF_CONFIGFS_ATTRIB_SHOW(max_write_same_len);
579
579
DEF_CONFIGFS_ATTRIB_SHOW(emulate_rsoc);
580
580
DEF_CONFIGFS_ATTRIB_SHOW(submit_type);
581
+
DEF_CONFIGFS_ATTRIB_SHOW(complete_type);
581
582
DEF_CONFIGFS_ATTRIB_SHOW(atomic_max_len);
582
583
DEF_CONFIGFS_ATTRIB_SHOW(atomic_alignment);
583
584
DEF_CONFIGFS_ATTRIB_SHOW(atomic_granularity);
···
1270
1269
return count;
1271
1270
}
1272
1271
1272
+
static ssize_t complete_type_store(struct config_item *item, const char *page,
1273
+
size_t count)
1274
+
{
1275
+
struct se_dev_attrib *da = to_attrib(item);
1276
+
int ret;
1277
+
u8 val;
1278
+
1279
+
ret = kstrtou8(page, 0, &val);
1280
+
if (ret < 0)
1281
+
return ret;
1282
+
1283
+
if (val > TARGET_QUEUE_COMPL)
1284
+
return -EINVAL;
1285
+
1286
+
da->complete_type = val;
1287
+
return count;
1288
+
}
1289
+
1273
1290
CONFIGFS_ATTR(, emulate_model_alias);
1274
1291
CONFIGFS_ATTR(, emulate_dpo);
1275
1292
CONFIGFS_ATTR(, emulate_fua_write);
···
1324
1305
CONFIGFS_ATTR(, alua_support);
1325
1306
CONFIGFS_ATTR(, pgr_support);
1326
1307
CONFIGFS_ATTR(, submit_type);
1308
+
CONFIGFS_ATTR(, complete_type);
1327
1309
CONFIGFS_ATTR_RO(, atomic_max_len);
1328
1310
CONFIGFS_ATTR_RO(, atomic_alignment);
1329
1311
CONFIGFS_ATTR_RO(, atomic_granularity);
···
1373
1353
&attr_pgr_support,
1374
1354
&attr_emulate_rsoc,
1375
1355
&attr_submit_type,
1356
+
&attr_complete_type,
1376
1357
&attr_atomic_alignment,
1377
1358
&attr_atomic_max_len,
1378
1359
&attr_atomic_granularity,
···
1397
1376
&attr_alua_support,
1398
1377
&attr_pgr_support,
1399
1378
&attr_submit_type,
1379
+
&attr_complete_type,
1400
1380
NULL,
1401
1381
};
1402
1382
EXPORT_SYMBOL(passthrough_attrib_attrs);
+1
drivers/target/target_core_device.c
+1
drivers/target/target_core_device.c
···
813
813
DA_UNMAP_ZEROES_DATA_DEFAULT;
814
814
dev->dev_attrib.max_write_same_len = DA_MAX_WRITE_SAME_LEN;
815
815
dev->dev_attrib.submit_type = TARGET_FABRIC_DEFAULT_SUBMIT;
816
+
dev->dev_attrib.submit_type = TARGET_FABRIC_DEFAULT_COMPL;
816
817
817
818
/* Skip allocating lun_stats since we can't export them. */
818
819
xcopy_lun = &dev->xcopy_lun;
+24
drivers/target/target_core_fabric_configfs.c
+24
drivers/target/target_core_fabric_configfs.c
···
1066
1066
CONFIGFS_ATTR(target_fabric_wwn_, cmd_completion_affinity);
1067
1067
1068
1068
static ssize_t
1069
+
target_fabric_wwn_default_complete_type_show(struct config_item *item,
1070
+
char *page)
1071
+
{
1072
+
struct se_wwn *wwn = container_of(to_config_group(item), struct se_wwn,
1073
+
param_group);
1074
+
return sysfs_emit(page, "%u\n",
1075
+
wwn->wwn_tf->tf_ops->default_compl_type);
1076
+
}
1077
+
CONFIGFS_ATTR_RO(target_fabric_wwn_, default_complete_type);
1078
+
1079
+
static ssize_t
1080
+
target_fabric_wwn_direct_complete_supported_show(struct config_item *item,
1081
+
char *page)
1082
+
{
1083
+
struct se_wwn *wwn = container_of(to_config_group(item), struct se_wwn,
1084
+
param_group);
1085
+
return sysfs_emit(page, "%u\n",
1086
+
wwn->wwn_tf->tf_ops->direct_compl_supp);
1087
+
}
1088
+
CONFIGFS_ATTR_RO(target_fabric_wwn_, direct_complete_supported);
1089
+
1090
+
static ssize_t
1069
1091
target_fabric_wwn_default_submit_type_show(struct config_item *item,
1070
1092
char *page)
1071
1093
{
···
1111
1089
1112
1090
static struct configfs_attribute *target_fabric_wwn_param_attrs[] = {
1113
1091
&target_fabric_wwn_attr_cmd_completion_affinity,
1092
+
&target_fabric_wwn_attr_default_complete_type,
1093
+
&target_fabric_wwn_attr_direct_complete_supported,
1114
1094
&target_fabric_wwn_attr_default_submit_type,
1115
1095
&target_fabric_wwn_attr_direct_submit_supported,
1116
1096
NULL,
+49
-11
drivers/target/target_core_transport.c
+49
-11
drivers/target/target_core_transport.c
···
902
902
return false;
903
903
}
904
904
905
+
static void target_complete(struct se_cmd *cmd, int success)
906
+
{
907
+
struct se_wwn *wwn = cmd->se_sess->se_tpg->se_tpg_wwn;
908
+
struct se_dev_attrib *da;
909
+
u8 compl_type;
910
+
int cpu;
911
+
912
+
if (!wwn) {
913
+
cpu = cmd->cpuid;
914
+
goto queue_work;
915
+
}
916
+
917
+
da = &cmd->se_dev->dev_attrib;
918
+
if (da->complete_type == TARGET_FABRIC_DEFAULT_COMPL)
919
+
compl_type = wwn->wwn_tf->tf_ops->default_compl_type;
920
+
else if (da->complete_type == TARGET_DIRECT_SUBMIT &&
921
+
wwn->wwn_tf->tf_ops->direct_compl_supp)
922
+
compl_type = TARGET_DIRECT_COMPL;
923
+
else
924
+
compl_type = TARGET_QUEUE_COMPL;
925
+
926
+
if (compl_type == TARGET_DIRECT_COMPL) {
927
+
/*
928
+
* Failure handling and processing secondary stages of
929
+
* complex commands can be too heavy to handle from the
930
+
* fabric driver so always defer.
931
+
*/
932
+
if (success && !cmd->transport_complete_callback) {
933
+
target_complete_ok_work(&cmd->work);
934
+
return;
935
+
}
936
+
937
+
compl_type = TARGET_QUEUE_COMPL;
938
+
}
939
+
940
+
queue_work:
941
+
INIT_WORK(&cmd->work, success ? target_complete_ok_work :
942
+
target_complete_failure_work);
943
+
944
+
if (!wwn || wwn->cmd_compl_affinity == SE_COMPL_AFFINITY_CPUID)
945
+
cpu = cmd->cpuid;
946
+
else
947
+
cpu = wwn->cmd_compl_affinity;
948
+
949
+
queue_work_on(cpu, target_completion_wq, &cmd->work);
950
+
}
951
+
905
952
/* May be called from interrupt context so must not sleep. */
906
953
void target_complete_cmd_with_sense(struct se_cmd *cmd, u8 scsi_status,
907
954
sense_reason_t sense_reason)
908
955
{
909
-
struct se_wwn *wwn = cmd->se_sess->se_tpg->se_tpg_wwn;
910
-
int success, cpu;
911
956
unsigned long flags;
957
+
int success;
912
958
913
959
if (target_cmd_interrupted(cmd))
914
960
return;
···
979
933
cmd->transport_state |= (CMD_T_COMPLETE | CMD_T_ACTIVE);
980
934
spin_unlock_irqrestore(&cmd->t_state_lock, flags);
981
935
982
-
INIT_WORK(&cmd->work, success ? target_complete_ok_work :
983
-
target_complete_failure_work);
984
-
985
-
if (!wwn || wwn->cmd_compl_affinity == SE_COMPL_AFFINITY_CPUID)
986
-
cpu = cmd->cpuid;
987
-
else
988
-
cpu = wwn->cmd_compl_affinity;
989
-
990
-
queue_work_on(cpu, target_completion_wq, &cmd->work);
936
+
target_complete(cmd, success);
991
937
}
992
938
EXPORT_SYMBOL(target_complete_cmd_with_sense);
993
939
+1
drivers/target/tcm_fc/tfc_conf.c
+1
drivers/target/tcm_fc/tfc_conf.c
+1
drivers/usb/gadget/function/f_tcm.c
+1
drivers/usb/gadget/function/f_tcm.c
+2
drivers/vhost/scsi.c
+2
drivers/vhost/scsi.c
···
2950
2950
.tfc_tpg_base_attrs = vhost_scsi_tpg_attrs,
2951
2951
.tfc_tpg_attrib_attrs = vhost_scsi_tpg_attrib_attrs,
2952
2952
2953
+
.default_compl_type = TARGET_QUEUE_COMPL,
2954
+
.direct_compl_supp = 1,
2953
2955
.default_submit_type = TARGET_QUEUE_SUBMIT,
2954
2956
.direct_submit_supp = 1,
2955
2957
};
+1
drivers/xen/xen-scsiback.c
+1
drivers/xen/xen-scsiback.c
+10
include/target/target_core_base.h
+10
include/target/target_core_base.h
···
111
111
/* Peripheral Device Text Identification Information */
112
112
#define PD_TEXT_ID_INFO_LEN 256
113
113
114
+
enum target_compl_type {
115
+
/* Use the fabric driver's default completion type */
116
+
TARGET_FABRIC_DEFAULT_COMPL,
117
+
/* Complete from the backend calling context */
118
+
TARGET_DIRECT_COMPL,
119
+
/* Defer completion to the LIO workqueue */
120
+
TARGET_QUEUE_COMPL,
121
+
};
122
+
114
123
enum target_submit_type {
115
124
/* Use the fabric driver's default submission type */
116
125
TARGET_FABRIC_DEFAULT_SUBMIT,
···
750
741
u32 atomic_granularity;
751
742
u32 atomic_max_with_boundary;
752
743
u32 atomic_max_boundary;
744
+
u8 complete_type;
753
745
u8 submit_type;
754
746
struct se_device *da_dev;
755
747
struct config_group da_group;
+9
-3
include/target/target_core_fabric.h
+9
-3
include/target/target_core_fabric.h
···
119
119
*/
120
120
unsigned int write_pending_must_be_called:1;
121
121
/*
122
+
* Set this if the driver does not require calling queue_data_in
123
+
* queue_status and check_stop_free from a worker thread when
124
+
* completing successful commands.
125
+
*/
126
+
unsigned int direct_compl_supp:1;
127
+
/*
122
128
* Set this if the driver supports submitting commands to the backend
123
129
* from target_submit/target_submit_cmd.
124
130
*/
125
131
unsigned int direct_submit_supp:1;
126
-
/*
127
-
* Set this to a target_submit_type value.
128
-
*/
132
+
/* Set this to a target_submit_type value. */
129
133
u8 default_submit_type;
134
+
/* Set this to the target_compl_type value. */
135
+
u8 default_compl_type;
130
136
};
131
137
132
138
int target_register_template(const struct target_core_fabric_ops *fo);