From: Alexei Potashnik <[email protected]>
If command didn't match a LUN and we're sending check condition, the
target_cmd_complete ftrace point will crash because it assumes that
cmd->t_task_cdb has been set.
The fix will temporarily set t_task_cdb to the se_cmd buffer
and copy first 6 bytes of cdb in there as soon as possible.
At a later point t_task_cdb is reset to the correct buffer,
but until then traces and printks don't cause a crash.
Signed-off-by: Alexei Potashnik <[email protected]>
Signed-off-by: Spencer Baugh <[email protected]>
---
drivers/target/iscsi/iscsi_target.c | 4 ++--
drivers/target/target_core_device.c | 11 +++++++++--
drivers/target/target_core_transport.c | 9 +++++----
include/target/target_core_fabric.h | 2 +-
4 files changed, 17 insertions(+), 9 deletions(-)
diff --git a/drivers/target/iscsi/iscsi_target.c b/drivers/target/iscsi/iscsi_target.c
index f615d75..98899f1 100644
--- a/drivers/target/iscsi/iscsi_target.c
+++ b/drivers/target/iscsi/iscsi_target.c
@@ -1003,8 +1003,8 @@ int iscsit_setup_scsi_cmd(struct iscsi_conn *conn, struct iscsi_cmd *cmd,
target_get_sess_cmd(&cmd->se_cmd, true);
- cmd->sense_reason = transport_lookup_cmd_lun(&cmd->se_cmd,
- scsilun_to_int(&hdr->lun));
+ cmd->sense_reason = transport_lookup_cmd_lun_cdb(&cmd->se_cmd,
+ scsilun_to_int(&hdr->lun), hdr->cdb);
if (cmd->sense_reason)
goto attach_cmd;
diff --git a/drivers/target/target_core_device.c b/drivers/target/target_core_device.c
index c4a8db6..acf84cf 100644
--- a/drivers/target/target_core_device.c
+++ b/drivers/target/target_core_device.c
@@ -56,13 +56,20 @@ static struct se_hba *lun0_hba;
struct se_device *g_lun0_dev;
sense_reason_t
-transport_lookup_cmd_lun(struct se_cmd *se_cmd, u64 unpacked_lun)
+transport_lookup_cmd_lun_cdb(struct se_cmd *se_cmd, u64 unpacked_lun, unsigned char *cdb)
{
struct se_lun *se_lun = NULL;
struct se_session *se_sess = se_cmd->se_sess;
struct se_node_acl *nacl = se_sess->se_node_acl;
struct se_dev_entry *deve;
+ /* Temporarily set t_task_cdb to the se_cmd buffer and save a portion
+ * of cdb in there (fabrics must provide at least 6 bytes). t_task_cdb
+ * will be correctly replaced in target_setup_cmd_from_cdb. Until then
+ * tracing and printks can access t_task_cdb without causing a crash. */
+ se_cmd->t_task_cdb = se_cmd->__t_task_cdb;
+ memcpy(se_cmd->t_task_cdb, cdb, 6);
+
rcu_read_lock();
deve = target_nacl_find_deve(nacl, unpacked_lun);
if (deve) {
@@ -142,7 +149,7 @@ transport_lookup_cmd_lun(struct se_cmd *se_cmd, u64 unpacked_lun)
return 0;
}
-EXPORT_SYMBOL(transport_lookup_cmd_lun);
+EXPORT_SYMBOL(transport_lookup_cmd_lun_cdb);
int transport_lookup_tmr_lun(struct se_cmd *se_cmd, u64 unpacked_lun)
{
diff --git a/drivers/target/target_core_transport.c b/drivers/target/target_core_transport.c
index f6626bb..1f761a3 100644
--- a/drivers/target/target_core_transport.c
+++ b/drivers/target/target_core_transport.c
@@ -1210,15 +1210,16 @@ target_setup_cmd_from_cdb(struct se_cmd *cmd, unsigned char *cdb)
* setup the pointer from __t_task_cdb to t_task_cdb.
*/
if (scsi_command_size(cdb) > sizeof(cmd->__t_task_cdb)) {
- cmd->t_task_cdb = kzalloc(scsi_command_size(cdb),
- GFP_KERNEL);
- if (!cmd->t_task_cdb) {
+ unsigned char *ptr = kzalloc(scsi_command_size(cdb),
+ GFP_KERNEL);
+ if (!ptr) {
pr_err("Unable to allocate cmd->t_task_cdb"
" %u > sizeof(cmd->__t_task_cdb): %lu ops\n",
scsi_command_size(cdb),
(unsigned long)sizeof(cmd->__t_task_cdb));
return TCM_OUT_OF_RESOURCES;
}
+ cmd->t_task_cdb = ptr;
} else
cmd->t_task_cdb = &cmd->__t_task_cdb[0];
/*
@@ -1404,7 +1405,7 @@ int target_submit_cmd_map_sgls(struct se_cmd *se_cmd, struct se_session *se_sess
/*
* Locate se_lun pointer and attach it to struct se_cmd
*/
- rc = transport_lookup_cmd_lun(se_cmd, unpacked_lun);
+ rc = transport_lookup_cmd_lun_cdb(se_cmd, unpacked_lun, cdb);
if (rc) {
transport_send_check_condition_and_sense(se_cmd, rc, 0);
target_put_sess_cmd(se_cmd);
diff --git a/include/target/target_core_fabric.h b/include/target/target_core_fabric.h
index 18afef9..bfa6368 100644
--- a/include/target/target_core_fabric.h
+++ b/include/target/target_core_fabric.h
@@ -116,7 +116,7 @@ void transport_deregister_session(struct se_session *);
void transport_init_se_cmd(struct se_cmd *,
const struct target_core_fabric_ops *,
struct se_session *, u32, int, int, unsigned char *);
-sense_reason_t transport_lookup_cmd_lun(struct se_cmd *, u64);
+sense_reason_t transport_lookup_cmd_lun_cdb(struct se_cmd *, u64, unsigned char *);
sense_reason_t target_setup_cmd_from_cdb(struct se_cmd *, unsigned char *);
int target_submit_cmd_map_sgls(struct se_cmd *, struct se_session *,
unsigned char *, unsigned char *, u64, u32, int, int, int,
--
2.5.0.rc3
On Thu, Jul 23, 2015 at 03:19:32PM -0700, Spencer Baugh wrote:
> From: Alexei Potashnik <[email protected]>
>
> If command didn't match a LUN and we're sending check condition, the
> target_cmd_complete ftrace point will crash because it assumes that
> cmd->t_task_cdb has been set.
>
> The fix will temporarily set t_task_cdb to the se_cmd buffer
> and copy first 6 bytes of cdb in there as soon as possible.
> At a later point t_task_cdb is reset to the correct buffer,
> but until then traces and printks don't cause a crash.
This is too ugly to live. Just dropping the t_task_cdb dereference
from the trace point sounds like the simples quick fix for now,
and removing the crazy layering violation in iSCSI that opencode
target_submit_cmd is the proper long term fix.
On Fri, 2015-07-24 at 12:52 +0200, Christoph Hellwig wrote:
> On Thu, Jul 23, 2015 at 03:19:32PM -0700, Spencer Baugh wrote:
> > From: Alexei Potashnik <[email protected]>
> >
> > If command didn't match a LUN and we're sending check condition, the
> > target_cmd_complete ftrace point will crash because it assumes that
> > cmd->t_task_cdb has been set.
> >
> > The fix will temporarily set t_task_cdb to the se_cmd buffer
> > and copy first 6 bytes of cdb in there as soon as possible.
> > At a later point t_task_cdb is reset to the correct buffer,
> > but until then traces and printks don't cause a crash.
>
> This is too ugly to live. Just dropping the t_task_cdb dereference
> from the trace point sounds like the simples quick fix for now,
Yes, that is what I'd prefer as well.
> and removing the crazy layering violation in iSCSI that opencode
> target_submit_cmd is the proper long term fix.
We've already been through this discussion a couple of years back when
target_submit_cmd() first came into existence.
The reason iscsi/iser-target continues to be a special case is due to
immediate data vs. non immediate data and their respective command
sequence number ordering requirements.
--nab
On Fri, Jul 24, 2015 at 01:32:14PM -0700, Nicholas A. Bellinger wrote:
> We've already been through this discussion a couple of years back when
> target_submit_cmd() first came into existence.
>
> The reason iscsi/iser-target continues to be a special case is due to
> immediate data vs. non immediate data and their respective command
> sequence number ordering requirements.
I don't see how immediate data plays into this, the write_pending
callbacks can simply skip the data transfer path, similar to what
Bart's port of the latest SRP target to lio does as well.
On Sat, 2015-07-25 at 08:48 +0200, Christoph Hellwig wrote:
> On Fri, Jul 24, 2015 at 01:32:14PM -0700, Nicholas A. Bellinger wrote:
> > We've already been through this discussion a couple of years back when
> > target_submit_cmd() first came into existence.
> >
> > The reason iscsi/iser-target continues to be a special case is due to
> > immediate data vs. non immediate data and their respective command
> > sequence number ordering requirements.
>
> I don't see how immediate data plays into this, the write_pending
> callbacks can simply skip the data transfer path, similar to what
> Bart's port of the latest SRP target to lio does as well.
iscsit_execute_cmd() is using iscsit_transport->iscsit_get_dataout() for
any remaining solicited data-out (R2T/RDMA_READ) payload when immediate
write data is smaller than total EDTL.