Re: [PATCH v3] scsi: tcmu: avoid cmd/qfull timers updated whenever a new cmd comes
On 2018/11/22 0:37, Mike Christie wrote:
On 11/20/2018 11:37 PM, Xiubo Li wrote:
[...]
-is_running = list_empty(&cmd->cmdr_queue_entry);
+is_running = test_bit(TCMU_CMD_BIT_INFLIGHT, &cmd->flags);
se_cmd = cmd->se_cmd;
if (is_running) {
@@ -1289,7 +1319,6 @@ static int tcmu_check_expired_cmd(int id, void
*p, void *data)
scsi_status = SAM_STAT_CHECK_CONDITION;
} else {
list_del_init(&cmd->cmdr_queue_entry);
Move this list_del_init call to outside the if/else.
You need do delete it from the cmdr_inflight_queue if that is how it
timed out, or if you later call tcmu_get_next_deadline it will still
show up and possibly be used to set the next time out which already
happened.
Firstly, this is in the timeout routine, if this cmd was already timed
out and it must be time_after(jiffies, cmd->deadline), so it won't be
used again.
It could be stuck a long time. What about jiffies rollover?
Okay, will update this.
Thanks.
Re: [PATCH v3] scsi: tcmu: avoid cmd/qfull timers updated whenever a new cmd comes
On 11/20/2018 11:37 PM, Xiubo Li wrote:
> [...]
>>> -is_running = list_empty(&cmd->cmdr_queue_entry);
>>> +is_running = test_bit(TCMU_CMD_BIT_INFLIGHT, &cmd->flags);
>>> se_cmd = cmd->se_cmd;
>>> if (is_running) {
>>> @@ -1289,7 +1319,6 @@ static int tcmu_check_expired_cmd(int id, void
>>> *p, void *data)
>>> scsi_status = SAM_STAT_CHECK_CONDITION;
>>> } else {
>>> list_del_init(&cmd->cmdr_queue_entry);
>> Move this list_del_init call to outside the if/else.
>>
>> You need do delete it from the cmdr_inflight_queue if that is how it
>> timed out, or if you later call tcmu_get_next_deadline it will still
>> show up and possibly be used to set the next time out which already
>> happened.
>
> Firstly, this is in the timeout routine, if this cmd was already timed
> out and it must be time_after(jiffies, cmd->deadline), so it won't be
> used again.
It could be stuck a long time. What about jiffies rollover?
Re: [PATCH v3] scsi: tcmu: avoid cmd/qfull timers updated whenever a new cmd comes
On 2018/11/21 11:19, Mike Christie wrote:
On 10/17/2018 02:54 AM, xiu...@redhat.com wrote:
From: Xiubo Li
[...]
diff --git a/drivers/target/target_core_user.c
b/drivers/target/target_core_user.c
index 9cd404a..00ed7bb 100644
--- a/drivers/target/target_core_user.c
+++ b/drivers/target/target_core_user.c
@@ -148,7 +148,7 @@ struct tcmu_dev {
size_t ring_size;
struct mutex cmdr_lock;
- struct list_head cmdr_queue;
+ struct list_head cmdr_qfull_queue;
We can just rename to qfull_queue and inflight_queue and then rename the
function names you modified below.
yeah, will fix it.
[...]
+static unsigned long tcmu_get_next_deadline(struct list_head *queue)
+{
+ struct tcmu_cmd *tcmu_cmd, *tmp_cmd;
+
+ list_for_each_entry_safe(tcmu_cmd, tmp_cmd, queue, cmdr_queue_entry) {
+ if (!time_after(jiffies, tcmu_cmd->deadline))
+ return tcmu_cmd->deadline;
+ }
+
+ return 0;
I think you can just pass the timer to this function, rename it to
tcmu_set_next_deadline() and move the mod/del_timer calls in here too
since it is always the same behavior.
Will fix it.
[...]
@@ -1230,6 +1249,9 @@ static unsigned int tcmu_handle_completions(struct
tcmu_dev *udev)
break;
}
+ clear_bit(TCMU_CMD_BIT_INFLIGHT, &cmd->flags);
+ list_del_init(&cmd->cmdr_queue_entry);
Move these 2 lines to tcmu_handle_completion.
Will fix it.
[...]
- is_running = list_empty(&cmd->cmdr_queue_entry);
+ is_running = test_bit(TCMU_CMD_BIT_INFLIGHT, &cmd->flags);
se_cmd = cmd->se_cmd;
if (is_running) {
@@ -1289,7 +1319,6 @@ static int tcmu_check_expired_cmd(int id, void *p, void
*data)
scsi_status = SAM_STAT_CHECK_CONDITION;
} else {
list_del_init(&cmd->cmdr_queue_entry);
Move this list_del_init call to outside the if/else.
You need do delete it from the cmdr_inflight_queue if that is how it
timed out, or if you later call tcmu_get_next_deadline it will still
show up and possibly be used to set the next time out which already
happened.
Firstly, this is in the timeout routine, if this cmd was already timed
out and it must be time_after(jiffies, cmd->deadline), so it won't be
used again.
Secondly, For the timed out cmds, it will be deleted in the
tcmu_handle_completion() later as above mentioned.
And last is that currently maybe only the qfull queue timer is enabled.
Then there is no need to delete it for the inflight queue.
-
idr_remove(&udev->commands, id);
tcmu_free_cmd(cmd);
scsi_status = SAM_STAT_TASK_SET_FULL;
@@ -1372,7 +1401,8 @@ static struct se_device *tcmu_alloc_device(struct se_hba
*hba, const char *name)
[...]
@@ -1437,13 +1468,18 @@ static bool run_cmdr_queue(struct tcmu_dev *udev, bool
fail)
* cmd was requeued, so just put all cmds back in
* the queue
*/
- list_splice_tail(&cmds, &udev->cmdr_queue);
+ list_splice_tail(&cmds, &udev->cmdr_qfull_queue);
drained = false;
goto done;
I think this needs to be a break to fix the bug where we fail to add to
the ring after the initial command (if we fail on the initial command
then the qfull timeout should be the same as before). If say the first
command is put on the ring and we run out of space for the 2nd command
then the second might have a deadline of 1 but the first only had a
deadline of 1.
Yeah, that's right, here maybe some cmds from the qfull queue have been
added to the ring successfully, so we need to update the qfull timer for
the reset ones.
Will fix it.
Thanks,
BRs
Xiubo
Re: [PATCH v3] scsi: tcmu: avoid cmd/qfull timers updated whenever a new cmd comes
On 10/17/2018 02:54 AM, xiu...@redhat.com wrote:
> From: Xiubo Li
>
> Currently there has one cmd timeout timer and one qfull timer for
> each udev, and whenever there has any new coming cmd it will update
> the cmd timer or qfull timer. And for some corner case the timers
> are always working only for the ringbuffer's and full queue's newest
> cmd. That's to say the timer won't be fired even there has one cmd
> stuck for a very long time and the deadline is reached.
>
> This fix will keep the cmd/qfull timers to be pended for the oldest
> cmd in ringbuffer and full queue, and will update them with the next
> cmd's deadline only when the old cmd's deadline is reached or removed
> from the ringbuffer and full queue.
>
> Signed-off-by: Xiubo Li
> ---
>
> Changed in V3:
> - Add inflight queue support, thanks Mike.
>
> Changed in V2:
> - Add qfull timer
> - Addressed all the comments from Mike, thanks.
>
> drivers/target/target_core_user.c | 89
> ++-
> 1 file changed, 70 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/target/target_core_user.c
> b/drivers/target/target_core_user.c
> index 9cd404a..00ed7bb 100644
> --- a/drivers/target/target_core_user.c
> +++ b/drivers/target/target_core_user.c
> @@ -148,7 +148,7 @@ struct tcmu_dev {
> size_t ring_size;
>
> struct mutex cmdr_lock;
> - struct list_head cmdr_queue;
> + struct list_head cmdr_qfull_queue;
We can just rename to qfull_queue and inflight_queue and then rename the
function names you modified below.
>
> uint32_t dbi_max;
> uint32_t dbi_thresh;
> @@ -159,6 +159,7 @@ struct tcmu_dev {
>
> struct timer_list cmd_timer;
> unsigned int cmd_time_out;
> + struct list_head cmdr_inflight_queue;
>
> struct timer_list qfull_timer;
> int qfull_time_out;
> @@ -192,6 +193,7 @@ struct tcmu_cmd {
> unsigned long deadline;
>
> #define TCMU_CMD_BIT_EXPIRED 0
> +#define TCMU_CMD_BIT_INFLIGHT 1
> unsigned long flags;
> };
> /*
> @@ -915,11 +917,13 @@ static int tcmu_setup_cmd_timer(struct tcmu_cmd
> *tcmu_cmd, unsigned int tmo,
> return 0;
>
> tcmu_cmd->deadline = round_jiffies_up(jiffies + msecs_to_jiffies(tmo));
> - mod_timer(timer, tcmu_cmd->deadline);
> + if (!timer_pending(timer))
> + mod_timer(timer, tcmu_cmd->deadline);
> +
> return 0;
> }
>
> -static int add_to_cmdr_queue(struct tcmu_cmd *tcmu_cmd)
> +static int add_to_cmdr_qfull_queue(struct tcmu_cmd *tcmu_cmd)
> {
> struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
> unsigned int tmo;
> @@ -942,7 +946,7 @@ static int add_to_cmdr_queue(struct tcmu_cmd *tcmu_cmd)
> if (ret)
> return ret;
>
> - list_add_tail(&tcmu_cmd->cmdr_queue_entry, &udev->cmdr_queue);
> + list_add_tail(&tcmu_cmd->cmdr_queue_entry, &udev->cmdr_qfull_queue);
> pr_debug("adding cmd %u on dev %s to ring space wait queue\n",
>tcmu_cmd->cmd_id, udev->name);
> return 0;
> @@ -999,7 +1003,7 @@ static sense_reason_t queue_cmd_ring(struct tcmu_cmd
> *tcmu_cmd, int *scsi_err)
> base_command_size = tcmu_cmd_get_base_cmd_size(tcmu_cmd->dbi_cnt);
> command_size = tcmu_cmd_get_cmd_size(tcmu_cmd, base_command_size);
>
> - if (!list_empty(&udev->cmdr_queue))
> + if (!list_empty(&udev->cmdr_qfull_queue))
> goto queue;
>
> mb = udev->mb_addr;
> @@ -1096,13 +1100,16 @@ static sense_reason_t queue_cmd_ring(struct tcmu_cmd
> *tcmu_cmd, int *scsi_err)
> UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
> tcmu_flush_dcache_range(mb, sizeof(*mb));
>
> + list_add_tail(&tcmu_cmd->cmdr_queue_entry, &udev->cmdr_inflight_queue);
> + set_bit(TCMU_CMD_BIT_INFLIGHT, &tcmu_cmd->flags);
> +
> /* TODO: only if FLUSH and FUA? */
> uio_event_notify(&udev->uio_info);
>
> return 0;
>
> queue:
> - if (add_to_cmdr_queue(tcmu_cmd)) {
> + if (add_to_cmdr_qfull_queue(tcmu_cmd)) {
> *scsi_err = TCM_OUT_OF_RESOURCES;
> return -1;
> }
> @@ -1194,9 +1201,22 @@ static void tcmu_handle_completion(struct tcmu_cmd
> *cmd, struct tcmu_cmd_entry *
> tcmu_free_cmd(cmd);
> }
>
> +static unsigned long tcmu_get_next_deadline(struct list_head *queue)
> +{
> + struct tcmu_cmd *tcmu_cmd, *tmp_cmd;
> +
> + list_for_each_entry_safe(tcmu_cmd, tmp_cmd, queue, cmdr_queue_entry) {
> + if (!time_after(jiffies, tcmu_cmd->deadline))
> + return tcmu_cmd->deadline;
> + }
> +
> + return 0;
I think you can just pass the timer to this function, rename it to
tcmu_set_next_deadline() and move the mod/del_timer calls in here too
since it is always the same behavior.
> +}
> +
> static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
> {
> struct tcmu_mailbox *mb;
> + struct tcmu_cmd *cmd;
> int handled = 0;
>
> if (test_bit(TCMU_DEV
