Re: [PATCHv2 1/4] nvme: trace nvme queue identifiers

From: Christoph Hellwig
Date: Fri Jun 29 2018 - 03:57:31 EST


On Thu, Jun 28, 2018 at 03:49:53PM -0600, Keith Busch wrote:
> We can not match a command to its completion based on the command id
> alone. We need the submitting queue identifier to pair the completion,
> so this patch adds that to the trace buffer.
>
> This patch is also collapsing the admin and io submission traces into
> a single one since we don't need to duplicate this and creating
> unnecessary code branches.
>
> Signed-off-by: Keith Busch <keith.busch@xxxxxxxxx>
> ---
> drivers/nvme/host/core.c | 7 +++----
> drivers/nvme/host/trace.h | 41 ++++++++++-------------------------------
> 2 files changed, 13 insertions(+), 35 deletions(-)
>
> diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> index 46df030b2c3f..398a5fb26603 100644
> --- a/drivers/nvme/host/core.c
> +++ b/drivers/nvme/host/core.c
> @@ -97,6 +97,8 @@ static dev_t nvme_chr_devt;
> static struct class *nvme_class;
> static struct class *nvme_subsys_class;
>
> +static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
> + unsigned nsid);

Huh?

> static void nvme_ns_remove(struct nvme_ns *ns);
> static int nvme_revalidate_disk(struct gendisk *disk);
> static void nvme_put_subsystem(struct nvme_subsystem *subsys);
> @@ -652,10 +654,7 @@ blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req,
> }
>
> cmd->common.command_id = req->tag;
> - if (ns)
> - trace_nvme_setup_nvm_cmd(req->q->id, cmd);
> - else
> - trace_nvme_setup_admin_cmd(cmd);
> + trace_nvme_setup_nvm_cmd(req, cmd);
> return ret;
> }
> EXPORT_SYMBOL_GPL(nvme_setup_cmd);
> diff --git a/drivers/nvme/host/trace.h b/drivers/nvme/host/trace.h
> index 01390f0e1671..42c99445dd96 100644
> --- a/drivers/nvme/host/trace.h
> +++ b/drivers/nvme/host/trace.h
> @@ -75,34 +75,9 @@ const char *nvme_trace_parse_nvm_cmd(struct trace_seq *p, u8 opcode,
> #define __parse_nvme_cmd(opcode, cdw10) \
> nvme_trace_parse_nvm_cmd(p, opcode, cdw10)
>
> -TRACE_EVENT(nvme_setup_admin_cmd,
> - TP_PROTO(struct nvme_command *cmd),
> - TP_ARGS(cmd),
> - TP_STRUCT__entry(
> - __field(u8, opcode)
> - __field(u8, flags)
> - __field(u16, cid)
> - __field(u64, metadata)
> - __array(u8, cdw10, 24)
> - ),
> - TP_fast_assign(
> - __entry->opcode = cmd->common.opcode;
> - __entry->flags = cmd->common.flags;
> - __entry->cid = cmd->common.command_id;
> - __entry->metadata = le64_to_cpu(cmd->common.metadata);
> - memcpy(__entry->cdw10, cmd->common.cdw10,
> - sizeof(__entry->cdw10));
> - ),
> - TP_printk(" cmdid=%u, flags=0x%x, meta=0x%llx, cmd=(%s %s)",
> - __entry->cid, __entry->flags, __entry->metadata,
> - show_admin_opcode_name(__entry->opcode),
> - __parse_nvme_admin_cmd(__entry->opcode, __entry->cdw10))
> -);
> -
> -
> TRACE_EVENT(nvme_setup_nvm_cmd,
> - TP_PROTO(int qid, struct nvme_command *cmd),
> - TP_ARGS(qid, cmd),
> + TP_PROTO(struct request *req, struct nvme_command *cmd),
> + TP_ARGS(req, cmd),
> TP_STRUCT__entry(
> __field(int, qid)
> __field(u8, opcode)
> @@ -113,7 +88,7 @@ TRACE_EVENT(nvme_setup_nvm_cmd,
> __array(u8, cdw10, 24)
> ),
> TP_fast_assign(
> - __entry->qid = qid;
> + __entry->qid = blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req)) + !!req->rq_disk;

This looks pretty ugly. I'd much prefer:

if (req->rq_disk)
__entry->qid = blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req));
else
__entry->qid = 0;

Also the indentation in the existing code seems weird and not based
on tabs. We should fix that up.

> + __entry->qid ?
> + show_opcode_name(__entry->opcode) :
> + show_admin_opcode_name(__entry->opcode),
> + __entry->qid ?
> + __parse_nvme_cmd(__entry->opcode, __entry->cdw10) :
> + __parse_nvme_admin_cmd(__entry->opcode, __entry->cdw10))
> );

Can we hide this? E.g. pass the qid to show_opcode_name and then
implement show_opcode_name as:

#define show_opcode_name(qid, opcode) \
(qid ? show_admin_opcode_name(opcode) : show_nvm_opcode_name(opcode))

Same for __parse_nvme_cmd.

> + __entry->qid = blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req)) + !!req->rq_disk;

Same as above.

In fact I guess we should just have a helper:

static inline u16 nvme_req_to_qid(struct request *req)
{
if (!req->rq_disk)
return 0;
return blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req));
}