[permalink] [raw]

Subject: Re: [RFC 02/11] drivers: Add HTE subsystem

Hi Jonathan,

I got some time to implement RFC version 2 while doing so I have a follow up comment

inline regarding clock source comment of yours.

Best Regards,

Dipen Patel

On 8/1/21 9:13 AM, Jonathan Cameron wrote:
> On Tue, 27 Jul 2021 21:38:45 -0700
> Dipen Patel <[email protected]> wrote:
>
>> On 7/4/21 1:15 PM, Jonathan Cameron wrote:
>>> On Fri, 25 Jun 2021 16:55:23 -0700
>>> Dipen Patel <[email protected]> wrote:
>>>
>>>> Some devices can timestamp system lines/signals/Buses in real-time
>>>> using the hardware counter or other hardware means which can give
>>>> finer granularity and help avoid jitter introduced by software means
>>>> of timestamping. To utilize such functionality there has to be
>>>> framework where such devices can register themselves as producers or
>>>> providers so that the consumers or clients devices can request specific
>>>> line from the providers. This patch introduces such subsystem as
>>>> hardware timestamping engine (HTE).
>>>>
>>>> It provides below APIs for the provider:
>>>> - hte_register_chip() -- To register the HTE chip.
>>>> - hte_unregister_chip() -- To unregister the HTE chip.
>>>> - hte_push_ts_ns_atomic() -- To push timestamp data into HTE subsystem.
>>>>
>>>> It provides below APIs for the consumer:
>>>> - of_hte_request_ts() -- To request timestamp functionality.
>>>> - devm_of_hte_request_ts() -- Managed version of the above.
>>>> - hte_req_ts_by_dt_node() -- To request timestamp functionality by
>>>> using HTE provider dt node.
>>>> - devm_hte_release_ts() -- The managed version to release timestamp
>>>> functionality and associated resources.
>>>> - hte_retrieve_ts_ns() -- To retrieve timestamps.
>>>> - hte_retrieve_ts_ns_wait() -- Same as above but blocking version.
>>>> - hte_enable_ts() -- To disable timestamp functionality.
>>>> - hte_disable_ts() -- To enable timestamp functionality.
>>>> - hte_available_ts() -- To query available timestamp data.
>>>> - hte_release_ts() -- To release timestamp functionality and its
>>>> associated resources.
>>>> - hte_get_clk_src_info() -- To query clock source information from
>>>> the provider
>>>>
>>>> It provides centralized software buffer management per requested id to
>>>> store the timestamp data for the consumers as below:
>>>> - hte_set_buf_len() -- To set the buffer length.
>>>> - hte_get_buf_len() -- To get the buffer length.
>>>> - hte_set_buf_watermark() -- To set the software threshold/watermark.
>>>> - hte_get_buf_watermark() -- To get the software threshold/watermark.
>>>>
>>>> The detail about parameters and API usage are described in each
>>>> functions definitions in drivers/hte/hte.c file.
>>>>
>>>> The patch adds compilation support in Makefile and menu options in
>>>> Kconfig.
>>>>
>>>> Signed-off-by: Dipen Patel <[email protected]>
>>> Hi Dipen, this isn't a particularly thorough review as I'm still getting my head
>>> around what this is doing + it is an RFC :)
>> Thanks for the review comments. My responses inline.
> You are welcome, some follow up responses inline.
> I've tried to crop this down a bit so only kept the bits we are discussing.
>
>>>> +
>>>> +static int hte_ts_dis_en_common(struct hte_ts_desc *desc, bool en)
>>>> +{
>>>> + u32 ts_id;
>>>> + struct hte_device *gdev;
>>>> + struct hte_ts_info *ei;
>>>> + int ret;
>>>> +
>>>> + if (!desc)
>>>> + return -EINVAL;
>>>> +
>>>> + ei = (struct hte_ts_info *)desc->data_subsys;
>>> As above, no need to cast - though it rather implies the type of data_subsys
>>> should not be void *.
>> desc is public facing structure, I wanted to make subsystem related
>>
>> information opaque that is why I had it void *.
>>
> you can keep it opaque, just have a forwards definition of
> struct hte_ts_desc;
> which just means it is defined somewhere. You can have that in the header with
> the definition hidden away.
>
> It will only need to have a visible complete definition when you dereference it inside
> the the core.
>
> Mind you, I'm suggesting allowing it to be embedded in another structure anyway which
> would require you to have it exposed. Perhaps this desire to keep it opaque is
> a reason to not take that suggestion but it isn't relevant for this one.
>
>
>>>> + */
>>>> +struct hte_ts_desc *devm_of_hte_request_ts(struct device *dev,
>>>> + const char *label,
>>>> + void (*cb)(enum hte_notify n))
>>>> +{
>>>> +
>>>> + struct hte_ts_desc **ptr, *desc;
>>>> +
>>>> + ptr = devres_alloc(__devm_hte_release_ts, sizeof(*ptr), GFP_KERNEL);
>>> Superficially looks like you might get way with just calling dev_add_action_or_reset() in here
>>> and avoid this boilerplate. A lot of cases that looked like this got cleaned up in the
>>> last kernel cycle.
>> I based my patches from linux-next/master. Not sure if that has
>>
>> dev_add_action_or_reset
> typo on my part was meant to be
>
> devm_add_action_or_reset()
>
>>>> +
>>>> +/**
>>>> + * hte_req_ts_by_dt_node() - Request entity to monitor by passing HTE device
>>>> + * node directly, where meaning of the entity is provider specific, for example
>>>> + * lines, signals, GPIOs, buses etc...
>>>> + *
>>>> + * @of_node: HTE provider device node.
>>>> + * @id: entity id to monitor, this id belongs to HTE provider of_node.
>>>> + * @cb: Optional callback to notify.
>>>> + *
>>>> + * Context: Holds mutex lock, can not be called from atomic context.
>>> What mutex and why? If it is one you can check is held even better.
>> ___hte_req_ts holds the mutex lock to serialize multiple consumers
>>
>> requesting same entity.
> Add that detail to the comment.
>
>>>
>>>> + * Returns: ts descriptor on success or error pointers.
>>>> + */
>>>> +struct hte_ts_desc *hte_req_ts_by_dt_node(struct device_node *of_node,
>>>> + unsigned int id,
>>>> + void (*cb)(enum hte_notify n))
>>>> +{
>>>> + struct hte_device *gdev;
>>>> + struct hte_ts_desc *desc;
>>>> + int ret;
>>>> + u32 xlated_id;
>>>> +
>>>> + gdev = of_node_to_htedevice(of_node);
>>>> + if (IS_ERR(gdev))
>>>> + return ERR_PTR(-ENOTSUPP);
>>>> +
>>>> + if (!gdev->chip || !gdev->chip->ops)
>>>> + return ERR_PTR(-ENOTSUPP);
>>>> +
>>>> + desc = kzalloc(sizeof(*desc), GFP_KERNEL);
>>>> + if (!desc) {
>>>> + ret = -ENOMEM;
>>>> + goto out_put_device;
>>>> + }
>>> Pass a desc pointer into this function rather than allocating the structure
>>> in here. That lets the caller embed that structure inside one of it's own
>>> structures if it wants to, resulting in fewer small allocations which is always good.
>>>
>>> It's far from obvious that the caller needs to free desc.
>> Are you suggesting to shift burden of allocation/deallocation (static or dynamic)
>>
>> at client/consumer side?
> It's been a while so I've forgotten how this works, but 'probably' yes...
> If a function creates some sort of record (of fixed known size and type) then
> letting that be passed in + filled in by the function is normally more efficient
> than having an allocation in here. Chances are the consumer will just have
> it embedded in an existing state structure and not need to do any explicit
> allocation / deallocation. Disadvantage is you can't keep it opaque.
>
>>>
>>>> +
>>>> + desc->con_id = id;
>>>> + ret = gdev->chip->xlate(gdev->chip, NULL, desc, &xlated_id);
>>>> + if (ret < 0) {
>>>> + dev_err(gdev->chip->dev,
>>>> + "failed to xlate id: %d\n", id);
>>>> + goto out_free_desc;
>>>> + }
>>>> +
>>>> + ret = ___hte_req_ts(gdev, desc, xlated_id, cb);
>>>> + if (ret < 0) {
>>>> + dev_err(gdev->chip->dev,
>>>> + "failed to request id: %d\n", id);
>>>> + goto out_free_desc;
>>>> + }
>>>> +
>>>> + return desc;
>>>> +
>>>> +out_free_desc:
>>>> + kfree(desc);
>>>> +
>>>> +out_put_device:
>>>> + return ERR_PTR(ret);
>>>> +}
>>>> +EXPORT_SYMBOL_GPL(hte_req_ts_by_dt_node);
>>>> +
>>>> +/**
>>>> + * hte_get_clk_src_info() - Consumer calls this API to query clock source
>>>> + * information of the desc.
>>>> + *
>>>> + * @desc: ts descriptor, same as returned from request API.
>>>> + *
>>>> + * Context: Any context.
>>>> + * Returns: 0 on success else negative error code on failure.
>>>> + */
>>>> +int hte_get_clk_src_info(const struct hte_ts_desc *desc,
>>>> + struct hte_clk_info *ci)
>>>> +{
>>>> + struct hte_chip *chip;
>>>> + struct hte_ts_info *ei;
>>>> +
>>>> + if (!desc || !desc->data_subsys || !ci) {
>>>> + pr_debug("%s:%d\n", __func__, __LINE__);
>>>> + return -EINVAL;
>>>> + }
>>>> +
>>>> + ei = desc->data_subsys;
>>>> + if (!ei || !ei->gdev || !ei->gdev->chip)
>>>> + return -EINVAL;
>>>> +
>>>> + chip = ei->gdev->chip;
>>>> + if (!chip->ops->get_clk_src_info)
>>>> + return -ENOTSUPP;
>>>> +
>>>> + return chip->ops->get_clk_src_info(chip, ci);
>>>> +}
>>>> +EXPORT_SYMBOL_GPL(hte_get_clk_src_info);
>>>> +
>>>> +static inline void hte_add_to_device_list(struct hte_device *gdev)
>>>> +{
>>>> + struct hte_device *prev;
>>> Needs to take an appropriate lock as you may have concurrent calls.
>> There is spin_lock held from register API from where this gets
>> called.
> Great. I'd missed that.
>
>>>
>>>> +
>>>> + if (list_empty(&hte_devices)) {
>>>> + list_add_tail(&gdev->list, &hte_devices);
>>> Needs a comment. I've no idea why you might want to only add it if there were
>>> no other hte_devices already there.
>>>
>>>> + return;
>>>> + }
>>>> +
>>>> + prev = list_last_entry(&hte_devices, struct hte_device, list);
>>> Why woud you do this?
>> Thanks for pointing out. I definitely missed cleaning this up. Now, I will
>>
>> remove this function in next RFC version as one line can be added directly
>>
>> in register API.
>>
>>>
>>>> + list_add_tail(&gdev->list, &hte_devices);
>>>> +}
>>>> +
>>>> +/**
>>>> + * hte_push_ts_ns_atomic() - Used by the provider to push timestamp in nano
>>>> + * seconds i.e data->tsc will be in ns, it is assumed that provider will be
>>>> + * using this API from its ISR or atomic context.
>>>> + *
>>>> + * @chip: The HTE chip, used during the registration.
>>>> + * @xlated_id: entity id understood by both subsystem and provider, usually this
>>>> + * is obtained from xlate callback during request API.
>>>> + * @data: timestamp data.
>>>> + * @n: Size of the data.
>>>> + *
>>>> + * Context: Atomic.
>>>> + * Returns: 0 on success or a negative error code on failure.
>>>> + */
>>>> +int hte_push_ts_ns_atomic(const struct hte_chip *chip, u32 xlated_id,
>>>> + struct hte_ts_data *data, size_t n)
>>>> +{
>>>> + unsigned int ret;
>>>> + bool notify;
>>>> + size_t el_avail;
>>>> + struct hte_ts_buf *buffer;
>>>> + struct hte_ts_info *ei;
>>>> +
>>>> + if (!chip || !data || !chip->gdev)
>>>> + return -EINVAL;
>>>> +
>>>> + if (xlated_id > chip->nlines)
>>>> + return -EINVAL;
>>>> +
>>>> + ei = &chip->gdev->ei[xlated_id];
>>>> +
>>>> + if (!test_bit(HTE_TS_REGISTERED, &ei->flags) ||
>>>> + test_bit(HTE_TS_DISABLE, &ei->flags)) {
>>>> + dev_dbg(chip->dev, "Unknown timestamp push\n");
>>>> + return -EINVAL;
>>>> + }
>>>> +
>>>> + /* timestamp sequence counter, start from 0 */
>>>> + data->seq = ei->seq++;
>>>> +
>>>> + buffer = ei->buf;
>>>> + el_avail = buffer->access->el_available(buffer);
>>>> + ret = buffer->access->store(buffer, data, n);
>>> If we are doing this from the hte core, why is buffer definition in the scope of the
>>> drivers rather than the core? That seems backwards to me.
>> I do not understand this comment. The buffer definition is in scope of hte core
>>
>> as it is the only entity that manages it.
> I think I figured that out later and forgot to come back and edit this comment.
> However...
>
> In that case, why is it an ops function? Don't introduce abstraction
> until you need it. Will be simpler and easier to review if you just
> call those functions directly for now. e.g.
>
> ret = hs_ts_store_to_buf(buffer, data, n);
>
> Chances are you'll never introduce another buffer choice.
> For a long time I thought we'd have both fifo and ring options in IIO
> but it turned out no one really cared. We do have an ops structure, but
> that's because in IIO the buffer interface is used for two things:
> 1) Pushing to a kfifo that is going to userspace.
> 2) Pushing to a callback function owned by a consumer.
> and there is a rather fiddly data demux on the front end to ensure each
> of those only gets the data requested via that path - at least with timestamps
> there is only one type of data!
>
> Hmm, thinking about this raises an interesting question.
> Why do we want a kfifo here at all for HTE? You could
> just call a callback function registered by the consumer of that
> kfifo directly. If that consumer then wants to buffer then of
> course it can, but it not (perhaps it only cares about the latest
> value and will drop the rest) then it can chose not to. Maybe
> it's just gathering stats rather than caring about individual
> timestamps? Probably lots of other things that might happen in
> the consumer that I've not thought of. We need a buffer if
> userspace becomes involved, but here IIRC that's not (yet) true.
>
>>>
>>>> + if (ret != n) {
>>>> + atomic_inc(&ei->dropped_ts);
>>>> + if (ei->cb)
>>>> + ei->cb(HTE_TS_DROPPED);
>>>> + return -ENOMEM;
>>>> + }
> ...
>
>>>
>>>> +
>>>> +/**
>>>> + * struct hte_ts_data - HTE timestamp data.
>>>> + * The provider uses and fills timestamp related details during push_timestamp
>>>> + * API call. The consumer uses during retrieve_timestamp API call.
>>>> + *
>>>> + * @tsc: Timestamp value.
>>>> + * @seq: Sequence counter of the timestamps.
>>>> + * @dir: Direction of the event at the time of timestamp.
>>>> + */
>>>> +struct hte_ts_data {
>>>> + u64 tsc;
>>>> + u64 seq;
>>>> + int dir;
>>>> +};
>>>> +
>>>> +/**
>>>> + * struct hte_clk_info - Clock source info that HTE provider uses.
>>>> + * The provider uses hardware clock as a source to timestamp real time. This
>>>> + * structure presents the clock information to consumers.
>>>> + *
>>>> + * @hz: Clock rate in HZ, for example 1KHz clock = 1000.
>>>> + * @type: Clock type. CLOCK_* types.
>>> So this is something we got a it wrong in IIO. It's much better to define
>>> a subset of clocks that can be potentially used. There are some that make
>>> absolutely no sense and consumers really don't want to have to deal with them.
>> Is there anything I have to change here?
> Yes - specify which clocks would make sense. You might not need to explicitly
> allow only those, but that might also be worthwhile. Otherwise, the chances are
> you'll end up with a bunch of special purpose code in consumers on the basis
> they might get CLOCK_TAI or similar and have to deal with it.
> As for exactly which clocks do make sense, that's one which may take some figuring
> out. Probably REALTIME, MONOTONIC and BOOTTIME depending on whether you care
> what happens when the time of the system gets adjusted, or whether it carries
> on measuring time across suspend. Very application dependent but there are some
> you can definitely rule out. Don't repeat my mistake of leaving it vague
> (which incidentally was a follow up to picking a silly clock to use for timestamps
> before we allowed it to be configured).

I believe your comment is under assumption that providers have choice in selecting

clock source to timestamp in turns clients have it as well. For now, the provider

I have implemented has single clock source and hence I only implemented get_clock*

hook that provider implement and client can retrieve that information. I guess I can

always implement set_clock* hook as well for the future providers which support

multiple clock sources. Please let me if I missed your point.

>>>
>>>> + */
>>>> +struct hte_clk_info {
>>>> + u64 hz;
>>>> + clockid_t type;
>>>> +};
>>>> +
>>>> +/**
>>>> + * HTE subsystem notifications for the consumers.
>>>> + *
>>>> + * @HTE_TS_AVAIL: Timestamps available notification.
>>>> + * @HTE_TS_DROPPED: Timestamps dropped notification.
>>> Something I've missed so far is whether drops are in a kfifo or a ring
>>> fashion. I'm guess that's stated somewhere, but it might be useful to have
>>> it here.
>> Dropped are from kfifo if kfifo does not have space.
> Ok, perhaps expand the comment?
>
> ...
>
>>>
>>>> + *
>>>> + * xlated_id parameter is used to communicate between HTE subsystem and the
>>>> + * providers. It is the same id returned during xlate API call and translated
>>>> + * by the provider. This may be helpful as both subsystem and provider locate
>>>> + * the requested entity in constant time, where entity could be anything from
>>>> + * lines, signals, events, buses etc.. that providers support.
>>>> + */
>>>> +struct hte_ops {
>>>> + int (*request)(struct hte_chip *chip, u32 xlated_id);
>>>> + int (*release)(struct hte_chip *chip, u32 xlated_id);
>>>> + int (*enable)(struct hte_chip *chip, u32 xlated_id);
>>>> + int (*disable)(struct hte_chip *chip, u32 xlated_id);
>>>> + int (*get_clk_src_info)(struct hte_chip *chip,
>>>> + struct hte_clk_info *ci);
>>>> +};
>>>> +
>>>> +/**
>>>> + * struct hte_chip - Abstract HTE chip structure.
>>>> + * @name: functional name of the HTE IP block.
>>>> + * @dev: device providing the HTE.
>>> Unclear naming. Is this the parent device, or one associated with the HTE itself?
>>> I'm guessing today you don't have one associated with the HTE, but it is plausible you
>>> might gain on in future to make it fit nicely in the device model as a function of another
>>> device.
>> This is provider's device, could be &pdev->dev or any dev provider deems fit hence the
>>
>> generic name.
> Ok, for now this works as a name, but I wonder if you will end up growing another
> layer in the device model as would happen for majority of subsystems.
> You may end up doing so when adding support to query the provider via a handle
> in the dt of the consumer. It could probably be avoided, but throwing this into
> a class might make your life easier as you can use more standard infrastructure.
>
>>>
>>>> + * @ops: callbacks for this HTE.
>>>> + * @nlines: number of lines/signals supported by this chip.
>>>> + * @xlate: Callback which translates consumer supplied logical ids to
>>>> + * physical ids, return from 0 for the success and negative for the
>>>> + * failures. It stores (0 to @nlines) in xlated_id parameter for the success.
>>>> + * @of_hte_n_cells: Number of cells used to form the HTE specifier.
>>>> + * @gdev: HTE subsystem abstract device, internal to the HTE subsystem.
>>>> + * @data: chip specific private data.
>>>> + */
>>>> +struct hte_chip {
>>>> + const char *name;
>>>> + struct device *dev;
>>>> + const struct hte_ops *ops;
>>>> + u32 nlines;
>>>> + int (*xlate)(struct hte_chip *gc,
>>>> + const struct of_phandle_args *args,
>>>> + struct hte_ts_desc *desc, u32 *xlated_id);
>>>> + u8 of_hte_n_cells;
>>>> +
>>>> + /* only used internally by the HTE framework */
>>>> + struct hte_device *gdev;
>>>> + void *data;
>>>> +};
> ...
>
> Jonathan

2021-09-26 15:41:07

by Jonathan Cameron

[permalink] [raw]

Subject: Re: [RFC 02/11] drivers: Add HTE subsystem

On Mon, 13 Sep 2021 22:43:02 -0700
Dipen Patel <[email protected]> wrote:

> Hi Jonathan,
>
> I got some time to implement RFC version 2 while doing so I have a follow up comment
>
> inline regarding clock source comment of yours.
>
> Best Regards,
>
> Dipen Patel
>
...

> >>>> +/**
> >>>> + * struct hte_clk_info - Clock source info that HTE provider uses.
> >>>> + * The provider uses hardware clock as a source to timestamp real time. This
> >>>> + * structure presents the clock information to consumers.
> >>>> + *
> >>>> + * @hz: Clock rate in HZ, for example 1KHz clock = 1000.
> >>>> + * @type: Clock type. CLOCK_* types.
> >>> So this is something we got a it wrong in IIO. It's much better to define
> >>> a subset of clocks that can be potentially used. There are some that make
> >>> absolutely no sense and consumers really don't want to have to deal with them.
> >> Is there anything I have to change here?
> > Yes - specify which clocks would make sense. You might not need to explicitly
> > allow only those, but that might also be worthwhile. Otherwise, the chances are
> > you'll end up with a bunch of special purpose code in consumers on the basis
> > they might get CLOCK_TAI or similar and have to deal with it.
> > As for exactly which clocks do make sense, that's one which may take some figuring
> > out. Probably REALTIME, MONOTONIC and BOOTTIME depending on whether you care
> > what happens when the time of the system gets adjusted, or whether it carries
> > on measuring time across suspend. Very application dependent but there are some
> > you can definitely rule out. Don't repeat my mistake of leaving it vague
> > (which incidentally was a follow up to picking a silly clock to use for timestamps
> > before we allowed it to be configured).
>
> I believe your comment is under assumption that providers have choice in selecting
>
> clock source to timestamp in turns clients have it as well. For now, the provider
>
> I have implemented has single clock source and hence I only implemented get_clock*
>
> hook that provider implement and client can retrieve that information. I guess I can
>
> always implement set_clock* hook as well for the future providers which support
>
> multiple clock sources. Please let me if I missed your point.

I'll be honest I can't really remember :( too many sleeps.

Sorry - if it is still relevant perhaps it'll come back to me on v2.

Thanks,

Jonathan