[Ovmsdev] Transition to new esp-idf
Mark Webb-Johnson
mark at webb-johnson.net
Mon Jan 22 13:04:21 HKT 2018
Ok, understood. I thought it was running under console task.
I guess the issue is primarily SCP? A general command output should be reasonably small, but a file could by MB in size? What happens if the send_mbuf is full?
Regards, Mark.
> On 22 Jan 2018, at 12:53 PM, Stephen Casner <casner at acm.org> wrote:
>
> Umm, Mark, we're not running a separate task. If we were, this would
> not be a problem. With one task, sleeping would do no good because
> that would be causing Mongoose to sleep. The only way we can save the
> state of the "user" code that is doing the output would be to recurse
> into Mongoose so it can actually send, but that is likely to lead to a
> stack overflow due to repeated recursion. After pondering many ideas,
> that's why I made the change to mg_send.
>
> It is still not difficult to use one of several options make a
> solution that allows SSH to send immediately while MQTT does not, but
> it's just that some mods to mongoose.c will be required.
>
> -- Steve
>
>
> On Mon, 22 Jan 2018, Mark Webb-Johnson wrote:
>
>>> Perhaps check add a loop to check if ctx->send_mbuf.len and ctx->send_mbuf.size to see how much space is free. If not enough, then sleep the task (10ms or 100ms?) and try again? Or use MG_EV_SEND to set a semaphore, and pickup on that in the SendCallback? Rely on the fact that mongoose is running in a separate task and will empty the buffer when it can.
>>
>> To be hopefully clear(er):
>>
>> For event driven systems sending a big file, the usual approach is to send a block, wait for SENT callback, then send the next block. Repeat until no more file left. That approach minimises the buffer usage.
>>
>> We are trying to shoe-horn SSH into this event driven system, but the wolfssh system is expecting a normal blocking API.
>>
>> But, we are running in a separate task, so with a semaphore / poll we can convert the events into a blocking API.
>>
>> Two approaches I can see:
>>
>> Simple is to just check if ctx->send_mbuf has enough space. If not, sleep for a while, and check again. Rely on the mongoose task emptying the buffer.
>>
>> More complex is to use the mongoose MG_EV_SEND callback (which signifies that some data has been sent), and a semaphore to signal data has been sent. The OvmsSSH::EventHandler and SendCallback could then use that to co-ordinate and avoid sleeping. This is the preferred approach.
>>
>> Perhaps this is general enough to be put into a library? An object that could be used to encapsulate the semaphore (initialised to indicate data has been sent), method to indicate data has been sent, and method to wait for that indication. I have a similar problem (although the reverse - receive rather than transmit) in ovms_ota now, and perhaps a generic solution could solve both our problems?
>>
>> Regards, Mark.
>>
>>> On 22 Jan 2018, at 12:34 PM, Mark Webb-Johnson <mark at webb-johnson.net> wrote:
>>>
>>> It doesn’t seem as if there is a good solution. I can see two approaches:
>>>
>>> Use a MG_F_USER_? flag to mean ‘immediate write’ and extend mg_send to honour that.
>>>
>>> Add a separate mg_flush() call (used after mg_send) to flush the fd.
>>>
>>> That static function is going to be a pain to workaround. Perhaps a #include for our C code in mongoose.c?
>>>
>>> All of this is going to be fighting the event-driven mechanism of Mongoose. Is there another way of doing this?
>>>
>>> For console_ssh, I think this is where it is:
>>>
>>> int SendCallback(WOLFSSH* ssh, void* data, uint32_t size, void* ctx)
>>> {
>>> mg_send((mg_connection*)ctx, (char*)data, size);
>>> return size;
>>> }
>>>
>>> Perhaps check add a loop to check if ctx->send_mbuf.len and ctx->send_mbuf.size to see how much space is free. If not enough, then sleep the task (10ms or 100ms?) and try again? Or use MG_EV_SEND to set a semaphore, and pickup on that in the SendCallback? Rely on the fact that mongoose is running in a separate task and will empty the buffer when it can.
>>>
>>> Regards, Mark.
>>>
>>>> On 22 Jan 2018, at 3:17 AM, Stephen Casner <casner at acm.org <mailto:casner at acm.org>> wrote:
>>>>
>>>> Mark,
>>>>
>>>> Well, in turn, I'm sorry for making an API change that was driving you
>>>> crazy. It would have been smarter to add it as a new function even
>>>> though that would be duplicating more code.
>>>>
>>>> As the code currently stands, telnet and SSH will work so long as no
>>>> operation does more contiguous output than the amount of available
>>>> free memory can hold, otherwise an out-of-memory crash will occur. I
>>>> don't know if we consider that an acceptable risk. Maybe with v3.1
>>>> hardware it will be.
>>>>
>>>> Your suggestion to put a new funtion into a separate module is a fine
>>>> idea, but that function needs to access some functions in mongoose.c
>>>> that are scoped static. That means we can't entirely avoid modifying
>>>> mongoose.
>>>>
>>>> -- Steve
>>>>
>>>> On Fri, 19 Jan 2018, Mark Webb-Johnson wrote:
>>>>
>>>>> Oops. Sorry. That change broke MQTT. I couldn’t understand what was going on (as mg_send was sending immediately). MQTT works like this:
>>>>>
>>>>> void mg_mqtt_publish(struct mg_connection *nc, const char *topic,
>>>>> uint16_t message_id, int flags, const void *data,
>>>>> size_t len) {
>>>>> size_t old_len = nc->send_mbuf.len;
>>>>>
>>>>> uint16_t topic_len = htons((uint16_t) strlen(topic));
>>>>> uint16_t message_id_net = htons(message_id);
>>>>>
>>>>> mg_send(nc, &topic_len, 2);
>>>>> mg_send(nc, topic, strlen(topic));
>>>>> if (MG_MQTT_GET_QOS(flags) > 0) {
>>>>> mg_send(nc, &message_id_net, 2);
>>>>> }
>>>>> mg_send(nc, data, len);
>>>>>
>>>>> mg_mqtt_prepend_header(nc, MG_MQTT_CMD_PUBLISH, flags,
>>>>> nc->send_mbuf.len - old_len);
>>>>> }
>>>>>
>>>>> It uses mg_send a bunch of times, then goes back and modifies the send_mbuf by inserting a header, then finishes so that the actual transmission can occur. Seems a really dumb way to do it, but such is life.
>>>>>
>>>>> It was driving me crazy last night, so in the end I just updated mongoose this morning and hey! everything worked. Now I know why :-(
>>>>>
>>>>> I see that mg_send_dns_query() does the same (it calls mg_dns_insert_header, which then calls mbuf_insert). Making mg_send transmit immediately would break that as well.
>>>>>
>>>>> How about introducing a new mg_send_now() that calls mg_send() then sends the data immediately? Perhaps it could be a separate .h/.c file mongoose_extensions to avoid the change getting overwritten?
>>>>>
>>>>> Regards, Mark.
>>>>>
>>>>>> On 19 Jan 2018, at 2:36 PM, Stephen Casner <casner at acm.org <mailto:casner at acm.org>> wrote:
>>>>>>
>>>>>> Mark,
>>>>>>
>>>>>> The update of Mongoose to v6.10 removed the change I had made so that
>>>>>> the mg_send() call would transmit on the network immediately if the
>>>>>> socket was ready. I needed to make that change because we would
>>>>>> otherwise run out of RAM with SSH because mg_send() would just buffer
>>>>>> everything until the next poll.
>>>>>>
>>>>>> -- Steve
>>>>>>
>>>>>> On Fri, 19 Jan 2018, Mark Webb-Johnson wrote:
>>>>>>
>>>>>>> I re-worked the ovms_server_* framework, and v2 implementation, to use MONGOOSE.
>>>>>>>
>>>>>>> It seems to be _basically_ working. It can connect/disconnect/etc. Some slight memory saving, but standardising the networking throughout on mongoose should simplify things.
>>>>>>>
>>>>>>> I am seeing problems with transmitting the FEATURES and PARAMETERS sometimes - particularly in low memory situations. I’m still trying to find out why.
>>>>>>>
>>>>>>> Regards, Mark.
>>>>>>>
>>>>>>>> On 17 Jan 2018, at 8:33 AM, Mark Webb-Johnson <mark at webb-johnson.net <mailto:mark at webb-johnson.net>> wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>> This is the issue Michael pointed out. The 'server response is incomplete’ problem with select(). Apologies for this; I am not sure why I didn’t notice it before.
>>>>>>>>
>>>>>>>> Gory details are here:
>>>>>>>>
>>>>>>>> https://github.com/espressif/esp-idf/issues/1510 <https://github.com/espressif/esp-idf/issues/1510> <https://github.com/espressif/esp-idf/issues/1510 <https://github.com/espressif/esp-idf/issues/1510>>
>>>>>>>>
>>>>>>>> I think Espressif implemented requirement this in a bizarre way, likely to break compatibility, but such is life. They did point it out as a ‘breaking change’ (at the bottom of the release notes for 3.0b1):
>>>>>>>>
>>>>>>>> https://github.com/espressif/esp-idf/releases/tag/v3.0-rc1 <https://github.com/espressif/esp-idf/releases/tag/v3.0-rc1> <https://github.com/espressif/esp-idf/releases/tag/v3.0-rc1 <https://github.com/espressif/esp-idf/releases/tag/v3.0-rc1>>
>>>>>>>>
>>>>>>>> LWIP socket file descriptors now take higher numeric values (via the LWIP LWIP_SOCKET_OFFSET macro). BSD sockets code should mostly work as expected (and, new in V3.0, some standard POSIX functions can now be used with sockets). However any code which assumes a socket file descriptor is always a low numbered integer may need modifying to account for LWIP_SOCKET_OFFSET.
>>>>>>>>
>>>>>>>> It sure broke us.
>>>>>>>>
>>>>>>>> I’ve made a one-line workaround fix (to ovms_buffer.cpp), and ovms server v2 connections are working again for me. That is committed and pushed already.
>>>>>>>>
>>>>>>>> It is kind of messy to have all these different networking implementations in our code base; I intend to move ovms_server_* to mongoose networking over the next few days. That will mean we won’t need a separate task/stack for server connections, and should save us 7KB internal RAM for each connection. Also ovms_ota. But that will have to wait, as I need to get the hardware complete first (some issues with 1.8v vs 3.3v logic on VDD_SDIO of the wrover module and some of our GPIOs), and that is top priority.
>>>>>>>>
>>>>>>>> Regards, Mark.
>>>>>>>>
>>>>>>>>> On 17 Jan 2018, at 7:05 AM, Greg D. <gregd2350 at gmail.com <mailto:gregd2350 at gmail.com> <mailto:gregd2350 at gmail.com <mailto:gregd2350 at gmail.com>>> wrote:
>>>>>>>>>
>>>>>>>>> But, I'm not getting much love out of the v2 server. The connection doesn't appear to be working - "server response is incomplete". Same error whether on wifi or modem.
>>>>>>>>
>>>>>>>
>>>>>> _______________________________________________
>>>>>> OvmsDev mailing list
>>>>>> OvmsDev at lists.teslaclub.hk <mailto:OvmsDev at lists.teslaclub.hk>
>>>>>> http://lists.teslaclub.hk/mailman/listinfo/ovmsdev
>>>>>
>>>> _______________________________________________
>>>> OvmsDev mailing list
>>>> OvmsDev at lists.teslaclub.hk <mailto:OvmsDev at lists.teslaclub.hk>
>>>> http://lists.teslaclub.hk/mailman/listinfo/ovmsdev
>>>
>>> _______________________________________________
>>> OvmsDev mailing list
>>> OvmsDev at lists.teslaclub.hk
>>> http://lists.teslaclub.hk/mailman/listinfo/ovmsdev
>>
> _______________________________________________
> OvmsDev mailing list
> OvmsDev at lists.teslaclub.hk
> http://lists.teslaclub.hk/mailman/listinfo/ovmsdev
More information about the OvmsDev
mailing list