Re: [Xen-devel] [PATCH 5/6] xenbus: process be_watch events in xenwatch multithreading

[Juergen Gross <jgross@suse.com>]

On 19/09/18 14:27, Dongli Zhang wrote:
> Hi Juergen,
> 
> On 09/19/2018 04:01 PM, Juergen Gross wrote:
>> On 19/09/18 08:15, Dongli Zhang wrote:
>>> Hi Juergen,
>>>
>>> On 09/14/2018 10:44 PM, Juergen Gross wrote:
>>>> On 14/09/18 16:29, Dongli Zhang wrote:
>>>>> Hi Juergen,
>>>>>
>>>>> On 09/14/2018 10:26 PM, Juergen Gross wrote:
>>>>>> On 14/09/18 16:18, Dongli Zhang wrote:
>>>>>>> Hi Juergen,
>>>>>>>
>>>>>>> On 09/14/2018 05:12 PM, Juergen Gross wrote:
>>>>>>>> On 14/09/18 09:34, Dongli Zhang wrote:
>>>>>>>>> This is the 5th patch of a (6-patch) patch set.
>>>>>>>>>
>>>>>>>>> With this patch, watch event in relative path pattern
>>>>>>>>> 'backend/<pvdev>/<domid>i/...' can be processed in per-domU xenwatch
>>>>>>>>
>>>>>>>> superfluous "i" ----------^
>>>>>>>>
>>>>>>>>> thread.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com>
>>>>>>>>> ---
>>>>>>>>>  drivers/xen/xenbus/xenbus_probe.c         |  2 +-
>>>>>>>>>  drivers/xen/xenbus/xenbus_probe_backend.c | 32 +++++++++++++++++++++++++++++++
>>>>>>>>>  include/xen/xenbus.h                      |  2 ++
>>>>>>>>>  3 files changed, 35 insertions(+), 1 deletion(-)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/xen/xenbus/xenbus_probe.c b/drivers/xen/xenbus/xenbus_probe.c
>>>>>>>>> index ba0644c..aa1b15a 100644
>>>>>>>>> --- a/drivers/xen/xenbus/xenbus_probe.c
>>>>>>>>> +++ b/drivers/xen/xenbus/xenbus_probe.c
>>>>>>>>> @@ -552,7 +552,7 @@ int xenbus_probe_devices(struct xen_bus_type *bus)
>>>>>>>>>  }
>>>>>>>>>  EXPORT_SYMBOL_GPL(xenbus_probe_devices);
>>>>>>>>>  
>>>>>>>>> -static unsigned int char_count(const char *str, char c)
>>>>>>>>> +unsigned int char_count(const char *str, char c)
>>>>>>>>
>>>>>>>> Please change the name of the function when making it globally
>>>>>>>> visible, e.g. by prefixing "xenbus_".
>>>>>>>>
>>>>>>>> Generally I think you don't need to use it below.
>>>>>>>>
>>>>>>>>>  {
>>>>>>>>>  	unsigned int i, ret = 0;
>>>>>>>>>  
>>>>>>>>> diff --git a/drivers/xen/xenbus/xenbus_probe_backend.c b/drivers/xen/xenbus/xenbus_probe_backend.c
>>>>>>>>> index b0bed4f..50df86a 100644
>>>>>>>>> --- a/drivers/xen/xenbus/xenbus_probe_backend.c
>>>>>>>>> +++ b/drivers/xen/xenbus/xenbus_probe_backend.c
>>>>>>>>> @@ -211,9 +211,41 @@ static void backend_changed(struct xenbus_watch *watch,
>>>>>>>>>  	xenbus_dev_changed(path, &xenbus_backend);
>>>>>>>>>  }
>>>>>>>>>  
>>>>>>>>> +static domid_t path_to_domid(const char *path)
>>>>>>>>> +{
>>>>>>>>> +	const char *p = path;
>>>>>>>>> +	domid_t domid = 0;
>>>>>>>>> +
>>>>>>>>> +	while (*p) {
>>>>>>>>> +		if (*p < '0' || *p > '9')
>>>>>>>>> +			break;
>>>>>>>>> +		domid = (domid << 3) + (domid << 1) + (*p - '0');
>>>>>>>>
>>>>>>>> reinventing atoi()?
>>>>>>>>
>>>>>>>> Please don't do that. kstrtou16() seems to be a perfect fit.
>>>>>>>
>>>>>>> I did use kstrtou*() in the early prototype and realized kstrtou16() returns 0
>>>>>>> if the input string contains non-numerical characters.
>>>>>>>
>>>>>>> E.g., the example of input can be "1/0/state", where 1 is fotherend_id
>>>>>>> (frontend_id) and 0 is handle.
>>>>>>>
>>>>>>> When "1/0/state" is used at input, kstrtou16() returns 0 (returned integer) and
>>>>>>> -22 (error).
>>>>>>
>>>>>> Aah, okay. Then simple_strtoul()?
>>>>>
>>>>> I did consider simple_strtoul() initially. Unfortunately, it is obsolete (below
>>>>> line 81). AFAIR, the patch would not be able to pass the check_patch script when
>>>>> this function is used.
>>>>
>>>> Better use that than open coding a new instance of it.
>>>>
>>>> Another variant would be to use sscanf() or similar. Then you could even
>>>> drop using strchr() by adding that in the format string:
>>>>
>>>> return (sscanf(path, "%*u/%u/", &domid) == 1) ? domid : DOMID_SELF;
>>>
>>> I recall what was happened.
>>>
>>> Suppose one sample of path is "backend/vif/19/3/state". (we would like to obtain
>>> domid=19)
>>>
>>> Initially I would like to use sscanf(path, "backend/%*[a-z]/%hu/%*u") to obtain
>>> the domid from xenstore path in one call.
>>>
>>> Unfortunately, unlike userspace sscanf(), the version in linux kernel does not
>>> support '[' so that I would not be able to use "%*[a-z]" in sscanf() in linux
>>> kernel.
>>
>> That is not correct. It doesn't support ranges in [], but it is
>> perfectly fine to use %[^/]. This requires a temporary buffer, as
>> %*[ isn't supported.
>>
>> Why don't you use:
>>
>> char temp[16];
>>
>> ...
>>
>> /* kernel sscanf() %[] doesn't support '*' modifier and needs length. */
>> sscanf(path, "backend/%15[^/]%hu/%*u", temp, &domid)
>>
> 
> One '\' between "%15[^/]" and "%hu" is missing.
> 
> We should use "backend/%15[^/]/%hu/%*u" instead.

Yes.

> Seems this is supported since commit f9310b2f9a19b7f16c7b1c1558f8b649b9b933c1.
> Only tag since 4.6 support this feature.
> 
> I should avoid using old ubuntu 4.4.0 kernel to test such features the next time :(

Right. I suggest to always use a very recent upstream kernel for
testing.

> 
> Thank you very much for your help!
> 
> As the "devicetype[32]" in struct xenbus_device_id is of size 32, should I use
> temp[32] instead of temp[16]?

Seems to be a good idea.


Juergen