Re: [PATCH v3 3/3] char drivers: ramoops debugfs entry

[Marco Stornelli <marco.stornelli@gmail.com>]

Il 29/07/2011 02:15, Sergiu Iordache ha scritto:
> On Mon, Jul 11, 2011 at 11:41 PM, Marco Stornelli
> <marco.stornelli@gmail.com>  wrote:
>> 2011/7/11 Sergiu Iordache<sergiu@chromium.org>:
>>> On Thu, Jul 7, 2011 at 11:43 PM, Marco Stornelli
>>> <marco.stornelli@gmail.com>  wrote:
>>>> 2011/7/8 Greg KH<greg@kroah.com>:
>>>>> On Thu, Jul 07, 2011 at 04:27:27PM -0700, Andrew Morton wrote:
>>>>>> On Thu, 7 Jul 2011 16:16:43 -0700
>>>>>> Sergiu Iordache<sergiu@google.com>  wrote:
>>>>>>
>>>>>>> Ramoops currently dumps the log of a panic/oops in a memory area which
>>>>>>> is known not to be overwritten on restart (for example 1MB starting at
>>>>>>> 15MB). The way it works is by dividing the memory area in records of a
>>>>>>> set size (fixed at 4K before my patches, configurable after) and by
>>>>>>> dumping a record there for each oops/panic. The problem is that right
>>>>>>> now you have to access that memory area through other means, such as
>>>>>>> /dev/mem, which is not always possible.
>>>>>>>
>>>>>>> What my patch did was to add a debugfs entry which returns a valid
>>>>>>> record each time (a single dump done by ramoops). The first call
>>>>>>> returns the first dump. The first call after the last valid dump
>>>>>>> returns an empty buffer. .
>>>>>>
>>>>>> Please fully describe this "record" in the v2 patch changelog.  We'll
>>>>>> want to review it for endianness, 32/64-bit compat issues,
>>>>>> maintainability, extensibility, etc.
>>>>>>
>>>>>>> After it has returned nothing, the next
>>>>>>> calls return records from the start again.
>>>>>>
>>>>>> That sounds a bit weird.  One would expect it to keep returning zero,
>>>>>> requiring userspace to lseek or close/open.
>>>>>>
>>>>>>> The validity of a dump is
>>>>>>> checked by looking after the header. Any comments on this approach are
>>>>>>> welcome.
>>>>>>>
>>>>>>> Changing the entry from debugfs to sysfs wouldn't be a problem. If
>>>>>>> sysfs is a valid solution I'll come with a patch that updates the
>>>>>>> documentation as well along with the sysfs entry.
>>>>>>
>>>>>> sysfs sounds OK to me.  Then again, sysfs is supposed to be
>>>>>> one-value-per-file, so using it would be naughty.
>>>>>>
>>>>>> I dunno, I'd be inclined to abuse the sysfs rule and hope that nobody
>>>>>> notices rather than create a fake char device.  But there's certainly
>>>>>> plenty of precedent for the fake char driver.
>>>>>
>>>>> No, please don't abuse sysfs that way.
>>>>>
>>>>> Use debugfs or a char device node.
>>>>>
>>>>> thanks,
>>>>>
>>>>> greg k-h
>>>>>
>>>>
>>>> I agree with Greg. I asked to not break the existent way to read data
>>>> via /dev/mem because for me it's the right way to do this thing.
>>>> However to do an easy *debug* a debugfs entry can be useful. IMHO, a
>>>> "production" script/application that use debugfs instead of /dev/mem
>>>> in this case is simply broken because the debugfs can't be like a
>>>> system call or other kernel interaction mechanism. Debugfs should be
>>>> used only for debug.
>>>>
>>>> Marco
>>>
>>> Any consensus/decision on how to go on with this patch idea?
>>>
>>> The options that I see right now are:
>>> - keep access through /dev/mem only (but access to /dev/mem is
>>> sometimes restricted);
>>> - keep the debugfs entry as well(as in the patch);
>>> - remove the debugfs entry and add a char driver to access the memory
>>> using read and seek operations.
>>>
>>> + the rejected(?) options from before
>>>
>>> Sergiu
>>>
>>
>> For me the best option it's to use a sysfs/proc entry to export
>> (read-only) the memory address, record size etc. At that point we can
>> use a generic script/program to access via /dev/mem. However I let
>> Andrew/Greg say the last word.
>
> Well, since the only method to read the dump data is /dev/mem,
> exporting the record size/address/etc is needed in order to parse it
> properly. But as far as I can see the data is already exported through
> sysfs in /sys/module/ramoops/parameters/.
> The current module still needs a patch to write the variables of the
> module parameters from the platform data in case that is used, but is
> there any reason why we would need other sysfs entries except these?

I'd say no. I think it's sufficient.

Marco