Received-SPF: pass (google.com: best guess record for domain of linux-wireless-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [RFCv3 2/3] nl80211: Support >4096 byte NEW_WIPHY event nlmsg
To:     Johannes Berg <johannes@sipsolutions.net>,
        linux-wireless@vger.kernel.org
References: <20190906154303.9303-1-denkenz@gmail.com>
 <20190906154303.9303-2-denkenz@gmail.com>
 <4397b67b63d1b1a332afa9010e7f48abd54b49b2.camel@sipsolutions.net>
From:   Denis Kenzior <denkenz@gmail.com>
Message-ID: <5bd58103-bdb7-b72c-0b64-76c8573ca380@gmail.com>
Date:   Wed, 11 Sep 2019 07:41:34 -0500
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101
 Thunderbird/60.8.0
MIME-Version: 1.0
In-Reply-To: <4397b67b63d1b1a332afa9010e7f48abd54b49b2.camel@sipsolutions.net>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-wireless-owner@vger.kernel.org
Precedence: bulk

Hi Johannes,

On 9/11/19 4:44 AM, Johannes Berg wrote:
> Hi,
> 
> The first patch looks good, couple of nits/comments on this one below.
> 
> On Fri, 2019-09-06 at 10:43 -0500, Denis Kenzior wrote:
>> For historical reasons, NEW_WIPHY messages generated by dumps or
>> GET_WIPHY commands were limited to 4096 bytes.  This was due to the
>> fact that the kernel only allocated 4k buffers prior to commit
>> 9063e21fb026 ("netlink: autosize skb lengthes").  Once the sizes of
>> NEW_WIPHY messages exceeded these sizes, split dumps were introduced.
> 
> Actually, userspace prior to around the same time *also* only used 4k
> buffers (old libnl), and so even with that kernel we still could
> possibly have to deal with userspace that had 4k messages only ... but
> we could have solved that part trivially instead of adding code to split
> it, just the kernel part was still in the way then.
> 
> Anyway, I can reword this per my understanding (but will have to reread
> all my messages I guess).
> 

Sure

>> - The code in case '3' is quite complex, but it does try to support a
>>    message running out of room in the middle of a channel dump and
>>    restarting from where it left off in the next split message.  Perhaps
>>    this can be simplified, but it seems this capability is useful.
>>    Please take extra care when reviewing this.
> 
> Is it useful? You say it basically all fits today, and that means the
> channels will either fit into a single message or not ... Then again, if
> we add a lot of channels or a lot more data to each channel. Hmm. OK, I
> guess better if we do have it.
> 

For my dual band cards the channel dump all fits into a ~1k attribute if 
I recall correctly.  So there may not really be a need for this.  Or at 
the very least we could keep things simple(r) and only split at the band 
level, not at the individual channel level.

All the cards I tried would split well after case 9 with 4096 byte 
buffers anyway.  The channel dump is quite early in the message and it 
would really need to become bloated for this code path to be triggered...

> 
>> +	void *last_good_pos = 0;
> 
> Use NULL.

Will fix

> 
>> +		last_good_pos = nlmsg_get_pos(msg);
>>   		state->split_start++;
> 
> Maybe we're better off having a local macro for these two lines? That
> way, we don't risk updating one without the other, which would be
> confusing.

Yep, will do that.

> 
>> @@ -2004,81 +2004,78 @@ static int nl80211_send_wiphy(struct cfg80211_registered_device *rdev,
>>   		if (!nl_bands)
>>   			goto nla_put_failure;
>>   
>> -		for (band = state->band_start;
>> -		     band < NUM_NL80211_BANDS; band++) {
>> +		/* Position in the buffer if we added a set of channel info */
>> +		last_channel_pos = 0;
> 
> NULL

Will fix

> 
>> [snip]
> 
>> +chan_put_failure:
>> +			if (!last_channel_pos)
>> +				goto nla_put_failure;
>> +
>> +			nlmsg_trim(msg, last_channel_pos);
>> +			nla_nest_end(msg, nl_freqs);
>>   			nla_nest_end(msg, nl_band);
>>   
>> -			if (state->split) {
>> -				/* start again here */
>> -				if (state->chan_start)
>> -					band--;
>> +			if (state->chan_start < sband->n_channels)
>>   				break;
>> -			}
>> +
>> +			state->chan_start = 0;
>> +			state->band_start++;
>>   		}
>> -		nla_nest_end(msg, nl_bands);
>>   
>> -		if (band < NUM_NL80211_BANDS)
>> -			state->band_start = band + 1;
>> -		else
>> -			state->band_start = 0;
>> +band_put_failure:
>> +		if (!last_channel_pos)
>> +			goto nla_put_failure;
>> +
>> +		nla_nest_end(msg, nl_bands);
>>   
>> -		/* if bands & channels are done, continue outside */
>> -		if (state->band_start == 0 && state->chan_start == 0)
>> -			state->split_start++;
>> -		if (state->split)
>> +		if (state->band_start < NUM_NL80211_BANDS)
>>   			break;
> 
> Thinking out loud, maybe we could simplify this by just having a small
> "stack" of nested attributes to end?
> 
> I mean, essentially, you have here similar code to the nla_put_failure
> label, in that it finishes and sends out the message, except here you
> have to end a bunch of nested attributes.
> 
> What if we did something like
> 
> #define dump_nest_start(msg, attr) ({ 				\
> 	struct nlattr r = nla_nest_start_noflag(msg, attr);	\
> 	BUG_ON(nest_stack_depth >= ARRAY_SIZE(nest_stack);	\
> 	nest_stack[nest_stack_depth++] = r;			\
> 	r;							\
> })
> 
> #define dump_nest_end(msg, r) do {				\
> 	BUG_ON(nest_stack_depth > 0);				\
> 	nest_stack_depth--;					\
> 	BUG_ON(nest_stack[nest_stack_depth] == r);		\
> 	nla_nest_end(msg, r);					\
> } while (0)
> 
> 
> or something like that (we probably don't want to use
> nla_nest_start_noflag() for future attributes, etc. but anyway).
> 
> Then we could unwind any nesting at the end in the common code at the
> nla_put_failure label very easily, I think?

I see where you're going with this, I think I do anyway...

The current logic uses last_channel_pos for some of the trickery in 
addition to last_good_pos.  So nla_put_failure would have to be made 
aware of that.  Perhaps we can store last_good_pos in the stack, but the 
split mechanism only allows the splits to be done at certain points... 
I think the above could be doable, but the code would be even more complex.

Right now only the channel dump uses this (and I'm still not fully 
convinced we should go to all the trouble), so one argument would be not 
to introduce something this generic until another user of it manifests 
itself?

Regards,
-Denis