Wireshark-dev: Re: [Wireshark-dev] [Wireshark-commits] rev 39149: /trunk/epan/dissectors/ /trun
Date Index · Thread Index · Other Months · All Mailing Lists
Date Prev · Date Next · Thread Prev · Thread Next
From: Jeff Morriss <jeff.morriss.ws@xxxxxxxxx>
Date: Wed, 28 Sep 2011 16:46:29 -0400
Anders Broman wrote:

Jeff Morriss skrev 2011-09-27 21:56:

Anders Broman wrote:

Anders Broman skrev 2011-09-26 18:25:

Stig Bjørlykke skrev 2011-09-26 18:14:

On Mon, Sep 26, 2011 at 4:51 PM,<etxrab@xxxxxxxxxxxxx>  wrote:

Log:
  Get rid of check_col, while at it set ENC.

Directory: /trunk/epan/dissectors/
  Changes    Path                      Action
  +35 -37    packet-dmp.c              Modified

The remaining check_col in packet-dmp.c was left on purpose, because
we do not need this calculations if not having COL_INFO.

Should we really remove all occurrences of check_col?



Perhaps not, but how often isn't the info column present?

Thinking about it this raises the question
- is check_col deprecated or not?
Should lengthy code to compute column info be protected in some other manner? fd->visited? (!tree)? We only fill in the column info on the first pass before the tree is built, right?
Actually a quick test indicates that we only build the column info on the 2nd pass (for those frames which are currently displayed).
In other words, if you've got 10,000 frames of which 25 are displayed in your Wireshark window, your dissector is called 10,000 times with check_col(COL_INFO)==0 and 25 times with check_col(COL_INFO)==1. As you scroll down the dissector is called again for the newly-visible frames with check_col(COL_INFO)==1. (It does appear that Wireshark does cache the column information; scrolling back or clicking on earlier packets does not result in the column info being rebuilt, even if the dissector is called again.)
Actually another case when column info won't be present is tshark when doing statistics, etc., but not printing the packet details.
This suggests that check_col()'s should be left in--at least if they are protecting us from doing some work, including val_to_str()'s. 
Even though the check_col is built into the function calls them self s?
My point is that if you've got a bunch of code which runs BEFORE getting into the col_*() routines (and this would include simple--but potentially expensive if the value_string list is big--things like val_to_str()) then it could be saved by leaving the check_col() in the calling routine (i.e., the dissector). 

IOW, in this code sequence:


        col_add_fstr(pinfo->cinfo, COL_INFO, "%s ",
                     val_to_str(h1, chm_h1_message_type_acro_values, "Unknown"));
val_to_str() always does work, even if col_add_fstr() immediately bails out because the column isn't being built.
I don't know how many value_strings are out there which are big enough for this to matter. But I suspect--and sounds like Stig does too--that there are times where the amount of work wasted might matter. This just implies that check_col()'s shouldn't be completely deprecated but rather made optional. Certainly there is no point in guarding this code with a check_col() ;-) : 


            col_set_str(pinfo->cinfo, COL_INFO, "Unknown ");