The current implementation of the packet list has some serious drawbacks especially when working with large capture files.
It uses a GTKCList which is a bit inflexible, e.g. it needs to keep all the content text in memory even for the content currently not being displayed.
huge memory consumption: we must keep (at least) the info column texts in memory (which makes around half of the memory consumption of a packet!)
coloring rules for all packets: coloring rules are processed for all packets, but would only be required for the displayed ones
columns can't be changed: the GTKCList don't allow columns to be added, removed, or moved without having to exit and restart Wireshark (or at least destroy the widget)
ADNS name resolution not shown for all packets: name resolution not shown in packets already displayed while resolution is still done (and won't be updated afterwards)
changing the time stamp format is slow: this requires the GTKCList (column?) to be regenerated
known bugs with the vertical scroll bar: see http://bugs.wireshark.org/bugzilla/show_bug.cgi?id=220
The first two points are very serious for performance when working with large capture files (>10MB). Allocating massive amounts of memory takes it's time and processing the coloring rules for packets that are not displayed is simply useless.
Optimizing these things will help improving the overall performance while working with large capture files a lot.
The GTKCList doesn't provide us with the features we require to fix the problems listed above. So it's clear we have to use a different GTK widget or write a new one on our own.
Beside the current functionality, it should:
- generate the column contents with a callback, rather than saving the text of the columns when a row is added
- allow to dynamically change columns
According to the GTK+ API reference, GTKClist (and GTKCTree) are deprecated in favor of GTKTreeView, which supports everything above. Unfortunately, GTKTreeView doesn't exist in GTK+ 1.2.
Replacing the GTKCList by a GTKTreeView seemed to be straightforward, as the complete functionality is encapsulated in the file gtk/packet_list.c
I've done a prototype implementation using the GTKTreeView which wasn't very hard to do, but then I was running into some serious problems:
- rows using more display space: GTKTreeView uses 1-2px more for each row, ending up with a few lines less on the screen. I've even tried to implement my own GTKTreeCellRenderer but couldn't limit the space usage
- appending rows is significantly slower compared to the GTKCList (doubles / triples the time!). I've had a look at the append function, but didn't saw a way to improve speed here
The files of this prototype implementation are attached (beware, ugly prototype code ahead) in case we might need it later …
Conclusion: this won't be GTK1.x compatible and faces some significant problems which I wasn't able to solve. As the packet list is very performance critical, it might even be a good idea to not use a standard widget here so we can fine tune the behaviour … - UlfLamping
The "new" packet list is now the default!
Using our own custom TreeView most problems have been addressed. With the new packet list rows do use more display space but that may be a price we have to pay. Adding rows to the list is fast as very little data is copied.
Loading a 19Mb file(203950 packets) time is reduced from ~14s to < 4s.
Memory usage From ~170 Mb to ~113 Mb.
Applying coloring is fast, from 22s to less than 1s.
Changing time format is fast, from >4min 30s less than 1s.
Changing time reference is fast, from 22s to less than 1s.
- Uses more display space.
- Sorting is mabye slower(make a measure).
Not yet implemented feature:
- Recreate packet list when adding/deleting columns Implementation sketch/idea:
- Use a GpointerArray for col_text then it should be possible to insert a new index(column) on the fly. g_ptr_array_sort could be used to order the list(?). Marking the packets as undissected and calling gtk_widget_queue_draw should redraw the packet list with the new column. We currently have new_packet_list_recreate() which destroys the packet list and creates a new. This must be rewritten.
- Update internal structures (cfile.cinfo.*) when DnD moving columns in the packet list.
- Resize columns correctly, according to both title and content length.
- Add missing sorting indicator in the No. column on startup.
- Select the next packet when removing a Time Reference from a packet not matching the display filter.
[WMeier] SVN #29971 <Tab> and then <Shift Tab> between the summary pane and the
details pane doesn't quite work the same ….
After <Tab> from summary to details and then <Shift Tab> back to summary it takes an extra down-arrow to move down 1 row.
Were in the code is <Tab> handled?
jyoung Based on some google searches it looks like <Tab> processing is handled natively by the core GTK widgets:
jyoung SVN #30025; Understand issue with <Tab> and <Shift><Tab>.
<Tab> and <Shift><Tab> will move focus to the label of the selected column in the summary pane.
A sequence of six (6) <Tab> keypresses moves focus sequentially from selected frame in packet list (summary pane); to selected element in details pane, to hex pane, to leftmost item in toolbar 1 (the list capture interfaces button), to the leftmost item in toolbar 2 (the Filter: label) to the label of selected column in the summary pane, and back to the originally selected frame. So with Tab forward we can get back where we started (with one extra <tab> keypress).
But on TAB backwards (<shift><tab>) the focus initially moves from the selected frame in the packet list (summary pane) to the label of the selected column in the summary pane; to the rightmost item in toolbar 2 ("Apply"); to the rightmost item in toolbar 1 ("Show some help…"); to the hex pane, to the selected item in the details pane, and back to the label of the selected column in the summary pane. On TAB Backwards focus always skips over the originally selected item in the summary pane.
I suspect slipping in focus = gtk_window_get_focus(GTK_WINDOW(top_level)); and gtk_window_set_focus(GTK_WINDOW(top_level), focus); within an appropriate callback will solve the <Tab> and <Shift><Tab> issue. The trick is to figure out which callback (or callbacks) to add them to.
jyoung SVN #30083; Attempted to replicate problem reported in bug 4035: "Application crash when changing real-time option". I could NOT replicate this bug. But I did find some issues with cursor placement/movement within the summary portal when a busy active capture was in progess (there were lots of packets on the wire). With the old packet list I am able to stop the auto-scroll feature by clicking on the summary portal and pressing up arrow once. But that does not happen in the new packet list; the list keeps scrolling. Also when I did disable the "Auto Scroll in Live Capture" (either by the Menu option or by clicking on the "Go back in packet history" button) I could NOT actually scroll through the packet list while new packets are added to the list. The cursor would temporarily show up on frame 1 or perhaps 2 and then disappear until I pressed down arrow again which would cause the cursor to momentarily appear back on frame 1 or frame 2 only to disappear again.
ABroman It would be nice to get a grip on the remaining issues, going trough the bugs I have found the following remaining bugs: see list below. Some of these may have been fixed in recent commits. It would also be nice to know what remains of the other issues mentioned on this page.
Bug 4056 - new_packet_list: Hex pane display issue after startup. (Partly fixed)
Imported from https://wiki.wireshark.org/Development/OptimizePacketList on 2020-08-11 23:12:56 UTC