WEBVTT 00:00.000 --> 00:10.000 How many of you are using an IP here? 00:10.000 --> 00:14.000 So who's annoyed by the message once in a while that you get about? 00:14.000 --> 00:19.000 You need to restart, not need to continue using it. 00:19.000 --> 00:20.000 Nobody's annoyed. 00:20.000 --> 00:23.000 Oh, shit, you're working for nothing. 00:23.000 --> 00:26.000 And so Alex is going to talk about that. 00:26.000 --> 00:29.000 So please sit, remember sit in the middle. 00:29.000 --> 00:31.000 So people that come late can sit. 00:31.000 --> 00:32.000 Thank you very much. 00:32.000 --> 00:34.000 Oh, one last thing for the newcomers. 00:34.000 --> 00:37.000 We exit by the door that are downstairs. 00:37.000 --> 00:38.000 There are stickers on the tables. 00:38.000 --> 00:40.000 Here, here and there. 00:40.000 --> 00:43.000 Please feel free to take them. 00:43.000 --> 00:44.000 Hello. 00:44.000 --> 00:45.000 Welcome to this talk. 00:45.000 --> 00:52.000 So I'm Alex and I have been working for Amazilaf for more than 11 years now. 00:52.000 --> 00:57.000 And I'm here to present you something we call FOX over. 00:57.000 --> 01:02.000 There is not my name on the slide because it has been a walk. 01:02.000 --> 01:05.000 Not just done by me, but it also involved. 01:05.000 --> 01:10.000 Think early, jet Davies, Nikolay Zerl and other people. 01:10.000 --> 01:15.000 So this is what he do with what he was talking about. 01:15.000 --> 01:17.000 How many people have been annoyed by that? 01:17.000 --> 01:20.000 On the daily basis. 01:21.000 --> 01:25.000 It's mostly people on the Linux or on the release. 01:25.000 --> 01:28.000 And let's explain why you see this. 01:28.000 --> 01:33.000 So the reason is that when we create a new process in Firefox, 01:33.000 --> 01:36.000 we have to read from this. 01:36.000 --> 01:41.000 And whatever is on this may have changed because your package manager applied an update. 01:41.000 --> 01:48.000 And so when we start the process, the title says that you need to view a website 01:48.000 --> 01:50.000 and just open the tab. 01:50.000 --> 01:53.000 Then we have the child that is going to communicate with the parent process, 01:53.000 --> 01:55.000 although IPC. 01:55.000 --> 02:01.000 And the first thing that we are going to do is to verify the version that we are talking about of the IPC protocol. 02:01.000 --> 02:04.000 Because if we are not choosing the same version, 02:04.000 --> 02:09.000 we are going to run into very weird issues that we had in the past, 02:09.000 --> 02:12.000 that we are triggering super complicated bugs, 02:12.000 --> 02:14.000 as well wasting our time. 02:14.000 --> 02:19.000 So we decided to say if the version on this does not match the version in memory, 02:19.000 --> 02:22.000 between the parent and the child, then we stop everything, 02:22.000 --> 02:23.000 and we are no people. 02:23.000 --> 02:27.000 Because it's much fun for you to be annoyed, 02:27.000 --> 02:30.000 and for us to be annoyed. 02:30.000 --> 02:35.000 That's the main motivation for the user facing value of FOX server. 02:35.000 --> 02:41.000 So first I will try to explain to you how we go into the situation with so much processes, 02:41.000 --> 02:44.000 and why you run into that issue so often. 02:44.000 --> 02:47.000 Then we will explain a little bit in detail, 02:47.000 --> 02:49.000 but not to mention what is FOX server. 02:49.000 --> 02:54.000 And I will look at the improvements outside of this dialogue, 02:54.000 --> 02:57.000 which should not come anymore. 02:57.000 --> 02:59.000 Let's do a little bit of a story. 02:59.000 --> 03:03.000 A long time ago, the web was very simple. 03:03.000 --> 03:09.000 You had no audio, you had no video, you had computers with only one processor, 03:09.000 --> 03:13.000 as the security was not so much of a big issue. 03:13.000 --> 03:17.000 The attack surface was much simpler. 03:17.000 --> 03:23.000 So the original design of a browser was a single process, doing everything. 03:23.000 --> 03:31.000 Doing UI, rendering of the web pages, doing the disk access, everything, everything. 03:31.000 --> 03:36.000 And there was no good reason to split that into many processes. 03:37.000 --> 03:42.000 Because you had no CPU within our processes to take advantage of it. 03:42.000 --> 03:48.000 Then a few years later, we introduced audio, we introduced video. 03:48.000 --> 03:55.000 And we did that, for example, using Flash, using MPAPI, which some of you might remember. 03:55.000 --> 04:01.000 And those were very complex pieces of software running within the process. 04:01.000 --> 04:09.000 So adding a lot of attack surface to the process, adding a lot of stability issues. 04:09.000 --> 04:13.000 So there was a good incentive to have a separate process for them. 04:13.000 --> 04:18.000 Because, for example, if Flash was behaving badly, we could kill it and restart it. 04:18.000 --> 04:21.000 We started having the world browser to crash. 04:21.000 --> 04:27.000 So it was convenient for user and first convenient for us. 04:27.000 --> 04:31.000 So there was a bug that we would be blamed about. 04:31.000 --> 04:35.000 And the story continues a little bit over. 04:35.000 --> 04:39.000 We had GMP, which introduced some video codecs. 04:39.000 --> 04:42.000 And at some point, the web became a little bit more complicated. 04:42.000 --> 04:45.000 We started to have to isolate stuff. 04:45.000 --> 04:50.000 And with more complexity, we introduced even more processes. 04:50.000 --> 04:53.000 Like GPU content, web extension. 04:53.000 --> 05:00.000 And as of today, everything involved is a process type in Firefox. 05:00.000 --> 05:07.000 So it means that you have around 10 process types that you can have on your computer. 05:07.000 --> 05:12.000 If you go into about processes, you will see a lot of them. 05:12.000 --> 05:19.000 And for example, content processes can be insulated for each type on the browser. 05:19.000 --> 05:22.000 So each time you load the new website, you have a new process. 05:22.000 --> 05:28.000 It's more or less that we have processes to manage access to the GPU, 05:28.000 --> 05:37.000 to take care of the video codecs, to perform audio and video recording, 05:37.000 --> 05:39.000 to perform network access. 05:39.000 --> 05:48.000 And one of the main reason we have that way to put things in two different processes 05:48.000 --> 05:54.000 is security, mostly, because we can perform sandboxing at the process level. 05:54.000 --> 05:57.000 On Linux we perform sandboxing using a second. 05:57.000 --> 06:00.000 So it means that we are filtering system codes. 06:00.000 --> 06:06.000 And one very simple example is that content process does not have access to the file system. 06:06.000 --> 06:15.000 It does do a broker request to the parent process that can verify the file system access. 06:15.000 --> 06:18.000 So it means that we can control even not directly security. 06:18.000 --> 06:22.000 We can for being accessing some stuff for security purposes. 06:22.000 --> 06:28.000 We also use that for fingerprint processing focus. 06:28.000 --> 06:36.000 And recently we also introduced a new kind of process utility process that has a unique property 06:36.000 --> 06:40.000 to allow us to sandbox even more simply things. 06:40.000 --> 06:44.000 Because in the past I think a new type of process was a little bit complicated. 06:44.000 --> 06:51.000 And we wanted to be able to say to sandbox a little bit more, especially for example. 06:51.000 --> 06:56.000 Audio recording that was in the past living within the RDD process. 06:56.000 --> 07:03.000 So it means that we had two shares the same sandbox between video decoders and audio decoders. 07:03.000 --> 07:12.000 And video decoders may require our other access to a specific area of, for example, your GPU to perform audio decoding. 07:12.000 --> 07:18.000 And doing that often means that you have to open holes in the sandbox. 07:18.000 --> 07:21.000 And sandbox you want it to be as tight as possible. 07:21.000 --> 07:26.000 And the only way to sandbox differently is to have different kind of process. 07:26.000 --> 07:31.000 Let's just explain a little bit how we create a process. 07:31.000 --> 07:37.000 So for those of you who are a little bit used to that only mix, you do clone and fork. 07:37.000 --> 07:45.000 And then you do exact and exact is a good thing because it results everything from scratch, which is also a bad thing. 07:45.000 --> 07:48.000 Because it reloads everything from scratch. 07:48.000 --> 07:52.000 So everything we did in one process we have to redo every time. 07:52.000 --> 07:58.000 And we lose all the nice things that we have on fork like copium right, 07:58.000 --> 08:06.000 which means that when you fork a process you should be able to share memory that has not been modified by the child. 08:06.000 --> 08:09.000 And we cannot do that anymore. 08:09.000 --> 08:14.000 So each time we create a new process it has a ghost in CPU time and in memory. 08:14.000 --> 08:18.000 And as you so now we have a lot of processes in Firefox. 08:18.000 --> 08:21.000 And this is why we are for example. 08:22.000 --> 08:27.000 So how do we create a new process with box server? We ask another process. 08:27.000 --> 08:35.000 On Chrome they already have Zigotty. On Firefox OS a long time ago we had a process called new work. 08:35.000 --> 08:38.000 And this strategy is very simple. We start our process. 08:38.000 --> 08:42.000 We get it as far as we can in the startup process. 08:42.000 --> 08:46.000 And we will fork from that point later. 08:46.000 --> 08:50.000 So this way it's faster. We don't have to redo some of the installation. 08:50.000 --> 08:53.000 We can share more memory between processes. 08:53.000 --> 08:56.000 The risk is that now we don't have exact. 08:56.000 --> 09:02.000 So we may get into a state where we are sharing steps that we don't want to share. 09:02.000 --> 09:06.000 Like for example some random generators, which is not a good thing. 09:06.000 --> 09:12.000 Newer was one of the solutions we had on Firefox OS. 09:12.000 --> 09:14.000 It was very complicated. 09:14.000 --> 09:17.000 But completely started content processes. 09:17.000 --> 09:20.000 So the ones that we used to run our web pages. 09:20.000 --> 09:22.000 And it froze them. 09:22.000 --> 09:31.000 And it did a lot of magic to be able to improve them and put them back into a state where it can be used as a new process. 09:31.000 --> 09:33.000 And it was very efficient. 09:33.000 --> 09:35.000 But it was also super fragile. 09:35.000 --> 09:39.000 And I have been myself the victim of a nasty regression. 09:39.000 --> 09:42.000 And I just added a test for Firefox OS. 09:42.000 --> 09:46.000 And we had a very bad execution time regression. 09:46.000 --> 09:53.000 And people blame me when it was in fact a regression on newer. 09:53.000 --> 09:54.000 We spoke several. 09:54.000 --> 09:58.000 We have decided to go with a more simple design. 09:58.000 --> 10:02.000 We do not start a whole content process. 10:02.000 --> 10:04.000 We just start a very minimal process. 10:04.000 --> 10:08.000 But that is not minimal enough to share enough of data. 10:09.000 --> 10:16.000 We do a communication of our socket to perform the call to say, 10:16.000 --> 10:18.000 hey, I want a new process. 10:18.000 --> 10:20.000 And we do not start a JS. 10:20.000 --> 10:23.000 And that is already a bit of changes. 10:23.000 --> 10:27.000 We are to deal with chain model states. 10:27.000 --> 10:32.000 If we do not do that, we completely regressed web content performances. 10:32.000 --> 10:36.000 And I added a bit of with PGO and PIDs. 10:36.000 --> 10:41.000 Basically, when we learned it, alpha of July, we got back to it. 10:41.000 --> 10:45.000 Because we have completely destroyed performances. 10:45.000 --> 10:47.000 Because we just broke PGO. 10:47.000 --> 10:49.000 But that is fine. 10:49.000 --> 10:50.000 And why is it fine? 10:50.000 --> 10:53.000 Because we have been able to relent. 10:53.000 --> 10:57.000 And now it is enabled by default on Linux, 10:57.000 --> 11:00.000 unlike it since end of October. 11:00.000 --> 11:05.000 There is still some work to perform, to be able to enable that on all of the channels. 11:05.000 --> 11:11.000 But as much as I know, there has been nobody complaining about user creation, 11:11.000 --> 11:13.000 which is nice. 11:13.000 --> 11:17.000 So if you are using Nikely on Linux, please have a look. 11:17.000 --> 11:20.000 And maybe we have missed some issues. 11:20.000 --> 11:23.000 And please report them so that we can fix it. 11:23.000 --> 11:26.000 Because this is the wins that we have. 11:27.000 --> 11:34.000 So on those two shots, the first one is the memory usage of content process. 11:34.000 --> 11:41.000 You don't see the figures, but stack of points on the other side is around 15 megabytes. 11:41.000 --> 11:44.000 That's before we are able to focus out there. 11:44.000 --> 11:49.000 And when we are enabled it, we go to 75 megabytes. 11:50.000 --> 11:54.000 So you can also see the times we enabled and got back out. 11:54.000 --> 11:56.000 So 50% improvement. 11:56.000 --> 12:03.000 And here we have a metric that we used to see as a start-up time of a process. 12:03.000 --> 12:08.000 So we went from around 60 milliseconds to around 52 milliseconds. 12:08.000 --> 12:10.000 So 75% improvement. 12:10.000 --> 12:12.000 That's not bad. 12:12.000 --> 12:14.000 And I guess that's all. 12:14.000 --> 12:15.000 If you have question. 12:15.000 --> 12:16.000 Yes. 12:16.000 --> 12:26.000 So any questions? 12:26.000 --> 12:28.000 And I guess we don't have questions. 12:28.000 --> 12:30.000 Thank you Alexander.