On Mon, Nov 11, 2024 at 4:52 PM Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Mon, 11 Nov 2024 at 12:19, Josef Bacik <josef@xxxxxxxxxxxxxx> wrote: > > > > --- a/fs/namei.c > > +++ b/fs/namei.c > > @@ -3782,7 +3782,15 @@ static int do_open(struct nameidata *nd, > > + /* > > + * This permission hook is different than fsnotify_open_perm() hook. > > + * This is a pre-content hook that is called without sb_writers held > > + * and after the file was truncated. > > + */ > > + return fsnotify_file_area_perm(file, MAY_OPEN, &file->f_pos, 0); > > } > > Stop adding sh*t like this to the VFS layer. > > Seriously. I spend time and effort looking at profiles, and then > people who do *not* seem to spend the time and effort just willy nilly > add fsnotify and security events and show down basic paths. > > I'm going to NAK any new fsnotify and permission hooks unless people > show that they don't add any overhead. > > Because I'm really really tired of having to wade through various > permission hooks in the profiles that I can not fix or optimize, > because those hoosk have no sane defined semantics, just "let user > space know". > > Yes, right now it's mostly just the security layer. But this really > looks to me like now fsnotify will be the same kind of pain. > > And that location is STUPID. Dammit, it is even *documented* to be > stupid. It's a "pre-content" hook that happens after the contents have > already been truncated. WTF? That's no "pre". > > I tried to follow the deep chain of inlines to see what actually ends > up happening, and it looks like if the *whole* filesystem has no > notify events at all, the fsnotify_sb_has_watchers() check will make > this mostly go away, except for all the D$ accesses needed just to > check for it. > > But even *one* entirely unrelated event will now force every single > open to typically call __fsnotify_parent() (or possibly "just" > fsnotify), because there's no sane "nobody cares about this dentry" > kind of thing. > > So effectively this is a new hook that gets called on every single > open call that nobody else cares about than you, and that people have > lived without for three decades. > > Stop it, or at least add the code to not do this all completely pointlessly. > > Because otherwise I will not take this kind of stuff any more. I just > spent time trying to figure out how to avoid the pointless cache > misses we did for every path component traversal. > > So I *really* don't want to see another pointless stupid fsnotify hook > in my profiles. Did you see the patch that added the fsnotify_file_has_pre_content_watches() thing? It'll look at the inode/sb/mnt to see if there are any watches and carry on if there's nothing. This will be the cheapest thing open/write/whatever does. If there's no watches, nothing happens. I'm not entirely sure what else you want me to do here, other than not add the feature, which I suppose is a position to take. The overhead here is purely for people who use HSM. I'm telling you that in production, with real world workloads, the overhead is far outweighed by having to copy bytes around we'll never use. Your argument is similar to the argument against adding compression to btrfs, "but it costs CPU!?", sure, but the benefit of writing significantly less waaaaay outweighed the CPU cost and was a huge net positive. This is going to cost something if it's on, it costs nothing if it's off. If you want performance numbers that's reasonable, I'll run fsperf tomorrow and get you a side by side comparison. Thanks, Josef
