Re: [PATCH] tracing/user_events: Run BPF program if attached

On Thu, May 18, 2023 at 09:36:00AM -0400, Steven Rostedt wrote:

On Wed, 17 May 2023 20:14:31 -0700
Alexei Starovoitov [off-list ref] wrote:

quoted

On Wed, May 17, 2023 at 7:08 PM Steven Rostedt [off-list ref] wrote:

quoted

quoted

The delete IOCTL is different than reg/unreg. I don't see a problem with
adding a CAP_SYSADMIN check on the delete IOCTL (and other delete paths)
to prevent this. It shouldn't affect anything we are doing to add this
and it makes it so non-admins cannot delete any events if they are given
write access to the user_events_data file.  

sysadmin for delete is a pointless.
user_events_ioctl_reg() has the same issue.
Two different processes using you fancy TRACELOGGING_DEFINE_PROVIDER()
macro and picking the same name will race.

TRACELOGGING_DEFINE_PROVIDER( // defines the MyProvider symbol
    MyProvider, // Name of the provider symbol to define
    "MyCompany_MyComponent_MyProvider", // Human-readable provider
name, no ' ' or ':' chars.
    // {d5b90669-1aad-5db8-16c9-6286a7fcfe33} // Provider guid
(ignored on Linux)
    (0xd5b90669,0x1aad,0x5db8,0x16,0xc9,0x62,0x86,0xa7,0xfc,0xfe,0x33));

I totally get it that Beau is copy pasting these ideas from windows,
but windows is likely similarly broken if it's registering names
globally.

FD should be the isolation boundary.
fd = open("/sys/kernel/tracing/user_event")
and make sure all events are bound to that file.
when file is closed the events _should be auto deleted_.

That's another issue I just spotted.
Looks like user_events_release() is leaking memory.
user_event_refs are just lost.

tbh the more I look into the code the more I want to suggest to mark it
depends on BROKEN
and go back to redesign.

I don't think these changes require a redesign. I do like the idea that
the events live with the fd. That is, when the fd dies, so does the event.

Although, we may keep it around for a bit (no new events, but being
able to parse it. That is, the event itself isn't deleted until the fd
is closed, and so is the tracing files being read are closed.

Beau,

How hard is it to give the event an owner, but not for task or user,
but with the fd. That way you don't need to worry about other tasks
deleting the event. And it also automatically cleans itself up. If we
leave it to the sysadmin to clean up, it's going to cause leaks,
because it's not something the sysadmin will want to do, as they will
need to keep track of what events are created.

We need to ensure that multiple processes can use the same event name:
Example we have shared libraries that processes use to publish events.

We need to ensure that if the original FD closes and other processes/FDs
are using it, those don't get ripped out from underneath it:
Example we have perf attached and then the process exits.

I think we can accomodate that all neatly if we just make the event
self-delete upon the last ref-count decrement. That way no one needs
the delete IOCTL and we can prevent things piling up.

We have flags in the struct, so we could either make this optional or
default. I like this approach Steven.

Thanks,
-Beau

-- Steve

PS. I missed my connection due to unseasonal freezing temperatures, and
my little airport didn't have a driver for the deicer, making my flight
2 hours delayed (had to wait for the sun to come up and deice the
plane!). Thus, instead of enjoying myself by the pool, I'm in an
airport lounge without much to do.