On Mon, Sep 30, 2024 at 03:32:25PM GMT, Lorenzo Stoakes wrote:
On Mon, Sep 30, 2024 at 04:21:23PM GMT, Aleksa Sarai wrote:
On 2024-09-30, Lorenzo Stoakes lorenzo.stoakes@oracle.com wrote:
On Mon, Sep 30, 2024 at 02:34:33PM GMT, Christian Brauner wrote:
On Mon, Sep 30, 2024 at 11:39:49AM GMT, Lorenzo Stoakes wrote:
On Mon, Sep 30, 2024 at 12:33:18PM GMT, Florian Weimer wrote:
- Lorenzo Stoakes:
> If you wish to utilise a pidfd interface to refer to the current process > (from the point of view of userland - from the kernel point of view - the > thread group leader), it is rather cumbersome, requiring something like: > > int pidfd = pidfd_open(getpid(), 0); > > ... > > close(pidfd); > > Or the equivalent call opening /proc/self. It is more convenient to use a > sentinel value to indicate to an interface that accepts a pidfd that we > simply wish to refer to the current process.
The descriptor will refer to the current thread, not process, right?
No it refers to the current process (i.e. thread group leader from kernel perspective). Unless you specify PIDFD_THREAD, this is the same if you did the above.
The distinction matters for pidfd_getfd if a process contains multiple threads with different file descriptor tables, and probably for pidfd_send_signal as well.
You mean if you did a strange set of flags to clone()? Otherwise these are shared right?
Again, we are explicitly looking at process not thread from userland perspective. A PIDFD_SELF_THREAD might be possible, but this series doesn't try to implement that.
Florian raises a good point. Currently we have:
(1) int pidfd_tgid = pidfd_open(getpid(), 0); (2) int pidfd_thread = pidfd_open(getpid(), PIDFD_THREAD);
and this instructs:
pidfd_send_signal() pidfd_getfd()
to do different things. For pidfd_send_signal() it's whether the operation has thread-group scope or thread-scope for pidfd_send_signal() and for pidfd_getfd() it determines the fdtable to use.
The thing is that if you pass:
pidfd_getfd(PDIFD_SELF)
and you have:
TGID
T1 { clone(CLONE_THREAD) unshare(CLONE_FILES) }
T2 { clone(CLONE_THREAD) unshare(CLONE_FILES) }
You have 3 threads in the same thread-group that all have distinct file descriptor tables from each other.
So if T1 did:
pidfd_getfd(PIDFD_SELF, ...)
and we mirror the PIDTYPE_TGID behavior then T1 will very likely expect to get the fd from its file descriptor table. IOW, its reasonable to expect that T1 is interested in their very own resource, not someone else's even if it is the thread-group leader.
But what T1 will get in reality is an fd from TGID's file descriptor table (and similar for T2).
Iirc, yes that confusion exists already with /proc/self. But the question is whether we should add the same confusion to the pidfd api or whether we make PIDFD_SELF actually mean PIDTYPE_PID aka the actual calling thread.
My thinking is that if you have the reasonable suspicion that you're multi-threaded and that you're interested in the thread-group resource then you should be using:
int pidfd = pidfd_open(getpid(), 0)
and hand that thread-group leader pidfd around since you're interested in another thread. But if you're really just interested in your own resource then pidfd_open(getpid(), 0) makes no sense and you would want PIDFD_SELF.
Thoughts?
I mean from my perspective, my aim is to get current->mm for process_madvise() so both work for me :) however you both raise a very good point here (sorry Florian, perhaps I was a little too dismissive as to your point, you're absolutely right).
My intent was for PIDFD_SELF to simply mirror the pidfd_open(getpid(), 0) behaviour, but you and Florian make a strong case that you'd _probably_ find this very confusing had you unshared in this fashion.
I mean in general this confusion already exists, and is for what PIDFD_THREAD was created, but I suspect ideally if you could go back you might actually do this by default Christian + let the TGL behaviour be the optional thing?
For most users this will not be an issue, but for those they'd get the same result whichever they used, but yes actually I think you're both right - PIDFD_SELF should in effect imply PIDFD_THREAD.
Funnily enough we ran into issues with this when running Go code in runc that did precisely this -- /proc/self gave you the wrong fd table in very specific circumstances that were annoying to debug. For languages with green-threading you can't turn off (like Go) these kinds of issues pop up surprisingly often.
Yeah, damn, useful insight that such things do happen in the wild.
We can adjust the pidfd_send_signal() call to infer the correct scope (actually nicely we can do that without any change there, by having __pidfd_get_pid() set f_flags accordingly).
So TL;DR: I agree, I will respin with PIDFD_SELF referring to the thread.
My question in return here then is - should we introduce PIDFD_SELF_PROCESS also (do advise if you feel this naming isn't quite right) - to provide thread group leader behaviour?
Sorry to bike-shed, but to match /proc/self and /proc/thread-self, maybe they should be called PIDFD_SELF (for tgid) and PIDFD_THREAD_SELF (for current's tid)? In principle I guess users might use PIDFD_SELF by accident but if we mirror the naming with /proc/{,thread-}self that might not be that big of an issue?
Lol, you know I wasn't even aware /proc/thread-self existed...
Wait until you learn that /proc/$TID thread entries exist but aren't shown when you do ls -al /proc, only when you explicitly access them.
Yeah, that actually makes sense and is consistent, though obviously the concern is people will reflexively use PIDFD_SELF and end up with potentially confusing results.
I will obviously be doing a manpage patch for this so we can spell it out there very clearly and also in the header - so I'd actually lean towards doing this myself.
Christian, Florian? Thoughts?
I think adding both would be potentially useful. How about:
#define PIDFD_SELF -100 #define PIDFD_THREAD_GROUP -200
This will make PIDFD_SELF mean current and PIDFD_THREAD_GROUP mean current->pid_links[PIDTYPE_TGID]. I don't think we need to or should mirror procfs in any way. pidfds are intended to be usable without procfs at all.
I want to leave one comment on a bit of quirkiness in the api when we add this. I don't consider it a big deal but it should be pointed out.
It can be useful to allow PIDFD_SELF or PIDFD_THREAD_GROUP to refer to the calling thread even for pidfd_open() to avoid an additional getpid() system call:
(1) pidfd_open(PIDFD_SELF, PIDFD_THREAD) (2) pidfd_open(PIDFD_SELF, 0)
So if we allow this (Should we allow it?) then (1) is just redundant but whathever. But (2) is at least worth discussing: Should we reject (2) on the grounds of contradictory requests or allow it and document that it's equivalent to pidfd_open(getpid(), PIDFD_THREAD)? It feels like the latter would be ok.
Similar for pidfd_send_signal():
// redundant but ok: (1) pidfd_send_signal(PIDFD_SELF, PIDFD_SIGNAL_THREAD)
// redundant but ok: (2) pidfd_send_signal(PIDFD_THREAD_GROUP, PIDFD_SIGNAL_THREAD_GROUP)
// weird way to spell pidfd_send_signal(PIDFD_THREAD_GROUP, 0) (3) pidfd_send_signal(PIDFD_SELF, PIDFD_SIGNAL_THREAD_GROUP)
// weird way to spell pidfd_send_signal(PIDFD_SELF, 0) (4) pidfd_send_signal(PIDFD_THREAD_GROUP, PIDFD_SIGNAL_THREAD)
I think all of this is ok but does anyone else have a strong opinion?