On Tue, Feb 05, 2019 at 12:08:25PM -0700, Jens Axboe wrote:
> Proof is in the pudding, here's the main commit introducing io_uring
> and now wiring it up to the AF_UNIX garbage collection:
>
> http://git.kernel.dk/cgit/linux-block/commit/?h=io_uring&id=158e6f42b67d0abe9ee84886b96ca8c4b3d3dfd5
>
> How does that look?
In a word - wrong. Some theory: garbage collector assumes that there is
a subset of file references such that
* for all files with such references there's an associated unix_sock.
* all such references are stored in SCM_RIGHTS datagrams that can be
found by the garbage collector (currently: for data-bearing AF_UNIX sockets -
queued SCM_RIGHTS datagrams, for listeners - SCM_RIGHTS datagrams sent via
yet-to-be-accepted connections).
* there is an efficient way to count those references for given file
(->inflight of the corresponding unix_sock).
* removal of those references would render the graph acyclic.
* file can _NOT_ be subject to syscalls unless there are references
to it outside of that subset.
unix_inflight() moves a reference into the subset
unix_notinflight() moves a reference out of the subset
activity that might add such references ought to call wait_for_unix_gc() first
(basically, to stall the massive insertions when gc is running).
Note that unix_gc() does *NOT* work in terms of dropping file references -
the primary effect is locating the SCM_RIGHTS datagrams that can be disposed
of and taking them out. It simply won't do anything to your file references,
no matter what. Add a printk into your ->release() and try to register io_uring
descriptor into itself, then close it. And observe ->release() not being
called for that object. Ever.
PS: The algorithm used by unix_gc() is basically this -
grab unix_gc_lock (giving exclusion with unix_inflight/unix_notinflight
and stabilizing ->inflight counters)
Candidates = {}
for all unix_sock u such that u->inflight > 0
if file corresponding to u has no other references
Candidates += u
/* everything else already is reachable; due to unix_gc_lock these
can't die or get syscall-visible references under us */
Might_Die = Candidates
/* invariant to maintain: for u in Candidates u->inflight will be equal
to the number of references from SCM_RIGHTS datagrams *except*
those immediately reachable from elements of Might_Die */
for all u in Candidates
for each file reference v in SCM_RIGHTS datagrams
immediately reachable from u
if v in Candidates
v->inflight--
To_Scan = () // stuff reachable from those must live
for all u in Might_Die
if u->inflight > 0
queue u into To_Scan
while To_Scan is non-empty
u = dequeue(To_Scan)
Might_Die -= u
for each file reference v in SCM_RIGHTS datagrams
immediately reachable from u
if v in Candidates
v->inflight++ // maintain the invariant
if v in Might_Die
queue v into To_Scan
/* at that point nothing in Might_Die is reachable from the outside */
/* restore the original values of ->inflight */
for all u in Might_Die
for each file reference v in SCM_RIGHTS datagrams
immediately reachable from u
if v in Candidates
v->inflight++
hitlist = ()
for all u in Might_Die
for each SCM_RIGHTS datagram D immediately reachable from u
if D contains references to something in Candidates
move D to hitlist
/* all those datagrams would've never become reachable */
drop unix_gc_lock
discard all datagrams in hitlist.
