[release-branch.go1.7] runtime: fix corruption crash/race between select and stack growth
Fixes #18708 (backport) To implement the blocking of a select, a goroutine builds a list of offers to communicate (pseudo-g's, aka sudog), one for each case, queues them on the corresponding channels, and waits for another goroutine to complete one of those cases and wake it up. Obviously it is not OK for two other goroutines to complete multiple cases and both wake the goroutine blocked in select. To make sure that only one branch of the select is chosen, all the sudogs contain a pointer to a shared (single) 'done uint32', which is atomically cas'ed by any interested goroutines. The goroutine that wins the cas race gets to wake up the select. A complication is that 'done uint32' is stored on the stack of the goroutine running the select, and that stack can move during the select due to stack growth or stack shrinking. The relevant ordering to block and unblock in select is: 1. Lock all channels. 2. Create list of sudogs and queue sudogs on all channels. 3. Switch to system stack, mark goroutine as asleep, unlock all channels. 4. Sleep until woken. 5. Wake up on goroutine stack. 6. Lock all channels. 7. Dequeue sudogs from all channels. 8. Free list of sudogs. 9. Unlock all channels. There are two kinds of stack moves: stack growth and stack shrinking. Stack growth happens while the original goroutine is running. Stack shrinking happens asynchronously, during garbage collection. While a channel listing a sudog is locked by select in this process, no other goroutine can attempt to complete communication on that channel, because that other goroutine doesn't hold the lock and can't find the sudog. If the stack moves while all the channel locks are held or when the sudogs are not yet or no longer queued in the channels, no problem, because no goroutine can get to the sudogs and therefore to selectdone. We only need to worry about the stack (and 'done uint32') moving with the sudogs queued in unlocked channels. Stack shrinking can happen any time the goroutine is stopped. That code already acquires all the channel locks before doing the stack move, so it avoids this problem. Stack growth can happen essentially any time the original goroutine is running on its own stack (not the system stack). In the first half of the select, all the channels are locked before any sudogs are queued, and the channels are not unlocked until the goroutine has stopped executing on its own stack and is asleep, so that part is OK. In the second half of the select, the goroutine wakes up on its own goroutine stack and immediately locks all channels. But the actual call to lock might grow the stack, before acquiring any locks. In that case, the stack is moving with the sudogs queued in unlocked channels. Not good. One goroutine has already won a cas on the old stack (that goroutine woke up the selecting goroutine, moving it out of step 4), and the fact that done = 1 now should prevent any other goroutines from completing any other select cases. During the stack move, however, sudog.selectdone is moved from pointing to the old done variable on the old stack to a new memory location on the new stack. Another goroutine might observe the moved pointer before the new memory location has been initialized. If the new memory word happens to be zero, that goroutine might win a cas on the new location, thinking it can now complete the select (again). It will then complete a second communication (reading from or writing to the goroutine stack incorrectly) and then attempt to wake up the selecting goroutine, which is already awake. The scribbling over the goroutine stack unexpectedly is already bad, but likely to go unnoticed, at least immediately. As for the second wakeup, there are a variety of ways it might play out. * The goroutine might not be asleep. That will produce a runtime crash (throw) like in #17007: runtime: gp: gp=0xc0422dcb60, goid=2299, gp->atomicstatus=8 runtime: g: g=0xa5cfe0, goid=0, g->atomicstatus=0 fatal error: bad g->status in ready Here, atomicstatus=8 is copystack; the second, incorrect wakeup is observing that the selecting goroutine is in state "Gcopystack" instead of "Gwaiting". * The goroutine might be sleeping in a send on a nil chan. If it wakes up, it will crash with 'fatal error: unreachable'. * The goroutine might be sleeping in a send on a non-nil chan. If it wakes up, it will crash with 'fatal error: chansend: spurious wakeup'. * The goroutine might be sleeping in a receive on a nil chan. If it wakes up, it will crash with 'fatal error: unreachable'. * The goroutine might be sleeping in a receive on a non-nil chan. If it wakes up, it will silently (incorrectly!) continue as if it received a zero value from a closed channel, leaving a sudog queued on the channel pointing at that zero vaue on the goroutine's stack; that space will be reused as the goroutine executes, and when some other goroutine finally completes the receive, it will do a stray write into the goroutine's stack memory, which may cause problems. Then it will attempt the real wakeup of the goroutine, leading recursively to any of the cases in this list. * The goroutine might have been running a select in a finalizer (I hope not!) and might now be sleeping waiting for more things to finalize. If it wakes up, as long as it goes back to sleep quickly (before the real GC code tries to wake it), the spurious wakeup does no harm (but the stack was still scribbled on). * The goroutine might be sleeping in gcParkAssist. If it wakes up, that will let the goroutine continue executing a bit earlier than we would have liked. Eventually the GC will attempt the real wakeup of the goroutine, leading recursively to any of the cases in this list. * The goroutine cannot be sleeping in bgsweep, because the background sweepers never use select. * The goroutine might be sleeping in netpollblock. If it wakes up, it will crash with 'fatal error: netpollblock: corrupted state'. * The goroutine might be sleeping in main as another thread crashes. If it wakes up, it will exit(0) instead of letting the other thread crash with a non-zero exit status. * The goroutine cannot be sleeping in forcegchelper, because forcegchelper never uses select. * The goroutine might be sleeping in an empty select - select {}. If it wakes up, it will return to the next line in the program! * The goroutine might be sleeping in a non-empty select (again). In this case, it will wake up spuriously, with gp.param == nil (no reason for wakeup), but that was fortuitously overloaded for handling wakeup due to a closing channel and the way it is handled is to rerun the select, which (accidentally) handles the spurious wakeup correctly: if cas == nil { // This can happen if we were woken up by a close(). // TODO: figure that out explicitly so we don't need this loop. goto loop } Before looping, it will dequeue all the sudogs on all the channels involved, so that no other goroutine will attempt to wake it. Since the goroutine was blocked in select before, being blocked in select again when the spurious wakeup arrives may be quite likely. In this case, the spurious wakeup does no harm (but the stack was still scribbled on). * The goroutine might be sleeping in semacquire (mutex slow path). If it wakes up, that is taken as a signal to try for the semaphore again, not a signal that the semaphore is now held, but the next iteration around the loop will queue the sudog a second time, causing a cycle in the wakeup list for the given address. If that sudog is the only one in the list, when it is eventually dequeued, it will (due to the precise way the code is written) leave the sudog on the queue inactive with the sudog broken. But the sudog will also be in the free list, and that will eventually cause confusion. * The goroutine might be sleeping in notifyListWait, for sync.Cond. If it wakes up, (*Cond).Wait returns. The docs say "Unlike in other systems, Wait cannot return unless awoken by Broadcast or Signal," so the spurious wakeup is incorrect behavior, but most callers do not depend on that fact. Eventually the condition will happen, attempting the real wakeup of the goroutine and leading recursively to any of the cases in this list. * The goroutine might be sleeping in timeSleep aka time.Sleep. If it wakes up, it will continue running, leaving a timer ticking. When that time bomb goes off, it will try to ready the goroutine again, leading to any one of the cases in this list. * The goroutine cannot be sleeping in timerproc, because timerproc never uses select. * The goroutine might be sleeping in ReadTrace. If it wakes up, it will print 'runtime: spurious wakeup of trace reader' and return nil. All future calls to ReadTrace will print 'runtime: ReadTrace called from multiple goroutines simultaneously'. Eventually, when trace data is available, a true wakeup will be attempted, leading to any one of the cases in this list. None of these fatal errors appear in any of the trybot or dashboard logs. The 'bad g->status in ready' that happens if the goroutine is running (the most likely scenario anyway) has happened once on the dashboard and eight times in trybot logs. Of the eight, five were atomicstatus=8 during net/http tests, so almost certainly this bug. The other three were atomicstatus=2, all near code in select, but in a draft CL by Dmitry that was rewriting select and may or may not have had its own bugs. This bug has existed since Go 1.4. Until then the select code was implemented in C, 'done uint32' was a C stack variable 'uint32 done', and C stacks never moved. I believe it has become more common recently because of Brad's work to run more and more tests in net/http in parallel, which lengthens race windows. The fix is to run step 6 on the system stack, avoiding possibility of stack growth. Fixes #17007 and possibly other mysterious failures. Change-Id: I9d6575a51ac96ae9d67ec24da670426a4a45a317 Reviewed-on: https://go-review.googlesource.com/34835 Run-TryBot: Russ Cox <rsc@golang.org> Reviewed-by: Austin Clements <austin@google.com> Reviewed-on: https://go-review.googlesource.com/35637 Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
Showing
Please
register
or
sign in
to comment