INVARIANTS kernels may trigger a KASSERT panic from sndbuf_acquire(),
when fuzzing write(2) using stress2, because of a race in chn_write().
In the case of chn_write(), what sndbuf_acquire() does is extend the
ready-to-read area of the buffer by a specified amount of bytes. The
KASSERT in question makes sure the number of bytes we want to extend the
ready area by, is less than or equal to the number of free bytes in the
buffer. This makes sense, because we cannot extend the ready area to
something larger than what is available (i.e., free) in the first place.
What chn_write() currently does for every write is; calculate the
appropriate write size, let's say X, unlock the channel, uiomove() X
bytes to the channel's buffer, lock the channel, and call
sndbuf_acquire() to extend the ready area by X bytes. The problem with
this approach, however, is the following.
Suppose an empty channel buffer with a length of 1024 bytes, and 2
threads, (A) and (B), where (B) is a higher-priority one. Suppose thread
(A) wants to write 1024 bytes. It unlocks the channel and uiomove()s
1024 bytes to the channel buffer. At the same time, thread (B) picks up
the lock, and because it is higher priority, it keeps dominating the
lock for a few iterations. By the time thread (A) picks up the lock
again, it tries to call sndbuf_acquire() with an size of 1024 bytes,
which was calculated before it performed the uiomove(). In this case,
there is a very high chance that the buffer will not be empty, that is,
have a free area of 1024 bytes, as was the case when thread (A) started
executing, and so the KASSERT will trigger a panic because the condition
(bytes <= free) is not met.
Another scenario that can trigger a panic is the following: suppose a
buffer with a size of 4 bytes, and two threads: (A) and (B). In the
first iteration, thread (A) wants to write 2 bytes, while the buffer is
empty, BUT the pointer (sndbuf_getfreeptr()) is at the end (i.e.,
buf[3]). In the first iteration of the loop, because of the way we
calculate t, we'll end up writing only 1 byte, so after sz -= t, sz will
be 1, and so we'll need one more iteration in the inner loop, to write
the remaining 1 byte. Now we're at the end of the first loop, thread (A)
unlocks the channel, it has written 1 byte, it needs to write 1 more,
and the buffer is left with 3 empty slots. Now thread (B) picks up the
lock, and it wants to write 3 (or more) bytes. Eventually it writes the
3 bytes, and it leaves the buffer with 0 free slots. By the time thread
(A) picks up the lock again, and continues with the second iteration of
the inner loop, it will try to write the last byte, but sndbuf_acquire()
will panic because there is no free space anymore.
To fix this, get rid of the inner loop and calculate the write size on
each iteration. That means that in the scenarios explained above, we'll
end up entering the chn_sleep() case. Modify it as well, so that we do
not kill the channel if we need to sleep more.
Do the same for chn_read() to avoid possible similar panics from
sndbuf_dispose().
Reported by: pho
Tested by: christos, pho
Sponsored by: The FreeBSD Foundation
MFC after: 1 week