[-cc Alan Cox] Sebastian, please re-test your g_nokia+dummy_hcd testcase with this series. Sasha and Dave, my trinity testbeds die in other areas right now; I would really appreciate if you would please re-test this series. Michael and Shawn, I'd appreciate if you test with this series although I know it won't WARN because this patchset removes it. Back in early December I realized that a classic recursive read/write semaphore with writer priority was the ideal mechanism for handling the line discipline referencing. Line discipline references act as "readers"; closing or changing the line discipline is prevented while these references are outstanding. Conversely, line discipline references should not be granted while the line discipline is closing or changing; these are the "writers". Unfortunately, the existing rwsem does not allowed recursive read acquire, uses a FIFO priority for waiting threads, and does not support timeouts. So this implements just that: a read-recursive, writer-priority read/write semaphore with timed waits. Read recursion is handled with a small (192-bit) bitmap indexed by a prime modulo-reduced 'current' value. When a read lock is obtained, the corresponding bit is set. If a write lock is waiting when a subsequent read lock is tried, before waiting the read lock, the bitmap is tested to determine if this thread already owns a read lock. This approach may allow a non-recursive read lock (if, for example, 2 thread ptrs hash to the same value) which is not an error; it simply delays the waiting writer further. More importantly, this approach will not mistakenly wait a recursive read, which will end in deadlock. Note for the following discussion: the existing code livelocks under _any_ recursive tty_ldisc_ref_wait() call when the line discipline is halted. This is the hangup livelock first observed by Sasha Levin. Two weaknesses exist in this implementation of recursion testing: 1) the bitmap may not be large enough to avoid a 'thundering herd' livelock. For example, say 1000 processes all have the same tty open and are always reading from it. These 1000 processes will saturate the bitmap so it will appear as if every process is recursively reading so a read lock will always be granted. Thus the write lock will never be granted. 2) the bitmap is only cleared once a write lock is granted. So a tty that never closes will, over time, develop a saturated bitmap of apparent recursive reads even though those read locks were released. This needs to be addressed, perhaps by periodically attempting a dummy write lock. Initially, it was my intention to have this in 2 patchsets but since the v2 and v3 patchsets were pushed back, this ends up being an all-in-one :) Other changes in v4 from v3: - From Jiri's review of v3: 'tty: Add diagnostic for halted line discipline' was split: the function relocation is now in a separate patch, 'tty: Relocate tty_ldisc_halt() to avoid forward declaration' 'n_tty: Factor packet mode status change for reuse': packet_mode_flush() was renamed n_tty_packet_mode_flush() 'tty: Refactor wait for ldisc refs out of tty_ldisc_hangup()': the parentheses were removed in the return value conversion. 'tty: Strengthen no-subsequent-use guarantee of tty_ldisc_halt()': although addressed in the follow-on patch, the timeout values are now identical so as to avoid confusion in reviewing. 'tty: Kick waiters _after_ the ldisc is locked' was dropped: although my logic is solid, this change made the conversion to a r/w semaphore awkward-looking. I think it makes sense to fix this separately in n_tty anyway. 'tty: Remove unnecessary buffer work flush' was merged with new patch 'tty: Complete ownership transfer of flip buffers' Patches 19-31 implement the switch to ldsem. Patch 32 removes the 'tty is NULL' diagnostic. The logic supporting this change is in the commit message but I'll repeat it here: Now that the driver i/o path is separate from tty lifetimes (implemented in Jiri's last patch series, soon to be in 3.9-rc1), a driver may unknowingly submit i/o to a tty that no longer exists. There is little sense in WARNing about an expected outcome. Patch 14/32 'tty: Complete ownership transfer of flip buffers' ensures that no bad will come of the superfluous work -- other than that work for no-good-reason was submitted in the first place -- by waiting for work that may have retrieved what will soon be a stale tty value and by cancelling outstanding work before the port is destroyed (the work is owned by the port and contained within its structure). As before, this series passes the stress tests that Ilya wrote plus some new ones that I have written. Unfortunately, there is no platform that I can run trinity on that doesn't blow up on other 'next' bugs. Peter Hurley (32): tty: Add diagnostic for halted line discipline n_tty: Factor packet mode status change for reuse n_tty: Don't flush buffer when closing ldisc tty: Refactor wait for ldisc refs out of tty_ldisc_hangup() tty: Remove unnecessary re-test of ldisc ref count tty: Fix ldisc halt sequence on hangup tty: Relocate tty_ldisc_halt() to avoid forward declaration tty: Strengthen no-subsequent-use guarantee of tty_ldisc_halt() tty: Halt both ldiscs concurrently tty: Wait for SAK work before waiting for hangup work n_tty: Correct unthrottle-with-buffer-flush comments n_tty: Fully initialize ldisc before restarting buffer work tty: Don't reenable already enabled ldisc tty: Complete ownership transfer of flip buffers tty: Make core responsible for synchronizing its work tty: Fix 'deferred reopen' ldisc comment tty: Bracket ldisc release with TTY_DEBUG_HANGUP messages tty: Add ldisc hangup debug messages tty: Don't protect atomic operation with mutex tty: Separate release semantics of ldisc reference tty: Document unsafe ldisc reference acquire tty: Fold one-line assign function into callers tty: Locate get/put ldisc functions together tty: Remove redundant tty_wait_until_sent() tty: Add read-recursive, writer-prioritized rw semaphore tty: Add lock/unlock ldisc pair functions tty: Replace ldisc locking with ldisc_sem tty: Clarify ldisc variable tty: Fix hangup race with TIOCSETD ioctl tty: Clarify multiple-references comment in TIOCSETD ioctl tty: Fix tty_ldisc_lock name collision tty: Drop "tty is NULL" flip buffer diagnostic drivers/tty/Makefile | 2 +- drivers/tty/n_tty.c | 62 +++--- drivers/tty/tty_buffer.c | 4 +- drivers/tty/tty_io.c | 33 ++- drivers/tty/tty_ldisc.c | 551 ++++++++++++++++------------------------------ drivers/tty/tty_ldsem.c | 507 ++++++++++++++++++++++++++++++++++++++++++ drivers/tty/tty_port.c | 1 + include/linux/tty.h | 7 +- include/linux/tty_ldisc.h | 50 ++++- 9 files changed, 816 insertions(+), 401 deletions(-) create mode 100644 drivers/tty/tty_ldsem.c -- 1.8.1.2 -- To unsubscribe from this list: send the line "unsubscribe linux-serial" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
