and flags with an sxlock. This leads to a significant and measurable
performance improvement as a result of access to shared locking for
frequent lookup operations, reduced general overhead, and reduced overhead
in the event of contention. All of these are imported for threaded
applications where simultaneous access to a shared file descriptor array
occurs frequently. Kris has reported 2x-4x transaction rate improvements
on 8-core MySQL benchmarks; smaller improvements can be expected for many
workloads as a result of reduced overhead.
- Generally eliminate the distinction between "fast" and regular
acquisisition of the filedesc lock; the plan is that they will now all
be fast. Change all locking instances to either shared or exclusive
locks.
- Correct a bug (pointed out by kib) in fdfree() where previously msleep()
was called without the mutex held; sx_sleep() is now always called with
the sxlock held exclusively.
- Universally hold the struct file lock over changes to struct file,
rather than the filedesc lock or no lock. Always update the f_ops
field last. A further memory barrier is required here in the future
(discussed with jhb).
- Improve locking and reference management in linux_at(), which fails to
properly acquire vnode references before using vnode pointers. Annotate
improper use of vn_fullpath(), which will be replaced at a future date.
In fcntl(), we conservatively acquire an exclusive lock, even though in
some cases a shared lock may be sufficient, which should be revisited.
The dropping of the filedesc lock in fdgrowtable() is no longer required
as the sxlock can be held over the sleep operation; we should consider
removing that (pointed out by attilio).
Tested by: kris
Discussed with: jhb, kris, attilio, jeff
sleep lock missed the witness code, and the system will panic
immediately on boot if WITNESS is enabled.
Changed the witness definition to the new type.
wait (time waited to acquire) and hold times for *all* kernel locks. If
the architecture has a system synchronized TSC, the profiling code will
use that - thereby minimizing profiling overhead. Large chunks of profiling
code have been moved out of line, the overhead measured on the T1 for when
it is compiled in but not enabled is < 1%.
Approved by: scottl (standing in for mentor rwatson)
Reviewed by: des and jhb
specific privilege names to a broad range of privileges. These may
require some future tweaking.
Sponsored by: nCircle Network Security, Inc.
Obtained from: TrustedBSD Project
Discussed on: arch@
Reviewed (at least in part) by: mlaier, jmg, pjd, bde, ceri,
Alex Lyashkov <umka at sevcity dot net>,
Skip Ford <skip dot ford at verizon dot net>,
Antoine Brodin <antoine dot brodin at laposte dot net>
in syscons. This replaces a simple access semaphore that was assumed to be
protected by Giant but often was not. If two threads that were otherwise
SMP-safe called printf at the same time, there was a high likelyhood that
the semaphore would get corrupted and result in a permanently frozen video
console. This is similar to what is already done in the serial console
drivers.
placeholder similar to KTR_DEV. Explain the use of KTR_DEV and
KTR_SUBSYS in a comment as well.
- Retire KTR_WITNESS and instead have KTR_WITNESS default to off but use
KTR_SUBSYS if it is enabled.
lock_obj objects:
- Add new lock_init() and lock_destroy() functions to setup and teardown
lock_object objects including KTR logging and registering with WITNESS.
- Move all the handling of LO_INITIALIZED out of witness and the various
lock init functions into lock_init() and lock_destroy().
- Remove the constants for static indices into the lock_classes[] array
and change the code outside of subr_lock.c to use LOCK_CLASS to compare
against a known lock class.
- Move the 'show lock' ddb function and lock_classes[] array out of
kern_mutex.c over to subr_lock.c.
struct sx). Instead of storing a direct pointer to a our lock_class
struct in lock_object, reserve 4 bits in the lo_flags field to serve as an
index into a global lock_classes array that contains pointers to the lock
classes. Only debugging code such as WITNESS or INVARIANTS checks and KTR
logging need to access the lock_class member, so this shouldn't add any
overhead to production kernels. It might add some slight overhead to
kernels using those debug options however.
As with the previous set of changes to lock_object, this is going to
completely obliterate the kernel ABI, so be sure to recompile all your
modules.
spin locks that are not in the static order list. It is not safe to call
printf while holding the witness spin mutex since the console drivers that
back printf may need to use their own spin locks which would try to talk
to witness when they were locked. Given this, it is possible for one
CPU to lock a console driver lock (such as sio) which then tries to lock
the witness lock while another CPU is doing the printf while holding the
witness lock. Fix this by moving the printf outside of the witness lock.
All other printf's in witness are already correct.
MFC after: 3 days
lock object (and thus off of each mutex and sx lock):
- Rename the all_locks list to pending_locks and only put locks initialized
before SI_SUB_WITNESS on the list so that the SI_SUB_WITNESS can add them
to witness once it starts up.
- Now that pending_locks is only used during early startup, change it from
a TAILQ to an STAILQ. This removes a pointer from the STAILQ_ENTRY in
struct lock_object.
- Since the pending_locks list is only used during the single-threaded
early boot it no longer needs to be protected by a mutex, so remove
all_mtx.
- Since the lo_list member of struct lock_object is now only used during
early boot before witness is running, collapse lo_list and lo_witness
into a union. This shaves the second pointer off of struct lock_object.
- Axe lock_cur_cnt and lock_max_cnt.
With these changes, struct mtx shrinks from 36 to 28 bytes on 32-bit
platforms and from 72 to 56 bytes on 64-bit platforms. Note that this
commit will completely and utterly destroy the kernel ABI, so no MFC.
Tested on: alpha, amd64, i386, sparc64
and increase flexibility to allow various different approaches to be tried
in the future.
- Split struct ithd up into two pieces. struct intr_event holds the list
of interrupt handlers associated with interrupt sources.
struct intr_thread contains the data relative to an interrupt thread.
Currently we still provide a 1:1 relationship of events to threads
with the exception that events only have an associated thread if there
is at least one threaded interrupt handler attached to the event. This
means that on x86 we no longer have 4 bazillion interrupt threads with
no handlers. It also means that interrupt events with only INTR_FAST
handlers no longer have an associated thread either.
- Renamed struct intrhand to struct intr_handler to follow the struct
intr_foo naming convention. This did require renaming the powerpc
MD struct intr_handler to struct ppc_intr_handler.
- INTR_FAST no longer implies INTR_EXCL on all architectures except for
powerpc. This means that multiple INTR_FAST handlers can attach to the
same interrupt and that INTR_FAST and non-INTR_FAST handlers can attach
to the same interrupt. Sharing INTR_FAST handlers may not always be
desirable, but having sio(4) and uhci(4) fight over an IRQ isn't fun
either. Drivers can always still use INTR_EXCL to ask for an interrupt
exclusively. The way this sharing works is that when an interrupt
comes in, all the INTR_FAST handlers are executed first, and if any
threaded handlers exist, the interrupt thread is scheduled afterwards.
This type of layout also makes it possible to investigate using interrupt
filters ala OS X where the filter determines whether or not its companion
threaded handler should run.
- Aside from the INTR_FAST changes above, the impact on MD interrupt code
is mostly just 's/ithread/intr_event/'.
- A new MI ddb command 'show intrs' walks the list of interrupt events
dumping their state. It also has a '/v' verbose switch which dumps
info about all of the handlers attached to each event.
- We currently don't destroy an interrupt thread when the last threaded
handler is removed because it would suck for things like ppbus(8)'s
braindead behavior. The code is present, though, it is just under
#if 0 for now.
- Move the code to actually execute the threaded handlers for an interrrupt
event into a separate function so that ithread_loop() becomes more
readable. Previously this code was all in the middle of ithread_loop()
and indented halfway across the screen.
- Made struct intr_thread private to kern_intr.c and replaced td_ithd
with a thread private flag TDP_ITHREAD.
- In statclock, check curthread against idlethread directly rather than
curthread's proc against idlethread's proc. (Not really related to intr
changes)
Tested on: alpha, amd64, i386, sparc64
Tested on: arm, ia64 (older version of patch by cognet and marcel)
any other non-sleepable lock. In plain English: Giant comes before all
other mutexes.
- Add some extra description to the lock order reversal printf's to indicate
when a reversal is triggered by a hard-coded implicit rule.
Requested by: truckman (2)
MFC after: 1 week
link proctree and allproc to Giant since that order is already implicitly
enforced.
- Use a goto to handle the case where we want to enforce a reversal before
calling isitmydescendant() in witness_checkorder() so that the logic is
easier to follow and so that it is easier to add more forced-reversal
cases in the future.
MFC after: 3 days
if an indirect relationship exists (keep both A->B->C and A->C).
This allows witness_checkorder() to use isitmychild() instead of
the much more expensive isitmydescendant() to check for valid lock
ordering.
Don't do an expensive tree walk to update the w_level values when
the tree is updated. Only update the w_level values when using the
debugger to display the tree.
Nuke the experimental "witness_watch > 1" mode that only compared
w_level for the two locks. This information is no longer maintained
at run time, and the use of isitmychild() in witness_checkorder
should bring performance close enough to the acceptable level that
this hack is not needed.
Report witness data structure allocation statistics under the
debug.witness sysctl.
Reviewed by: jhb
MFC after: 30 days
list lock, as there has been a report that an alternative lock order
is getting introduced. This should help ferret it out.
Reported by: Ed Maste <emaste at phaedrus dot sandvine dot ca>
lists, as well as accessor macros. For now, this is a recursive mutex
due code sequences where IPv4 multicast calls into IGMP calls into
ip_output(), which then tests for a multicast forwarding case.
For support macros in in_var.h to check multicast address lists, assert
that in_multi_mtx is held.
Acquire in_multi_mtx around iteration over the IPv4 multicast address
lists, such as in ip_input() and ip_output().
Acquire in_multi_mtx when manipulating the IPv4 layer multicast addresses,
as well as over the manipulation of ifnet multicast address lists in order
to keep the two layers in sync.
Lock down accesses to IPv4 multicast addresses in IGMP, or assert the
lock when performing IGMP join/leave events.
Eliminate spl's associated with IPv4 multicast addresses, portions of
IGMP that weren't previously expunged by IGMP locking.
Add in_multi_mtx, igmp_mtx, and if_addr_mtx lock order to hard-coded
lock order in WITNESS, in that order.
Problem reported by: Ed Maste <emaste at phaedrus dot sandvine dot ca>
MFC after: 10 days
mutex instead of a MTX_DEF one in order to defer preemption while
reading the date and time registers. If we don't manage to read them
within the time slot where we are guaranteed that no updates occur we
might actually read them during an update in which case the output is
undefined.
3ware's 9xxx series controllers. This corresponds to
the 9.2 release (for FreeBSD 5.2.1) on the 3ware website.
Highlights of this release are:
1. The driver has been re-architected to use a "Common Layer"
(all tw_cl* files), which is a consolidation of all OS-independent
parts of the driver. The FreeBSD OS specific portions of the
driver go into an "OS Layer" (all tw_osl* files).
This re-architecture is to achieve better maintainability, consistency
of behavior across OS's, and better portability to new OS's (drivers
for new OS's can be written by just adding an OS Layer that's specific
to the OS, by complying to a "Common Layer Programming Interface" API.
2. The driver takes advantage of multiple processors.
3. The driver has a new firmware image bundled, the new features of which
include Online Capacity Expansion and multi-lun support, among others.
More details about 3ware's 9.2 release can be found here:
http://www.3ware.com/download/Escalade9000Series/9.2/9.2_Release_Notes_Web.pdf
Since the Common Layer is used across OS's, the FreeBSD specific include
path for header files (/sys/dev/twa) is not part of the #include pre-processor
directive in any of the source files. For being able to integrate twa into
the kernel despite this, Makefile.<arch> has been changed to add the include
path to CFLAGS.
Reviewed by: scottl
at some point result in a status event being triggered (it should
be a link down event: the Microsoft driver design guide says you
should generate one when the NIC is initialized). Some drivers
generate the event during MiniportInitialize(), such that by the
time MiniportInitialize() completes, the NIC is ready to go. But
some drivers, in particular the ones for Atheros wireless NICs,
don't generate the event until after a device interrupt occurs
at some point after MiniportInitialize() has completed.
The gotcha is that you have to wait until the link status event
occurs one way or the other before you try to fiddle with any
settings (ssid, channel, etc...). For the drivers that set the
event sycnhronously this isn't a problem, but for the others
we have to pause after calling ndis_init_nic() and wait for the event
to arrive before continuing. Failing to wait can cause big trouble:
on my SMP system, calling ndis_setstate_80211() after ndis_init_nic()
completes, but _before_ the link event arrives, will lock up or
reset the system.
What we do now is check to see if a link event arrived while
ndis_init_nic() was running, and if it didn't we msleep() until
it does.
Along the way, I discovered a few other problems:
- Defered procedure calls run at PASSIVE_LEVEL, not DISPATCH_LEVEL.
ntoskrnl_run_dpc() has been fixed accordingly. (I read the documentation
wrong.)
- Similarly, the NDIS interrupt handler, which is essentially a
DPC, also doesn't need to run at DISPATCH_LEVEL. ndis_intrtask()
has been fixed accordingly.
- MiniportQueryInformation() and MiniportSetInformation() run at
DISPATCH_LEVEL, and each request must complete before another
can be submitted. ndis_get_info() and ndis_set_info() have been
fixed accordingly.
- Turned the sleep lock that guards the NDIS thread job list into
a spin lock. We never do anything with this lock held except manage
the job list (no other locks are held), so it's safe to do this,
and it's possible that ndis_sched() and ndis_unsched() can be
called from DISPATCH_LEVEL, so using a sleep lock here is
semantically incorrect. Also updated subr_witness.c to add the
lock to the order list.
witness_proc_has_locks(), as they are unused, which results in a compiler
error. This problem was introduced with the implementation of "show
alllocks".
Spotted by: Artem Kuchin <matrix at itlegion dot ru>
of lock types in the kernel. This results in an increase of witness
data usage from ~145k to ~280k on i386 for kernels with
'options WITNESS'.
- Remove the unused witness malloc bucket.
Submitted by: Michal Mertl mime at traveller dot cz (1)
and threads currently holding sleep mutexes (and spin mutexes for
curthread). This can be quite useful in looking for a lock condition
summary for a system, as it avoids manually iterating through threads
and processes to find all the interesting locks.
NB: "alllocks" is up there with "lockedvnods" for a bad argument for
show.
MFC after: 2 weeks
remove previous entropy harvesting mutex names as they are no longer
present. Commit to this file was ommitted when randomdev_soft.c:1.5
was made.
Feet shot: Robert Huff <roberthuff at rcn dot com>
when the spin lock in question isn't -- it's the critical_enter() that
KDB set. No more panic in DDB for console -> syscons -> tty -> knote
operations.
