This avoids the need for sched_bind() in the default case so that you
can start up the NDIS subsystem at boot time when only CPU 0 is running.
There are potentially ways to fix it so that the DPC threads aren't
started until after the other CPUs are launched, but doing it correctly
is tricky. You need to defer the startup of the ntoskrnl subsystem
(ntoskrnl_libinit()), not just defer ndis_attach().
For now, I don't think it will make much difference having just the
single DPC thread (I started out with just one anyway). Note that this
turns the KeSetTargetProcessorDpc() routine into a no-op, since the
CPU number in struct kdpc is now ignored.
is KeRaiseIrql(newirql, &oldirql), not oldirql = KeRaiseIrql(newirql).
(The macro ultimately translates to KfRaiseIrql() which does use
the latter API, so this has no effect on generated code.)
Also, wait for thread termination the right way: kthread_exit()
will ultimately do a wakeup(td->td_proc). This is the event we
should wait on. Eliminate the previous synchronization machinery
for this since it was never guaranteed to work correctly.
processor, to insure DPC thread 0 runs on CPU0, DPC thread 1 runs on
CPU1, and so on.
Elevate the priority of the workitem threads, though don't use as
high a priority as the DPC threads.
- Change ndis_return() from a DPC to a workitem so that it doesn't
run at DISPATCH_LEVEL (with the dispatcher lock held).
- In if_ndis.c, submit packets to the stack via (*ifp->if_input)() in
a workitem instead of doing it directly in ndis_rxeof(), because
ndis_rxeof() runs in a DPC, and hence at DISPATCH_LEVEL. This
implies that the 'dispatch level' mutex for the current CPU is
being held, and we don't want to call if_input while holding
any locks.
- Reimplement IoConnectInterrupt()/IoDisconnectInterrupt(). The original
approach I used to track down the interrupt resource (by scanning
the device tree starting at the nexus) is prone to problems when
two devices share an interrupt. (E.g removing ndis1 might disable
interrupts for ndis0.) The new approach is to multiplex all the
NDIS interrupts through a common internal dispatcher (ntoskrnl_intr())
and allow IoConnectInterrupt()/IoDisconnectInterrupt() to add or
remove interrupts from the dispatch list.
- Implement KeAcquireInterruptSpinLock() and KeReleaseInterruptSpinLock().
- Change the DPC and workitem threads to use the KeXXXSpinLock
API instead of mtx_lock_spin()/mtx_unlock_spin().
- Simplify the NdisXXXPacket routines by creating an actual
packet pool structure and using the InterlockedSList routines
to manage the packet queue.
- Only honor the value returned by OID_GEN_MAXIMUM_SEND_PACKETS
for serialized drivers. For deserialized drivers, we now create
a packet array of 64 entries. (The Microsoft DDK documentation
says that for deserialized miniports, OID_GEN_MAXIMUM_SEND_PACKETS
is ignored, and the driver for the Marvell 8335 chip, which is
a deserialized miniport, returns 1 when queried.)
- Clean up timer handling in subr_ntoskrnl.
- Add the following conditional debugging code:
NTOSKRNL_DEBUG_TIMERS - add debugging and stats for timers
NDIS_DEBUG_PACKETS - add extra sanity checking for NdisXXXPacket API
NTOSKRNL_DEBUG_SPINLOCKS - add test for spinning too long
- In kern_ndis.c, always start the HAL first and shut it down last,
since Windows spinlocks depend on it. Ntoskrnl should similarly be
started second and shut down next to last.
First and most importantly, I threw out the thread priority-twiddling
implementation of KeRaiseIrql()/KeLowerIrq()/KeGetCurrentIrql() in
favor of a new scheme that uses sleep mutexes. The old scheme was
really very naughty and sought to provide the same behavior as
Windows spinlocks (i.e. blocking pre-emption) but in a way that
wouldn't raise the ire of WITNESS. The new scheme represents
'DISPATCH_LEVEL' as the acquisition of a per-cpu sleep mutex. If
a thread on cpu0 acquires the 'dispatcher mutex,' it will block
any other thread on the same processor that tries to acquire it,
in effect only allowing one thread on the processor to be at
'DISPATCH_LEVEL' at any given time. It can then do the 'atomic sit
and spin' routine on the spinlock variable itself. If a thread on
cpu1 wants to acquire the same spinlock, it acquires the 'dispatcher
mutex' for cpu1 and then it too does an atomic sit and spin to try
acquiring the spinlock.
Unlike real spinlocks, this does not disable pre-emption of all
threads on the CPU, but it does put any threads involved with
the NDISulator to sleep, which is just as good for our purposes.
This means I can now play nice with WITNESS, and I can safely do
things like call malloc() when I'm at 'DISPATCH_LEVEL,' which
you're allowed to do in Windows.
Next, I completely re-wrote most of the event/timer/mutex handling
and wait code. KeWaitForSingleObject() and KeWaitForMultipleObjects()
have been re-written to use condition variables instead of msleep().
This allows us to use the Windows convention whereby thread A can
tell thread B "wake up with a boosted priority." (With msleep(), you
instead have thread B saying "when I get woken up, I'll use this
priority here," and thread A can't tell it to do otherwise.) The
new KeWaitForMultipleObjects() has been better tested and better
duplicates the semantics of its Windows counterpart.
I also overhauled the IoQueueWorkItem() API and underlying code.
Like KeInsertQueueDpc(), IoQueueWorkItem() must insure that the
same work item isn't put on the queue twice. ExQueueWorkItem(),
which in my implementation is built on top of IoQueueWorkItem(),
was also modified to perform a similar test.
I renamed the doubly-linked list macros to give them the same names
as their Windows counterparts and fixed RemoveListTail() and
RemoveListHead() so they properly return the removed item.
I also corrected the list handling code in ntoskrnl_dpc_thread()
and ntoskrnl_workitem_thread(). I realized that the original logic
did not correctly handle the case where a DPC callout tries to
queue up another DPC. It works correctly now.
I implemented IoConnectInterrupt() and IoDisconnectInterrupt() and
modified NdisMRegisterInterrupt() and NdisMDisconnectInterrupt() to
use them. I also tried to duplicate the interrupt handling scheme
used in Windows. The interrupt handling is now internal to ndis.ko,
and the ndis_intr() function has been removed from if_ndis.c. (In
the USB case, interrupt handling isn't needed in if_ndis.c anyway.)
NdisMSleep() has been rewritten to use a KeWaitForSingleObject()
and a KeTimer, which is how it works in Windows. (This is mainly
to insure that the NDISulator uses the KeTimer API so I can spot
any problems with it that may arise.)
KeCancelTimer() has been changed so that it only cancels timers, and
does not attempt to cancel a DPC if the timer managed to fire and
queue one up before KeCancelTimer() was called. The Windows DDK
documentation seems to imply that KeCantelTimer() will also call
KeRemoveQueueDpc() if necessary, but it really doesn't.
The KeTimer implementation has been rewritten to use the callout API
directly instead of timeout()/untimeout(). I still cheat a little in
that I have to manage my own small callout timer wheel, but the timer
code works more smoothly now. I discovered a race condition using
timeout()/untimeout() with periodic timers where untimeout() fails
to actually cancel a timer. I don't quite understand where the race
is, using callout_init()/callout_reset()/callout_stop() directly
seems to fix it.
I also discovered and fixed a bug in winx32_wrap.S related to
translating _stdcall calls. There are a couple of routines
(i.e. the 64-bit arithmetic intrinsics in subr_ntoskrnl) that
return 64-bit quantities. On the x86 arch, 64-bit values are
returned in the %eax and %edx registers. However, it happens
that the ctxsw_utow() routine uses %edx as a scratch register,
and x86_stdcall_wrap() and x86_stdcall_call() were only preserving
%eax before branching to ctxsw_utow(). This means %edx was getting
clobbered in some cases. Curiously, the most noticeable effect of this
bug is that the driver for the TI AXC110 chipset would constantly drop
and reacquire its link for no apparent reason. Both %eax and %edx
are preserved on the stack now. The _fastcall and _regparm
wrappers already handled everything correctly.
I changed if_ndis to use IoAllocateWorkItem() and IoQueueWorkItem()
instead of the NdisScheduleWorkItem() API. This is to avoid possible
deadlocks with any drivers that use NdisScheduleWorkItem() themselves.
The unicode/ansi conversion handling code has been cleaned up. The
internal routines have been moved to subr_ntoskrnl and the
RtlXXX routines have been exported so that subr_ndis can call them.
This removes the incestuous relationship between the two modules
regarding this code and fixes the implementation so that it honors
the 'maxlen' fields correctly. (Previously it was possible for
NdisUnicodeStringToAnsiString() to possibly clobber memory it didn't
own, which was causing many mysterious crashes in the Marvell 8335
driver.)
The registry handling code (NdisOpen/Close/ReadConfiguration()) has
been fixed to allocate memory for all the parameters it hands out to
callers and delete whem when NdisCloseConfiguration() is called.
(Previously, it would secretly use a single static buffer.)
I also substantially updated if_ndis so that the source can now be
built on FreeBSD 7, 6 and 5 without any changes. On FreeBSD 5, only
WEP support is enabled. On FreeBSD 6 and 7, WPA-PSK support is enabled.
The original WPA code has been updated to fit in more cleanly with
the net80211 API, and to eleminate the use of magic numbers. The
ndis_80211_setstate() routine now sets a default authmode of OPEN
and initializes the RTS threshold and fragmentation threshold.
The WPA routines were changed so that the authentication mode is
always set first, followed by the cipher. Some drivers depend on
the operations being performed in this order.
I also added passthrough ioctls that allow application code to
directly call the MiniportSetInformation()/MiniportQueryInformation()
methods via ndis_set_info() and ndis_get_info(). The ndis_linksts()
routine also caches the last 4 events signalled by the driver via
NdisMIndicateStatus(), and they can be queried by an application via
a separate ioctl. This is done to allow wpa_supplicant to directly
program the various crypto and key management options in the driver,
allowing things like WPA2 support to work.
Whew.
as a part of the GENERIC kernel with INVARIANT* and WITNESS*
turned off.
(For non GENERIC kernel KTR and MUTEX_PROFILING should be also
off).
Submitted by: Eygene A. Ryabinkin <rea at rea dot mbslab dot kiae dot ru>
Approved by: re (scottl)
PR: 81767
struct ifnet or the layer 2 common structure it was embedded in have
been replaced with a struct ifnet pointer to be filled by a call to the
new function, if_alloc(). The layer 2 common structure is also allocated
via if_alloc() based on the interface type. It is hung off the new
struct ifnet member, if_l2com.
This change removes the size of these structures from the kernel ABI and
will allow us to better manage them as interfaces come and go.
Other changes of note:
- Struct arpcom is no longer referenced in normal interface code.
Instead the Ethernet address is accessed via the IFP2ENADDR() macro.
To enforce this ac_enaddr has been renamed to _ac_enaddr.
- The second argument to ether_ifattach is now always the mac address
from driver private storage rather than sometimes being ac_enaddr.
Reviewed by: sobomax, sam
We can't call KeFlushQueuedDpcs() during bootstrap (cold == 1), since
the flush operation sleeps to wait for completion, and we can't sleep
here (clowns will eat us).
On an i386 SMP system, if we're loaded/probed/attached during bootstrap,
smp_rendezvous() won't run us anywhere except CPU 0 (since the other CPUs
aren't launched until later), which means we won't be able to set up
the GDTs anywhere except CPU 0. To deal with this case, ctxsw_utow()
now checks to see if the TID for the current processor has been properly
initialized and sets up the GTD for the current CPU if not.
Lastly, in if_ndis.c:ndis_shutdown(), do an ndis_stop() to insure we
really halt the NIC and stop interrupts from happening.
Note that loading a driver during bootstrap is, unfortunately, kind of
a hit or miss sort of proposition. In Windows, the expectation is that
by the time a given driver's MiniportInitialize() method is called,
the system is already in 'multiuser' state, i.e. it's up and running
enough to support all the stuff specified in the NDIS API, which includes
the underlying OS-supplied facilities it implicitly depends on, such as
having all CPUs running, having the DPC queues initialized, WorkItem
threads running, etc. But in UNIX, a lot of that stuff won't work during
bootstrap. This causes a problem since we need to call MiniportInitialize()
at least once during ndis_attach() in order to find out what kind of NIC
we have and learn its station address.
What this means is that some cards just plain won't work right if
you try to pre-load the driver along with the kernel: they'll only be
probed/attach correctly if the driver is kldloaded _after_ the system
has reached multiuser. I can't really think of a way around this that
would still preserve the ability to use an NDIS device for diskless
booting.
prevent anything from making calls to the NIC while it's being shut down.
This is yet another attempt to stop things like mdnsd from trying to
poke at the card while it's not properly initialized and panicking
the system.
Also, remove unneeded debug message from if_ndis.c.
works again.
This driver uses NdisScheduleWorkItem(), and we have to take special steps
to insure that its workitems don't collide with any of the other workitems
used by the NDISulator. In particular, if one of the driver's work jobs
blocks, it can prevent NdisMAllocateSharedMemoryAsync() from completing
when expected.
The original hack to fix this was to have NdisMAllocateSharedMemoryAsync()
defer its work to the DPC queue instead of the general task queue. To
fix it now, I decided to add some additional workitem threads. (There's
supposed to be a pool of worker threads in Windows anyway.) Currently,
there are 4. There should be at least 2. One is reserved for the legacy
ExQueueWorkItem() API, while the others are used in round-robin by the
IoQueueWorkItem() API. NdisMAllocateSharedMemoryAsync() uses the latter
API while NdisScheduleWorkItem() uses the former, so the deadlock is
avoided.
Fixed NdisMRegisterDevice()/NdisMDeregisterDevice() to work a little
more sensibly with the new driver_object/device_object framework. It
doesn't really register a working user-mode interface, but the existing
code was completely wrong for the new framework.
Fixed a couple of bugs dealing with the cancellation of events and
DPCs. When cancelling an event that's still on the timer queue (i.e.
hasn't expired yet), reset dh_inserted in its dispatch header to FALSE.
Previously, it was left set to TRUE, which would make a cancelled
timer appear to have not been cancelled. Also, when removing a DPC
from a queue, reset its list pointers, otherwise a cancelled DPC
might mistakenly be treated as still pending.
Lastly, fix the behavior of ntoskrnl_wakeup() when dealing with
objects that have nobody waiting on them: sync event objects get
their signalled state reset to FALSE, but notification objects
should still be set to TRUE.
The Ralink RT2500 driver uses this API instead of NdisMIndicateReceivePacket().
Drivers use NdisMEthIndicateReceive() when they know they support
802.3 media and expect to hand their packets only protocols that want
to deal with that particular media type. With this API, the driver does
not manage its own NDIS_PACKET/NDIS_BUFFER structures. Instead, it
lets bound protocols have a peek at the data, and then they supply
an NDIS_PACKET/NDIS_BUFFER combo to the miniport driver, into which
it copies the packet data.
Drivers use NdisMIndicateReceivePacket() to allow their packets to
be read by any protocol, not just those bound to 802.3 media devices.
To make this work, we need an internal pool of NDIS_PACKETS for
receives. Currently, we check to see if the driver exports a
MiniportTransferData() method in its characteristics structure,
and only allocate the pool for drivers that have this method.
This should allow the RT2500 driver to work correctly, though I
still have to fix ndiscvt(8) to parse its .inf file properly.
Also, change kern_ndis.c:ndis_halt_nic() to reap timers before
acquiring NDIS_LOCK(), since the reaping process might entail sleeping
briefly (and we can't sleep with a lock held).
the same time.
Fix if_ndis_pccard.c so that it sets sc->ndis_dobj and sc->ndis_regvals.
Correct IMPORT_SFUNC() macros for the READ_PORT_BUFFER_xxx() routines,
which take 3 arguments, not 2.
This fixes it so that the Windows driver for my Cisco Aironet 340 PCMCIA
card works again. (Yes, I know the an(4) driver supports this card natively,
but it's the only PCMCIA device I have with a Windows XP driver.)
routines (_alldiv(), _allmul(), _alludiv(), _aullmul(), etc...)
that use the _stdcall calling convention.
These routines all take two arguments, but the arguments are 64 bits wide.
On the i386 this means they each consume two 32-bit slots on the stack.
Consequently, when we specify the argument count in the IMPORT_SFUNC()
macro, we have to lie and claim there are 4 arguments instead of two.
This will cause the resulting i386 assembly wrapper to push the right
number of longwords onto the stack.
This fixes a crash I discovered with the RealTek 8180 driver, which
uses these routines a lot during initialization.
technically a no-op since uintmax_t is uint64_t on all currently
supported architectures, but we should use an explicit cast instead
of depending on this obscure coincidence.
Remove unused fields from ndis_miniport_block.
Fix a bug in KeFlushQueuedDpcs() (we weren't calculating the kq pointer
correctly).
In if_ndis.c, clear the IFF_RUNNING flag before calling ndis_halt_nic().
Add some guards in kern_ndis.c to avoid letting anyone invoke ndis_get_info()
or ndis_set_info() if the NIC isn't fully initialized. Apparently, mdnsd
will sometimes try to invoke the ndis_ioctl() routine at exactly the
wrong moment (to futz with its multicast filters) when the interface
comes up, and can trigger a crash unless we guard against it.
Oh, one additional change I forgot to mention in the last commit:
NdisOpenFile() was broken in the case for firmware files that were
pre-loaded as modules. When searching for the module in NdisOpenFile(),
we would match against a symbol name, which would contain the string
we were looking for, then save a pointer to the linker file handle.
Later, in NdisMapFile(), we would refer to the filename hung off
this handle when trying to find the starting address symbol. Only
problem is, this filename is different from the embedded symbol
name we're searching for, so the mapping would fail. I found this
problem while testing the AirGo driver, which requires a small
firmware file.
- Remove the old task threads from kern_ndis.c and reimplement them in
subr_ntoskrnl.c, in order to more properly emulate the Windows DPC
API. Each CPU gets its own DPC queue/thread, and each queue can
have low, medium and high importance DPCs. New APIs implemented:
KeSetTargetProcessorDpc(), KeSetImportanceDpc() and KeFlushQueuedDpcs().
(This is the biggest change.)
- Fix a bug in NdisMInitializeTimer(): the k_dpc pointer in the
nmt_timer embedded in the ndis_miniport_timer struct must be set
to point to the DPC, also embedded in the struct. Failing to do
this breaks dequeueing of DPCs submitted via timers, and in turn
breaks cancelling timers.
- Fix a bug in KeCancelTimer(): if the timer is interted in the timer
queue (i.e. the timeout callback is still pending), we have to both
untimeout() the timer _and_ call KeRemoveQueueDpc() to nuke the DPC
that might be pending. Failing to do this breaks cancellation of
periodic timers, which always appear to be inserted in the timer queue.
- Make use of the nmt_nexttimer field in ndis_miniport_timer: keep a
queue of pending timers and cancel them all in ndis_halt_nic(), prior
to calling MiniportHalt(). Also call KeFlushQueuedDpcs() to make sure
any DPCs queued by the timers have expired.
- Modify NdisMAllocateSharedMemory() and NdisMFreeSharedMemory() to keep
track of both the virtual and physical addresses of the shared memory
buffers that get handed out. The AirGo MIMO driver appears to have a bug
in it: for one of the segments is allocates, it returns the wrong
virtual address. This would confuse NdisMFreeSharedMemory() and cause
a crash. Why it doesn't crash Windows too I have no idea (from reading
the documentation for NdisMFreeSharedMemory(), it appears to be a violation
of the API).
- Implement strstr(), strchr() and MmIsAddressValid().
- Implement IoAllocateWorkItem(), IoFreeWorkItem(), IoQueueWorkItem() and
ExQueueWorkItem(). (This is the second biggest change.)
- Make NdisScheduleWorkItem() call ExQueueWorkItem(). (Note that the
ExQueueWorkItem() API is deprecated by Microsoft, but NDIS still uses
it, since NdisScheduleWorkItem() is incompatible with the IoXXXWorkItem()
API.)
- Change if_ndis.c to use the NdisScheduleWorkItem() interface for scheduling
tasks.
With all these changes and fixes, the AirGo MIMO driver for the Belkin
F5D8010 Pre-N card now works. Special thanks to Paul Robinson
(paul dawt robinson at pwermedia dawt net) for the loan of a card
for testing.
here on in, if_ndis.ko will be pre-built as a module, and can be built
into a static kernel (though it's not part of GENERIC). Drivers are
created using the new ndisgen(8) script, which uses ndiscvt(8) under
the covers, along with a few other tools. The result is a driver module
that can be kldloaded into the kernel.
A driver with foo.inf and foo.sys files will be converted into
foo_sys.ko (and foo_sys.o, for those who want/need to make static
kernels). This module contains all of the necessary info from the
.INF file and the driver binary image, converted into an ELF module.
You can kldload this module (or add it to /boot/loader.conf) to have
it loaded automatically. Any required firmware files can be bundled
into the module as well (or converted/loaded separately).
Also, add a workaround for a problem in NdisMSleep(). During system
bootstrap (cold == 1), msleep() always returns 0 without actually
sleeping. The Intel 2200BG driver uses NdisMSleep() to wait for
the NIC's firmware to come to life, and fails to load if NdisMSleep()
doesn't actually delay. As a workaround, if msleep() (and hence
ndis_thsuspend()) returns 0, use a hard DELAY() to sleep instead).
This is not really the right thing to do, but we can't really do much
else. At the very least, this makes the Intel driver happy.
There are probably other drivers that fail in this way during bootstrap.
Unfortunately, the only workaround for those is to avoid pre-loading
them and kldload them once the system is running instead.
layer, but with a twist.
The twist has to do with the fact that Microsoft supports structured
exception handling in kernel mode. On the i386 arch, exception handling
is implemented by hanging an exception registration list off the
Thread Environment Block (TEB), and the TEB is accessed via the %fs
register. The problem is, we use %fs as a pointer to the pcpu stucture,
which means any driver that tries to write through %fs:0 will overwrite
the curthread pointer and make a serious mess of things.
To get around this, Project Evil now creates a special entry in
the GDT on each processor. When we call into Windows code, a context
switch routine will fix up %fs so it points to our new descriptor,
which in turn points to a fake TEB. When the Windows code returns,
or calls out to an external routine, we swap %fs back again. Currently,
Project Evil makes use of GDT slot 7, which is all 0s by default.
I fully expect someone to jump up and say I can't do that, but I
couldn't find any code that makes use of this entry anywhere. Sadly,
this was the only method I could come up with that worked on both
UP and SMP. (Modifying the LDT works on UP, but becomes incredibly
complicated on SMP.) If necessary, the context switching stuff can
be yanked out while preserving the convention calling wrappers.
(Fortunately, it looks like Microsoft uses some special epilog/prolog
code on amd64 to implement exception handling, so the same nastiness
won't be necessary on that arch.)
The advantages are:
- Any driver that uses %fs as though it were a TEB pointer won't
clobber pcpu.
- All the __stdcall/__fastcall/__regparm stuff that's specific to
gcc goes away.
Also, while I'm here, switch NdisGetSystemUpTime() back to using
nanouptime() again. It turns out nanouptime() is way more accurate
than just using ticks(). On slower machines, the Atheros drivers
I tested seem to take a long time to associate due to the loss
in accuracy.
ndis_timercall() in NdisMInitializeTimer(), we can't use the raw
function pointer. This is because ntoskrnl_run_dpc() expects to
invoke a function with Microsoft calling conventions. On i386,
this works because ndis_timercall() is declared with the __stdcall
attribute, but this is a no-op on amd64. To do it correctly, we
have to generate a wrapper for ndis_timercall() and us the wrapper
instead of of the raw function pointer.
Fix this by adding ndis_timercall() to the funcptr table in subr_ndis.c,
and create ndis_findwrap() to extract the wrapped function from the
table in NdisMInitializeTimer() instead of just passing ndis_timercall()
to KeInitializeDpc() directly.
ExAllocatePoolWithTag(), not malloc(), so it should be released
with ExFreePool(), not free(). Fix a couple if instances of
free(fh, ...) that got overlooked.
- On amd64, InterlockedPushEntrySList() and InterlockedPopEntrySList()
are mapped to ExpInterlockedPushEntrySList and
ExpInterlockedPopEntrySList() via macros (which do the same thing).
Add IMPORT_FUNC_MAP()s for these.
- Implement ExQueryDepthSList().
alloc and free routine pointers in the lookaside list with pointers
to ExAllocatePoolWithTag() and ExFreePool() (in the case where the
driver does not provide its own alloc and free routines). For amd64,
this is wrong: we have to use pointers to the wrapped versions of these
functions, not the originals.
nll_obsoletelock field in the lookaside list structure is only defined
for the i386 arch. For amd64, the field is gone, and different list
update routines are used which do their locking internally. Apparently
the Inprocomm amd64 driver uses lookaside lists. I'm not positive this
will make it work yet since I don't have an Inprocomm NIC to test, but
this needs to be fixed anyway.
work on SMP" saga. After several weeks and much gnashing of teeth,
I have finally tracked down all the problems, despite their best
efforts to confound and annoy me.
Problem nunmber one: the Atheros windows driver is _NOT_ a de-serialized
miniport! It used to be that NDIS drivers relied on the NDIS library
itself for all their locking and serialization needs. Transmit packet
queues were all handled internally by NDIS, and all calls to
MiniportXXX() routines were guaranteed to be appropriately serialized.
This proved to be a performance problem however, and Microsoft
introduced de-serialized miniports with the NDIS 5.x spec. Microsoft
still supports serialized miniports, but recommends that all new drivers
written for Windows XP and later be deserialized. Apparently Atheros
wasn't listening when they said this.
This means (among other things) that we have to serialize calls to
MiniportSendPackets(). We also have to serialize calls to MiniportTimer()
that are triggered via the NdisMInitializeTimer() routine. It finally
dawned on me why NdisMInitializeTimer() takes a special
NDIS_MINIPORT_TIMER structure and a pointer to the miniport block:
the timer callback must be serialized, and it's only by saving the
miniport block handle that we can get access to the serialization
lock during the timer callback.
Problem number two: haunted hardware. The thing that was _really_
driving me absolutely bonkers for the longest time is that, for some
reason I couldn't understand, my test machine would occasionally freeze
or more frustratingly, reset completely. That's reset and in *pow!*
back to the BIOS startup. No panic, no crashdump, just a reset. This
appeared to happen most often when MiniportReset() was called. (As
to why MiniportReset() was being called, see problem three below.)
I thought maybe I had created some sort of horrible deadlock
condition in the process of adding the serialization, but after three
weeks, at least 6 different locking implementations and heroic efforts
to debug the spinlock code, the machine still kept resetting. Finally,
I started single stepping through the MiniportReset() routine in
the driver using the kernel debugger, and this ultimately led me to
the source of the problem.
One of the last things the Atheros MiniportReset() routine does is
call NdisReadPciSlotInformation() several times to inspect a portion
of the device's PCI config space. It reads the same chunk of config
space repeatedly, in rapid succession. Presumeably, it's polling
the hardware for some sort of event. The reset occurs partway through
this process. I discovered that when I single-stepped through this
portion of the routine, the reset didn't occur. So I inserted a 1
microsecond delay into the read loop in NdisReadPciSlotInformation().
Suddenly, the reset was gone!!
I'm still very puzzled by the whole thing. What I suspect is happening
is that reading the PCI config space so quickly is causing a severe
PCI bus error. My test system is a Sun w2100z dual Opteron system,
and the NIC is a miniPCI card mounted in a miniPCI-to-PCI carrier card,
plugged into a 100Mhz PCI slot. It's possible that this combination of
hardware causes a bus protocol violation in this scenario which leads
to a fatal machine check. This is pure speculation though. Really all I
know for sure is that inserting the delay makes the problem go away.
(To quote Homer Simpson: "I don't know how it works, but fire makes
it good!")
Problem number three: NdisAllocatePacket() needs to make sure to
initialize the npp_validcounts field in the 'private' section of
the NDIS_PACKET structure. The reason if_ndis was calling the
MiniportReset() routine in the first place is that packet transmits
were sometimes hanging. When sending a packet, an NDIS driver will
call NdisQueryPacket() to learn how many physical buffers the packet
resides in. NdisQueryPacket() is actually a macro, which traverses
the NDIS_BUFFER list attached to the NDIS_PACKET and stashes some
of the results in the 'private' section of the NDIS_PACKET. It also
sets the npp_validcounts field to TRUE To indicate that the results are
now valid. The problem is, now that if_ndis creates a pool of transmit
packets via NdisAllocatePacketPool(), it's important that each time
a new packet is allocated via NdisAllocatePacket() that validcounts
be initialized to FALSE. If it isn't, and a previously transmitted
NDIS_PACKET is pulled out of the pool, it may contain stale data
from a previous transmission which won't get updated by NdisQueryPacket().
This would cause the driver to miscompute the number of fragments
for a given packet, and botch the transmission.
Fixing these three problems seems to make the Atheros driver happy
on SMP, which hopefully means other serialized miniports will be
happy too.
And there was much rejoicing.
Other stuff fixed along the way:
- Modified ndis_thsuspend() to take a mutex as an argument. This
allows KeWaitForSingleObject() and KeWaitForMultipleObjects() to
avoid any possible race conditions with other routines that
use the dispatcher lock.
- Fixed KeCancelTimer() so that it returns the correct value for
'pending' according to the Microsoft documentation
- Modfied NdisGetSystemUpTime() to use ticks and hz rather than
calling nanouptime(). Also added comment that this routine wraps
after 49.7 days.
- Added macros for KeAcquireSpinLock()/KeReleaseSpinLock() to hide
all the MSCALL() goop.
- For x86, KeAcquireSpinLockRaiseToDpc() needs to be a separate
function. This is because it's supposed to be _stdcall on the x86
arch, whereas KeAcquireSpinLock() is supposed to be _fastcall.
On amd64, all routines use the same calling convention so we can
just map KeAcquireSpinLockRaiseToDpc() directly to KfAcquireSpinLock()
and it will work. (The _fastcall attribute is a no-op on amd64.)
- Implement and use IoInitializeDpcRequest() and IoRequestDpc() (they're
just macros) and use them for interrupt handling. This allows us to
move the ndis_intrtask() routine from if_ndis.c to kern_ndis.c.
- Fix the MmInitializeMdl() macro so that is uses sizeof(vm_offset_t)
when computing mdl_size instead of uint32_t, so that it matches the
MmSizeOfMdl() routine.
- Change a could of M_WAITOKs to M_NOWAITs in the unicode routines in
subr_ndis.c.
- Use the dispatcher lock a little more consistently in subr_ntoskrnl.c.
- Get rid of the "wait for link event" hack in ndis_init(). Now that
I fixed NdisReadPciSlotInformation(), it seems I don't need it anymore.
This should fix the witness panic a couple of people have reported.
- Use MSCALL1() when calling the MiniportHangCheck() function in
ndis_ticktask(). I accidentally missed this one when adding the
wrapping for amd64.
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.
both consuming 1K of stack space. This is unfriendly. Allocate the buffers
off the heap instead. It's a little slower, but these aren't performance
critical routines.
Also, add a spinlock to NdisAllocatePacketPool(), NdisAllocatePacket(),
NdisFreePacketPool() and NdisFreePacket(). The pool is maintained as a
linked list. I don't know for a fact that it can be corrupted, but why
take chances.
