certain instructions in a function prologue or epilogue. DTrace has a
hook into the invalid opcode fault handler that checks whether the fault
was due to an probe and if so, runs the DTrace magic.
Upon returning from an invalid opcode fault caused by a probe, DTrace must
emulate the instruction that was replaced with the invalid opcode and then
return control to the instruction following the invalid opcode.
There were a pair of related bugs in the emulation for the leave
instruction. The leave instruction is used to pop off a stack frame prior
to returning from a function. The emulation for this instruction must
move the trap frame for the invalid opcode fault down the stack to the
bottom of the stack frame that is being removed, and then execute an iret.
At two points in this process, the emulation code was storing values above
the current value of the stack pointer. This opened up a window in which
if we were two take an interrupt, the trap frame for the interrupt would
overwrite the values stored on the stack, causing the system to panic
later.
The first bug was that at one point the emulation code saves the new value
for $esp above the current stack pointer value. The fix is to save this
value instead inside of the original trap frame. At this point we do
not need the original trap frame so this is safe.
The second bug is that when the emulate code loads $esp from the stack, it
points part-way through the new trap frame instead of at its beginning.
The emulation code adjusts the stack pointer to the correct value
immediately afterwards, but this still leaves a one instruction window in
which an interrupt would corrupt this trap frame. Fix this by adjusting
the stack frame value before loading it into $esp.
This fixes panics in invop_leave on i386 when using fbt return probes.
Reviewed by: rpaulo, attilio
MFC after: 1 week
Now in the case when one-shot timers are used cyclic events should fire
closer to theier scheduled times. As the cyclic is currently used only
to drive DTrace profile provider, this is the area where the change
makes a difference.
Reviewed by: mav (earlier version, a while ago)
X-MFC after: clocksource/eventtimer subsystem
cpuset_t objects.
That is going to offer the underlying support for a simple bump of
MAXCPU and then support for number of cpus > 32 (as it is today).
Right now, cpumask_t is an int, 32 bits on all our supported architecture.
cpumask_t on the other side is implemented as an array of longs, and
easilly extendible by definition.
The architectures touched by this commit are the following:
- amd64
- i386
- pc98
- arm
- ia64
- XEN
while the others are still missing.
Userland is believed to be fully converted with the changes contained
here.
Some technical notes:
- This commit may be considered an ABI nop for all the architectures
different from amd64 and ia64 (and sparc64 in the future)
- per-cpu members, which are now converted to cpuset_t, needs to be
accessed avoiding migration, because the size of cpuset_t should be
considered unknown
- size of cpuset_t objects is different from kernel and userland (this is
primirally done in order to leave some more space in userland to cope
with KBI extensions). If you need to access kernel cpuset_t from the
userland please refer to example in this patch on how to do that
correctly (kgdb may be a good source, for example).
- Support for other architectures is going to be added soon
- Only MAXCPU for amd64 is bumped now
The patch has been tested by sbruno and Nicholas Esborn on opteron
4 x 12 pack CPUs. More testing on big SMP is expected to came soon.
pluknet tested the patch with his 8-ways on both amd64 and i386.
Tested by: pluknet, sbruno, gianni, Nicholas Esborn
Reviewed by: jeff, jhb, sbruno
safer for i386 because it can be easily over 4 GHz now. More worse, it can
be easily changed by user with 'machdep.tsc_freq' tunable (directly) or
cpufreq(4) (indirectly). Note it is intentionally not used in performance
critical paths to avoid performance regression (but we should, in theory).
Alternatively, we may add "virtual TSC" with lower frequency if maximum
frequency overflows 32 bits (and ignore possible incoherency as we do now).
Add systrace_linux32 and systrace_freebsd32 modules which provide
support for tracing compat system calls in addition to native system
call tracing provided by systrace module.
Provided that all the systrace modules are loaded now you can select
what syscalls to trace in the following manner:
syscall::xxx:yyy - work on all system calls that match the specification
syscall:freebsd:xxx:yyy - only native system calls
syscall:linux32:xxx:yyy - linux32 compat system calls
syscall:freebsd32:xxx:yyy - freebsd32 compat system calls on amd64
PR: kern/152822
Submitted by: Artem Belevich <fbsdlist@src.cx>
Reviewed by: jhb (earlier version)
MFC after: 3 weeks
In this case we call target function only on a single CPU and do not
need any synchronization at the setup stage.
It's a bit non-obvious but setup function of NULL means that
smp_rendezvous_cpus waits for all CPUs to arrive at the rendezvous
point, but without doing any actual setup. While using
smp_no_rendevous_barrier means that each CPU proceeds on its own
schedule without any synchronization whatsoever.
MFC after: 3 weeks
The dealock was caused in the following way:
- thread T1 on CPU C1 holds a spin mutex, IPIs CPU C2 and waits for the
IPI to be handled
- C2 executes timer interrupt filter, thus has interrupts disabled, and
gets blocked on the spin mutex held by T1
The problem seems to have been introduced by simplifications made to
OpenSolaris code during porting.
The problem is fixed by reorganizing the code to more closely resemble
the upstream version. Interrupt filter (cyclic_fire) now doesn't
acquire any locks, all per-CPU data accesses are performed on a
target CPU with preemption and interrupts disabled thus precluding
concurrent access to the data.
cyp_mtx spin mutex is used to disable preemtion and interrupts; it's not
used for classical mutual exclusion, because xcall already serializes
calls to a CPU. It's an emulation of OpenSolaris
cyb_set_level(CY_HIGH_LEVEL) call, the spin mutexes could probably be
reduced to just a spinlock_enter()/_exit() pair.
Diff with upstream version is now reduced by ~500 lines, however it still
remains quite large - many things that are not needed (at the moment) or
are irrelevant on FreeBSD were simply ripped out during porting.
Examples of such things:
- support for CPU onlining/offlining
- support for suspend/resume
- support for running callouts at soft interrupt levels
- support for callout rebinding from CPU to CPU
- support for CPU partitions
Tested by: Artem Belevich <fbsdlist@src.cx>
MFC after: 3 weeks
X-MFC with: r216252
Extend struct sysvec with three new elements:
sv_fetch_syscall_args - the method to fetch syscall arguments from
usermode into struct syscall_args. The structure is machine-depended
(this might be reconsidered after all architectures are converted).
sv_set_syscall_retval - the method to set a return value for usermode
from the syscall. It is a generalization of
cpu_set_syscall_retval(9) to allow ABIs to override the way to set a
return value.
sv_syscallnames - the table of syscall names.
Use sv_set_syscall_retval in kern_sigsuspend() instead of hardcoding
the call to cpu_set_syscall_retval().
The new functions syscallenter(9) and syscallret(9) are provided that
use sv_*syscall* pointers and contain the common repeated code from
the syscall() implementations for the architecture-specific syscall
trap handlers.
Syscallenter() fetches arguments, calls syscall implementation from
ABI sysent table, and set up return frame. The end of syscall
bookkeeping is done by syscallret().
Take advantage of single place for MI syscall handling code and
implement ptrace_lwpinfo pl_flags PL_FLAG_SCE, PL_FLAG_SCX and
PL_FLAG_EXEC. The SCE and SCX flags notify the debugger that the
thread is stopped at syscall entry or return point respectively. The
EXEC flag augments SCX and notifies debugger that the process address
space was changed by one of exec(2)-family syscalls.
The i386, amd64, sparc64, sun4v, powerpc and ia64 syscall()s are
changed to use syscallenter()/syscallret(). MIPS and arm are not
converted and use the mostly unchanged syscall() implementation.
Reviewed by: jhb, marcel, marius, nwhitehorn, stas
Tested by: marcel (ia64), marius (sparc64), nwhitehorn (powerpc),
stas (mips)
MFC after: 1 month
Currently dtrace_gethrtime uses formula similar to the following for
converting TSC ticks to nanoseconds:
rdtsc() * 10^9 / tsc_freq
The dividend overflows 64-bit type and wraps-around every 2^64/10^9 =
18446744073 ticks which is just a few seconds on modern machines.
Now we instead use precalculated scaling factor of
10^9*2^N/tsc_freq < 2^32 and perform TSC value multiplication separately
for each 32-bit half. This allows to avoid overflow of the dividend
described above.
The idea is taken from OpenSolaris.
This has an added feature of always scaling TSC with invariant value
regardless of TSC frequency changes. Thus the timestamps will not be
accurate if TSC actually changes, but they are always proportional to
TSC ticks and thus monotonic. This should be much better than current
formula which produces wildly different non-monotonic results on when
tsc_freq changes.
Also drop write-only 'cp' variable from amd64 dtrace_gethrtime_init()
to make it identical to the i386 twin.
PR: kern/127441
Tested by: Thomas Backman <serenity@exscape.org>
Reviewed by: jhb
Discussed with: current@, bde, gnn
Silence from: jb
Approved by: re (gnn)
MFC after: 1 week
On amd64 KERNBASE/kernbase does not mean start of kernel memory.
This should fix a KASSERT panic in dtrace_copycheck when copyin*()
is used in D program.
Also make checks for user memory a bit stricter.
Reported by: Thomas Backman <serenity@exscape.org>
Submitted by: wxs (kaddr part)
Tested by: Thomas Backman (prototype), wxs
Reviewed by: alc (concept), jhb, current@
Aprroved by: jb (concept)
MFC after: 2 weeks
PR: kern/134408
adds probes for mutexes, reader/writer and shared/exclusive locks to
gather contention statistics and other locking information for
dtrace scripts, the lockstat(1M) command and other potential
consumers.
Reviewed by: attilio jhb jb
Approved by: gnn (mentor)
events are:
nfsclient:accesscache:flush:done
nfsclient:accesscache:get:hit
nfsclient:accesscache:get:miss
nfsclient:accesscache:load:done
They pass the vnode, uid, and requested or loaded access mode (if any);
the load event may also report a load error if the RPC fails.
The attribute cache events are:
nfsclient:attrcache:flush:done
nfsclient:attrcache:get:hit
nfsclient:attrcache:get:miss
nfsclient:attrcache:load:done
They pass the vnode, optionally the vattr if one is present (hit or load),
and in the case of a load event, also a possible RPC error.
MFC after: 1 month
Sponsored by: Google, Inc.
provider. The NFS client exposes 'start' and 'done' probes for NFSv2
and NFSv3 RPCs when using the new RPC implementation, passing in the
vnode, mbuf chain, credential, and NFSv2 or NFSv3 procedure number.
For 'done' probes, the error number is also available.
Probes are named in the following way:
...
nfsclient:nfs2:write:start
nfsclient:nfs2:write:done
...
nfsclient:nfs3:access:start
nfsclient:nfs3:access:done
...
Access to the unmarshalled arguments is not easily available at this
point in the stack, but the passed probe arguments are sufficient to
to a lot of interesting things in practice. Technically, these probes
may cover multiple RPC retransmits, and even transactions if the
transaction ID change as a result of authentication failure or a
jukebox error from the server, but usefully capture the intent of a
single NFS request, such as access, getattr, write, etc.
Typical use might involve profiling RPC latency by system call, number
of RPCs, how often a getattr leads to a call to access, when failed
access control checks occur, etc. More detailed RPC information might
best be provided by adding a krpc provider. It would also be useful
to add NFS client probes for events such as the access cache or
attribute cache satisfying requests without an RPC.
Sponsored by: Google, Inc.
MFC after: 1 month
When I changed kern_conf.c three months ago I made device unit numbers
equal to (unneeded) device minor numbers. We used to require
bitshifting, because there were eight bits in the middle that were
reserved for a device major number. Not very long after I turned
dev2unit(), minor(), unit2minor() and minor2unit() into macro's.
The unit2minor() and minor2unit() macro's were no-ops.
We'd better not remove these four macro's from the kernel, because there
is a lot of (external) code that may still depend on them. For now it's
harmless to remove all invocations of unit2minor() and minor2unit().
Reviewed by: kib
