An older PVHVM guest (v3.0 based) crashed during vCPU hot-plug with:
kernel BUG at drivers/xen/events.c:1328!
RCU has detected that a CPU has not entered a quiescent state within the
grace period. It needs to send the CPU a reschedule IPI if it is not
offline. rcu_implicit_offline_qs() does this check:
/*
* If the CPU is offline, it is in a quiescent state. We can
* trust its state not to change because interrupts are disabled.
*/
if (cpu_is_offline(rdp->cpu)) {
rdp->offline_fqs++;
return 1;
}
Else the CPU is online. Send it a reschedule IPI.
The CPU is in the middle of being hot-plugged and has been marked online
(!cpu_is_offline()). See start_secondary():
set_cpu_online(smp_processor_id(), true);
...
per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
start_secondary() then waits for the CPU bringing up the hot-plugged CPU to
mark it as active:
/*
* Wait until the cpu which brought this one up marked it
* online before enabling interrupts. If we don't do that then
* we can end up waking up the softirq thread before this cpu
* reached the active state, which makes the scheduler unhappy
* and schedule the softirq thread on the wrong cpu. This is
* only observable with forced threaded interrupts, but in
* theory it could also happen w/o them. It's just way harder
* to achieve.
*/
while (!cpumask_test_cpu(smp_processor_id(), cpu_active_mask))
cpu_relax();
/* enable local interrupts */
local_irq_enable();
The CPU being hot-plugged will be marked active after it has been fully
initialized by the CPU managing the hot-plug. In the Xen PVHVM case
xen_smp_intr_init() is called to set up the hot-plugged vCPU's
XEN_RESCHEDULE_VECTOR.
The hot-plugging CPU is marked online, not marked active and does not have
its IPI vectors set up. rcu_implicit_offline_qs() sees the hot-plugging
cpu is !cpu_is_offline() and tries to send it a reschedule IPI:
This will lead to:
kernel BUG at drivers/xen/events.c:1328!
xen_send_IPI_one()
xen_smp_send_reschedule()
rcu_implicit_offline_qs()
rcu_implicit_dynticks_qs()
force_qs_rnp()
force_quiescent_state()
__rcu_process_callbacks()
rcu_process_callbacks()
__do_softirq()
call_softirq()
do_softirq()
irq_exit()
xen_evtchn_do_upcall()
because xen_send_IPI_one() will attempt to use an uninitialized IRQ for
the XEN_RESCHEDULE_VECTOR.
There is at least one other place that has caused the same crash:
xen_smp_send_reschedule()
wake_up_idle_cpu()
add_timer_on()
clocksource_watchdog()
call_timer_fn()
run_timer_softirq()
__do_softirq()
call_softirq()
do_softirq()
irq_exit()
xen_evtchn_do_upcall()
xen_hvm_callback_vector()
clocksource_watchdog() uses cpu_online_mask to pick the next CPU to handle
a watchdog timer:
/*
* Cycle through CPUs to check if the CPUs stay synchronized
* to each other.
*/
next_cpu = cpumask_next(raw_smp_processor_id(), cpu_online_mask);
if (next_cpu >= nr_cpu_ids)
next_cpu = cpumask_first(cpu_online_mask);
watchdog_timer.expires += WATCHDOG_INTERVAL;
add_timer_on(&watchdog_timer, next_cpu);
This resulted in an attempt to send an IPI to a hot-plugging CPU that
had not initialized its reschedule vector. One option would be to make
the RCU code check to not check for CPU offline but for CPU active.
As becoming active is done after a CPU is online (in older kernels).
But Srivatsa pointed out that "the cpu_active vs cpu_online ordering has been
completely reworked - in the online path, cpu_active is set *before* cpu_online,
and also, in the cpu offline path, the cpu_active bit is reset in the CPU_DYING
notification instead of CPU_DOWN_PREPARE." Drilling in this the bring-up
path: "[brought up CPU].. send out a CPU_STARTING notification, and in response
to that, the scheduler sets the CPU in the cpu_active_mask. Again, this mask
is better left to the scheduler alone, since it has the intelligence to use it
judiciously."
The conclusion was that:
"
1. At the IPI sender side:
It is incorrect to send an IPI to an offline CPU (cpu not present in
the cpu_online_mask). There are numerous places where we check this
and warn/complain.
2. At the IPI receiver side:
It is incorrect to let the world know of our presence (by setting
ourselves in global bitmasks) until our initialization steps are complete
to such an extent that we can handle the consequences (such as
receiving interrupts without crashing the sender etc.)
" (from Srivatsa)
As the native code enables the interrupts at some point we need to be
able to service them. In other words a CPU must have valid IPI vectors
if it has been marked online.
It doesn't need to handle the IPI (interrupts may be disabled) but needs
to have valid IPI vectors because another CPU may find it in cpu_online_mask
and attempt to send it an IPI.
This patch will change the order of the Xen vCPU bring-up functions so that
Xen vectors have been set up before start_secondary() is called.
It also will not continue to bring up a Xen vCPU if xen_smp_intr_init() fails
to initialize it.
Orabug 13823853
Signed-off-by Chuck Anderson <chuck.anderson@oracle.com>
Acked-by: Srivatsa S. Bhat <srivatsa.bhat@linux.vnet.ibm.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
When a event is being bound to a VCPU there is a window between the
EVTCHNOP_bind_vpcu call and the adjustment of the local per-cpu masks
where an event may be lost. The hypervisor upcalls the new VCPU but
the kernel thinks that event is still bound to the old VCPU and
ignores it.
There is even a problem when the event is being bound to the same VCPU
as there is a small window beween the clear_bit() and set_bit() calls
in bind_evtchn_to_cpu(). When scanning for pending events, the kernel
may read the bit when it is momentarily clear and ignore the event.
Avoid this by masking the event during the whole bind operation.
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
CC: stable@vger.kernel.org
The sizeof() argument in init_evtchn_cpu_bindings() is incorrect
resulting in only the first 64 (or 32 in 32-bit guests) ports having
their bindings being initialized to VCPU 0.
In most cases this does not cause a problem as request_irq() will set
the irq affinity which will set the correct local per-cpu mask.
However, if the request_irq() is called on a VCPU other than 0, there
is a window between the unmasking of the event and the affinity being
set were an event may be lost because it is not locally unmasked on
any VCPU. If request_irq() is called on VCPU 0 then local irqs are
disabled during the window and the race does not occur.
Fix this by initializing all NR_EVENT_CHANNEL bits in the local
per-cpu masks.
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
CC: stable@vger.kernel.org
If there are UNUSABLE regions in the machine memory map, dom0 will
attempt to map them 1:1 which is not permitted by Xen and the kernel
will crash.
There isn't anything interesting in the UNUSABLE region that the dom0
kernel needs access to so we can avoid making the 1:1 mapping and
treat it as RAM.
We only do this for dom0, as that is where tboot case shows up.
A PV domU could have an UNUSABLE region in its pseudo-physical map
and would need to be handled in another patch.
This fixes a boot failure on hosts with tboot.
tboot marks a region in the e820 map as unusable and the dom0 kernel
would attempt to map this region and Xen does not permit unusable
regions to be mapped by guests.
(XEN) 0000000000000000 - 0000000000060000 (usable)
(XEN) 0000000000060000 - 0000000000068000 (reserved)
(XEN) 0000000000068000 - 000000000009e000 (usable)
(XEN) 0000000000100000 - 0000000000800000 (usable)
(XEN) 0000000000800000 - 0000000000972000 (unusable)
tboot marked this region as unusable.
(XEN) 0000000000972000 - 00000000cf200000 (usable)
(XEN) 00000000cf200000 - 00000000cf38f000 (reserved)
(XEN) 00000000cf38f000 - 00000000cf3ce000 (ACPI data)
(XEN) 00000000cf3ce000 - 00000000d0000000 (reserved)
(XEN) 00000000e0000000 - 00000000f0000000 (reserved)
(XEN) 00000000fe000000 - 0000000100000000 (reserved)
(XEN) 0000000100000000 - 0000000630000000 (usable)
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
[v1: Altered the patch and description with domU's with UNUSABLE regions]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
When CONFIG_PREEMPT is enabled, Linux will not be able to boot and warn:
[ 4.127825] ------------[ cut here ]------------
[ 4.133376] WARNING: at init/main.c:699 do_one_initcall+0x150/0x158()
[ 4.140738] initcall xen_init_events+0x0/0x10c returned with preemption imbalance
This is because xen_percpu_init uses get_cpu but doesn't have the corresponding
put_cpu.
Signed-off-by: Julien Grall <julien.grall@linaro.org>
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
tmem is not supported on arm or arm64 yet. Will revert this
once the Xen hypervisor supports it.
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Unbinding an event channel (either with the ioctl or when the evtchn
device is closed) may deadlock because disable_irq() is called with
port_user_lock held which is also locked by the interrupt handler.
Think of the IOCTL_EVTCHN_UNBIND is being serviced, the routine has
just taken the lock, and an interrupt happens. The evtchn_interrupt
is invoked, tries to take the lock and spins forever.
A quick glance at the code shows that the spinlock is a local IRQ
variant. Unfortunately that does not help as "disable_irq() waits for
the interrupt handler on all CPUs to stop running. If the irq occurs
on another VCPU, it tries to take port_user_lock and can't because
the unbind ioctl is holding it." (from David). Hence we cannot
depend on the said spinlock to protect us. We could make it a system
wide IRQ disable spinlock but there is a better way.
We can piggyback on the fact that the existence of the spinlock is
to make get_port_user() checks be up-to-date. And we can alter those
checks to not depend on the spin lock (as it's protected by u->bind_mutex
in the ioctl) and can remove the unnecessary locking (this is
IOCTL_EVTCHN_UNBIND) path.
In the interrupt handler we cannot use the mutex, but we do not
need it.
"The unbind disables the irq before making the port user stale, so when
you clear it you are guaranteed that the interrupt handler that might
use that port cannot be running." (from David).
Hence this patch removes the spinlock usage on the teardown path
and piggybacks on disable_irq happening before we muck with the
get_port_user() data. This ensures that the interrupt handler will
never run on stale data.
Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
[v1: Expanded the commit description a bit]
On ARM64, when CONFIG_XEN=y, the compilation will fail because CPU hotplug is
not yet supported with XEN. For now, disable it.
Signed-off-by: Julien Grall <julien.grall@linaro.org>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Only create the delayed resume workqueue if we are running in the same domain
as xenstored and issue a warning if the workqueue creation fails.
Move the work initialization to the device probe so it is done only once.
Signed-off-by: Aurelien Chartier <aurelien.chartier@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: David Vrabel <david.vrabel@citrix.com>
... because the "clock_event_device framework" already accounts for idle
time through the "event_handler" function pointer in
xen_timer_interrupt().
The patch is intended as the completion of [1]. It should fix the double
idle times seen in PV guests' /proc/stat [2]. It should be orthogonal to
stolen time accounting (the removed code seems to be isolated).
The approach may be completely misguided.
[1] https://lkml.org/lkml/2011/10/6/10
[2] http://lists.xensource.com/archives/html/xen-devel/2010-08/msg01068.html
John took the time to retest this patch on top of v3.10 and reported:
"idle time is correctly incremented for pv and hvm for the normal
case, nohz=off and nohz=idle." so lets put this patch in.
CC: stable@vger.kernel.org
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Signed-off-by: John Haxby <john.haxby@oracle.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Convert printks to pr_<level> (excludes printk(KERN_DEBUG...)
to be more consistent throughout the xen subsystem.
Add pr_fmt with KBUILD_MODNAME or "xen:" KBUILD_MODNAME
Coalesce formats and add missing word spaces
Add missing newlines
Align arguments and reflow to 80 columns
Remove DRV_NAME from formats as pr_fmt adds the same content
This does change some of the prefixes of these messages
but it also does make them more consistent.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
xen_hvm_post_suspend, xen_pre_suspend, xen_post_suspend are only used if
CONFIG_HIBERNATE_CALLBACKS is defined, resulting in:
drivers/xen/manage.c:46:13: warning: ‘xen_hvm_post_suspend’ defined but not used [-Wunused-function]
drivers/xen/manage.c:52:13: warning: ‘xen_pre_suspend’ defined but not used [-Wunused-function]
drivers/xen/manage.c:59:13: warning: ‘xen_post_suspend’ defined but not used [-Wunused-function]
If the kernel config is missing CONFIG_HIBERNATE_CALLBACKS.
Simply ifdef CONFIG_HIBERNATE_CALLBACKS the three functions.
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
There are two tool-stack that can instruct the Xen PCI frontend
and backend to change states: 'xm' (Python code with a daemon),
and 'xl' (C library - does not keep state changes).
With the 'xm', the path to disconnect a single PCI device (xm pci-detach
<guest> <BDF>) is:
4(Connected)->7(Reconfiguring*)-> 8(Reconfigured)-> 4(Connected)->5(Closing*).
The * is for states that the tool-stack sets. For 'xl', it is similar:
4(Connected)->7(Reconfiguring*)-> 8(Reconfigured)-> 4(Connected)
Both of them also tear down the XenBus structure, so the backend
state ends up going in the 3(Initialised) and calls pcifront_xenbus_remove.
When a PCI device is plugged back in (xm pci-attach <guest> <BDF>)
both of them follow the same pattern:
2(InitWait*), 3(Initialized*), 4(Connected*)->4(Connected).
[xen-pcifront ignores the 2,3 state changes and only acts when
4 (Connected) has been reached]
Note that this is for a _single_ PCI device. If there were two
PCI devices and only one was disconnected 'xm' would show the same
state changes.
The problem is that git commit 3d925320e9
("xen/pcifront: Use Xen-SWIOTLB when initting if required") introduced
a mechanism to initialize the SWIOTLB when the Xen PCI front moves to
Connected state. It also had some aggressive seatbelt code check that
would warn the user if one tried to change to Connected state without
hitting first the Closing state:
pcifront pci-0: PCI frontend already installed!
However, that code can be relaxed and we can continue on working
even if the frontend is instructed to be the 'Connected' state with
no devices and then gets tickled to be in 'Connected' state again.
In other words, this 4(Connected)->5(Closing)->4(Connected) state
was expected, while 4(Connected)->.... anything but 5(Closing)->4(Connected)
was not. This patch removes that aggressive check and allows
Xen pcifront to work with the 'xl' toolstack (for one or more
PCI devices) and with 'xm' toolstack (for more than two PCI
devices).
Acked-by: Bjorn Helgaas <bhelgaas@google.com>
Cc: linux-pci@vger.kernel.org
Cc: stable@vger.kernel.org
[v2: Added in the description about two PCI devices]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
If the per-cpu time data structure has been onlined already and
we are trying to online it again, then free the previous copy
before blindly over-writting it.
A developer naturally should not call this function multiple times
but just in case.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
We don't check whether the per_cpu data structure has actually
been freed in the past. This checks it and if it has been freed
in the past then just continues on without double-freeing.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
We don't do any code movement. We just encapsulate the struct clock_event_device
in a new structure which contains said structure and a pointer to
a char *name. The 'name' will be used in 'xen/time: Don't leak interrupt
name when offlining'.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
When the user does:
echo 0 > /sys/devices/system/cpu/cpu1/online
echo 1 > /sys/devices/system/cpu/cpu1/online
kmemleak reports:
kmemleak: 7 new suspected memory leaks (see /sys/kernel/debug/kmemleak)
unreferenced object 0xffff88003fa51260 (size 32):
comm "swapper/0", pid 1, jiffies 4294667339 (age 1027.789s)
hex dump (first 32 bytes):
73 70 69 6e 6c 6f 63 6b 31 00 00 00 00 00 00 00 spinlock1.......
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
backtrace:
[<ffffffff81660721>] kmemleak_alloc+0x21/0x50
[<ffffffff81190aac>] __kmalloc_track_caller+0xec/0x2a0
[<ffffffff812fe1bb>] kvasprintf+0x5b/0x90
[<ffffffff812fe228>] kasprintf+0x38/0x40
[<ffffffff81663789>] xen_init_lock_cpu+0x61/0xbe
[<ffffffff816633a6>] xen_cpu_up+0x66/0x3e8
[<ffffffff8166bbf5>] _cpu_up+0xd1/0x14b
[<ffffffff8166bd48>] cpu_up+0xd9/0xec
[<ffffffff81ae6e4a>] smp_init+0x4b/0xa3
[<ffffffff81ac4981>] kernel_init_freeable+0xdb/0x1e6
[<ffffffff8165ce39>] kernel_init+0x9/0xf0
[<ffffffff8167edfc>] ret_from_fork+0x7c/0xb0
[<ffffffffffffffff>] 0xffffffffffffffff
Instead of doing it like the "xen/smp: Don't leak interrupt name when offlining"
patch did (which has a per-cpu structure which contains both the
IRQ number and char*) we use a per-cpu pointers to a *char.
The reason is that the "__this_cpu_read(lock_kicker_irq);" macro
blows up with "__bad_size_call_parameter()" as the size of the
returned structure is not within the parameters of what it expects
and optimizes for.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
When we free it we want to make sure to set it to a default
value of -1 so that we don't double-free it (in case somebody
calls us twice).
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
This patch adds a new structure to contain the common two things
that each of the per-cpu interrupts need:
- an interrupt number,
- and the name of the interrupt (to be added in 'xen/smp: Don't leak
interrupt name when offlining').
This allows us to carry the tuple of the per-cpu interrupt data structure
and expand it as we need in the future.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
There are two functions that do a bunch of 'free_irq' on
the per_cpu IRQ. Instead of having duplicate code just move
it to one function.
This is just code movement.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Fix to return -ENOENT in the pcistub_device_find() and pci_get_drvdata()
error handling case instead of 0(overwrite to 0 by str_to_slot()), as done
elsewhere in this function.
Acked-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Chuck Anderson
David Vrabel
Julien Grall
Stefano Stabellini
Aurelien Chartier
Laszlo Ersek
Joe Perches
Konrad Rzeszutek Wilk
Wei Yongjun