Xiaohui Xin and some other folks at Intel have been looking into what's
behind the performance hit of paravirt_ops when running native.
It appears that the hit is entirely due to the paravirtualized
spinlocks introduced by:
| commit 8efcbab674
| Date: Mon Jul 7 12:07:51 2008 -0700
|
| paravirt: introduce a "lock-byte" spinlock implementation
The extra call/return in the spinlock path is somehow
causing an increase in the cycles/instruction of somewhere around 2-7%
(seems to vary quite a lot from test to test). The working theory is
that the CPU's pipeline is getting upset about the
call->call->locked-op->return->return, and seems to be failing to
speculate (though I haven't seen anything definitive about the precise
reasons). This doesn't entirely make sense, because the performance
hit is also visible on unlock and other operations which don't involve
locked instructions. But spinlock operations clearly swamp all the
other pvops operations, even though I can't imagine that they're
nearly as common (there's only a .05% increase in instructions
executed).
If I disable just the pv-spinlock calls, my tests show that pvops is
identical to non-pvops performance on native (my measurements show that
it is actually about .1% faster, but Xiaohui shows a .05% slowdown).
Summary of results, averaging 10 runs of the "mmperf" test, using a
no-pvops build as baseline:
nopv Pv-nospin Pv-spin
CPU cycles 100.00% 99.89% 102.18%
instructions 100.00% 100.10% 100.15%
CPI 100.00% 99.79% 102.03%
cache ref 100.00% 100.84% 100.28%
cache miss 100.00% 90.47% 88.56%
cache miss rate 100.00% 89.72% 88.31%
branches 100.00% 99.93% 100.04%
branch miss 100.00% 103.66% 107.72%
branch miss rt 100.00% 103.73% 107.67%
wallclock 100.00% 99.90% 102.20%
The clear effect here is that the 2% increase in CPI is
directly reflected in the final wallclock time.
(The other interesting effect is that the more ops are
out of line calls via pvops, the lower the cache access
and miss rates. Not too surprising, but it suggests that
the non-pvops kernel is over-inlined. On the flipside,
the branch misses go up correspondingly...)
So, what's the fix?
Paravirt patching turns all the pvops calls into direct calls, so
_spin_lock etc do end up having direct calls. For example, the compiler
generated code for paravirtualized _spin_lock is:
<_spin_lock+0>: mov %gs:0xb4c8,%rax
<_spin_lock+9>: incl 0xffffffffffffe044(%rax)
<_spin_lock+15>: callq *0xffffffff805a5b30
<_spin_lock+22>: retq
The indirect call will get patched to:
<_spin_lock+0>: mov %gs:0xb4c8,%rax
<_spin_lock+9>: incl 0xffffffffffffe044(%rax)
<_spin_lock+15>: callq <__ticket_spin_lock>
<_spin_lock+20>: nop; nop /* or whatever 2-byte nop */
<_spin_lock+22>: retq
One possibility is to inline _spin_lock, etc, when building an
optimised kernel (ie, when there's no spinlock/preempt
instrumentation/debugging enabled). That will remove the outer
call/return pair, returning the instruction stream to a single
call/return, which will presumably execute the same as the non-pvops
case. The downsides arel 1) it will replicate the
preempt_disable/enable code at eack lock/unlock callsite; this code is
fairly small, but not nothing; and 2) the spinlock definitions are
already a very heavily tangled mass of #ifdefs and other preprocessor
magic, and making any changes will be non-trivial.
The other obvious answer is to disable pv-spinlocks. Making them a
separate config option is fairly easy, and it would be trivial to
enable them only when Xen is enabled (as the only non-default user).
But it doesn't really address the common case of a distro build which
is going to have Xen support enabled, and leaves the open question of
whether the native performance cost of pv-spinlocks is worth the
performance improvement on a loaded Xen system (10% saving of overall
system CPU when guests block rather than spin). Still it is a
reasonable short-term workaround.
[ Impact: fix pvops performance regression when running native ]
Analysed-by: "Xin Xiaohui" <xiaohui.xin@intel.com>
Analysed-by: "Li Xin" <xin.li@intel.com>
Analysed-by: "Nakajima Jun" <jun.nakajima@intel.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Acked-by: H. Peter Anvin <hpa@zytor.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Xen-devel <xen-devel@lists.xensource.com>
LKML-Reference: <4A0B62F7.5030802@goop.org>
[ fixed the help text ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Now that x86-64 has directly accessible percpu variables, it can also
implement the direct versions of these operations, which operate on a
vcpu_info structure directly embedded in the percpu area.
In fact, the 64-bit versions are more or less identical, and so can be
shared. The only two differences are:
1. xen_restore_fl_direct takes its argument in eax on 32-bit, and rdi on 64-bit.
Unfortunately it isn't possible to directly refer to the 2nd lsb of rdi directly
(as you can with %ah), so the code isn't quite as dense.
2. check_events needs to variants to save different registers.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Due to confusion between the ftrace infrastructure and the gcc profiling
tracer "ftrace", this patch renames the config options from FTRACE to
FUNCTION_TRACER. The other two names that are offspring from FTRACE
DYNAMIC_FTRACE and FTRACE_MCOUNT_RECORD will stay the same.
This patch was generated mostly by script, and partially by hand.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add support for exporting statistics on mmu updates, multicall
batching and pv spinlocks into debugfs. The base path is xen/ and
each subsystem adds its own directory: mmu, multicalls, spinlocks.
In each directory, writing 1 to "zero_stats" will cause the
corresponding stats to be zeroed the next time they're updated.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Acked-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
For some reason I managed to miss a bunch of irq-related functions
which also need to be compiled without -pg when using ftrace. This
patch moves them into their own file, and starts a cleanup process
I've been meaning to do anyway.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Sam Ravnborg <sam@ravnborg.org>
Cc: "Alex Nixon (Intern)" <Alex.Nixon@eu.citrix.com>
Cc: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
ftrace requires certain low-level code, like spinlocks and timestamps,
to be compiled without -pg in order to avoid infinite recursion. This
patch splits out the core paravirt spinlocks and the Xen spinlocks
into separate files which can be compiled without -pg.
Also do xen/time.c while we're about it. As a result, we can now use
ftrace within a Xen domain.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Split xen-asm into 32- and 64-bit files, and implement the 64-bit
variants.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Stephen Tweedie <sct@redhat.com>
Cc: Eduardo Habkost <ehabkost@redhat.com>
Cc: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch implements Xen save/restore and migration.
Saving is triggered via xenbus, which is polled in
drivers/xen/manage.c. When a suspend request comes in, the kernel
prepares itself for saving by:
1 - Freeze all processes. This is primarily to prevent any
partially-completed pagetable updates from confusing the suspend
process. If CONFIG_PREEMPT isn't defined, then this isn't necessary.
2 - Suspend xenbus and other devices
3 - Stop_machine, to make sure all the other vcpus are quiescent. The
Xen tools require the domain to run its save off vcpu0.
4 - Within the stop_machine state, it pins any unpinned pgds (under
construction or destruction), performs canonicalizes various other
pieces of state (mostly converting mfns to pfns), and finally
5 - Suspend the domain
Restore reverses the steps used to save the domain, ending when all
the frozen processes are thawed.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
move arch/x86/xen/manage.c under drivers/xen/to share codes
with x86 and ia64.
ia64/xen also uses manage.c
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
split out x86 specific part from grant-table.c and
allow ia64/xen specific initialization.
ia64/xen grant table is based on pseudo physical address
(guest physical address) unlike x86/xen. On ia64 init_mm
doesn't map identity straight mapped area.
ia64/xen specific grant table initialization is necessary.
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
move arch/x86/xen/events.c undedr drivers/xen to share codes
with x86 and ia64. And minor adjustment to compile.
ia64/xen also uses events.c
Signed-off-by: Yaozu (Eddie) Dong <eddie.dong@intel.com>
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
ia64/xen also uses it too. Move it into common place so that
ia64/xen can share the code.
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This patchs adds the mechanism to allow us to patch inline versions of
common operations.
The implementations of the direct-access versions save_fl, restore_fl,
irq_enable and irq_disable are now in assembler, and the same code is
used for both out of line and inline uses.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: Chris Wright <chrisw@sous-sol.org>
Cc: Keir Fraser <keir@xensource.com>
The guest domain can be asked to shutdown or reboot itself, or have a
sysrq key injected, via xenbus. This patch adds a watcher for those
events, and does the appropriate action.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: Chris Wright <chrisw@sous-sol.org>
This is a fairly straightforward Xen implementation of smp_ops.
Xen has its own IPI mechanisms, and has no dependency on any
APIC-based IPI. The smp_ops hooks and the flush_tlb_others pv_op
allow a Xen guest to avoid all APIC code in arch/i386 (the only apic
operation is a single apic_read for the apic version number).
One subtle point which needs to be addressed is unpinning pagetables
when another cpu may have a lazy tlb reference to the pagetable. Xen
will not allow an in-use pagetable to be unpinned, so we must find any
other cpus with a reference to the pagetable and get them to shoot
down their references.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Andi Kleen <ak@suse.de>
Xen maintains a base clock which measures nanoseconds since system
boot. This is provided to guests via a shared page which contains a
base time in ns, a tsc timestamp at that point and tsc frequency
parameters. Guests can compute the current time by reading the tsc
and using it to extrapolate the current time from the basetime. The
hypervisor makes sure that the frequency parameters are updated
regularly, paricularly if the tsc changes rate or stops.
This is implemented as a clocksource, so the interface to the rest of
the kernel is a simple clocksource which simply returns the current
time directly in nanoseconds.
Xen also provides a simple timer mechanism, which allows a timeout to
be set in the future. When that time arrives, a timer event is sent
to the guest. There are two timer interfaces:
- An old one which also delivers a stream of (unused) ticks at 100Hz,
and on the same event, the actual timer events. The 100Hz ticks
cause a lot of spurious wakeups, but are basically harmless.
- The new timer interface doesn't have the 100Hz ticks, and can also
fail if the specified time is in the past.
This code presents the Xen timer as a clockevent driver, and uses the
new interface by preference.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Xen implements interrupts in terms of event channels. Each guest
domain gets 1024 event channels which can be used for a variety of
purposes, such as Xen timer events, inter-domain events,
inter-processor events (IPI) or for real hardware IRQs.
Within the kernel, we map the event channels to IRQs, and implement
the whole interrupt handling using a Xen irq_chip.
Rather than setting NR_IRQ to 1024 under PARAVIRT in order to
accomodate Xen, we create a dynamic mapping between event channels and
IRQs. Ideally, Linux will eventually move towards dynamically
allocating per-irq structures, and we can use a 1:1 mapping between
event channels and irqs.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Xen pagetable handling, including the machinery to implement direct
pagetables.
Xen presents the real CPU's pagetables directly to guests, with no
added shadowing or other layer of abstraction. Naturally this means
the hypervisor must maintain close control over what the guest can put
into the pagetable.
When the guest modifies the pte/pmd/pgd, it must convert its
domain-specific notion of a "physical" pfn into a global machine frame
number (mfn) before inserting the entry into the pagetable. Xen will
check to make sure the domain is allowed to create a mapping of the
given mfn.
Xen also requires that all mappings the guest has of its own active
pagetable are read-only. This is relatively easy to implement in
Linux because all pagetables share the same pte pages for kernel
mappings, so updating the pte in one pagetable will implicitly update
the mapping in all pagetables.
Normally a pagetable becomes active when you point to it with cr3 (or
the Xen equivalent), but when you do so, Xen must check the whole
pagetable for correctness, which is clearly a performance problem.
Xen solves this with pinning which keeps a pagetable effectively
active even if its currently unused, which means that all the normal
update rules are enforced. This means that it need not revalidate the
pagetable when loading cr3.
This patch has a first-cut implementation of pinning, but it is more
fully implemented in a later patch.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
This patch is a rollup of all the core pieces of the Xen
implementation, including:
- booting and setup
- pagetable setup
- privileged instructions
- segmentation
- interrupt flags
- upcalls
- multicall batching
BOOTING AND SETUP
The vmlinux image is decorated with ELF notes which tell the Xen
domain builder what the kernel's requirements are; the domain builder
then constructs the address space accordingly and starts the kernel.
Xen has its own entrypoint for the kernel (contained in an ELF note).
The ELF notes are set up by xen-head.S, which is included into head.S.
In principle it could be linked separately, but it seems to provoke
lots of binutils bugs.
Because the domain builder starts the kernel in a fairly sane state
(32-bit protected mode, paging enabled, flat segments set up), there's
not a lot of setup needed before starting the kernel proper. The main
steps are:
1. Install the Xen paravirt_ops, which is simply a matter of a
structure assignment.
2. Set init_mm to use the Xen-supplied pagetables (analogous to the
head.S generated pagetables in a native boot).
3. Reserve address space for Xen, since it takes a chunk at the top
of the address space for its own use.
4. Call start_kernel()
PAGETABLE SETUP
Once we hit the main kernel boot sequence, it will end up calling back
via paravirt_ops to set up various pieces of Xen specific state. One
of the critical things which requires a bit of extra care is the
construction of the initial init_mm pagetable. Because Xen places
tight constraints on pagetables (an active pagetable must always be
valid, and must always be mapped read-only to the guest domain), we
need to be careful when constructing the new pagetable to keep these
constraints in mind. It turns out that the easiest way to do this is
use the initial Xen-provided pagetable as a template, and then just
insert new mappings for memory where a mapping doesn't already exist.
This means that during pagetable setup, it uses a special version of
xen_set_pte which ignores any attempt to remap a read-only page as
read-write (since Xen will map its own initial pagetable as RO), but
lets other changes to the ptes happen, so that things like NX are set
properly.
PRIVILEGED INSTRUCTIONS AND SEGMENTATION
When the kernel runs under Xen, it runs in ring 1 rather than ring 0.
This means that it is more privileged than user-mode in ring 3, but it
still can't run privileged instructions directly. Non-performance
critical instructions are dealt with by taking a privilege exception
and trapping into the hypervisor and emulating the instruction, but
more performance-critical instructions have their own specific
paravirt_ops. In many cases we can avoid having to do any hypercalls
for these instructions, or the Xen implementation is quite different
from the normal native version.
The privileged instructions fall into the broad classes of:
Segmentation: setting up the GDT and the GDT entries, LDT,
TLS and so on. Xen doesn't allow the GDT to be directly
modified; all GDT updates are done via hypercalls where the new
entries can be validated. This is important because Xen uses
segment limits to prevent the guest kernel from damaging the
hypervisor itself.
Traps and exceptions: Xen uses a special format for trap entrypoints,
so when the kernel wants to set an IDT entry, it needs to be
converted to the form Xen expects. Xen sets int 0x80 up specially
so that the trap goes straight from userspace into the guest kernel
without going via the hypervisor. sysenter isn't supported.
Kernel stack: The esp0 entry is extracted from the tss and provided to
Xen.
TLB operations: the various TLB calls are mapped into corresponding
Xen hypercalls.
Control registers: all the control registers are privileged. The most
important is cr3, which points to the base of the current pagetable,
and we handle it specially.
Another instruction we treat specially is CPUID, even though its not
privileged. We want to control what CPU features are visible to the
rest of the kernel, and so CPUID ends up going into a paravirt_op.
Xen implements this mainly to disable the ACPI and APIC subsystems.
INTERRUPT FLAGS
Xen maintains its own separate flag for masking events, which is
contained within the per-cpu vcpu_info structure. Because the guest
kernel runs in ring 1 and not 0, the IF flag in EFLAGS is completely
ignored (and must be, because even if a guest domain disables
interrupts for itself, it can't disable them overall).
(A note on terminology: "events" and interrupts are effectively
synonymous. However, rather than using an "enable flag", Xen uses a
"mask flag", which blocks event delivery when it is non-zero.)
There are paravirt_ops for each of cli/sti/save_fl/restore_fl, which
are implemented to manage the Xen event mask state. The only thing
worth noting is that when events are unmasked, we need to explicitly
see if there's a pending event and call into the hypervisor to make
sure it gets delivered.
UPCALLS
Xen needs a couple of upcall (or callback) functions to be implemented
by each guest. One is the event upcalls, which is how events
(interrupts, effectively) are delivered to the guests. The other is
the failsafe callback, which is used to report errors in either
reloading a segment register, or caused by iret. These are
implemented in i386/kernel/entry.S so they can jump into the normal
iret_exc path when necessary.
MULTICALL BATCHING
Xen provides a multicall mechanism, which allows multiple hypercalls
to be issued at once in order to mitigate the cost of trapping into
the hypervisor. This is particularly useful for context switches,
since the 4-5 hypercalls they would normally need (reload cr3, update
TLS, maybe update LDT) can be reduced to one. This patch implements a
generic batching mechanism for hypercalls, which gets used in many
places in the Xen code.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Cc: Ian Pratt <ian.pratt@xensource.com>
Cc: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
Cc: Adrian Bunk <bunk@stusta.de>
Jeremy Fitzhardinge