When operating on an unpinned pagetable (ie, one under construction or
destruction), it isn't necessary to use a hypercall to update a
pud/pmd entry. Jan Beulich observed that a similar optimisation
avoided many thousands of hypercalls while doing a kernel build.
One tricky part is that early in the kernel boot there's no page
structure, so we can't check to see if the page is pinned. In that
case, we just always use the hypercall.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
-tip tree auto-testing found the following early bootup hang:
-------------->
get_memcfg_from_srat: assigning address to rsdp
RSD PTR v0 [Nvidia]
BUG: Int 14: CR2 ffd00040
EDI 8092fbfe ESI ffd00040 EBP 80b0aee8 ESP 80b0aed0
EBX 000f76f0 EDX 0000000e ECX 00000003 EAX ffd00040
err 00000000 EIP 802c055a CS 00000060 flg 00010006
Stack: ffd00040 80bc78d0 80b0af6c 80b1dbfe 8093d8ba 00000008 80b42810 80b4ddb4
80b42842 00000000 80b0af1c 801079c8 808e724e 00000000 80b42871 802c0531
00000100 00000000 0003fff0 80b0af40 80129999 00040100 00040100 00000000
Pid: 0, comm: swapper Not tainted 2.6.26-rc4-sched-devel.git #570
[<802c055a>] ? strncmp+0x11/0x25
[<80b1dbfe>] ? get_memcfg_from_srat+0xb4/0x568
[<801079c8>] ? mcount_call+0x5/0x9
[<802c0531>] ? strcmp+0xa/0x22
[<80129999>] ? printk+0x38/0x3a
[<80129999>] ? printk+0x38/0x3a
[<8011b122>] ? memory_present+0x66/0x6f
[<80b216b4>] ? setup_memory+0x13/0x40c
[<80b16b47>] ? propagate_e820_map+0x80/0x97
[<80b1622a>] ? setup_arch+0x248/0x477
[<80129999>] ? printk+0x38/0x3a
[<80b11759>] ? start_kernel+0x6e/0x2eb
[<80b110fc>] ? i386_start_kernel+0xeb/0xf2
=======================
<------
with this config:
http://redhat.com/~mingo/misc/config-Wed_May_28_01_33_33_CEST_2008.bad
The thing is, the crash makes little sense at first sight. We crash on a
benign-looking printk. The code around it got changed in -tip but
checking those topic branches individually did not reproduce the bug.
Bisection led to this commit:
| d5edbc1f75 is first bad commit
| commit d5edbc1f75
| Author: Jeremy Fitzhardinge <jeremy@goop.org>
| Date: Mon May 26 23:31:22 2008 +0100
|
| xen: add p2m mfn_list_list
Which is somewhat surprising, as on native hardware Xen client side
should have little to no side-effects.
After some head scratching, it turns out the following happened:
randconfig enabled the following Xen options:
CONFIG_XEN=y
CONFIG_XEN_MAX_DOMAIN_MEMORY=8
# CONFIG_XEN_BLKDEV_FRONTEND is not set
# CONFIG_XEN_NETDEV_FRONTEND is not set
CONFIG_HVC_XEN=y
# CONFIG_XEN_BALLOON is not set
which activated this piece of code in arch/x86/xen/mmu.c:
> @@ -69,6 +69,13 @@
> __attribute__((section(".data.page_aligned"))) =
> { [ 0 ... TOP_ENTRIES - 1] = &p2m_missing[0] };
>
> +/* Arrays of p2m arrays expressed in mfns used for save/restore */
> +static unsigned long p2m_top_mfn[TOP_ENTRIES]
> + __attribute__((section(".bss.page_aligned")));
> +
> +static unsigned long p2m_top_mfn_list[TOP_ENTRIES / P2M_ENTRIES_PER_PAGE]
> + __attribute__((section(".bss.page_aligned")));
The problem is, you must only put variables into .bss.page_aligned that
have a _size_ that is _exactly_ page aligned. In this case the size of
p2m_top_mfn_list is not page aligned:
80b8d000 b p2m_top_mfn
80b8f000 b p2m_top_mfn_list
80b8f008 b softirq_stack
80b97008 b hardirq_stack
80b9f008 b bm_pte
So all subsequent variables get unaligned which, depending on luck,
breaks the kernel in various funny ways. In this case what killed the
kernel first was the misaligned bootmap pte page, resulting in that
creative crash above.
Anyway, this was a fun bug to track down :-)
I think the moral is that .bss.page_aligned is a dangerous construct in
its current form, and the symptoms of breakage are very non-trivial, so
i think we need build-time checks to make sure all symbols in
.bss.page_aligned are truly page aligned.
The Xen fix below gets the kernel booting again.
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>
When saving a domain, the Xen tools need to remap all our mfns to
portable pfns. In order to remap our p2m table, it needs to know
where all its pages are, so maintain the references to the p2m table
for it to use.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
When using sparsemem and memory hotplug, the kernel's pseudo-physical
address space can be discontigious. Previously this was dealt with by
having the upper parts of the radix tree stubbed off. Unfortunately,
this is incompatible with save/restore, which requires a complete p2m
table.
The solution is to have a special distinguished all-invalid p2m leaf
page, which we can point all the hole areas at. This allows the tools
to see a complete p2m table, but it only costs a page for all memory
holes.
It also simplifies the code since it removes a few special cases.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Add a config option to set the max size of a Xen domain. This is used
to scale the size of the physical-to-machine array; it ends up using
around 1 page/GByte, so there's no reason to be very restrictive.
For a 32-bit guest, the default value of 8GB is probably sufficient;
there's not much point in giving a 32-bit machine much more memory
than that.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
We now support the use of memory hotplug, so the physical to machine
page mapping structure must be dynamic. This is implemented as a
two-level radix tree structure, which allows us to efficiently
incrementally allocate memory for the p2m table as new pages are
added.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Non-PAE operation has been deprecated in Xen for a while, and is
rarely tested or used. xen-unstable has now officially dropped
non-PAE support. Since Xen/pvops' non-PAE support has also been
broken for a while, we may as well completely drop it altogether.
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>
The usual pagetable locking protocol doesn't seem to apply to updates
to init_mm, so don't rely on preemption being disabled in xen_set_pte_at
on init_mm.
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>
We can fold the essentially common pte functions together now.
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>
pte_t always contains a "pte" field for the whole pte value, so make
use of it.
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>
Adjust the definition of lookup_address to take an unsigned long
level argument. Adjust callers in xen/mmu.c that pass in a
dummy variable.
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
based on this patch from Andi Kleen:
| Subject: CPA: Return the page table level in lookup_address()
| From: Andi Kleen <ak@suse.de>
|
| Needed for the next change.
|
| And change all the callers.
and ported it to x86.git.
Signed-off-by: Andi Kleen <ak@suse.de>
Acked-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Looks like a mismerge/misapply dropped one of the cases of pte flag
masking for Xen. Also, only mask the flags for present ptes.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The hypervisor doesn't allow PCD or PWT to be set on guest ptes, so
make sure they're masked out. Also, fix up some previous mispatching.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Make sure pte_t, whatever its definition, has a pte element with type
pteval_t. This allows common code to access it without needing to be
specifically parameterised on what pagetable mode we're compiling for.
For 32-bit, this means that pte_t becomes a union with "pte" and "{
pte_low, pte_high }" (PAE) or just "pte_low" (non-PAE).
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
_PAGE_PCD maps a page with caching disabled, which is typically used for
mapping harware registers. Xen never allows it to be set on a mapping, and
unprivileged guests never need it since they can't see the real underlying
hardware. However, some uncached mappings are made early when probing the
(non-existent) APIC, and its OK to mask off the PCD flag in these cases.
This became necessary because Xen started checking for this bit, rather
than silently masking it off.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When a pagetable is created, it is made globally visible in the rmap
prio tree before it is pinned via arch_dup_mmap(), and remains in the
rmap tree while it is unpinned with arch_exit_mmap().
This means that other CPUs may race with the pinning/unpinning
process, and see a pte between when it gets marked RO and actually
pinned, causing any pte updates to fail with write-protect faults.
As a result, all pte pages must be properly locked, and only unlocked
once the pinning/unpinning process has finished.
In order to avoid taking spinlocks for the whole pagetable - which may
overflow the PREEMPT_BITS portion of preempt counter - it locks and pins
each pte page individually, and then finally pins the whole pagetable.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Hugh Dickens <hugh@veritas.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andi Kleen <ak@suse.de>
Cc: Keir Fraser <keir@xensource.com>
Cc: Jan Beulich <jbeulich@novell.com>
When a pagetable is no longer in use, it must be unpinned so that its
pages can be freed. However, this is only possible if there are no
stray uses of the pagetable. The code currently deals with all the
usual cases, but there's a rare case where a vcpu is changing cr3, but
is doing so lazily, and the change hasn't actually happened by the time
the pagetable is unpinned, even though it appears to have been completed.
This change adds a second per-cpu cr3 variable - xen_current_cr3 -
which tracks the actual state of the vcpu cr3. It is only updated once
the actual hypercall to set cr3 has been completed. Other processors
wishing to unpin a pagetable can check other vcpu's xen_current_cr3
values to see if any cross-cpu IPIs are needed to clean things up.
[ Stable folks: 2.6.23 bugfix ]
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: Stable Kernel <stable@kernel.org>
This patch cleans up duplicate includes in
arch/i386/xen/
Signed-off-by: Jesper Juhl <jesper.juhl@gmail.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Currently, the set_lazy_mode pv_op is overloaded with 5 functions:
1. enter lazy cpu mode
2. leave lazy cpu mode
3. enter lazy mmu mode
4. leave lazy mmu mode
5. flush pending batched operations
This complicates each paravirt backend, since it needs to deal with
all the possible state transitions, handling flushing, etc. In
particular, flushing is quite distinct from the other 4 functions, and
seems to just cause complication.
This patch removes the set_lazy_mode operation, and adds "enter" and
"leave" lazy mode operations on mmu_ops and cpu_ops. All the logic
associated with enter and leaving lazy states is now in common code
(basically BUG_ONs to make sure that no mode is current when entering
a lazy mode, and make sure that the mode is current when leaving).
Also, flush is handled in a common way, by simply leaving and
re-entering the lazy mode.
The result is that the Xen, lguest and VMI lazy mode implementations
are much simpler.
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: Andi Kleen <ak@suse.de>
Cc: Zach Amsden <zach@vmware.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Avi Kivity <avi@qumranet.com>
Cc: Anthony Liguory <aliguori@us.ibm.com>
Cc: "Glauber de Oliveira Costa" <glommer@gmail.com>
Cc: Jun Nakajima <jun.nakajima@intel.com>
