[PATCH v4 0/2] kexec: Use hypercall_create_continuation to protect KEXEC ops

[PATCH v4 0/2] kexec: Use hypercall_create_continuation to protect KEXEC ops

[Eric DeVolder]

During testing (using the script below) we found that multiple
invocations of kexec of unload/load are not safe.

This does not exist in classic Xen kernels in which the kexec-tools
did the kexec via Linux kernel syscall (which in turn made the
hypercall), as the Linux code has a mutex_trylock which would
inhibit multiple concurrent calls.

But with the kexec-tools utilizing xc_kexec_* that is no longer
the case and we need to protect against multiple concurrent
invocations.

Please see the patches and review at your convenience!

==== try-crash.pl from bhavesh.davda@oracle.com ====
#!/usr/bin/perl -w

use strict;
use warnings;
use threads;

sub threaded_task {
    threads->create(sub {
        my $thr_id = threads->self->tid;
        #print "Starting load thread $thr_id\n";
        system("/sbin/kexec  -p --command-line=\"placeholder root=/dev/mapper/nimbula-root ro rhbg console=tty0 console=hvc0 earlyprintk=xen nomodeset printk.time=1 irqpoll maxcpus=1 nr_cpus=1 reset_devices cgroup_disable=memory mce=off selinux=0 console=ttyS1,115200n8\" --initrd=/boot/initrd-4.1.12-61.1.9.el6uek.x86_64kdump.img /boot/vmlinuz-4.1.12-61.1.9.el6uek.x86_64");
        #print "Ending load thread $thr_id\n";
        threads->detach(); #End thread.
    });
    threads->create(sub {
        my $thr_id = threads->self->tid;
        #print "Starting unload thread $thr_id\n";
        system("/sbin/kexec  -p -u");
        #print "Ending unload thread $thr_id\n";
        threads->detach(); #End thread.
    });
}

for my $i (0..99)
{
    threaded_task();
}


Eric DeVolder (2):
  kexec: use hypercall_create_continuation to protect KEXEC ops
  kexec: remove spinlock now that all KEXEC hypercall ops are protected
    at the top-level

 xen/common/kexec.c | 17 ++++++++++-------
 1 file changed, 10 insertions(+), 7 deletions(-)

-- 
2.7.4


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

[PATCH v4 1/2] kexec: use hypercall_create_continuation to protect KEXEC ops

[Eric DeVolder]

When we concurrently try to unload and load crash
images we eventually get:

 Xen call trace:
    [<ffff82d08018b04f>] machine_kexec_add_page+0x3a0/0x3fa
    [<ffff82d08018b184>] machine_kexec_load+0xdb/0x107
    [<ffff82d080116e8d>] kexec.c#kexec_load_slot+0x11/0x42
    [<ffff82d08011724f>] kexec.c#kexec_load+0x119/0x150
    [<ffff82d080117c1e>] kexec.c#do_kexec_op_internal+0xab/0xcf
    [<ffff82d080117c60>] do_kexec_op+0xe/0x1e
    [<ffff82d08025c620>] pv_hypercall+0x20a/0x44a
    [<ffff82d080260116>] cpufreq.c#test_all_events+0/0x30

 Pagetable walk from ffff820040088320:
  L4[0x104] = 00000002979d1063 ffffffffffffffff
  L3[0x001] = 00000002979d0063 ffffffffffffffff
  L2[0x000] = 00000002979c7063 ffffffffffffffff
  L1[0x088] = 80037a91ede97063 ffffffffffffffff

The interesting thing is that the page bits (063) look legit.

The operation on which we blow up is us trying to write
in the L1 and finding that the L2 entry points to some
bizzare MFN. It stinks of a race, and it looks like
the issue is due to no concurrency locks when dealing
with the crash kernel space.

Specifically we concurrently call kimage_alloc_crash_control_page
which iterates over the kexec_crash_area.start -> kexec_crash_area.size
and once found:

  if ( page )
  {
      image->next_crash_page = hole_end;
      clear_domain_page(_mfn(page_to_mfn(page)));
  }

clears. Since the parameters of what MFN to use are provided
by the callers (and the area to search is bounded) the the 'page'
is probably the same. So #1 we concurrently clear the
'control_code_page'.

The next step is us passing this 'control_code_page' to
machine_kexec_add_page. This function requires the MFNs:
page_to_maddr(image->control_code_page).

And this would always return the same virtual address, as
the MFN of the control_code_page is inside of the
kexec_crash_area.start -> kexec_crash_area.size area.

Then machine_kexec_add_page updates the L1 .. which can be done
concurrently and on subsequent calls we mangle it up.

This is all a theory at this time, but testing reveals
that adding the hypercall_create_continuation() at the
kexec hypercall fixes the crash.

NOTE: This patch follows 5c5216 (kexec: clear kexec_image slot
when unloading kexec image) to prevent crashes during
simultaneous load/unloads.

NOTE: Consideration was given to using the existing flag
KEXEC_FLAG_IN_PROGRESS to denote a kexec hypercall in
progress. This, however, overloads the original intent of
the flag which is to denote that we are about-to/have made
the jump to the crash path. The overloading would lead to
failures in existing checks on this flag as the flag would
always be set at the top level in do_kexec_op_internal().
For this reason, the new flag KEXEC_FLAG_HC_IN_PROGRESS
was introduced.

While at it, fixed the #define mismatched spacing

Signed-off-by: Eric DeVolder <eric.devolder@oracle.com>
Reviewed-by: Bhavesh Davda <bhavesh.davda@oracle.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>
---
v4: 04/19/2017
 - Unchanged

v3: 04/19/2017
 - Incorporated feedback from Jan Beulich to change the
   name of the new flag to KEXEC_FLAG_IN_HYPERCALL

v2: 04/17/2017
 - Patch titled 'kexec: use hypercall_create_continuation to protect KEXEC ops'
 - Jan Beulich directed me to use a continuation method instead
   of spinlock.
 - Incorporated feedback from Daniel Kiper <daniel.kiper@oracle.com>
 - Incorporated feedback from Konrad Wilk <konrad.wilk@oracle.com>

v1: 04/10/2017
 - Patch titled 'kexec: Add spinlock for the whole hypercall'
 - Used spinlock in do_kexec_op_internal()
---
 xen/common/kexec.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/xen/common/kexec.c b/xen/common/kexec.c
index 072cc8e..253c204 100644
--- a/xen/common/kexec.c
+++ b/xen/common/kexec.c
@@ -50,9 +50,10 @@ static cpumask_t crash_saved_cpus;
 
 static struct kexec_image *kexec_image[KEXEC_IMAGE_NR];
 
-#define KEXEC_FLAG_DEFAULT_POS   (KEXEC_IMAGE_NR + 0)
-#define KEXEC_FLAG_CRASH_POS     (KEXEC_IMAGE_NR + 1)
-#define KEXEC_FLAG_IN_PROGRESS   (KEXEC_IMAGE_NR + 2)
+#define KEXEC_FLAG_DEFAULT_POS  (KEXEC_IMAGE_NR + 0)
+#define KEXEC_FLAG_CRASH_POS    (KEXEC_IMAGE_NR + 1)
+#define KEXEC_FLAG_IN_PROGRESS  (KEXEC_IMAGE_NR + 2)
+#define KEXEC_FLAG_IN_HYPERCALL (KEXEC_IMAGE_NR + 3)
 
 static unsigned long kexec_flags = 0; /* the lowest bits are for KEXEC_IMAGE... */
 
@@ -1193,6 +1194,9 @@ static int do_kexec_op_internal(unsigned long op,
     if ( ret )
         return ret;
 
+    if ( test_and_set_bit(KEXEC_FLAG_IN_HYPERCALL, &kexec_flags) )
+        return hypercall_create_continuation(__HYPERVISOR_kexec_op, "lh", op, uarg);
+
     switch ( op )
     {
     case KEXEC_CMD_kexec_get_range:
@@ -1227,6 +1231,8 @@ static int do_kexec_op_internal(unsigned long op,
         break;
     }
 
+    clear_bit(KEXEC_FLAG_IN_HYPERCALL, &kexec_flags);
+
     return ret;
 }
 
-- 
2.7.4


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

[PATCH v4 2/2] kexec: remove spinlock now that all KEXEC hypercall ops are protected at the top-level

[Eric DeVolder]

The spinlock in kexec_swap_images() was removed as
this function is only reachable on the kexec hypercall, which is
now protected at the top-level in do_kexec_op_internal(),
thus the local spinlock is no longer necessary.

Per recommendation from Jan Beulich and Andrew Cooper, I left
an ASSERT in place of the spin_lock().

Signed-off-by: Eric DeVolder <eric.devolder@oracle.com>
Reviewed-by: Bhavesh Davda <bhavesh.davda@oracle.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
v4: 04/19/2017
 - Fix ASSERT to work properly. Tested with Config.mk:debug=y

v3: 04/19/2017
 - Incorporated feedback from Jan Beulich and Andrew Cooper
   to leave an ASSERT where spin_lock() once was.

v2: 04/17/2017
 - Patch titled 'kexec: use hypercall_create_continuation to protect KEXEC ops'
 - Separated removal of spinlock in kexec_swap_images() into its own patch.

v1: 04/10/2017
 - Patch titled 'kexec: Add spinlock for the whole hypercall'
 - Removal of spinlock in kexec_swap_images() was part of other patch.
---
 xen/common/kexec.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/xen/common/kexec.c b/xen/common/kexec.c
index 253c204..96efa3b 100644
--- a/xen/common/kexec.c
+++ b/xen/common/kexec.c
@@ -820,7 +820,6 @@ static int kexec_exec(XEN_GUEST_HANDLE_PARAM(void) uarg)
 static int kexec_swap_images(int type, struct kexec_image *new,
                              struct kexec_image **old)
 {
-    static DEFINE_SPINLOCK(kexec_lock);
     int base, bit, pos;
     int new_slot, old_slot;
 
@@ -832,7 +831,7 @@ static int kexec_swap_images(int type, struct kexec_image *new,
     if ( kexec_load_get_bits(type, &base, &bit) )
         return -EINVAL;
 
-    spin_lock(&kexec_lock);
+    ASSERT(test_bit(KEXEC_FLAG_IN_HYPERCALL, &kexec_flags));
 
     pos = (test_bit(bit, &kexec_flags) != 0);
     old_slot = base + pos;
@@ -846,8 +845,6 @@ static int kexec_swap_images(int type, struct kexec_image *new,
     clear_bit(old_slot, &kexec_flags);
     *old = kexec_image[old_slot];
 
-    spin_unlock(&kexec_lock);
-
     return 0;
 }
 
-- 
2.7.4


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [PATCH v4 2/2] kexec: remove spinlock now that all KEXEC hypercall ops are protected at the top-level

[Jan Beulich]

>>> On 19.04.17 at 23:01, <eric.devolder@oracle.com> wrote:
> The spinlock in kexec_swap_images() was removed as
> this function is only reachable on the kexec hypercall, which is
> now protected at the top-level in do_kexec_op_internal(),
> thus the local spinlock is no longer necessary.
> 
> Per recommendation from Jan Beulich and Andrew Cooper, I left
> an ASSERT in place of the spin_lock().
> 
> Signed-off-by: Eric DeVolder <eric.devolder@oracle.com>
> Reviewed-by: Bhavesh Davda <bhavesh.davda@oracle.com>
> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>

Reviewed-by: Jan Beulich <jbeulich@suse.com>



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [PATCH v4 2/2] kexec: remove spinlock now that all KEXEC hypercall ops are protected at the top-level

[Andrew Cooper]

On 20/04/17 11:46, Jan Beulich wrote:
>>>> On 19.04.17 at 23:01, <eric.devolder@oracle.com> wrote:
>> The spinlock in kexec_swap_images() was removed as
>> this function is only reachable on the kexec hypercall, which is
>> now protected at the top-level in do_kexec_op_internal(),
>> thus the local spinlock is no longer necessary.
>>
>> Per recommendation from Jan Beulich and Andrew Cooper, I left
>> an ASSERT in place of the spin_lock().
>>
>> Signed-off-by: Eric DeVolder <eric.devolder@oracle.com>
>> Reviewed-by: Bhavesh Davda <bhavesh.davda@oracle.com>
>> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> Reviewed-by: Jan Beulich <jbeulich@suse.com>

Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>

Julien - can I get a 4.9 release ack for these two patches?

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [PATCH v4 2/2] kexec: remove spinlock now that all KEXEC hypercall ops are protected at the top-level

[Daniel Kiper]

On Thu, Apr 20, 2017 at 01:14:30PM +0100, Andrew Cooper wrote:
> On 20/04/17 11:46, Jan Beulich wrote:
> >>>> On 19.04.17 at 23:01, <eric.devolder@oracle.com> wrote:
> >> The spinlock in kexec_swap_images() was removed as
> >> this function is only reachable on the kexec hypercall, which is
> >> now protected at the top-level in do_kexec_op_internal(),
> >> thus the local spinlock is no longer necessary.
> >>
> >> Per recommendation from Jan Beulich and Andrew Cooper, I left
> >> an ASSERT in place of the spin_lock().
> >>
> >> Signed-off-by: Eric DeVolder <eric.devolder@oracle.com>
> >> Reviewed-by: Bhavesh Davda <bhavesh.davda@oracle.com>
> >> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> > Reviewed-by: Jan Beulich <jbeulich@suse.com>
>
> Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>

Reviewed-by: Daniel Kiper <daniel.kiper@oracle.com>

Daniel

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel

Re: [PATCH v4 2/2] kexec: remove spinlock now that all KEXEC hypercall ops are protected at the top-level

[Julien Grall]

Hi Andrew,

On 20/04/17 13:14, Andrew Cooper wrote:
> On 20/04/17 11:46, Jan Beulich wrote:
>>>>> On 19.04.17 at 23:01, <eric.devolder@oracle.com> wrote:
>>> The spinlock in kexec_swap_images() was removed as
>>> this function is only reachable on the kexec hypercall, which is
>>> now protected at the top-level in do_kexec_op_internal(),
>>> thus the local spinlock is no longer necessary.
>>>
>>> Per recommendation from Jan Beulich and Andrew Cooper, I left
>>> an ASSERT in place of the spin_lock().
>>>
>>> Signed-off-by: Eric DeVolder <eric.devolder@oracle.com>
>>> Reviewed-by: Bhavesh Davda <bhavesh.davda@oracle.com>
>>> Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
>> Reviewed-by: Jan Beulich <jbeulich@suse.com>
>
> Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
>
> Julien - can I get a 4.9 release ack for these two patches?

Release-acked-by: Julien Grall <julien.grall@arm.com>

Cheers,

-- 
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
https://lists.xen.org/xen-devel