On 6/16/2016 6:16 PM, Jan Beulich wrote:
On 16.06.16 at 16:12, <cz...@bitdefender.com> wrote:
Prepare for ARM implementation of control-register write vm-events by moving
X86-specific hvm_event_cr to the common-side.
Signed-off-by: Corneliu ZUZU <cz...@bitdefender.com>
---
xen/arch/x86/hvm/event.c | 30 ------------------------------
xen/arch/x86/hvm/hvm.c | 2 +-
xen/arch/x86/monitor.c | 37 -------------------------------------
xen/arch/x86/vm_event.c | 2 +-
xen/common/monitor.c | 40 ++++++++++++++++++++++++++++++++++++++++
xen/common/vm_event.c | 31 +++++++++++++++++++++++++++++++
xen/include/asm-x86/hvm/event.h | 13 ++++---------
xen/include/asm-x86/monitor.h | 2 --
xen/include/xen/monitor.h | 4 ++++
xen/include/xen/vm_event.h | 10 ++++++++++
10 files changed, 91 insertions(+), 80 deletions(-)
Considering that there's no ARM file getting altered here at all,
mentioning ARM in the subject is a little odd.
This patch and the following one should be meld together.
I only split them to ease reviewing, sorry I forgot to mention that in
the cover letter.
--- a/xen/common/monitor.c
+++ b/xen/common/monitor.c
@@ -62,6 +62,46 @@ int monitor_domctl(struct domain *d, struct
xen_domctl_monitor_op *mop)
switch ( mop->event )
{
+#if CONFIG_X86
#ifdef please.
Ack.
+ case XEN_DOMCTL_MONITOR_EVENT_WRITE_CTRLREG:
+ {
+ struct arch_domain *ad = &d->arch;
Peeking into the next patch I see that this stays there. Common code,
however, shouldn't access ->arch sub-structures - respective fields
should be moved out.
Then we need to find a resolution that avoids code duplication.
The code is the same, but those bits that are currently on the arch side
(arch.monitor.write_ctrlreg_*) cannot be moved to common as they are,
since their -number- might differ from arch-to-arch.
But we could:
- in public/vm_event.h, besides the VM_EVENT_X86_* and VM_EVENT_ARM_*
defines (wcr index), also have
#define VM_EVENT_X86_CR_COUNT 4
#define VM_EVENT_ARM_CR_COUNT 4
- move the 3 write_ctrlreg_{enabled,sync,onchangeonly} fields from
arch_domain to domain (common) and make them 8-bits wide each for now
(widened more in the future if the need arises)
- let monitor_ctrlreg_bitmask() macro to be architecture-dependent and
use the introduced VM_EVENT_<arch>_CR_COUNT
Tamas, we also talked on this matter @ some point (when I sent the
patches that moved vm-event parts to common). What do you think of the
above?
And looking at all the uses of this variable I get the impression that
you really want a shorthand for &d->arch.monitor (if any such
helper variable is worthwhile to have here in the first place).
Well, this was a simple cut-paste operation, not very old content aware :)
Personally I prefer the current shorthand (ad) (seems more intuitive and
is consistent with the other XEN_DOMCTL_MONITOR_EVENT_* cases), but if
you prefer I'll change that shorthand to am = &d->arch.monitor?
--- a/xen/include/asm-x86/monitor.h
+++ b/xen/include/asm-x86/monitor.h
@@ -24,8 +24,6 @@
#include <xen/sched.h>
-#define monitor_ctrlreg_bitmask(ctrlreg_index) (1U << (ctrlreg_index))
-
static inline
int arch_monitor_domctl_op(struct domain *d, struct xen_domctl_monitor_op
*mop)
{
--- a/xen/include/xen/monitor.h
+++ b/xen/include/xen/monitor.h
@@ -25,6 +25,10 @@
struct domain;
struct xen_domctl_monitor_op;
+#if CONFIG_X86
+#define monitor_ctrlreg_bitmask(ctrlreg_index) (1U << (ctrlreg_index))
+#endif
What's the point in removing this from the x86 header if then it
needs to be put in such a conditional? If the conditional gets
dropped in the next patch, then I think you have two options:
Leave it where it is here, and move it there. Or move it here,
but omit the conditional right away - I can't see this definition
being present to harm the ARM build in any way.
Meld comment above.
--- a/xen/include/xen/vm_event.h
+++ b/xen/include/xen/vm_event.h
@@ -96,6 +96,16 @@ void vm_event_vcpu_unpause(struct vcpu *v);
int vm_event_monitor_traps(struct vcpu *v, uint8_t sync,
vm_event_request_t *req);
+#if CONFIG_X86
+/*
+ * Called for the current vCPU on control-register changes by guest.
+ * The event might not fire if the client has subscribed to it in onchangeonly
+ * mode, hence the bool_t return type for control register write events.
+ */
+bool_t vm_event_monitor_cr(unsigned int index, unsigned long value,
+ unsigned long old);
+#endif
Same goes for the declaration here.
Jan
Meld comment above.
Corneliu.
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel
