On 14/01/2020 16:50, Jan Beulich wrote:
On 14.01.2020 17:26, Julien Grall wrote:
On 14/01/2020 16:08, Jan Beulich wrote:
On 13.01.2020 22:33, Julien Grall wrote:
--- a/xen/arch/x86/hvm/irq.c
+++ b/xen/arch/x86/hvm/irq.c
@@ -29,7 +29,8 @@
bool hvm_domain_use_pirq(const struct domain *d, const struct pirq *pirq)
-    return is_hvm_domain(d) && pirq && pirq->arch.hvm.emuirq != IRQ_UNBOUND;
+    return is_hvm_domain(d) && pirq &&
+        const_pirq_dpci(pirq)->emuirq != IRQ_UNBOUND;
/* Must be called with hvm_domain->irq_lock hold */
@@ -396,7 +397,7 @@ int hvm_inject_msi(struct domain *d, uint64_t addr, 
uint32_t data)
               struct pirq *info = pirq_info(d, pirq);
/* if it is the first time, allocate the pirq */
-            if ( !info || info->arch.hvm.emuirq == IRQ_UNBOUND )
+            if ( !info || pirq_dpci(info)->emuirq == IRQ_UNBOUND )
                   int rc;
@@ -409,7 +410,7 @@ int hvm_inject_msi(struct domain *d, uint64_t addr, uint32_t data)
                   if ( !info )
                       return -EBUSY;
-            else if ( info->arch.hvm.emuirq != IRQ_MSI_EMU )
+            else if ( pirq_dpci(info)->emuirq != IRQ_MSI_EMU )
                   return -EINVAL;
               send_guest_pirq(d, info);
               return 0;

All of these uses (and others further down) make pretty clear
that the emuirq field doesn't belong in the structure you put it
in - the 'd' in dpci stands for "direct" afaik, and the field is
for a certain variant of emulation of interrupt delivery into
guests, i.e. not really pass-through focused at all.

I am happy to keep emuirq in struct pirq if you are happy with slightly
increasing the size allocated on PV.

The main thing I want to get rid of is the weird allocation size we do

While I understand this, to be honest I'd rather not see the size
grow for no good (to PV) reason. I don't think the current model is
_this_ bad.

Well, I did lost two days debugging a problem because of the allocation (the memory were getting corrupted randomly). The comment you added may help to avoid this problem but I still think that trying to allocate half a pirq is a pretty bad idea.

But if you really want to push for it, why can't the
two parts continue to live in a wrapper HVM structure, just like
they do today?

I am not sure what you are suggesting here. Could you extend your thought?

@@ -171,8 +172,26 @@ struct hvm_pirq_dpci {
       struct hvm_gmsi_info gmsi;
       struct timer timer;
       struct list_head softirq_list;
+    int emuirq;
+    struct pirq pirq;
+#define pirq_dpci(p) \
+    ((p) ? container_of(p, struct hvm_pirq_dpci, pirq) : NULL)
+#define const_pirq_dpci(p)                                              \
+    ((p) ? container_of(p, const struct hvm_pirq_dpci, pirq) : NULL)
+#define dpci_pirq(pd) (&(pd)->pirq)
+#define domain_pirq_to_emuirq(d, p) ({                                  \
+    struct pirq *__pi = pirq_info(d, p);                                \
+    __pi ? pirq_dpci(__pi)->emuirq : IRQ_UNBOUND;                       \
+#define domain_emuirq_to_pirq(d, emuirq) ({                             \
+    void *__ret = radix_tree_lookup(&(d)->arch.hvm.emuirq_pirq, emuirq);\
+    __ret ? radix_tree_ptr_to_int(__ret) : IRQ_UNBOUND;                 \

While for the latter you merely move the bogus double-leading-
underscore macro local variable (which on this occasion I'd
like to ask anyway to be changed), you actively introduce a
new similar name space violation in the domain_pirq_to_emuirq().

AFAIK, there is nothing in the coding style forbidding your "bogus"
naming. So I just followed the rest of the code.

Our coding style document is not to re-iterate C standard rules,
I think, and hence yes, you won't find anything to this effect

The fact such code has been added in Xen in the past clearly shows that the coding style is not sufficient to back your point here.

So rather than complaining that I don't follow an unwritten rule, you could have suggested it. This would have came accross as less rude.

Anyway, I will update it.

@@ -133,17 +132,10 @@ DECLARE_PER_CPU(unsigned int, irq_count);
struct arch_pirq {
       int irq;
-    union {
-        struct hvm_pirq {
-            int emuirq;
-            struct hvm_pirq_dpci dpci;
-        } hvm;
-    };
+    /* Is the PIRQ associated to an HVM domain? */
+    bool hvm;

It looks like this field is needed for only arch_free_pirq_struct().
As it'll make a difference to struct pirq's size, can you not get
away without it? All (perhaps indirect) callers of the function
know the domain, after all.

The free is done through an RCU callback with no extra parameters to
tell how it can be freed.

The only way I can think of to get rid of the field is to introduce two
different callback for the free. We would use a different callback
depending on the domain type.

How does that sound?

That's exactly what I was thinking of as a possible solution.

I will have a look.


Julien Grall

Xen-devel mailing list

Reply via email to