diff mbox

[Xen-devel,RFC,15/22] xen/arm: p2m: Re-implement relinquish_p2m_mapping using p2m_get_entry

Message ID 1469717505-8026-16-git-send-email-julien.grall@arm.com
State New
Headers show

Commit Message

Julien Grall July 28, 2016, 2:51 p.m. UTC
The current implementation of relinquish_p2m_mapping is modifying the
page table to erase the entry one by one. However, this is not necessary
because the domain is not running anymore and therefore will speed up
the domain destruction.

The function relinquish_p2m_mapping can be re-implemented using
p2m_get_entry by iterating over the range mapped and using the mapping
order given by the callee.

Given that the preemption was chosen arbitrarily, it is no done on every
512 iterations. Meaning that Xen may check more often if the function is
preempted when there are no mappings.

Finally drop the operation RELINQUISH in apply_* as nobody is using it
anymore. Note that the functions could have been dropped in one go at
the end, however I find easier to drop the operations one by one
avoiding a big deletion in the patch that remove the last operation.

Signed-off-by: Julien Grall <julien.grall@arm.com>

---
    Further investigation needs to be done before applying this patch to
    check if someone could take advantage of this change (such
    modifying an entry which was relinquished).
---
 xen/arch/arm/p2m.c | 70 ++++++++++++++++++++++++++++++++++++++++--------------
 1 file changed, 52 insertions(+), 18 deletions(-)

Comments

Julien Grall Sept. 6, 2016, 3:05 p.m. UTC | #1
Hi Stefano,

On 05/09/16 22:58, Stefano Stabellini wrote:
> On Thu, 28 Jul 2016, Julien Grall wrote:
>> The current implementation of relinquish_p2m_mapping is modifying the
>> page table to erase the entry one by one. However, this is not necessary
>> because the domain is not running anymore and therefore will speed up
>> the domain destruction.
>
> Could you please elaborate on this? Who is going to remove the p2m
> entries if not this function?

The current version of relinquish is removing the reference on the page 
and then invalidate the entry (which may involve a cache flush).

As the page tables are not used anymore by the hardware, the latter 
action is not necessary. This is an optimization because flushing the 
cache can be expensive. However as mentioned later in the commit 
message, we need to have a think on how the other helpers interact with 
the page table to avoid return wrong entry.

I am thinking to defer this optimization for the next release (i.e Xen 
4.9) to avoid rushing on it.

>
>
>> The function relinquish_p2m_mapping can be re-implemented using
>> p2m_get_entry by iterating over the range mapped and using the mapping
>> order given by the callee.
>>
>> Given that the preemption was chosen arbitrarily, it is no done on every
>                                                           ^ now?

Yes, will fix it in the next version.

Regards,
Julien Grall Sept. 7, 2016, 7:37 a.m. UTC | #2
Hi Stefano,

On 06/09/2016 19:21, Stefano Stabellini wrote:
> On Tue, 6 Sep 2016, Julien Grall wrote:
>> Hi Stefano,
>>
>> On 05/09/16 22:58, Stefano Stabellini wrote:
>>> On Thu, 28 Jul 2016, Julien Grall wrote:
>>>> The current implementation of relinquish_p2m_mapping is modifying the
>>>> page table to erase the entry one by one. However, this is not necessary
>>>> because the domain is not running anymore and therefore will speed up
>>>> the domain destruction.
>>>
>>> Could you please elaborate on this? Who is going to remove the p2m
>>> entries if not this function?
>>
>> The current version of relinquish is removing the reference on the page and
>> then invalidate the entry (which may involve a cache flush).
>>
>> As the page tables are not used anymore by the hardware, the latter action is
>> not necessary. This is an optimization because flushing the cache can be
>> expensive. However as mentioned later in the commit message, we need to have a
>> think on how the other helpers interact with the page table to avoid return
>> wrong entry.
>
> The idea is that nobody will remove the p2m entries, until the whole p2m
> is torn down (p2m_teardown)?

That the idea but I hadn't had time to confirm it was the case.

>> I am thinking to defer this optimization for the next release (i.e Xen 4.9) to
>> avoid rushing on it.
>
> If we are sure that there are no interactions with the p2m between the
> domain_relinquish_resources and the p2m_teardown call, then this is
> acceptable. Otherwise delaying this optimization is wiser.

I will delay the optimization.

Regards,
diff mbox

Patch

diff --git a/xen/arch/arm/p2m.c b/xen/arch/arm/p2m.c
index e7697bb..d0aba5b 100644
--- a/xen/arch/arm/p2m.c
+++ b/xen/arch/arm/p2m.c
@@ -721,7 +721,6 @@  static int p2m_mem_access_radix_set(struct p2m_domain *p2m, gfn_t gfn,
 enum p2m_operation {
     INSERT,
     REMOVE,
-    RELINQUISH,
     MEMACCESS,
 };
 
@@ -908,7 +907,6 @@  static int apply_one_level(struct domain *d,
 
         break;
 
-    case RELINQUISH:
     case REMOVE:
         if ( !p2m_valid(orig_pte) )
         {
@@ -1092,17 +1090,6 @@  static int apply_p2m_changes(struct domain *d,
         {
             switch ( op )
             {
-            case RELINQUISH:
-                /*
-                 * Arbitrarily, preempt every 512 operations or 8192 nops.
-                 * 512*P2M_ONE_PROGRESS == 8192*P2M_ONE_PROGRESS_NOP == 0x2000
-                 * This is set in preempt_count_limit.
-                 *
-                 */
-                p2m->lowest_mapped_gfn = _gfn(addr >> PAGE_SHIFT);
-                rc = -ERESTART;
-                goto out;
-
             case MEMACCESS:
             {
                 /*
@@ -1508,16 +1495,63 @@  int p2m_init(struct domain *d)
     return rc;
 }
 
+/*
+ * The function will go through the p2m and remove page reference when it
+ * is required.
+ * The mapping are left intact in the p2m. This is fine because the
+ * domain will never run at that point.
+ *
+ * XXX: Check what does it mean for other part (such as lookup)
+ */
 int relinquish_p2m_mapping(struct domain *d)
 {
     struct p2m_domain *p2m = &d->arch.p2m;
-    unsigned long nr;
+    unsigned long count = 0;
+    p2m_type_t t;
+    int rc = 0;
+    unsigned int order;
+
+    /* Convenience alias */
+    gfn_t start = p2m->lowest_mapped_gfn;
+    gfn_t end = p2m->max_mapped_gfn;
 
-    nr = gfn_x(p2m->max_mapped_gfn) - gfn_x(p2m->lowest_mapped_gfn);
+    p2m_write_lock(p2m);
 
-    return apply_p2m_changes(d, RELINQUISH, p2m->lowest_mapped_gfn, nr,
-                             INVALID_MFN, 0, p2m_invalid,
-                             d->arch.p2m.default_access);
+    for ( ; gfn_x(start) < gfn_x(end); start = gfn_add(start, 1UL << order) )
+    {
+        mfn_t mfn = p2m_get_entry(p2m, start, &t, NULL, &order);
+
+        count++;
+        /*
+         * Arbitrarily preempt every 512 iterations.
+         */
+        if ( !(count % 512) && hypercall_preempt_check() )
+        {
+            rc = -ERESTART;
+            break;
+        }
+
+        /* Skip hole and any superpage */
+        if ( mfn_eq(mfn, INVALID_MFN) || order != 0 )
+            /*
+             * The order corresponds to the order of the mapping in the
+             * page table. So we need to align the GFN before
+             * incrementing.
+             */
+            start = _gfn(gfn_x(start) & ~((1UL << order) - 1));
+        else
+            p2m_put_l3_page(mfn, t);
+    }
+
+    /*
+     * Update lowest_mapped_gfn so on the next call we still start where
+     * we stopped.
+     */
+    p2m->lowest_mapped_gfn = start;
+
+    p2m_write_unlock(p2m);
+
+    return rc;
 }
 
 int p2m_cache_flush(struct domain *d, gfn_t start, unsigned long nr)