qemu-devel
[Top][All Lists]
Advanced
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [RFC 1/2] iova_tree: add an id member to DMAMap

From: Si-Wei Liu
Subject: Re: [RFC 1/2] iova_tree: add an id member to DMAMap
Date: Wed, 1 May 2024 16:13:52 -0700
User-agent: Mozilla Thunderbird 


On 4/30/2024 10:19 AM, Eugenio Perez Martin wrote:

On Tue, Apr 30, 2024 at 7:55 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote:



On 4/29/2024 1:14 AM, Eugenio Perez Martin wrote:

On Thu, Apr 25, 2024 at 7:44 PM Si-Wei Liu <si-wei.liu@oracle.com> wrote:


On 4/24/2024 12:33 AM, Eugenio Perez Martin wrote:

On Wed, Apr 24, 2024 at 12:21 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote:

On 4/22/2024 1:49 AM, Eugenio Perez Martin wrote:

On Sat, Apr 20, 2024 at 1:50 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote:

On 4/19/2024 1:29 AM, Eugenio Perez Martin wrote:

On Thu, Apr 18, 2024 at 10:46 PM Si-Wei Liu <si-wei.liu@oracle.com> wrote:

On 4/10/2024 3:03 AM, Eugenio Pérez wrote:

IOVA tree is also used to track the mappings of virtio-net shadow
virtqueue.  This mappings may not match with the GPA->HVA ones.

This causes a problem when overlapped regions (different GPA but same
translated HVA) exists in the tree, as looking them by HVA will return
them twice.  To solve this, create an id member so we can assign unique
identifiers (GPA) to the maps.

Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
---
       include/qemu/iova-tree.h | 5 +++--
       util/iova-tree.c         | 3 ++-
       2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/include/qemu/iova-tree.h b/include/qemu/iova-tree.h
index 2a10a7052e..34ee230e7d 100644
--- a/include/qemu/iova-tree.h
+++ b/include/qemu/iova-tree.h
@@ -36,6 +36,7 @@ typedef struct DMAMap {
           hwaddr iova;
           hwaddr translated_addr;
           hwaddr size;                /* Inclusive */
+    uint64_t id;
           IOMMUAccessFlags perm;
       } QEMU_PACKED DMAMap;
       typedef gboolean (*iova_tree_iterator)(DMAMap *map);
@@ -100,8 +101,8 @@ const DMAMap *iova_tree_find(const IOVATree *tree, const 
DMAMap *map);
        * @map: the mapping to search
        *
        * Search for a mapping in the iova tree that translated_addr overlaps 
with the
- * mapping range specified.  Only the first found mapping will be
- * returned.
+ * mapping range specified and map->id is equal.  Only the first found
+ * mapping will be returned.
        *
        * Return: DMAMap pointer if found, or NULL if not found.  Note that
        * the returned DMAMap pointer is maintained internally.  User should
diff --git a/util/iova-tree.c b/util/iova-tree.c
index 536789797e..0863e0a3b8 100644
--- a/util/iova-tree.c
+++ b/util/iova-tree.c
@@ -97,7 +97,8 @@ static gboolean iova_tree_find_address_iterator(gpointer key, 
gpointer value,

           needle = args->needle;
           if (map->translated_addr + map->size < needle->translated_addr ||
-        needle->translated_addr + needle->size < map->translated_addr) {
+        needle->translated_addr + needle->size < map->translated_addr ||
+        needle->id != map->id) {

It looks this iterator can also be invoked by SVQ from
vhost_svq_translate_addr() -> iova_tree_find_iova(), where guest GPA
space will be searched on without passing in the ID (GPA), and exact
match for the same GPA range is not actually needed unlike the mapping
removal case. Could we create an API variant, for the SVQ lookup case
specifically? Or alternatively, add a special flag, say skip_id_match to
DMAMap, and the id match check may look like below:

(!needle->skip_id_match && needle->id != map->id)

I think vhost_svq_translate_addr() could just call the API variant or
pass DMAmap with skip_id_match set to true to svq_iova_tree_find_iova().


I think you're totally right. But I'd really like to not complicate
the API of the iova_tree more.

I think we can look for the hwaddr using memory_region_from_host and
then get the hwaddr. It is another lookup though...

Yeah, that will be another means of doing translation without having to
complicate the API around iova_tree. I wonder how the lookup through
memory_region_from_host() may perform compared to the iova tree one, the
former looks to be an O(N) linear search on a linked list while the
latter would be roughly O(log N) on an AVL tree?

Even worse, as the reverse lookup (from QEMU vaddr to SVQ IOVA) is
linear too. It is not even ordered.

Oh Sorry, I misread the code and I should look for g_tree_foreach ()
instead of g_tree_search_node(). So the former is indeed linear
iteration, but it looks to be ordered?

https://github.com/GNOME/glib/blob/main/glib/gtree.c#L1115

The GPA / IOVA are ordered but we're looking by QEMU's vaddr.

If we have these translations:
[0x1000, 0x2000] -> [0x10000, 0x11000]
[0x2000, 0x3000] -> [0x6000, 0x7000]

We will see them in this order, so we cannot stop the search at the first node.

Yeah, reverse lookup is unordered indeed, anyway.


But apart from this detail you're right, I have the same concerns with
this solution too. If we see a hard performance regression we could go
to more complicated solutions, like maintaining a reverse IOVATree in
vhost-iova-tree too. First RFCs of SVQ did that actually.

Agreed, yeap we can use memory_region_from_host for now.  Any reason why
reverse IOVATree was dropped, lack of users? But now we have one!


No, it is just simplicity. We already have an user in the hot patch in
the master branch, vhost_svq_vring_write_descs. But I never profiled
enough to find if it is a bottleneck or not to be honest.

Right, without vIOMMU or a lot of virtqueues / mappings, it's hard to
profile and see the difference.

I'll send the new series by today, thank you for finding these issues!

Thanks! In case you don't have bandwidth to add back reverse IOVA tree,
Jonah (cc'ed) may have interest in looking into it.


Actually, yes. I've tried to solve it using:
memory_region_get_ram_ptr -> It's hard to get this pointer to work
without messing a lot with IOVATree.
memory_region_find -> I'm totally unable to make it return sections
that make sense
flatview_for_each_range -> It does not return the same
MemoryRegionsection as the listener, not sure why.

Ouch, thank you for the summary of attempts that were done earlier.

The only advance I have is that memory_region_from_host is able to
tell if the vaddr is from the guest or not.

Hmmm, then it won't be too useful without a direct means to identifying
the exact memory region associated with the iova that is being mapped.
And, this additional indirection seems introduce a tiny bit of more
latency in the reverse lookup routine (should not be a scalability issue
though if it's a linear search)?


I didn't measure, but I guess yes it might. OTOH these structs may be
cached because virtqueue_pop just looked for them.

Oh, right, that's a good point.



So I'm convinced there must be a way to do it with the memory
subsystem, but I think the best way to do it ATM is to store a
parallel tree with GPA-> SVQ IOVA translations. At removal time, if we
find the entry in this new tree, we can directly remove it by GPA. If
not, assume it is a host-only address like SVQ vrings, and remove by
iterating on vaddr as we do now.

Yeah, this could work I think. On the other hand, given that we are now
trying to improve it, I wonder if possible to come up with a fast
version for the SVQ (host-only address) case without having to look up
twice? SVQ callers should be able to tell apart from the guest case
where GPA -> IOVA translation doesn't exist? Or just maintain a parallel
tree with HVA -> IOVA translations for SVQ reverse lookup only? I feel
SVQ mappings may be worth a separate fast lookup path - unlike guest
mappings, the insertion, lookup and removal for SVQ mappings seem
unavoidable during the migration downtime path.


I think the ideal order is the opposite actually. So:
1) Try for the NIC to support _F_VRING_ASID, no translation needed by QEMU
Right, that's the case for _F_VRING_ASID, which is simple and easy to deal with. Though I think this is an edge case across all vendor devices, as most likely only those no-chip IOMMU parents may support it. It's a luxury for normal device to steal another VF for this ASID feature... 


2) Try reverse lookup from HVA to GPA. Since dataplane should fit
this, we should test this first
So instead of a direct lookup from HVA to IOVA, the purpose of the extra reverse lookup from HVA to GPA is to verify the validity of GPA (avoid from being mistakenly picked from the overlapped region)? But this would seem require scanning the entire GPA space to identify possible GPA ranges that are potentially overlapped? I wonder if there exists possibility to simplify this assumption, could we go this extra layer of GPA wide scan and validation, *only* when overlap is indeed detected during memory listerner's region_add (say during which we try to insert a duplicate / overlapped HVA into the HVA -> IOVA tree)? Or simply put, the first match on the reverse lookup would mostly suffice, since we know virtio driver can't use guest memory from these overlapped regions? You may say this assumption is too bold, but do we have other means to guarantee the first match will always hit under SVQ lookup? Given that we don't receive an instance of issue report until we move the memory listener registration upfront to device initialization, I guess there should be some point or under certain condition that the non-overlapped 1:1 translation and lookup can be satisfied. Don't you agree? 

Thanks,
-Siwei

3) Look in SVQ host-only entries (SVQ, current shadow CVQ). It is the
control VQ, speed is not so important.

Overlapping regions may return the wrong SVQ IOVA though. We should
take extra care to make sure these are correctly handled. I mean,
there are valid translations in the tree unless the driver is buggy,
just may need to span many translations.


   It is guaranteed the guest does not
translate to that vaddr and that that vaddr is unique in the tree
anyway.

Does it sound reasonable? Jonah, would you be interested in moving this forward?

My thought would be that the reverse IOVA tree stuff can be added as a
follow-up optimization right after for extended scalability, but for now
as the interim, we may still need some form of simple fix, so as to
quickly unblock the other dependent work built on top of this one and
the early pinning series [1]. With it said, I'm completely fine if
performing the reverse lookup through linear tree walk e.g.
g_tree_foreach(), that should suffice small VM configs with just a
couple of queues and limited number of memory regions. Going forward, to
address the scalability bottleneck, Jonah could just replace the
corresponding API call with the one built on top of reverse IOVA tree (I
presume the use of these iova tree APIs is kind of internal that only
limits to SVQ and vhost-vdpa subsystems) once he gets there, and then
eliminate the other API variants that will no longer be in use. What do
you think about this idea / plan?


Yeah it makes sense to me. Hopefully we can even get rid of the id member.


Thanks,
-Siwei

[1] https://lists.nongnu.org/archive/html/qemu-devel/2024-04/msg00079.html


Thanks!


-Siwei



Thanks,
-Siwei

Thanks!


Of course,
memory_region_from_host() won't search out of the guest memory space for
sure. As this could be on the hot data path I have a little bit
hesitance over the potential cost or performance regression this change
could bring in, but maybe I'm overthinking it too much...

Thanks,
-Siwei


Thanks,
-Siwei

               return false;
           }
reply via email to
[Prev in Thread] Current Thread [Next in Thread]