From: Jason Wang <jasowang@redhat.com>
To: Yongji Xie <xieyongji@bytedance.com>
Cc: "Michael S. Tsirkin" <mst@redhat.com>,
"Stefan Hajnoczi" <stefanha@redhat.com>,
"Stefano Garzarella" <sgarzare@redhat.com>,
"Parav Pandit" <parav@nvidia.com>,
"Christoph Hellwig" <hch@infradead.org>,
"Randy Dunlap" <rdunlap@infradead.org>,
"Matthew Wilcox" <willy@infradead.org>,
viro@zeniv.linux.org.uk, "Jens Axboe" <axboe@kernel.dk>,
bcrl@kvack.org, "Jonathan Corbet" <corbet@lwn.net>,
"Mika Penttilä" <mika.penttila@nextfour.com>,
"Dan Carpenter" <dan.carpenter@oracle.com>,
virtualization@lists.linux-foundation.org,
netdev@vger.kernel.org, kvm@vger.kernel.org,
linux-fsdevel@vger.kernel.org
Subject: Re: [PATCH v5 08/11] vduse: Implement an MMU-based IOMMU driver
Date: Fri, 26 Mar 2021 15:36:19 +0800 [thread overview]
Message-ID: <817336fa-c026-fd4d-dd2e-eb5f40c63ad4@redhat.com> (raw)
In-Reply-To: <CACycT3t+2MC9rQ7iWdWQ4=O3ojCXHvHZ-M7y7AjXoXYZUiAOzQ@mail.gmail.com>
在 2021/3/26 下午2:56, Yongji Xie 写道:
> On Fri, Mar 26, 2021 at 2:16 PM Jason Wang <jasowang@redhat.com> wrote:
>>
>> 在 2021/3/26 下午1:14, Yongji Xie 写道:
>>
>> + }
>> + map->bounce_page = page;
>> +
>> + /* paired with vduse_domain_map_page() */
>> + smp_mb();
>>
>> So this is suspicious. It's better to explain like, we need make sure A
>> must be done after B.
>>
>> OK. I see. It's used to protect this pattern:
>>
>> vduse_domain_alloc_bounce_page: vduse_domain_map_page:
>> write map->bounce_page write map->orig_phys
>> mb() mb()
>> read map->orig_phys read map->bounce_page
>>
>> Make sure there will always be a path to do bouncing.
>>
>> Ok.
>>
>>
>> And it looks to me the iotlb_lock is sufficnet to do the synchronization
>> here. E.g any reason that you don't take it in
>> vduse_domain_map_bounce_page().
>>
>> Yes, we can. But the performance in multi-queue cases will go down if
>> we use iotlb_lock on this critical path.
>>
>> And what's more, is there anyway to aovid holding the spinlock during
>> bouncing?
>>
>> Looks like we can't. In the case that multiple page faults happen on
>> the same page, we should make sure the bouncing is done before any
>> page fault handler returns.
>>
>> So it looks to me all those extra complexitiy comes from the fact that
>> the bounce_page and orig_phys are set by different places so we need to
>> do the bouncing in two places.
>>
>> I wonder how much we can gain from the "lazy" boucning in page fault.
>> The buffer mapped via dma_ops from virtio driver is expected to be
>> accessed by the userspace soon. It looks to me we can do all those
>> stuffs during dma_map() then things would be greatly simplified.
>>
>> If so, we need to allocate lots of pages from the pool reserved for
>> atomic memory allocation requests.
>>
>> This should be fine, a lot of drivers tries to allocate pages in atomic
>> context. The point is to simplify the codes to make it easy to
>> determince the correctness so we can add optimization on top simply by
>> benchmarking the difference.
>>
>> OK. I will use this way in the next version.
>>
>> E.g we have serveral places that accesses orig_phys:
>>
>> 1) map_page(), write
>> 2) unmap_page(), write
>> 3) page fault handler, read
>>
>> It's not clear to me how they were synchronized. Or if it was
>> synchronzied implicitly (via iova allocator?), we'd better document it.
>>
>> Yes.
>>
>> Or simply use spinlock (which is the preferrable way I'd like to go). We
>> probably don't need to worry too much about the cost of spinlock since
>> iova allocater use it heavily.
>>
>> Actually iova allocator implements a per-CPU cache to optimize it.
>>
>> Thanks,
>> Yongji
>>
>>
>> Right, but have a quick glance, I guess what you meant is that usually there's no lock contention unless cpu hot-plug. This can work but the problem is that such synchornization depends on the internal implementation of IOVA allocator which is kind of fragile. I still think we should do that on our own.
>>
> I might miss something. Looks like we don't need any synchronization
> if the page fault handler is removed as you suggested. We should not
> access the same orig_phys concurrently (in map_page() and
> unmap_page()) unless we free the iova before accessing.
>
> Thanks,
> Yongji
You're right. I overestimate the complexitiy that is required by the
synchronization.
Thanks
>
next prev parent reply other threads:[~2021-03-26 7:37 UTC|newest]
Thread overview: 38+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-03-15 5:37 [PATCH v5 00/11] Introduce VDUSE - vDPA Device in Userspace Xie Yongji
2021-03-15 5:37 ` [PATCH v5 01/11] file: Export __receive_fd() to modules Xie Yongji
2021-03-15 9:08 ` Christoph Hellwig
2021-03-15 9:46 ` Yongji Xie
2021-03-15 14:44 ` Christian Brauner
2021-03-16 2:48 ` Yongji Xie
2021-03-25 8:23 ` Christoph Hellwig
2021-03-25 11:04 ` Yongji Xie
2021-03-15 5:37 ` [PATCH v5 02/11] eventfd: Increase the recursion depth of eventfd_signal() Xie Yongji
2021-03-15 5:37 ` [PATCH v5 03/11] vhost-vdpa: protect concurrent access to vhost device iotlb Xie Yongji
2021-03-23 3:02 ` Jason Wang
2021-03-23 7:25 ` Yongji Xie
2021-03-25 11:08 ` Stefano Garzarella
2021-03-15 5:37 ` [PATCH v5 04/11] vhost-iotlb: Add an opaque pointer for vhost IOTLB Xie Yongji
2021-03-15 5:37 ` [PATCH v5 05/11] vdpa: Add an opaque pointer for vdpa_config_ops.dma_map() Xie Yongji
2021-03-15 5:37 ` [PATCH v5 06/11] vdpa: factor out vhost_vdpa_pa_map() Xie Yongji
2021-03-23 3:09 ` Jason Wang
2021-03-15 5:37 ` [PATCH v5 07/11] vdpa: Support transferring virtual addressing during DMA mapping Xie Yongji
2021-03-23 3:13 ` Jason Wang
2021-03-23 7:26 ` Yongji Xie
2021-03-15 5:37 ` [PATCH v5 08/11] vduse: Implement an MMU-based IOMMU driver Xie Yongji
2021-03-24 3:54 ` Jason Wang
2021-03-24 7:39 ` Yongji Xie
2021-03-25 4:52 ` Jason Wang
2021-03-25 7:38 ` Yongji Xie
2021-03-26 4:26 ` Jason Wang
2021-03-26 5:14 ` Yongji Xie
[not found] ` <75e3b941-dfd2-ebbc-d752-8f25c1f14cab@redhat.com>
2021-03-26 6:56 ` Yongji Xie
2021-03-26 7:36 ` Jason Wang [this message]
2021-03-15 5:37 ` [PATCH v5 09/11] vduse: Introduce VDUSE - vDPA Device in Userspace Xie Yongji
2021-03-24 4:43 ` Jason Wang
2021-03-24 8:55 ` Yongji Xie
2021-03-25 6:30 ` Jason Wang
2021-03-25 7:47 ` Yongji Xie
2021-03-15 5:37 ` [PATCH v5 10/11] vduse: Add config interrupt support Xie Yongji
2021-03-24 4:45 ` Jason Wang
2021-03-24 8:56 ` Yongji Xie
2021-03-15 5:37 ` [PATCH v5 11/11] Documentation: Add documentation for VDUSE Xie Yongji
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=817336fa-c026-fd4d-dd2e-eb5f40c63ad4@redhat.com \
--to=jasowang@redhat.com \
--cc=axboe@kernel.dk \
--cc=bcrl@kvack.org \
--cc=corbet@lwn.net \
--cc=dan.carpenter@oracle.com \
--cc=hch@infradead.org \
--cc=kvm@vger.kernel.org \
--cc=linux-fsdevel@vger.kernel.org \
--cc=mika.penttila@nextfour.com \
--cc=mst@redhat.com \
--cc=netdev@vger.kernel.org \
--cc=parav@nvidia.com \
--cc=rdunlap@infradead.org \
--cc=sgarzare@redhat.com \
--cc=stefanha@redhat.com \
--cc=viro@zeniv.linux.org.uk \
--cc=virtualization@lists.linux-foundation.org \
--cc=willy@infradead.org \
--cc=xieyongji@bytedance.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).