On 03/07/2018 10:16 AM, Wodkowski, PawelX wrote:
From: Maxime Coquelin [mailto:maxime.coque...@redhat.com]
Sent: Tuesday, March 6, 2018 5:27 PM
To: Kulasek, TomaszX <tomaszx.kula...@intel.com>; y...@fridaylinux.org
Cc: Verkamp, Daniel <daniel.verk...@intel.com>; Harris, James R
<james.r.har...@intel.com>; Wodkowski, PawelX
<pawelx.wodkow...@intel.com>; firstname.lastname@example.org; Stojaczyk, DariuszX
Subject: Re: [dpdk-dev] [PATCH] vhost: stop device before updating public vring
On 03/05/2018 05:11 PM, Tomasz Kulasek wrote:
For now DPDK assumes that callfd, kickfd and last_idx are being set just
once during vring initialization and device cannot be running while DPDK
receives SET_VRING_KICK, SET_VRING_CALL and SET_VRING_BASE messages.
However, that assumption is wrong. For Vhost SCSI messages might arrive
at any point of time, possibly multiple times, one after another.
QEMU issues SET_VRING_CALL once during device initialization, then again
during device start. The second message will close previous callfd,
which is still being used by the user-implementation of vhost device.
This results in writing to invalid (closed) callfd.
Other messages like SET_FEATURES, SET_VRING_ADDR etc also will change
internal state of VQ or device. To prevent race condition device should
also be stopped before updateing vring data.
Signed-off-by: Dariusz Stojaczyk<dariuszx.stojac...@intel.com>
Signed-off-by: Pawel Wodkowski<pawelx.wodkow...@intel.com>
Signed-off-by: Tomasz Kulasek<tomaszx.kula...@intel.com>
lib/librte_vhost/vhost_user.c | 40
1 file changed, 40 insertions(+)
In last release, we have introduced a per-virtqueue lock to protect
vring handling against asynchronous device changes.
I think that would solve the issue you are facing, but you would need
to export the VQs locking functions to the vhost-user lib API to be
able to use it.
I don't think your current patch is the right solution anyway, because
it destroys the device in case we don't want it to remain alive, like
set_log_base, or set_features when only the logging feature gets
Please correct me if I can't see something obvious, but how this lock protect
SET_MEM_TABLE message? Current flow you are thinking of is:
DPDK vhost-user thread
2.2. Process vrings and copy all data
Yes, it will synchronize access to virtio_net structure but what if the BACKEND
zero copy mode and/or pass buffers to physical device? The request will
not end in 2.2 and you unmap the memory regions in the middle of request.
Even worse, the physical device will just abort the request but BACKEND can
or write random memory because BACKEND try to use invalid memory address
(retrieved at request start).
Right, it doesn't work with zero-copy.
To use this per-virtqueue lock:
1. the lock need to be held from request start to the end - but this can starve
vhost-user thread as there might be many request on-the-fly and when one is done
the new one might be started.
2. Becouse we don't know if something changed between requst start and request
BACKEND need walk through all descriptors chain at the request end and do the
The SET_MEM_TABLE is most obvious message but the same is true for other like
VHOST_IOTLB_INVALIDATE or SET_FEATURES.
SET_FEATURE should never be sent as soon as the device is started,
except to enable logging.
For VHOST_IOTLB_INVALIDATE, the solution might be to have a ref counter
per entry, and to only remove it for the cache once it is zero and send
the reply-ack tothe master once this is done. But the cost would be huge
as with large entries, a lot of threads might increment/decrement the
same variable so there will be contention.
For all other cases, like SET_MEM_TABLE, maybe the solution is to
disable/enable all the queues using the existing ops.
The application or library would have to take care that no guest buffers
are in the wild before returning from the disable.
Do you think that would work?