Translating the start addresses of the rings is not enough, we need to
be sure all the ring is made available by the guest.
It depends on the size of the rings, which is not known on SET_VRING_ADDR
reception. Furthermore, we need to be be safe against vring pages
invalidates.
This patch introduces a new access_ok flag per virtqueue, which is set
when all the rings are mapped, and cleared as soon as a page used by a
ring is invalidated. The invalidation part is implemented in a following
patch.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
When IOMMU is enabled, the ring addresses set by the
VHOST_USER_SET_VRING_ADDR requests are guest's IO virtual addresses,
whereas Qemu virtual addresses when IOMMU is disabled.
When enabled and the required translation is not in the IOTLB cache,
an IOTLB miss request is sent, but being called by the vhost-user
socket handling thread, the function does not wait for the requested
IOTLB update.
The function will be called again on the next IOTLB update message
reception if matching the vring addresses.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
This patch postpones rings addresses translations and checks, as
addresses sent by the master shuld not be interpreted as long as
ring is not started and enabled[0].
When protocol features aren't negotiated, the ring is started in
enabled state, so the addresses translations are postponed to
vhost_user_set_vring_kick().
Otherwise, it is postponed to when ring is enabled, in
vhost_user_set_vring_enable().
[0]: http://lists.nongnu.org/archive/html/qemu-devel/2017-05/msg04355.html
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
numa_realloc() reallocates the virtio_net device structure and
updates the vhost_devices[] table with the new pointer if the rings
are allocated different NUMA node.
Problem is that vhost_user_msg_handler() still dereferences old
pointer afterward.
This patch prevents this by fetching again the dev pointer in
vhost_devices[] after messages have been handled.
Fixes: af295ad469 ("vhost: realloc device and queues to same numa node as vring desc")
Cc: stable@dpdk.org
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
When VHOST_USER_F_PROTOCOL_FEATURES is negotiated, the ring is not
enabled when started, but enabled through dedicated
VHOST_USER_SET_VRING_ENABLE request.
When not negotiated, the ring is started in enabled state, at
VHOST_USER_SET_VRING_KICK request time.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
Vhost-user device IOTLB protocol extension introduces
VHOST_USER_IOTLB message type. The associated payload is the
vhost_iotlb_msg struct defined in Kernel, which in this was can
be either an IOTLB update or invalidate message.
On IOTLB update, the virtqueues get notified of a new entry.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
Currently, only QEMU sends requests, the backend sends
replies. In some cases, the backend may need to send
requests to QEMU, like IOTLB miss events when IOMMU is
supported.
This patch introduces a new channel for such requests.
QEMU sends a file descriptor of a new socket using
VHOST_USER_SET_SLAVE_REQ_FD.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
send_vhost_message() is currently only used to send
replies, so it modifies message flags to perpare the
reply.
With upcoming channel for backend initiated request,
this function can be used to send requests.
This patch introduces a new send_vhost_reply() that
does the message flags modifications, and makes
send_vhost_message() generic.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
This patch adaptively batches the small guest memory copies.
By batching the small copies, the efficiency of executing the
memory LOAD instructions can be improved greatly, because the
memory LOAD latency can be effectively hidden by the pipeline.
We saw great performance boosts for small packets PVP test.
This patch improves the performance for small packets, and has
distinguished the packets by size. So although the performance
for big packets doesn't change, it makes it relatively easy to
do some special optimizations for the big packets too.
Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Signed-off-by: Zhiyong Yang <zhiyong.yang@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Acked-by: Yuanhan Liu <yliu@fridaylinux.org>
Since vhost_user_set_features failure is not handled in any way, a
single error log has been added to at least to let the user know that
something has gone wrong.
Signed-off-by: Dariusz Stojaczyk <dariuszx.stojaczyk@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
The queue allocation was changed, from allocating one queue-pair at a
time to one queue at a time. Most of the changes have been done, but
just with one being missed: the size of copying the old queue is still
based on queue-pair at numa_realloc(), which leads to overwritten issue.
As a result, crash may happen.
Fix it by specifying the right copy size. Also, the net queue macros
are not used any more. Remove them.
Fixes: ab4d7b9f1a ("vhost: turn queue pair to vring")
Cc: stable@dpdk.org
Reported-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Jens Freimann <jfreiman@redhat.com>
Tested-by: Ciara Loftus <ciara.loftus@intel.com>
Accessing fields of a packed struct through unaligned pointers is
undefined behavior. Instead of passing pointers to particular fields,
a pointer to the root struct should be used. This patch does exactly
that.
Signed-off-by: Daniel Verkamp <daniel.verkamp@intel.com>
Signed-off-by: Dariusz Stojaczyk <dariuszx.stojaczyk@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
This patch fixes a memory leak.
virtio_net::guest_pages is allocated in vhost_setup_mem_table(),
reallocated in add_one_guest_page(), but never freed.
Fixes: e246896178 ("vhost: get guest/host physical address mappings")
Cc: stable@dpdk.org
Signed-off-by: Dariusz Stojaczyk <dariuszx.stojaczyk@intel.com>
Reviewed-by: Jens Freimann <jfreiman@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
When we try to allocate guest pages we need to check the return value of
malloc(). Print an error message and return when it fails.
Signed-off-by: Jens Freimann <jfreiman@redhat.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
It doesn't make any sense to invoke destroy_device() callback at
while handling SET_MEM_TABLE message.
From the vhost-user spec, it's the GET_VRING_BASE message indicates
the end of a vhost device: the destroy_device() should be invoked
from there (luckily, we already did that).
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Features could be changed after the feature negotiation. For example,
VHOST_F_LOG_ALL will be set/cleared at the start/end of live migration,
respecitively. Thus, we need a new callback to inform the application
on such change.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
They are virtio-net specific and should be defined inside the virtio-net
driver.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Currently, we check vq->desc, vq->kickfd and vq->callfd to know whether
a virtio device is ready or not. However, we only do it when handling
SET_VRING_KICK message, which could be wrong if a vhost-user frontend
send SET_VRING_KICK first and SET_VRING_CALL later.
To work for all possible vhost-user frontend implementations, we could
move the ready check at the end of vhost-user message handler.
Meanwhile, since we do the check more often than before, the "virtio
not ready" message is dropped, to not flood the screen.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The queue pair is very virtio-net specific, other devices don't have
such concept. To make it generic, we should log the number of vrings
instead of the number of queue pairs.
This patch just does a simple convert, a later patch would export the
number of vrings to applications.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Some vhost-user driver may need this info to setup its own page tables
for GPA (guest physical addr) to HPA (host physical addr) translation.
SPDK (Storage Performance Development Kit) is one example.
Besides, by exporting this memory info, we could also export the
gpa_to_vva() as an inline function, which helps for performance.
Otherwise, it has to be referenced indirectly by a "vid".
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Assume there is an application both support vhost-user net and
vhost-user scsi, the callback should be different. Making notify
ops per vhost driver allow application define different set of
callbacks for different driver.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
This patch adds a new status flag indicating the Virtio device
is ready to operate.
This is required to be able to call rte_vhost_mtu_get() in the
.new_device() callback, as rte_vhost_mtu_get needs that the
negotiation is done, but it is too early to rely on running status
flag, which is set just after .new_device() returns.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
This patch implements the vhost-user MTU protocol feature support.
When VIRTIO_NET_F_MTU is negotiated, QEMU notifies the vhost-user
backend with the configured MTU if dedicated protocol feature is
supported.
The value can be used by the application to ensure consistency with
value set by the user.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
We used to allocate queues based on the index from SET_VRING_CALL
request: if corresponding queue hasn't been allocated, allocate it.
Though it's pratically right (it's the first per-vring request we
will get from QEMU for vhost-user negotiation), but it's not technically
right: it's not documented in the vhost-user spec that it will always
be the first per-vring request. For example, SET_VRING_ADDR could also
be the first per-vring request.
Thus, we should not depend the SET_VRING_CALL on queue allocation.
Instead, we could catch all the per-vring messages at the entrance of
request handler, and allocate one if it hasn't been allocated before.
By that, we could remove a hack.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Found with clang static analysis:
lib/librte_vhost/vhost_user.c:996:3: warning:
Value stored to 'ret' is never read
ret = vhost_user_get_vring_base(dev, &msg.payload.state);
^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Signed-off-by: Emmanuel Roullit <emmanuel.roullit@gmail.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Setting up the mapping from GPA (guest physical address) to HPA (guest
physical address) could be very time consuming when the guest memory is
backened with small pages (4K). The bigger the guest memory, the longer
it takes. This could lead a very long vhost-user negotiation.
Since the mapping is only needed in zero copy mode so far, we could
avoid such time consuming settup when zero copy is turned off (which is
the default case).
It's actually a workaround, a right fix might be to start a new thread,
and hide the big latency there.
Fixes: e246896178 ("vhost: get guest/host physical address mappings")
Cc: stable@dpdk.org
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
REPLY_ACK features provide a generic way for QEMU to ensure both
completion and success of a request.
As described in vhost-user spec in QEMU repository, QEMU sets
VHOST_USER_NEED_REPLY flag (bit 3) when expecting a reply_ack from
the backend. Backend must reply with 0 for success or non-zero
otherwise when flag is set.
Currently, only VHOST_USER_SET_MEM_TABLE request implements reply_ack,
in order to synchronize mapping updates.
This patch enables REPLY_ACK feature generally, but only checks error
code for VHOST_USER_SET_MEM_TABLE.
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
When reg_size < page_size the function read in
rte_mem_virt2phy would not return, because
host_user_addr is invalid.
Fixes: e246896178 ("vhost: get guest/host physical address mappings")
Cc: stable@dpdk.org
Signed-off-by: Haifeng Lin <haifeng.lin@huawei.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
The basic idea is to shadow the used ring update: update them into a
local buffer first, and then flush them all to the virtio used vring
at once in the end.
And since we do avail ring reservation before enqueuing data, we would
know which and how many descs will be used. Which means we could update
the shadow used ring at the reservation time. It also introduce another
slight advantage: we don't need access the desc->flag any more inside
copy_mbuf_to_desc_mergeable().
Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Jianbo Liu <jianbo.liu@linaro.org>
Tested-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The basic idea of dequeue zero copy is, instead of copying data from
the desc buf, here we let the mbuf reference the desc buf addr directly.
Doing so, however, has one major issue: we can't update the used ring
at the end of rte_vhost_dequeue_burst. Because we don't do the copy
here, an update of the used ring would let the driver to reclaim the
desc buf. As a result, DPDK might reference a stale memory region.
To update the used ring properly, this patch does several tricks:
- when mbuf references a desc buf, refcnt is added by 1.
This is to pin lock the mbuf, so that a mbuf free from the DPDK
won't actually free it, instead, refcnt is subtracted by 1.
- We chain all those mbuf together (by tailq)
And we check it every time on the rte_vhost_dequeue_burst entrance,
to see if the mbuf is freed (when refcnt equals to 1). If that
happens, it means we are the last user of this mbuf and we are
safe to update the used ring.
- "struct zcopy_mbuf" is introduced, to associate an mbuf with the
right desc idx.
Dequeue zero copy is introduced for performance reason, and some rough
tests show about 50% perfomance boost for packet size 1500B. For small
packets, (e.g. 64B), it actually slows a bit down (well, it could up to
15%). That is expected because this patch introduces some extra works,
and it outweighs the benefit from saving few bytes copy.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Qian Xu <qian.q.xu@intel.com>
So far, we retrieve both the used ring and avail ring idx by the var
last_used_idx; it won't be a problem because the used ring is updated
immediately after those avail entries are consumed.
But that's not true when dequeue zero copy is enabled, that used ring is
updated only when the mbuf is consumed. Thus, we need use another var to
note the last avail ring idx we have consumed.
Therefore, last_avail_idx is introduced.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Qian Xu <qian.q.xu@intel.com>
So that we can convert a guest physical address to host physical
address, which will be used in later Tx zero copy implementation.
MAP_POPULATE is set while mmaping guest memory regions, to make
sure the page tables are setup and then rte_mem_virt2phy() could
yield proper physical address.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Qian Xu <qian.q.xu@intel.com>
Due to history reason (that vhost-cuse comes before vhost-user), some
fields for maintaining the vhost-user memory mappings (such as mmapped
address and size, with those we then can unmap on destroy) are kept in
"orig_region_map" struct, a structure that is defined only in vhost-user
source file.
The right way to go is to remove the structure and move all those fields
into virtio_memory_region struct. But we simply can't do that before,
because it breaks the ABI.
Now, thanks to the ABI refactoring, it's never been a blocking issue
any more. And here it goes: this patch removes orig_region_map and
redefines virtio_memory_region, to include all necessary info.
With that, we can simplify the guest/host address convert a bit.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tested-by: Qian Xu <qian.q.xu@intel.com>
No need to use a pointer to store/retrieve features.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Invoke get_device() at the beginning of vhost_user_msg_handler, so that
we could check the return value once. Which could save tons of duplicate
get-and-check device.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Some functions are with prefix "user_", while others with "vhost_".
Making them all starting with "vhost_user_" to unify the function names.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Due to history reason (that we have 2 vhost implementations), some
messages are handled in two calls: vhost specific implementation
handles it first and then invoke the common one to do another handling.
We have one implementation only now, we could write one method for
each message. Here fold those common handles to corresponding vhost
user handler.
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
The code structure is a bit messy now. For example, vhost-user message
handling is spread to three different files:
vhost-net-user.c virtio-net.c virtio-net-user.c
Where, vhost-net-user.c is the entrance to handle all those messages
and then invoke the right method for a specific message. Some of them
are stored at virtio-net.c, while others are stored at virtio-net-user.c.
The truth is all of them should be in one file, vhost_user.c.
So this patch refactors the source code structure: mainly on renaming
files and moving code from one file to another file that is more suitable
for storing it. Thus, no functional changes are made.
After the refactor, the code structure becomes to:
- socket.c handles all vhost-user socket file related stuff, such
as, socket file creation for server mode, reconnection
for client mode.
- vhost.c mainly on stuff like vhost device creation/destroy/reset.
Most of the vhost API implementation are there, too.
- vhost_user.c all stuff about vhost-user messages handling goes there.
- virtio_net.c all stuff about virtio-net should go there. It has virtio
net Rx/Tx implementation only so far: it's just a rename
from vhost_rxtx.c
Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
