This patch tries to implement an device IOTLB for vhost. This could be
used with userspace(qemu) implementation of DMA remapping
to emulate an IOMMU for the guest.
The idea is simple, cache the translation in a software device IOTLB
(which is implemented as an interval tree) in vhost and use vhost_net
file descriptor for reporting IOTLB miss and IOTLB
update/invalidation. When vhost meets an IOTLB miss, the fault
address, size and access can be read from the file. After userspace
finishes the translation, it writes the translated address to the
vhost_net file to update the device IOTLB.
When device IOTLB is enabled by setting VIRTIO_F_IOMMU_PLATFORM all vq
addresses set by ioctl are treated as iova instead of virtual address and
the accessing can only be done through IOTLB instead of direct userspace
memory access. Before each round or vq processing, all vq metadata is
prefetched in device IOTLB to make sure no translation fault happens
during vq processing.
In most cases, virtqueues are contiguous even in virtual address space.
The IOTLB translation for virtqueue itself may make it a little
slower. We might add fast path cache on top of this patch.
Signed-off-by: Jason Wang <jasowang@redhat.com>
[mst: use virtio feature bit: VHOST_F_DEVICE_IOTLB -> VIRTIO_F_IOMMU_PLATFORM ]
[mst: fix build warnings ]
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
[ weiyj.lk: missing unlock on error ]
Signed-off-by: Wei Yongjun <weiyj.lk@gmail.com>
Current pre-sorted memory region array has some limitations for future
device IOTLB conversion:
1) need extra work for adding and removing a single region, and it's
expected to be slow because of sorting or memory re-allocation.
2) need extra work of removing a large range which may intersect
several regions with different size.
3) need trick for a replacement policy like LRU
To overcome the above shortcomings, this patch convert it to interval
tree which can easily address the above issue with almost no extra
work.
The patch could be used for:
- Extend the current API and only let the userspace to send diffs of
memory table.
- Simplify Device IOTLB implementation.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This patch introduces vhost memory accessors which were just wrappers
for userspace address access helpers. This is a requirement for vhost
device iotlb implementation which will add iotlb translations in those
accessors.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
We use spinlock to synchronize the work list now which may cause
unnecessary contentions. So this patch switch to use llist to remove
this contention. Pktgen tests shows about 5% improvement:
Before:
~1300000 pps
After:
~1370000 pps
Signed-off-by: Jason Wang <jasowang@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
We used to implement the work flushing through tracking queued seq,
done seq, and the number of flushing. This patch simplify this by just
implement work flushing through another kind of vhost work with
completion. This will be used by lockless enqueuing patch.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This patch tries to poll for new added tx buffer or socket receive
queue for a while at the end of tx/rx processing. The maximum time
spent on polling were specified through a new kind of vring ioctl.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This patch introduces a helper which will return true if we're sure
that the available ring is empty for a specific vq. When we're not
sure, e.g vq access failure, return false instead. This could be used
for busy polling code to exit the busy loop.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This path introduces a helper which can give a hint for whether or not
there's a work queued in the work list. This could be used for busy
polling code to exit the busy loop.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Looking at how callers use this, maybe we should just rename init_used
to vhost_vq_init_access. The _used suffix was a hint that we
access the vq used ring. But maybe what callers care about is
that it must be called after access_ok.
Also, this function manipulates the vq->is_le field which isn't related
to the vq used ring.
This patch simply renames vhost_init_used() to vhost_vq_init_access() as
suggested by Michael.
No behaviour change.
Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The default use case for vhost is when the host and the vring have the
same endianness (default native endianness). But there are cases where
they differ and vhost should byteswap when accessing the vring.
The first case is when the host is big endian and the vring belongs to
a virtio 1.0 device, which is always little endian.
This is covered by the vq->is_le field. This field is initialized when
userspace calls the VHOST_SET_FEATURES ioctl. It is reset when the device
stops.
We already have a vhost_init_is_le() helper, but the reset operation is
opencoded as follows:
vq->is_le = virtio_legacy_is_little_endian();
It isn't clear that we are resetting vq->is_le here.
This patch moves the code to a helper with a more explicit name.
The other case where we may have to byteswap is when the architecture can
switch endianness at runtime (bi-endian). If endianness differs in the host
and in the guest, then legacy devices need to be used in cross-endian mode.
This mode is available with CONFIG_VHOST_CROSS_ENDIAN_LEGACY=y, which
introduces a vq->user_be field. Userspace may enable cross-endian mode
by calling the SET_VRING_ENDIAN ioctl before the device is started. The
cross-endian mode is disabled when the device is stopped.
The current names of the helpers that manipulate vq->user_be are unclear.
This patch renames those helpers to clearly show that this is cross-endian
stuff and with explicit enable/disable semantics.
No behaviour change.
Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
We don't want side effects. If something fails, we rollback vq->is_le to
its previous value.
Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
commit 5d9a07b0de ("vhost: relax used
address alignment") fixed the alignment for the used virtual address,
but not for the physical address used for logging.
That's a mistake: alignment should clearly be the same for virtual and
physical addresses,
Cc: stable@vger.kernel.org
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
callers of vhost_kvzalloc() expect the same behaviour on
allocation error as from kmalloc/vmalloc i.e. NULL return
value. So just return vzmalloc() returned value instead of
returning ERR_PTR(-ENOMEM)
Fixes: 4de7255f7d ("vhost: extend memory regions allocation to vmalloc")
Spotted-by: Dan Carpenter <dan.carpenter@oracle.com>
Suggested-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Igor Mammedov <imammedo@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
While reviewing vhost log code, I found out that log_file is never
set. Note: I haven't tested the change (QEMU doesn't use LOG_FD yet).
Cc: stable@vger.kernel.org
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
it became possible to use a bigger amount of memory
slots, which is used by memory hotplug for
registering hotplugged memory.
However QEMU crashes if it's used with more than ~60
pc-dimm devices and vhost-net enabled since host kernel
in module vhost-net refuses to accept more than 64
memory regions.
Allow to tweak limit via max_mem_regions module paramemter
with default value set to 64 slots.
Signed-off-by: Igor Mammedov <imammedo@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
with large number of memory regions we could end up with
high order allocations and kmalloc could fail if
host is under memory pressure.
Considering that memory regions array is used on hot path
try harder to allocate using kmalloc and if it fails resort
to vmalloc.
It's still better than just failing vhost_set_memory() and
causing guest crash due to it when a new memory hotplugged
to guest.
I'll still look at QEMU side solution to reduce amount of
memory regions it feeds to vhost to make things even better,
but it doesn't hurt for kernel to behave smarter and don't
crash older QEMU's which could use large amount of memory
regions.
Signed-off-by: Igor Mammedov <imammedo@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
For default region layouts performance stays the same
as linear search i.e. it takes around 210ns average for
translate_desc() that inlines find_region().
But it scales better with larger amount of regions,
235ns BS vs 300ns LS with 55 memory regions
and it will be about the same values when allowed number
of slots is increased to 509 like it has been done in kvm.
Signed-off-by: Igor Mammedov <imammedo@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This patch brings cross-endian support to vhost when used to implement
legacy virtio devices. Since it is a relatively rare situation, the
feature availability is controlled by a kernel config option (not set
by default).
The vq->is_le boolean field is added to cache the endianness to be
used for ring accesses. It defaults to native endian, as expected
by legacy virtio devices. When the ring gets active, we force little
endian if the device is modern. When the ring is deactivated, we
revert to the native endian default.
If cross-endian was compiled in, a vq->user_be boolean field is added
so that userspace may request a specific endianness. This field is
used to override the default when activating the ring of a legacy
device. It has no effect on modern devices.
Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
virtio 1.0 only requires used address to be 4 byte aligned,
vhost required 8 bytes (size of vring_used_elem).
Fix up vhost to match that.
Additionally, while vhost correctly requires 8 byte
alignment for log, it's unconnected to used ring:
it's a consequence that log has u64 entries.
Tweak code to make that clearer.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Most places in vhost can use __get/__put_user rather than
get/put_user since addresses are pre-validated.
This should be good for performance, but this also
will help make code sparse-clean: get/put_user macros
don't play well with __virtioXX bitwise tags.
Switch to get/put_user to __ variants everywhere in vhost.
There's one exception - for consistency switch that
as well, and add an explicit access_ok check.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
commit 2ae76693b8bcabf370b981cd00c36cd41d33fabc
vhost: replace rcu with mutex
replaced rcu sync for memory accesses with VQ mutex locl/unlock.
This is correct since all accesses are under VQ mutex, but incomplete:
we still do useless rcu lock/unlock operations, someone might copy this
code into some other context where this won't be right.
This use of RCU is also non standard and hard to understand.
Let's copy the pointer to each VQ structure, this way
the access rules become straight-forward, and there's
no need for RCU anymore.
Reported-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Refactor code to make sure features are only accessed
under VQ mutex. This makes everything simpler, no need
for RCU here anymore.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
All memory accesses are done under some VQ mutex.
So lock/unlock all VQs is a faster equivalent of synchronize_rcu()
for memory access changes.
Some guests cause a lot of these changes, so it's helpful
to make them faster.
Reported-by: "Gonglei (Arei)" <arei.gonglei@huawei.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Since vhost_dev_init() forever return 0, some branches are never run,
therefore need to be removed.
Signed-off-by: Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
the wake_up_process func is included by spin_lock/unlock in
vhost_work_queue,
but it could be done outside the spin_lock.
I have test it with kernel 3.0.27 and guest suse11-sp2 using iperf,
the num as below.
original modified
thread_num tp(Gbps) vhost(%) | tp(Gbps) vhost(%)
1 9.59 28.82 | 9.59 27.49
8 9.61 32.92 | 9.62 26.77
64 9.58 46.48 | 9.55 38.99
256 9.6 63.7 | 9.6 52.59
Signed-off-by: Chuanyu Qin <qinchuanyu@huawei.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let vhost_add_used() to use vhost_add_used_n() to reduce the code
duplication. To avoid the overhead brought by __copy_to_user(). We will use
put_user() when one used need to be added.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
memcpy_fromiovec is moved from net/core/iovec.c to lib/iovec.c.
linux/uio.h provides the declaration for memcpy_fromiovec.
Include linux/uio.h instead of inux/socket.h for it.
Signed-off-by: Asias He <asias@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, vhost-net and vhost-scsi are sharing the vhost core code.
However, vhost-scsi shares the code by including the vhost.c file
directly.
Making vhost a separate module makes it is easier to share code with
other vhost devices.
Signed-off-by: Asias He <asias@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
If device has an owner, we shouldn't touch ubuf_info
since it might be in use.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
RESET_OWNER ioctl would leave the fd in a bad state if
memory allocation failed: device is stopped
but owner is not reset. Make state changes
after allocating memory, such that a failed
ioctl has no effect.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
On top of 'vhost: Allow device specific fields per vq', we can move device
specific fields to device virt queue from vhost virt queue.
Signed-off-by: Asias He <asias@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This is useful for any device who wants device specific fields per vq.
For example, tcm_vhost wants a per vq field to track requests which are
in flight on the vq. Also, on top of this we can add patches to move
things like ubufs from vhost.h out to net.c.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Asias He <asias@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
After commit 2b8b328b61 (vhost_net: handle polling
errors when setting backend), we in fact track the polling state through
poll->wqh, so there's no need to duplicate the work with an extra
vhost_net_polling_state. So this patch removes this and make the code simpler.
This patch also removes the all tx starting/stopping code in tx path according
to Michael's suggestion.
Netperf test shows almost the same result in stream test, but gets improvements
on TCP_RR tests (both zerocopy or copy) especially on low load cases.
Tested between multiqueue kvm guest and external host with two direct
connected 82599s.
zerocopy disabled:
sessions|transaction rates|normalize|
before/after/+improvements
1 | 9510.24/11727.29/+23.3% | 693.54/887.68/+28.0% |
25| 192931.50/241729.87/+25.3% | 2376.80/2771.70/+16.6% |
50| 277634.64/291905.76/+5% | 3118.36/3230.11/+3.6% |
zerocopy enabled:
sessions|transaction rates|normalize|
before/after/+improvements
1 | 7318.33/11929.76/+63.0% | 521.86/843.30/+61.6% |
25| 167264.88/242422.15/+44.9% | 2181.60/2788.16/+27.8% |
50| 272181.02/294347.04/+8.1% | 3071.56/3257.85/+6.1% |
Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the polling errors were ignored, which can lead following issues:
- vhost remove itself unconditionally from waitqueue when stopping the poll,
this may crash the kernel since the previous attempt of starting may fail to
add itself to the waitqueue
- userspace may think the backend were successfully set even when the polling
failed.
Solve this by:
- check poll->wqh before trying to remove from waitqueue
- report polling errors in vhost_poll_start(), tx_poll_start(), the return value
will be checked and returned when userspace want to set the backend
After this fix, there still could be a polling failure after backend is set, it
will addressed by the next patch.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
vring changes already do a flush internally where appropriate, so we do
not need a second flush.
It's currently not very expensive but a follow-up patch makes flush more
heavy-weight, so remove the extra flush here to avoid regressing
performance if call or kick fds are changed on data path.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
If a single descriptor crosses a region, the
second chunk length should be decremented
by size translated so far, instead it includes
the full descriptor length.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Zerocopy handling code is vhost-net specific.
Move it from vhost.c/vhost.h out to net.c
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This will be used to disable zerocopy when error rate
is high.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Better document macros for DMA tracking. Add an
explicit one for DMA in progress instead of
relying on user supplying len != 1.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Even if skb is marked for zero copy, net core might still decide
to copy it later which is somewhat slower than a copy in user context:
besides copying the data we need to pin/unpin the pages.
Add a parameter reporting such cases through zero copy callback:
if this happens a lot, device can take this into account
and switch to copying in user context.
This patch updates all users but ignores the passed value for now:
it will be used by follow-up patches.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
