Age | Commit message (Collapse) | Author | Files | Lines |
|
We allocate vring_spinlock for all vrings in the vui, not just the ones
being used. So when we free the vui, we have to free vring_spinlock for
all vrings, not just the one being used.
Type: fix
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I3951fda71ce6e11474b04302116ea9e08b404758
|
|
Type: refactor
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
Change-Id: I3788cc857023fafcc8eb6d6ff4524425026a75d8
|
|
Type: improvement
Signed-off-by: Damjan Marion <damarion@cisco.com>
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
Change-Id: Ie595e69af8657b0ee18a84ac71c5d433108d9ef8
|
|
Prep for supporting multiple callbacks, optional args, etc.
Type: improvement
Change-Id: I96244c098712e8213374678623f12527b0e7f387
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Make it shorter to type, easier to debug, make adding callbacks in
future simpler.
Type: improvement
Change-Id: I6cdd6375e36da23bd452a7c7273ff42789e94433
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Type: improvement
Change-Id: Iac01d7830b53819ace8f199554be10ab89ecdb97
Signed-off-by: Nathan Skrzypczak <nathan.skrzypczak@gmail.com>
|
|
Take advantage of the new TX infra and support manual thread placement
Type: improvement
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: Id8023846a2eb59125fcf2f80f4f11eb681cf14dc
|
|
vhost interrupt mode support with mq does not work when coalesce frame
is not configured to 0.
When packed-ring is configured, we were also setting the wrong flag
for want-interrupt.
No need to trigger an interrupt to RX infra if there is at least one
queue in the same thread that is doing polling.
Type: fix
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I8e64250f2aa6bf611a777a53e4d2c1076d83305f
|
|
When vhost and another interface, say tap, are configured for interrupt
mode, performance number may be very low from vhost.
Further analysis discovers that when vhost posts an interrupt to the
RX infra, there is a 10 msec delay in waking up its input routine.
The delay is due to vhost posts the interrupt from the main thread
which tries to wake up the worker thread.
The fix is for vhost to move the posting interrupt call to the
corresponding input worker thread by calling
vnet_hw_if_set_rx_queue_file_index() to set it up.
While at it, streamline the function vhost_user_kickfd_read_ready()
since it will be called from the worker thread.
Type: fix
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I9beedcd33e1558c8335da4ee7fadc51c29ee4589
|
|
Patch 24601 introduced description into template. In vhost_user_process,
template is reused inside the while loop to call clib_file_add. But
description is allocated outside of the loop only once. As a result, the
same storage for description is being referenced by all instances of call
to clib_file_add. As long as we don't call clib_file_del, we may be
good with multiple fds sharing the same storage for the description. When
one of the fds disconnects and frees the description, the other fds is
holding onto the free memory pointer. Bad news eventually happens when
another fd disconnects and frees the description that was already free
previously.
The fix is to move the allocation of description inside the loop to avoid
sharing.
Type: fix
Fixes: gerrit.fd.io/r/c/vpp/+/24601
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: Ie670931acdc2c7b851982d98fd0d837284a19036
|
|
Type: improvement
This patch adds flags to represent the modern NICs capabilities.
Change-Id: I96d38d9ab7eac55974d72795cd100d8337168e1e
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
|
|
The check args->hwaddr is always true and it always copies the
mac address from args->hwaddr even though none was set.
Check args->use_custom_mac instead.
Type: fix
Fixes: gerrit 29970
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I0c51bf1ea79b02c4fbdc3c52e694f186bdd96600
|
|
Type: improvement
Change-Id: I4008cadfd5141f921afbdc09a3ebcd1dcf88eb29
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
VPP only supports a poor man's approach for interrupt notification to the
driver. It uses a simple binary flag for "interrupt needed" or "interrupt
not needed". Most drivers support more sophisticated event index already.
This feature is to add the long due missing feature and make it configurable,
off by default.
Type: feature
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I68dab7dd07045cafb49af97b7f70db9b8131ae03
|
|
Add descriptions to clib_file_t template structures so that
sockets can be identified via the 'show unix file' cli command.
Type: fix
Change-Id: Ibf82d55aa6c7b1126bd252b76d0dc8b7076f5046
Signed-off-by: Paul Vinciguerra <pvinci@vinciconsulting.com>
|
|
Current vhost multi-queues support has a hard limit of 8 queue pairs
due to static vring array. This limit was raised in qemu. VPP should
support more than 8 queue pairs also.
Change static vring allocation to dynamic. When the interface is
created, we allocate 8 queue pairs to begin with. We also keep track
of how many queue pairs that the interface actually uses.
We reply VHOST_USER_GET_QUEUE_NUM with 128 as our maximum number of
support queue pair. When qemu starts initializing queue pair greater
than 8, we expand the vrings as needed on demand.
Type: improvement
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I4a02d987d52d1bbe601b00e71f650fe6ebfcc0d7
|
|
Type: refactor
Change-Id: Ie67dc579e88132ddb1ee4a34cb69f96920101772
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
- reduces number of instructions generated 4 times compared to old code
- adds pool_foreach2 which is more friendly to clang-format
Type: improvement
Change-Id: I51e9c7fb09655c60d883987dadf5b2666c12b3f7
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Type: fix
Change-Id: I7ca955882c0e263a9ace4b14021e51488564e411
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
Type: refactor
Signed-off-by: Neale Ranns <neale.ranns@cisco.com>
Change-Id: Id1801519638a9b97175847d7ed58824fb83433d6
|
|
This is part of bigger refactor.
Type: refactor
Change-Id: I6fc2c0a1e2d217a70952901bcf775b8485bd3c20
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Type: refactor
tap, virtio and vhost use virtio/vhost header files from linux
kernel. Different features are supported on different kernel
versions, making it difficult to use those in VPP. This patch
removes virtio/vhost based header dependencies to local header
files.
Change-Id: I064a8adb5cd9753c986b6f224bb075200b3856af
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
|
|
GCC-10 increases string truncations warnings. Refactor string copies
confusing it.
Type: refactor
Change-Id: I9720a0539059de00ab212ff2fc73055f04f5af1d
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
virtio 1.1 defines a number of new features. Packed ring is among the most
notable and important one. It combines used, available, and descripptor rings
into one.
This patch provides experimental support for packed ring. To avoid
regression, when packed ring is configured for the interface, it is branched
to a separate RX and TX driver. Non packed ring should continue to perform
as it was before.
Packed ring is tested using qemu4.2 and ubuntu focal fossa (kernel 5.4.0-12)
on the guess VM which supports packed ring.
To configure VPP with packed ring, just add the optional keyword "packed"
when creating the vhost interface. To bring up the guest VM with packed ring,
add "packed=on" in the qemu launch command.
To facilitate troubleshooting, also added "verbose" option in
show vhost desc CLI to include displaying the indirect descriptors.
Known qemu reconnect issue -
If VPP is restarted, guest VMs also need to be restarted. The problem
is kernel virtio-net-pci keeps track of the previous available and used
indices. For virtio 1.0, these indices are in shared memory and qemu can
easily copy them to pass to the backend for reconnect. For virio 1.1, these
indices are no longer in shared memory. Qemu needs a new mechanism to retrieve
them and it is not currently implemented. So when the protocol reconnects,
qemu does not have the correct available and used indices to pass to the
backend. As a result, after the reconnect, virtio-net-pci is reading the TX
ring from the wrong position in the ring, not the same position which the
backend is writing. Similar problem exists also in the RX.
Type: feature
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I5afc50b0bafab5a1de7a6dd10f399db3fafd144c
|
|
In a rare event, after the vhost protocol message exchange has finished and
the interface had been brought up successfully, the driver MAY still change
its mind about the memory regions by sending new memory maps via
SET_MEM_TABLE. Upon processing SET_MEM_TABLE, VPP invalidates the old memory
regions and the descriptor tables. But it does not re-compute the new
descriptor tables based on the new memory maps. Since VPP does not have the
descriptor tables, it does not read the packets from the vring.
In the normal working case, after SET_MEM_TABLE, the driver follows up with
SET_VRING_ADDRESS which VPP computes the descriptor tables.
The fix is to stash away the descriptor table addresses from
SET_VRING_ADDRESS. Re-compute the new descriptor tables when processing
SET_MEM_TABLE if descriptor table addresses are known.
Type: fix
Ticket: VPP-1784
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I3361f14c3a0372b8d07943eb6aa4b3a3f10708f9
(cherry picked from commit 61b8ba69f7a9540ed00576504528ce439f0286f5)
|
|
Set VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD for the interface
to skip checksum calculation if guest supports checksum offload.
Type: fix
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: Ie933c3462394f07580ef7f2bec1d2eb3b075bd0c
|
|
Add gso option in create vhost interface to support gso and checksum
offload.
Tested with the following startup options in qemu:
csum=on,gso=on,guest_csum=on,guest_tso4=on,guest_tso6=on,guest_ufo=on,
host_tso4=on,host_tso6=on,host_ufo=on
Type: feature
Change-Id: I9ba1ee33677a694c4a0dfe66e745b098995902b8
Signed-off-by: Steven Luong <sluong@cisco.com>
|
|
Multiple API message handlers call vnet_get_sup_hw_interface(...)
without checking the inbound sw_if_index. This can cause a
pool_elt_at_index ASSERT in a debug image, and major disorder in a
production image.
Given that a number of places are coded as follows, add an
"api_visible_or_null" variant of vnet_get_sup_hw_interface, which
returns NULL given an invalid sw_if_index, or a hidden sw interface:
- hw = vnet_get_sup_hw_interface (vnm, sw_if_index);
+ hw = vnet_get_sup_hw_interface_api_visible_or_null (vnm, sw_if_index);
if (hw == NULL || memif_device_class.index != hw->dev_class_index)
return clib_error_return (0, "not a memif interface");
Rename two existing xxx_safe functions -> xxx_or_null to make it
obvious what they return.
Type: fix
Change-Id: I29996e8d0768fd9e0c5495bd91ff8bedcf2c5697
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Some combinations of new qemu (2.11) and old dpdk (16.10) may
send VHOST_USER_SET_FEATURES at the end of the protocol exchange
which the vhost interface is already declared up and ready.
Unfortunately, the process of VHOST_USER_SET_FEATURES will cause
the interface to go down. Not sure if it is correct or needed.
Because there is no additional messages thereafter, the hardware
interface stays down.
The fix is to check the interface again at the end of processing
VHOST_USER_SET_FEATURES. If it is up and ready, we bring back
the hardware interface.
Type: fix
Change-Id: I490cd03820deacbd8b44d8f2cb38c26349dbe3b2
Signed-off-by: Steven Luong <sluong@cisco.com>
|
|
The vlib init function subsystem now supports a mix of procedural and
formally-specified ordering constraints. We should eliminate procedural
knowledge wherever possible.
The following schemes are *roughly* equivalent:
static clib_error_t *init_runs_first (vlib_main_t *vm)
{
clib_error_t *error;
... do some stuff...
if ((error = vlib_call_init_function (init_runs_next)))
return error;
...
}
VLIB_INIT_FUNCTION (init_runs_first);
and
static clib_error_t *init_runs_first (vlib_main_t *vm)
{
... do some stuff...
}
VLIB_INIT_FUNCTION (init_runs_first) =
{
.runs_before = VLIB_INITS("init_runs_next"),
};
The first form will [most likely] call "init_runs_next" on the
spot. The second form means that "init_runs_first" runs before
"init_runs_next," possibly much earlier in the sequence.
Please DO NOT construct sets of init functions where A before B
actually means A *right before* B. It's not necessary - simply combine
A and B - and it leads to hugely annoying debugging exercises when
trying to switch from ad-hoc procedural ordering constraints to formal
ordering constraints.
Change-Id: I5e4353503bf43b4acb11a45fb33c79a5ade8426c
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Symptom
-------
With NDR traffic blasting at VPP, bringing up a new VM with vhost
connection to VPP causes packet drops. I am able to recreate this
problem easily using a simple setup like this.
TREX-------------- switch ---- VPP
|---------------| |-------|
Cause
-----
The reason for the packet drops is due to vhost holding onto the worker
barrier lock for too long in vhost_user_socket_read(). There are quite a
few of system calls inside the routine. At the end of the routine, it
unconditionally calls vhost_user_update_iface_state() for all message
types. vhost_user_update_iface_state() also unconditionally calls
vhost_user_rx_thread_placement() and vhost_user_tx_thread_placement().
vhost_user_rx_thread_placement scraps out all existing cpu/queue mappings
for the interface and creates brand new cpu/queue mappings for the
interface. This process is very disruptive and very expensive. In my
opinion, this area of code needs a makeover.
Fixes
-----
* vhost_user_socket_read() is rewritten that it should not hold
onto the worker barrier lock for system calls, or at least minimize the
need for doing it.
* Remove the call to vhost_user_update_iface_state as a default route at
the end of vhost_user_socket_read(). There is only a couple of message
types which really need to call vhost_user_update_iface_state(). We put
the call to those message types which need it.
* Remove vhost_user_rx_thread_placement() and
vhost_user_tx_thread_placement from vhost_user_update_iface_state().
There is no need to repetatively change the cpu/queue mappings.
* vhost_user_rx_thread_placement() is actually quite expensive. It should
be called only once per queue for the interface. There is no need to
scrap the existing cpu/queue mappings and create new cpu/queue mappings
when the additional queues becomes active/enable.
* Change to create the cpu/queue mappings for the first RX when the
interface is created. Dont remove the cpu/queue mapping when the
interface is disconnected. Remove the cpu/queue mapping only when the
interface is deleted.
The create vhost user interface CLI also has some very expensive system
calls if the command is entered with the optional keyword "server"
As a bonus, This patch makes the create vhost user interface binary-api and
CLI thread safe. Do the protection for the small amount of code which is
thread unsafe.
Change-Id: I4a19cbf7e9cc37ea01286169882e5603e6d7eb77
Signed-off-by: Steven Luong <sluong@cisco.com>
|
|
Change-Id: Id4f37f5d4a03160572954a416efa1ef9b3d79ad1
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Change-Id: Ied34720ca5a6e6e717eea4e86003e854031b6eab
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Change-Id: I0af68f6b41d0024aa64b93a8b18e2d179bf939b0
Signed-off-by: Jerome Tollet <jtollet@cisco.com>
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Fix inconsistencies between admin and link interface states
Admin down should imply link down:
link_up = admin_up && link_ready
Change-Id: I4d668d82d035b5d2ae508727f34f1722a0c3e677
Signed-off-by: Juraj Sloboda <jsloboda@cisco.com>
|
|
Change-Id: I0caa5fd584e3785f237d08f3d3be23e9bfee7605
Signed-off-by: Juraj Sloboda <jsloboda@cisco.com>
|
|
DBGvpp# show vhost-user
Virtio vhost-user interfaces
Global:
coalesce frames 32 time 1e-3
number of rx virtqueues in interrupt mode: 0
Interface: VirtualEthernet0/0/0�?x�D (ifindex 3)
The fix is to use format_vnet_hw_if_index_name rather than hi->name. The former
format the name with %v rather than %s
Change-Id: If4d275e1eb249cf87b2d6b796b42f24769f9e3e3
Signed-off-by: Steven <sluong@cisco.com>
|
|
When VM is having mixed type of vhost-user and SRIOV ports, QEMU (RedHat
v2.10) will not send disconnect signal to VPP, and just gives the new
memory region directly. VPP is not able to handle new memory region
mapping without disconnect signal first, which will result in a SEGV.
The fix will handle the VM reboot scenario without explict disconnect
signal from QEMU.
The fix is to invalidate the avail, desc, and used pointers in the txvq
when the new memory regions are received. This is because these pointers
are not valid anymore with the new memory regions. In the input node, check
to make sure the avail pointer is valid and punt if not.
Change-Id: Ieb8b427b202f4442a58907dab1661d63a03650de
Signed-off-by: Yichen Wang <yicwang@cisco.com>
|
|
It also refactors the vhost code which was in one big file vhost-user.c.
Receive side code is in vhost_user_input.c and
Transmit side code is in vhost_user_output.c
Change-Id: I1b539b5008685889723e228265786a2a3e9f3a78
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
|