On 17/12/2024 10:24, Maxime Coquelin wrote:
> Hi Kevin,
> 
> On 12/6/24 21:26, Kevin Traynor wrote:
>> On 06/12/2024 15:58, Maxime Coquelin wrote:
>>> This patch uses the new rte_vhost_driver_set_max_queue_num
>>> API to set the maximum number of queue pairs supported by
>>> the Vhost-user port.
>>>
>>> This is required for VDUSE which needs to specify the
>>> maximum number of queue pairs at creation time. Without it
>>> 128 queue pairs metadata would be allocated.
>>>
>>> To configure it, a new 'vhost-max-queue-pairs' option is
>>> introduced.
>>>
>>> Signed-off-by: Maxime Coquelin <[email protected]>
>>> Acked-by: Eelco Chaudron <[email protected]>
>>> ---
>>>
>>> Note: depends on DPDK v24.11.
>>>
>>> Changes in v4:
>>> ==============
>>> - Change to vhost-queue-pairs-max to vhost-max-queue-pairs (Eelco)
>>> - Remove mention to DPDK version in documentation (Eelco)
>>> - Add missing parameter description in vswitch.xml (Eelco)
>>> - Define min and max values for the new option (Maxime)
>>> Changes in v3:
>>> ==============
>>> - Introduce a new option to set the number of max queue pairs (Kevin)
>>> - Add documentation for new option
>>>
>>> Changes in v2:
>>> ==============
>>> - Address checkpatch warnings.
>>> ---
>>
>> Hi Maxime,
>>
>> Testing with vhost-user backend with DPDK 24.11 worked well - in that it
>> called the API with the right number and didn't break anything :-)
>>
>> A few comments on the code/docs below.
>>
>> fyi - I initially tested on OVS main branch with 23.11 and I saw a loop
>> between the destroy_connection() callback triggering
>> netdev_dpdk_vhost_client_reconfigure(), repeat, etc when i started a VM
>> with 4 queues. So OVS having DPDK 24.11 support will need to be a
>> dependency I think (even aside from the experimental API issue).
>>
> 
> Yes, I confirm 24.11 is also a functional dependency, as v23.11.2 miss
> one fix that will be in upcoming v23.11.3:
> 

As the 24.11 update patch is under review, we can just keep your
patchset for the main branch and make sure we apply in the correct order
(all going well - fingers crossed).

> 
> commit bc578c07f03ec3943fab88a5d156f28b98e1e652
> Author: Maxime Coquelin <[email protected]>
> Date:   Thu Oct 3 10:11:10 2024 +0200
> 
>      vhost: restrict set max queue pair API to VDUSE
> 
>      [ upstream commit e1808999d36bb2e136a649f4651f36030aa468f1 ]
> 
>      In order to avoid breaking Vhost-user live-migration, we want the
>      rte_vhost_driver_set_max_queue_num API to only be effective with
>      VDUSE.
> 
>      Furthermore, this API is only really needed for VDUSE where the
>      device number of queues is defined by the backend. For Vhost-user,
>      this is defined by the frontend (e.g. QEMU), so the advantages of
>      restricting more the maximum number of queue pairs is limited to
>      a small memory gain (a handful of pointers).
> 
>      Fixes: 4aa1f88ac13d ("vhost: add API to set max queue pairs")
> 
>      Reported-by: Yu Jiang <[email protected]>
>      Signed-off-by: Maxime Coquelin <[email protected]>
>      Acked-by: David Marchand <[email protected]>
> 
> 
>> thanks,
>> Kevin.
>>
>>>   Documentation/topics/dpdk/vhost-user.rst | 15 ++++++++++++
>>>   lib/netdev-dpdk.c                        | 30 ++++++++++++++++++++++++
>>>   vswitchd/vswitch.xml                     | 10 ++++++++
>>>   3 files changed, 55 insertions(+)
>>>
>>> diff --git a/Documentation/topics/dpdk/vhost-user.rst 
>>> b/Documentation/topics/dpdk/vhost-user.rst
>>> index 7bba08ac2..656f7f69f 100644
>>> --- a/Documentation/topics/dpdk/vhost-user.rst
>>> +++ b/Documentation/topics/dpdk/vhost-user.rst
>>> @@ -375,6 +375,21 @@ Tx retries max can be set for vhost-user-client ports::
>>>   
>>>     Configurable vhost tx retries are not supported with vhost-user ports.
>>>   
>>> +vhost-user-client max queue pairs config
>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> +
>>> +For vhost-user-client interfaces using the VDUSE backend, the maximum 
>>> umber of
>>
>> typo number
>>
>>> +queue pairs the Virtio device will support can be set at port creation 
>>> time. If
>>> +not set, the default value is 1 queue pair. This value is ignored for
>>> +Vhost-user backends.
>>> +
>>> +Maximum number of queue pairs can be set for vhost-user-client-ports::
>>> +
>>> +    $ ovs-vsctl add-port br0 vduse0 \
>>> +        -- set Interface vduse0 type=dpdkvhostuserclient \
>>> +            options:vhost-server-path=/dev/vduse/vduse0 \
>>> +            options:vhost-max-queue-pairs=4
>>> +
>>>   .. _dpdk-testpmd:
>>>   
>>>   DPDK in the Guest
>>> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
>>> index dc52a2b56..a8b605113 100644
>>> --- a/lib/netdev-dpdk.c
>>> +++ b/lib/netdev-dpdk.c
>>> @@ -19,6 +19,7 @@
>>>   
>>>   #include <errno.h>
>>>   #include <signal.h>
>>> +#include <stdint.h>
>>>   #include <stdlib.h>
>>>   #include <string.h>
>>>   #include <unistd.h>
>>> @@ -153,6 +154,11 @@ typedef uint16_t dpdk_port_t;
>>>   /* Legacy default value for vhost tx retries. */
>>>   #define VHOST_ENQ_RETRY_DEF 8
>>>   
>>> +/* VDUSE-only, ignore for Vhost-user */
>>> +#define VHOST_MAX_QUEUE_PAIRS_MIN 1
>>> +#define VHOST_MAX_QUEUE_PAIRS_DEF VHOST_MAX_QUEUE_PAIRS_MIN
>>> +#define VHOST_MAX_QUEUE_PAIRS_MAX 128
>>> +
>>>   #define IF_NAME_SZ (PATH_MAX > IFNAMSIZ ? PATH_MAX : IFNAMSIZ)
>>>   
>>>   /* List of required flags advertised by the hardware that will be used
>>> @@ -497,6 +503,9 @@ struct netdev_dpdk {
>>>   
>>>           atomic_uint8_t vhost_tx_retries_max;
>>>   
>>> +        /* Ignored by DPDK for Vhost-user backends, only for VDUSE */
>>> +        uint32_t vhost_max_queue_pairs;
>>> +
>>
>> I noticed this added a cacheline - perhaps we could use something
>> smaller and squash it in ?
>>
>> /* --- cacheline 1 boundary (64 bytes) --- */
>> union {
>>      OVS_CACHE_LINE_MARKER cacheline1;        /*    64     1 */
>>      struct {
>>              struct ovs_mutex mutex;          /*    64    48 */
>>              struct dpdk_mp * dpdk_mp;        /*   112     8 */
>>              ovsrcu_index vid;                /*   120     4 */
>>              _Bool      vhost_reconfigured;   /*   124     1 */
>>              atomic_uint8_t vhost_tx_retries_max; /*   125     1 */
>>
>>              /* XXX 2 bytes hole, try to pack */
>>
>>              /* --- cacheline 2 boundary (128 bytes) --- */
>>              uint32_t   vhost_max_queue_pairs; /*   128     4 */
>>              uint8_t    virtio_features_state; /*   132     1 */
>>      };                                       /*    64    72 */
>>      uint8_t            pad55[128];           /*    64   128 */
>> };                                               /*    64   128 */
>> /* --- cacheline 3 boundary (192 bytes) --- */
> 
> 
> Good catch, David already mentioned it to me off-list.
> DPDK Vhost library only supports up to 128 queue pairs, so I think we
> could use uint8_t type and so it would fit into the 1 byte hole.
> 
> Do you agree with this suggestion?
> 

Yes, sounds good. It should be fine but you may want to confirm with
pahole as sometimes the comments can get stale.

thanks,
Kevin.

>>
>>>           /* Flags for virtio features recovery mechanism. */
>>>           uint8_t virtio_features_state;
>>>   
>>> @@ -1609,6 +1618,8 @@ vhost_common_construct(struct netdev *netdev)
>>>   
>>>       atomic_init(&dev->vhost_tx_retries_max, VHOST_ENQ_RETRY_DEF);
>>>   
>>> +    dev->vhost_max_queue_pairs = VHOST_MAX_QUEUE_PAIRS_DEF;
>>> +
>>>       return common_construct(netdev, DPDK_ETH_PORT_ID_INVALID,
>>>                               DPDK_DEV_VHOST, socket_id);
>>>   }
>>> @@ -2491,6 +2502,7 @@ netdev_dpdk_vhost_client_set_config(struct netdev 
>>> *netdev,
>>>       struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
>>>       const char *path;
>>>       int max_tx_retries, cur_max_tx_retries;
>>> +    uint32_t max_queue_pairs;
>>>   
>>>       ovs_mutex_lock(&dev->mutex);
>>>       if (!(dev->vhost_driver_flags & RTE_VHOST_USER_CLIENT)) {
>>> @@ -2498,6 +2510,15 @@ netdev_dpdk_vhost_client_set_config(struct netdev 
>>> *netdev,
>>>           if (!nullable_string_is_equal(path, dev->vhost_id)) {
>>>               free(dev->vhost_id);
>>>               dev->vhost_id = nullable_xstrdup(path);
>>> +
>>> +            max_queue_pairs = smap_get_int(args, "vhost-max-queue-pairs",
>>> +                                           VHOST_MAX_QUEUE_PAIRS_DEF);
>>> +            if (max_queue_pairs < VHOST_MAX_QUEUE_PAIRS_MIN
>>> +                || max_queue_pairs > VHOST_MAX_QUEUE_PAIRS_MAX) {
>>> +                max_queue_pairs = VHOST_MAX_QUEUE_PAIRS_DEF;
>>> +            }
>>> +            dev->vhost_max_queue_pairs = max_queue_pairs;
>>> +
>>>               netdev_request_reconfigure(netdev);
>>>           }
>>>       }
>>> @@ -2514,6 +2535,7 @@ netdev_dpdk_vhost_client_set_config(struct netdev 
>>> *netdev,
>>>           VLOG_INFO("Max Tx retries for vhost device '%s' set to %d",
>>>                     netdev_get_name(netdev), max_tx_retries);
>>>       }
>>> +
>>>       ovs_mutex_unlock(&dev->mutex);
>>>   
>>>       return 0;
>>> @@ -6400,6 +6422,14 @@ netdev_dpdk_vhost_client_reconfigure(struct netdev 
>>> *netdev)
>>>               goto unlock;
>>>           }
>>>   
>>> +        err = rte_vhost_driver_set_max_queue_num(dev->vhost_id,
>>> +                                                 
>>> dev->vhost_max_queue_pairs);
>>
>> The log below is printed:
>>
>> 2024-12-06T18:28:51Z|00100|dpdk|INFO|VHOST_CONFIG: (/tmp/vhost0) Setting
>> max queue pairs to 1
>>
>> It's kind of true, but it could be confusing when using vhost-user
>> backend. Maybe we should add an OVS info log before or after as a
>> reminder that the max queue pairs setting is only valid for VDUSE backend.
> 
> Yes, that's unfortunate that I added a log in
> rte_vhost_driver_set_max_queue_num(), but only at DEBUG level:
> 
> +       if (!vsocket->is_vduse) {
> +               VHOST_CONFIG_LOG(path, DEBUG,
> +                               "Keeping %u max queue pairs for 
> Vhost-user backend",
> +                               VHOST_MAX_QUEUE_PAIRS);
> +               goto unlock_exit;
> +       }
> 
> I can indeed a log before calling the API to avoid the confusion:
> 
> "Setting max queue pairs, only effective with VDUSE backends"
> 
>>
>>> +        if (err) {
>>> +            VLOG_ERR("rte_vhost_driver_set_max_queue_num failed for "
>>> +                    "vhost-user client port: %s\n", dev->up.name);
>>> +            goto unlock;
>>> +        }
>>> +
>>>           err = rte_vhost_driver_start(dev->vhost_id);
>>>           if (err) {
>>>               VLOG_ERR("rte_vhost_driver_start failed for vhost user "
>>> diff --git a/vswitchd/vswitch.xml b/vswitchd/vswitch.xml
>>> index 36cb4e495..0b5c5dcd6 100644
>>> --- a/vswitchd/vswitch.xml
>>> +++ b/vswitchd/vswitch.xml
>>> @@ -3520,6 +3520,16 @@ ovs-vsctl add-port br0 p0 -- set Interface p0 
>>> type=patch options:peer=p1 \
>>>           </p>
>>>         </column>
>>>   
>>> +      <column name="options" key="vhost-max-queue-pairs"
>>> +              type='{"type": "integer", "minInteger" : 1, "maxInteger": 
>>> 128}'>
> 
> I already mention the max value is 128, so uint8_t will be enough for
> vhost_max_queue_pairs field.
> 
>>> +        <p>
>>> +          The value specifies the maximum number of queue pairs supported 
>>> by
>>> +          a vHost device. This is ignored for Vhost-user backends, only 
>>> VDUSE
>>> +          is supported.
>>> +          Only supported by dpdkvhsotuserclient interfaces.
>>
>> typo dpdkvhostuserclient
> 
> good catch, will fix.
> 
>>
>>> +        </p>
>>
>> It would be good to state the default here (like tx-retries-max below)
> 
> Indeed, will add in nex revision.
> 
>>
>>> +      </column>
>>> +
>>>         <column name="options" key="tx-retries-max"
>>>                 type='{"type": "integer", "minInteger": 0, "maxInteger": 
>>> 32}'>
>>>           <p>
>>
> 
> Thanks for the review,
> Maxime
> 

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to