[dpdk-dev] [PATCH 0/1] vhost: fix iotlb mempool single-consumer flag

["Eugenio Pérez" <eperezma@redhat.com>]

Bugzilla bug: 523

Using testpmd as a vhost-user with iommu:

/home/dpdk/build/app/dpdk-testpmd -l 1,3 \
	--vdev net_vhost0,iface=/tmp/vhost-user1,queues=1,iommu-support=1 \
	-- --auto-start --stats-period 5 --forward-mode=txonly

And qemu with packed virtqueue:

    <interface type='vhostuser'>
      <mac address='88:67:11:5f:dd:02'/>
      <source type='unix' path='/tmp/vhost-user1' mode='client'/>
      <model type='virtio'/>
      <driver name='vhost' rx_queue_size='1024' iommu='on' ats='on'/>
      <address type='pci' domain='0x0000' bus='0x07' slot='0x00' function='0x0'/>
    </interface>
...

  <qemu:commandline>
    <qemu:arg value='-set'/>
    <qemu:arg value='device.net1.packed=on'/>
  </qemu:commandline>

--

Is it possible to consume the iotlb's entries of the mempoo from different
threads. Thread sanitizer example output (after change rwlocks to POSIX ones):

WARNING: ThreadSanitizer: data race (pid=76927)
  Write of size 8 at 0x00017ffd5628 by thread T5:
    #0 vhost_user_iotlb_cache_insert ../lib/librte_vhost/iotlb.c:181 (dpdk-testpmd+0x769343)
    #1 vhost_user_iotlb_msg ../lib/librte_vhost/vhost_user.c:2380 (dpdk-testpmd+0x78e4bf)
    #2 vhost_user_msg_handler ../lib/librte_vhost/vhost_user.c:2848 (dpdk-testpmd+0x78fcf8)
    #3 vhost_user_read_cb ../lib/librte_vhost/socket.c:311 (dpdk-testpmd+0x770162)
    #4 fdset_event_dispatch ../lib/librte_vhost/fd_man.c:286 (dpdk-testpmd+0x7591c2)
    #5 ctrl_thread_init ../lib/librte_eal/common/eal_common_thread.c:193 (dpdk-testpmd+0xa2890b)
    #6 <null> <null> (libtsan.so.0+0x2a68d)

  Previous read of size 8 at 0x00017ffd5628 by thread T3:
    #0 vhost_user_iotlb_cache_find ../lib/librte_vhost/iotlb.c:252 (dpdk-testpmd+0x76ee96)
    #1 __vhost_iova_to_vva ../lib/librte_vhost/vhost.c:42 (dpdk-testpmd+0x77488c)
    #2 vhost_iova_to_vva ../lib/librte_vhost/vhost.h:753 (dpdk-testpmd+0x7abeb3)
    #3 map_one_desc ../lib/librte_vhost/virtio_net.c:497 (dpdk-testpmd+0x7abeb3)
    #4 fill_vec_buf_packed ../lib/librte_vhost/virtio_net.c:751 (dpdk-testpmd+0x7abeb3)
    #5 vhost_enqueue_single_packed ../lib/librte_vhost/virtio_net.c:1170 (dpdk-testpmd+0x7abeb3)
    #6 virtio_dev_rx_single_packed ../lib/librte_vhost/virtio_net.c:1346 (dpdk-testpmd+0x7abeb3)
    #7 virtio_dev_rx_packed ../lib/librte_vhost/virtio_net.c:1384 (dpdk-testpmd+0x7abeb3)
    #8 virtio_dev_rx ../lib/librte_vhost/virtio_net.c:1435 (dpdk-testpmd+0x7b0654)
    #9 rte_vhost_enqueue_burst ../lib/librte_vhost/virtio_net.c:1465 (dpdk-testpmd+0x7b0654)
    #10 eth_vhost_tx ../drivers/net/vhost/rte_eth_vhost.c:470 (dpdk-testpmd+0x1ddfbd8)
    #11 rte_eth_tx_burst ../lib/librte_ethdev/rte_ethdev.h:4800 (dpdk-testpmd+0x505fdb)
    #12 pkt_burst_transmit ../app/test-pmd/txonly.c:365 (dpdk-testpmd+0x5106ad)
    #13 run_pkt_fwd_on_lcore ../app/test-pmd/testpmd.c:2080 (dpdk-testpmd+0x4f8951)
    #14 start_pkt_forward_on_core ../app/test-pmd/testpmd.c:2106 (dpdk-testpmd+0x4f89d7)
    #15 eal_thread_loop ../lib/librte_eal/linux/eal_thread.c:127 (dpdk-testpmd+0xa5b20a)
    #16 <null> <null> (libtsan.so.0+0x2a68d)

  Location is global '<null>' at 0x000000000000 (rtemap_0+0x00003ffd5628)

  Thread T5 'vhost-events' (tid=76933, running) created by main thread at:
    #0 pthread_create <null> (libtsan.so.0+0x2cd42)
    #1 rte_ctrl_thread_create ../lib/librte_eal/common/eal_common_thread.c:216 (dpdk-testpmd+0xa289e7)
    #2 rte_vhost_driver_start ../lib/librte_vhost/socket.c:1190 (dpdk-testpmd+0x7728ef)
    #3 vhost_driver_setup ../drivers/net/vhost/rte_eth_vhost.c:1028 (dpdk-testpmd+0x1de233d)
    #4 eth_dev_configure ../drivers/net/vhost/rte_eth_vhost.c:1126 (dpdk-testpmd+0x1de29cc)
    #5 rte_eth_dev_configure ../lib/librte_ethdev/rte_ethdev.c:1439 (dpdk-testpmd+0x991ce2)
    #6 start_port ../app/test-pmd/testpmd.c:2450 (dpdk-testpmd+0x4f9b45)
    #7 main ../app/test-pmd/testpmd.c:3777 (dpdk-testpmd+0x4fe1ac)

  Thread T3 'lcore-slave-3' (tid=76931, running) created by main thread at:
    #0 pthread_create <null> (libtsan.so.0+0x2cd42)
    #1 rte_eal_init ../lib/librte_eal/linux/eal.c:1244 (dpdk-testpmd+0xa46e2b)
    #2 main ../app/test-pmd/testpmd.c:3673 (dpdk-testpmd+0x4fdd75)

--

Or:
WARNING: ThreadSanitizer: data race (pid=76927)
  Write of size 1 at 0x00017ffd00f8 by thread T5:
    #0 vhost_user_iotlb_cache_insert ../lib/librte_vhost/iotlb.c:182 (dpdk-testpmd+0x769370)
    #1 vhost_user_iotlb_msg ../lib/librte_vhost/vhost_user.c:2380 (dpdk-testpmd+0x78e4bf)
    #2 vhost_user_msg_handler ../lib/librte_vhost/vhost_user.c:2848 (dpdk-testpmd+0x78fcf8)
    #3 vhost_user_read_cb ../lib/librte_vhost/socket.c:311 (dpdk-testpmd+0x770162)
    #4 fdset_event_dispatch ../lib/librte_vhost/fd_man.c:286 (dpdk-testpmd+0x7591c2)
    #5 ctrl_thread_init ../lib/librte_eal/common/eal_common_thread.c:193 (dpdk-testpmd+0xa2890b)
    #6 <null> <null> (libtsan.so.0+0x2a68d)

  Previous write of size 1 at 0x00017ffd00f8 by thread T3:
    #0 vhost_user_iotlb_pending_insert ../lib/librte_vhost/iotlb.c:86 (dpdk-testpmd+0x75eb0c)
    #1 __vhost_iova_to_vva ../lib/librte_vhost/vhost.c:58 (dpdk-testpmd+0x774926)
    #2 vhost_iova_to_vva ../lib/librte_vhost/vhost.h:753 (dpdk-testpmd+0x7a79d1)
    #3 virtio_dev_rx_batch_packed ../lib/librte_vhost/virtio_net.c:1295 (dpdk-testpmd+0x7a79d1)
    #4 virtio_dev_rx_packed ../lib/librte_vhost/virtio_net.c:1376 (dpdk-testpmd+0x7a79d1)
    #5 virtio_dev_rx ../lib/librte_vhost/virtio_net.c:1435 (dpdk-testpmd+0x7b0654)
    #6 rte_vhost_enqueue_burst ../lib/librte_vhost/virtio_net.c:1465 (dpdk-testpmd+0x7b0654)
    #7 eth_vhost_tx ../drivers/net/vhost/rte_eth_vhost.c:470 (dpdk-testpmd+0x1ddfbd8)
    #8 rte_eth_tx_burst ../lib/librte_ethdev/rte_ethdev.h:4800 (dpdk-testpmd+0x505fdb)
    #9 pkt_burst_transmit ../app/test-pmd/txonly.c:365 (dpdk-testpmd+0x5106ad)
    #10 run_pkt_fwd_on_lcore ../app/test-pmd/testpmd.c:2080 (dpdk-testpmd+0x4f8951)
    #11 start_pkt_forward_on_core ../app/test-pmd/testpmd.c:2106 (dpdk-testpmd+0x4f89d7)
    #12 eal_thread_loop ../lib/librte_eal/linux/eal_thread.c:127 (dpdk-testpmd+0xa5b20a)
    #13 <null> <null> (libtsan.so.0+0x2a68d)

--

As a consequence, the two threads can modify the same entry of the mempool.
Usually, this cause a loop in iotlb_pending_entries list.

This behavior is only observed on packed vq + virtio-net kernel driver in the
guest, so we could make the single-consumer flag optional. However, I have not
found why I cannot see the issue in split, so the safer option is to never set
it.

Any comments?

Thanks!

Eugenio Pérez (1):
  vhost: fix iotlb mempool single-consumer flag

 lib/librte_vhost/iotlb.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

-- 
2.18.1