net: sfc: add missing xdp queue reinitialization
authorTaehee Yoo <ap420073@gmail.com>
Wed, 30 Mar 2022 16:37:03 +0000 (16:37 +0000)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 13 Apr 2022 18:59:15 +0000 (20:59 +0200)
[ Upstream commit 059a47f1da93811d37533556d67e72f2261b1127 ]

After rx/tx ring buffer size is changed, kernel panic occurs when
it acts XDP_TX or XDP_REDIRECT.

When tx/rx ring buffer size is changed(ethtool -G), sfc driver
reallocates and reinitializes rx and tx queues and their buffer
(tx_queue->buffer).
But it misses reinitializing xdp queues(efx->xdp_tx_queues).
So, while it is acting XDP_TX or XDP_REDIRECT, it uses the uninitialized
tx_queue->buffer.

A new function efx_set_xdp_channels() is separated from efx_set_channels()
to handle only xdp queues.

Splat looks like:
   BUG: kernel NULL pointer dereference, address: 000000000000002a
   #PF: supervisor write access in kernel mode
   #PF: error_code(0x0002) - not-present page
   PGD 0 P4D 0
   Oops: 0002 [#4] PREEMPT SMP NOPTI
   RIP: 0010:efx_tx_map_chunk+0x54/0x90 [sfc]
   CPU: 2 PID: 0 Comm: swapper/2 Tainted: G      D           5.17.0+ #55 e8beeee8289528f11357029357cf
   Code: 48 8b 8d a8 01 00 00 48 8d 14 52 4c 8d 2c d0 44 89 e0 48 85 c9 74 0e 44 89 e2 4c 89 f6 48 80
   RSP: 0018:ffff92f121e45c60 EFLAGS: 00010297
   RIP: 0010:efx_tx_map_chunk+0x54/0x90 [sfc]
   RAX: 0000000000000040 RBX: ffff92ea506895c0 RCX: ffffffffc0330870
   RDX: 0000000000000001 RSI: 00000001139b10ce RDI: ffff92ea506895c0
   RBP: ffffffffc0358a80 R08: 00000001139b110d R09: 0000000000000000
   R10: 0000000000000001 R11: ffff92ea414c0088 R12: 0000000000000040
   R13: 0000000000000018 R14: 00000001139b10ce R15: ffff92ea506895c0
   FS:  0000000000000000(0000) GS:ffff92f121ec0000(0000) knlGS:0000000000000000
   CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
   Code: 48 8b 8d a8 01 00 00 48 8d 14 52 4c 8d 2c d0 44 89 e0 48 85 c9 74 0e 44 89 e2 4c 89 f6 48 80
   CR2: 000000000000002a CR3: 00000003e6810004 CR4: 00000000007706e0
   RSP: 0018:ffff92f121e85c60 EFLAGS: 00010297
   PKRU: 55555554
   RAX: 0000000000000040 RBX: ffff92ea50689700 RCX: ffffffffc0330870
   RDX: 0000000000000001 RSI: 00000001145a90ce RDI: ffff92ea50689700
   RBP: ffffffffc0358a80 R08: 00000001145a910d R09: 0000000000000000
   R10: 0000000000000001 R11: ffff92ea414c0088 R12: 0000000000000040
   R13: 0000000000000018 R14: 00000001145a90ce R15: ffff92ea50689700
   FS:  0000000000000000(0000) GS:ffff92f121e80000(0000) knlGS:0000000000000000
   CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
   CR2: 000000000000002a CR3: 00000003e6810005 CR4: 00000000007706e0
   PKRU: 55555554
   Call Trace:
    <IRQ>
    efx_xdp_tx_buffers+0x12b/0x3d0 [sfc 84c94b8e32d44d296c17e10a634d3ad454de4ba5]
    __efx_rx_packet+0x5c3/0x930 [sfc 84c94b8e32d44d296c17e10a634d3ad454de4ba5]
    efx_rx_packet+0x28c/0x2e0 [sfc 84c94b8e32d44d296c17e10a634d3ad454de4ba5]
    efx_ef10_ev_process+0x5f8/0xf40 [sfc 84c94b8e32d44d296c17e10a634d3ad454de4ba5]
    ? enqueue_task_fair+0x95/0x550
    efx_poll+0xc4/0x360 [sfc 84c94b8e32d44d296c17e10a634d3ad454de4ba5]

Fixes: 3990a8fffbda ("sfc: allocate channels for XDP tx queues")
Signed-off-by: Taehee Yoo <ap420073@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/net/ethernet/sfc/efx_channels.c

index 3dbea02..4753c0c 100644 (file)
@@ -763,6 +763,85 @@ void efx_remove_channels(struct efx_nic *efx)
        kfree(efx->xdp_tx_queues);
 }
 
+static int efx_set_xdp_tx_queue(struct efx_nic *efx, int xdp_queue_number,
+                               struct efx_tx_queue *tx_queue)
+{
+       if (xdp_queue_number >= efx->xdp_tx_queue_count)
+               return -EINVAL;
+
+       netif_dbg(efx, drv, efx->net_dev,
+                 "Channel %u TXQ %u is XDP %u, HW %u\n",
+                 tx_queue->channel->channel, tx_queue->label,
+                 xdp_queue_number, tx_queue->queue);
+       efx->xdp_tx_queues[xdp_queue_number] = tx_queue;
+       return 0;
+}
+
+static void efx_set_xdp_channels(struct efx_nic *efx)
+{
+       struct efx_tx_queue *tx_queue;
+       struct efx_channel *channel;
+       unsigned int next_queue = 0;
+       int xdp_queue_number = 0;
+       int rc;
+
+       /* We need to mark which channels really have RX and TX
+        * queues, and adjust the TX queue numbers if we have separate
+        * RX-only and TX-only channels.
+        */
+       efx_for_each_channel(channel, efx) {
+               if (channel->channel < efx->tx_channel_offset)
+                       continue;
+
+               if (efx_channel_is_xdp_tx(channel)) {
+                       efx_for_each_channel_tx_queue(tx_queue, channel) {
+                               tx_queue->queue = next_queue++;
+                               rc = efx_set_xdp_tx_queue(efx, xdp_queue_number,
+                                                         tx_queue);
+                               if (rc == 0)
+                                       xdp_queue_number++;
+                       }
+               } else {
+                       efx_for_each_channel_tx_queue(tx_queue, channel) {
+                               tx_queue->queue = next_queue++;
+                               netif_dbg(efx, drv, efx->net_dev,
+                                         "Channel %u TXQ %u is HW %u\n",
+                                         channel->channel, tx_queue->label,
+                                         tx_queue->queue);
+                       }
+
+                       /* If XDP is borrowing queues from net stack, it must
+                        * use the queue with no csum offload, which is the
+                        * first one of the channel
+                        * (note: tx_queue_by_type is not initialized yet)
+                        */
+                       if (efx->xdp_txq_queues_mode ==
+                           EFX_XDP_TX_QUEUES_BORROWED) {
+                               tx_queue = &channel->tx_queue[0];
+                               rc = efx_set_xdp_tx_queue(efx, xdp_queue_number,
+                                                         tx_queue);
+                               if (rc == 0)
+                                       xdp_queue_number++;
+                       }
+               }
+       }
+       WARN_ON(efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_DEDICATED &&
+               xdp_queue_number != efx->xdp_tx_queue_count);
+       WARN_ON(efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED &&
+               xdp_queue_number > efx->xdp_tx_queue_count);
+
+       /* If we have more CPUs than assigned XDP TX queues, assign the already
+        * existing queues to the exceeding CPUs
+        */
+       next_queue = 0;
+       while (xdp_queue_number < efx->xdp_tx_queue_count) {
+               tx_queue = efx->xdp_tx_queues[next_queue++];
+               rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue);
+               if (rc == 0)
+                       xdp_queue_number++;
+       }
+}
+
 int efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
 {
        struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
@@ -837,6 +916,7 @@ int efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
                efx_init_napi_channel(efx->channel[i]);
        }
 
+       efx_set_xdp_channels(efx);
 out:
        /* Destroy unused channel structures */
        for (i = 0; i < efx->n_channels; i++) {
@@ -872,26 +952,9 @@ rollback:
        goto out;
 }
 
-static inline int
-efx_set_xdp_tx_queue(struct efx_nic *efx, int xdp_queue_number,
-                    struct efx_tx_queue *tx_queue)
-{
-       if (xdp_queue_number >= efx->xdp_tx_queue_count)
-               return -EINVAL;
-
-       netif_dbg(efx, drv, efx->net_dev, "Channel %u TXQ %u is XDP %u, HW %u\n",
-                 tx_queue->channel->channel, tx_queue->label,
-                 xdp_queue_number, tx_queue->queue);
-       efx->xdp_tx_queues[xdp_queue_number] = tx_queue;
-       return 0;
-}
-
 int efx_set_channels(struct efx_nic *efx)
 {
-       struct efx_tx_queue *tx_queue;
        struct efx_channel *channel;
-       unsigned int next_queue = 0;
-       int xdp_queue_number;
        int rc;
 
        efx->tx_channel_offset =
@@ -909,61 +972,14 @@ int efx_set_channels(struct efx_nic *efx)
                        return -ENOMEM;
        }
 
-       /* We need to mark which channels really have RX and TX
-        * queues, and adjust the TX queue numbers if we have separate
-        * RX-only and TX-only channels.
-        */
-       xdp_queue_number = 0;
        efx_for_each_channel(channel, efx) {
                if (channel->channel < efx->n_rx_channels)
                        channel->rx_queue.core_index = channel->channel;
                else
                        channel->rx_queue.core_index = -1;
-
-               if (channel->channel >= efx->tx_channel_offset) {
-                       if (efx_channel_is_xdp_tx(channel)) {
-                               efx_for_each_channel_tx_queue(tx_queue, channel) {
-                                       tx_queue->queue = next_queue++;
-                                       rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue);
-                                       if (rc == 0)
-                                               xdp_queue_number++;
-                               }
-                       } else {
-                               efx_for_each_channel_tx_queue(tx_queue, channel) {
-                                       tx_queue->queue = next_queue++;
-                                       netif_dbg(efx, drv, efx->net_dev, "Channel %u TXQ %u is HW %u\n",
-                                                 channel->channel, tx_queue->label,
-                                                 tx_queue->queue);
-                               }
-
-                               /* If XDP is borrowing queues from net stack, it must use the queue
-                                * with no csum offload, which is the first one of the channel
-                                * (note: channel->tx_queue_by_type is not initialized yet)
-                                */
-                               if (efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_BORROWED) {
-                                       tx_queue = &channel->tx_queue[0];
-                                       rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue);
-                                       if (rc == 0)
-                                               xdp_queue_number++;
-                               }
-                       }
-               }
        }
-       WARN_ON(efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_DEDICATED &&
-               xdp_queue_number != efx->xdp_tx_queue_count);
-       WARN_ON(efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED &&
-               xdp_queue_number > efx->xdp_tx_queue_count);
 
-       /* If we have more CPUs than assigned XDP TX queues, assign the already
-        * existing queues to the exceeding CPUs
-        */
-       next_queue = 0;
-       while (xdp_queue_number < efx->xdp_tx_queue_count) {
-               tx_queue = efx->xdp_tx_queues[next_queue++];
-               rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue);
-               if (rc == 0)
-                       xdp_queue_number++;
-       }
+       efx_set_xdp_channels(efx);
 
        rc = netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
        if (rc)