Re:Re: [PATCH] IB/hfi1: Fix potential leaks in IPoIB TX queue initialization
From: 李广硕
Date: Thu Nov 19 2026 - 07:45:00 EST
Hi Leon,
Thanks for the review.
From: Leon Romanovsky <leon@xxxxxxxxxx>
Date: 2026-11-15 21:55:24
To: Guangshuo Li <liguangshuo@xxxxxxxxxxxxxxx>
Cc: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxxxxxxxxxxxxx>,Jason Gunthorpe <jgg@xxxxxxxx>,linux-rdma@xxxxxxxxxxxxxxx,linux-kernel@xxxxxxxxxxxxxxx
Subject: Re: [PATCH] IB/hfi1: Fix potential leaks in IPoIB TX queue initialization>On Wed, Nov 11, 2026 at 08:23:19PM +0800, Guangshuo Li wrote:
>> When an SDMA header allocation fails, rollback skips the current TX
>> queue. Its ring and already allocated headers are leaked. Free the
>> partially initialized ring before unwinding the completed queues.
>
>free_tx_ring already performs that, see "for (j = 0; j < tx_ring_size; j++)"
>loop.
>
>Thanks
>
>>
>> This issue was found by our static analysis tool under development.
>>
>> Fixes: c874ad879c2f ("IB/hifi1: add a null check of kzalloc_node in hfi1_ipoib_txreq_init")
>> Signed-off-by: Guangshuo Li <liguangshuo@xxxxxxxxxxxxxxx>
>> ---
>> drivers/infiniband/hw/hfi1/ipoib_tx.c | 6 +++++-
>> 1 file changed, 5 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/infiniband/hw/hfi1/ipoib_tx.c b/drivers/infiniband/hw/hfi1/ipoib_tx.c
>> index 8b9cc55db59d..2e9225b558b4 100644
>> --- a/drivers/infiniband/hw/hfi1/ipoib_tx.c
>> +++ b/drivers/infiniband/hw/hfi1/ipoib_tx.c
>> @@ -743,7 +743,7 @@ int hfi1_ipoib_txreq_init(struct hfi1_ipoib_dev_priv *priv)
>> kzalloc_node(sizeof(*tx->sdma_hdr),
>> GFP_KERNEL, priv->dd->node);
>> if (!hfi1_txreq_from_idx(tx_ring, j)->sdma_hdr)
>> - goto free_txqs;
>> + goto free_tx_ring;
>> }
>>
>> netif_napi_add_tx(dev, &txq->napi, hfi1_ipoib_poll_tx_ring);
>> @@ -751,6 +751,10 @@ int hfi1_ipoib_txreq_init(struct hfi1_ipoib_dev_priv *priv)
>>
>> return 0;
>>
>> +free_tx_ring:
>> + while (j--)
>> + kfree(hfi1_txreq_from_idx(tx_ring, j)->sdma_hdr);
>> + kvfree(tx_ring->items);
>> free_txqs:
>> for (i--; i >= 0; i--) {
>> struct hfi1_ipoib_txq *txq = &priv->txqs[i];
>>
>> base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
>>
>
The j loop frees all the headers in each ring it visits. As I read
free_txqs, the outer loop starts with i--, so it skips the queue where
the header allocation failed. If that happens on queue 0, the loop
never runs.
Would it make sense to reuse the ring cleanup for the partial queue
and keep netif_napi_del() only for queues where NAPI was added? If so,
I'll send a v2 that does that.
Thanks,
Guangshuo