Thread (10 messages) 10 messages, 6 authors, 2025-11-25

Re: [RFC blktests fix PATCH] tcp: use GFP_ATOMIC in tcp_disconnect

From: Nilay Shroff <hidden>
Date: 2025-11-25 11:14:26
Also in: linux-block, linux-nvme


On 11/25/25 12:58 PM, Chaitanya Kulkarni wrote:
quoted hunk ↗ jump to hunk
On 11/24/25 22:27, Christoph Hellwig wrote:
quoted
I don't think GFP_ATOMIC is right here, you want GFP_NOIO.

And just use the scope API so that you don't have to pass a gfp_t
several layers down.
are you saying something like this ?
diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
index 29ad4735fac6..56d0a3777a4d 100644
--- a/drivers/nvme/host/tcp.c
+++ b/drivers/nvme/host/tcp.c
@@ -1438,17 +1438,28 @@ static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
  	struct nvme_tcp_ctrl *ctrl = to_tcp_ctrl(nctrl);
  	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
  	unsigned int noreclaim_flag;
+	unsigned int noio_flag;
  
  	if (!test_and_clear_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
  		return;
  
  	page_frag_cache_drain(&queue->pf_cache);
  
+	/**
+	 * Prevent memory reclaim from triggering block I/O during socket
+	 * teardown. The socket release path fput -> tcp_close ->
+	 * tcp_disconnect -> tcp_send_active_reset may allocate memory, and
+	 * allowing reclaim to issue I/O could deadlock if we're being called
+	 * from block device teardown (e.g., del_gendisk -> elevator cleanup)
+	 * which holds locks that the I/O completion path needs.
+	 */
+	noio_flag = memalloc_noio_save();
  	noreclaim_flag = memalloc_noreclaim_save();
  	/* ->sock will be released by fput() */
  	fput(queue->sock->file);
  	queue->sock = NULL;
  	memalloc_noreclaim_restore(noreclaim_flag);
+	memalloc_noio_restore(noio_flag);
  
  	kfree(queue->pdu);
  	mutex_destroy(&queue->send_mutex);
The memalloc_noreclaim_save() above shall already prevent filesystem reclaim,
so if the goal is to avoid fs_reclaim, we should not need an additional
memalloc_noio_save() here. That makes me wonder whether we are looking at the
correct code path. If this teardown path (nvme_tcp_free_queue()) is indeed executed,
it should already be avoiding filesystem reclaim in the first place.

Thanks,
--Nilay
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help