Re: Kernel BUG when registering the ring

[Jens Axboe <axboe@kernel.dk>]

On 2/10/20 6:22 PM, Glauber Costa wrote:
> Hello my dear io_uring friends
> 
> Today I tried to run my tests on a different, more powerful hardware
> (70+ Hyperthreads) and it crashed on creating the ring.
> 
> I don't recall anything fancy in my code for creating the ring -
> except maybe that I size the cq ring differently than the sq ring.
> 
> The backtrace attached leads me to believe that we just accessed a
> null struct somewhere

Yeah, but it's in the alloc code, it's not in io-wq/io_uring.
Here's where it is crashing:

struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
{
	[...]
	for_each_node(node) {
		struct io_wqe *wqe;

		wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, node);

I guess the node isn't online, and that's why it's crashing. Try the
below for starters, it should get it going.


diff --git a/fs/io-wq.c b/fs/io-wq.c
index 182aa17dc2ca..3898165baccb 100644
--- a/fs/io-wq.c
+++ b/fs/io-wq.c
@@ -1115,8 +1116,11 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 
 	for_each_node(node) {
 		struct io_wqe *wqe;
+		int alloc_node = node;
 
-		wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, node);
+		if (!node_online(alloc_node))
+			alloc_node = NUMA_NO_NODE;
+		wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
 		if (!wqe)
 			goto err;
 		wq->wqes[node] = wqe;

-- 
Jens Axboe