Re: [PATCH v2 3/6] RDMA/cma: Brute force GFP_NOIO

[Zhu Yanjun <zyjzyj2000@gmail.com>]

On 15.05.24 14:53, Håkon Bugge wrote:
> In cma_init(), we call memalloc_noio_{save,restore} in a parenthetic
> fashion when enabled by the module parameter force_noio.
> 
> This in order to conditionally enable rdma_cm to work aligned with
> block I/O devices. Any work queued later on work-queues created during
> module initialization will inherit the PF_MEMALLOC_{NOIO,NOFS}
> flag(s), due to commit ("workqueue: Inherit NOIO and NOFS alloc
> flags").
> 
> We do this in order to enable ULPs using the RDMA stack to be used as
> a network block I/O device. This to support a filesystem on top of a
> raw block device which uses said ULP(s) and the RDMA stack as the
> network transport layer.
> 
> Under intense memory pressure, we get memory reclaims. Assume the
> filesystem reclaims memory, goes to the raw block device, which calls
> into the ULP in question, which calls the RDMA stack. Now, if
> regular GFP_KERNEL allocations in the ULP or the RDMA stack require
> reclaims to be fulfilled, we end up in a circular dependency.
> 
> We break this circular dependency by:
> 
> 1. Force all allocations in the ULP and the relevant RDMA stack to use
>     GFP_NOIO, by means of a parenthetic use of
>     memalloc_noio_{save,restore} on all relevant entry points.
> 
> 2. Make sure work-queues inherits current->flags
>     wrt. PF_MEMALLOC_{NOIO,NOFS}, such that work executed on the
>     work-queue inherits the same flag(s).
> 
> Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
> ---
>   drivers/infiniband/core/cma.c | 20 +++++++++++++++++---
>   1 file changed, 17 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
> index 1e2cd7c8716e8..23a50cc3e81cb 100644
> --- a/drivers/infiniband/core/cma.c
> +++ b/drivers/infiniband/core/cma.c
> @@ -50,6 +50,10 @@ MODULE_LICENSE("Dual BSD/GPL");
>   #define CMA_IBOE_PACKET_LIFETIME 16
>   #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
>   
> +static bool cma_force_noio;
> +module_param_named(force_noio, cma_force_noio, bool, 0444);
> +MODULE_PARM_DESC(force_noio, "Force the use of GFP_NOIO (Y/N)");
> +
>   static const char * const cma_events[] = {
>   	[RDMA_CM_EVENT_ADDR_RESOLVED]	 = "address resolved",
>   	[RDMA_CM_EVENT_ADDR_ERROR]	 = "address error",
> @@ -5424,6 +5428,10 @@ static struct pernet_operations cma_pernet_operations = {
>   static int __init cma_init(void)
>   {
>   	int ret;
> +	unsigned int noio_flags;

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/process/maintainer-netdev.rst?h=v6.9#n376

"
Netdev has a convention for ordering local variables in functions.
Order the variable declaration lines longest to shortest, e.g.::

   struct scatterlist *sg;
   struct sk_buff *skb;
   int err, i;

If there are dependencies between the variables preventing the ordering
move the initialization out of line.
"

Zhu Yanjun

> +
> +	if (cma_force_noio)
> +		noio_flags = memalloc_noio_save();
>   
>   	/*
>   	 * There is a rare lock ordering dependency in cma_netdev_callback()
> @@ -5439,8 +5447,10 @@ static int __init cma_init(void)
>   	}
>   
>   	cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
> -	if (!cma_wq)
> -		return -ENOMEM;
> +	if (!cma_wq) {
> +		ret = -ENOMEM;
> +		goto out;
> +	}
>   
>   	ret = register_pernet_subsys(&cma_pernet_operations);
>   	if (ret)
> @@ -5458,7 +5468,8 @@ static int __init cma_init(void)
>   	if (ret)
>   		goto err_ib;
>   
> -	return 0;
> +	ret = 0;
> +	goto out;
>   
>   err_ib:
>   	ib_unregister_client(&cma_client);
> @@ -5469,6 +5480,9 @@ static int __init cma_init(void)
>   	unregister_pernet_subsys(&cma_pernet_operations);
>   err_wq:
>   	destroy_workqueue(cma_wq);
> +out:
> +	if (cma_force_noio)
> +		memalloc_noio_restore(noio_flags);
>   	return ret;
>   }
>