Re: [PATCH v2 for-next 1/3] RDMA/core: Provide rdma_user_mmap_disassociate() to disassociate mmap pages

From: Junxian Huang
Date: Mon Aug 26 2024 - 09:21:38 EST




On 2024/8/26 21:16, Jason Gunthorpe wrote:
> On Mon, Aug 26, 2024 at 09:12:33PM +0800, Junxian Huang wrote:
>
>> diff --git a/drivers/infiniband/core/uverbs.h b/drivers/infiniband/core/uverbs.h
>> index 821d93c8f712..05b589aad5ef 100644
>> --- a/drivers/infiniband/core/uverbs.h
>> +++ b/drivers/infiniband/core/uverbs.h
>> @@ -160,6 +160,9 @@ struct ib_uverbs_file {
>> struct page *disassociate_page;
>>
>> struct xarray idr;
>> +
>> + struct mutex disassociation_lock;
>> + bool disassociated;
>> };
>>
>> struct ib_uverbs_event {
>> diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
>> index 6b4d5a660a2f..6503f9a23211 100644
>> --- a/drivers/infiniband/core/uverbs_main.c
>> +++ b/drivers/infiniband/core/uverbs_main.c
>> @@ -722,12 +722,12 @@ static void rdma_umap_open(struct vm_area_struct *vma)
>> return;
>>
>> /* We are racing with disassociation */
>> - if (!down_read_trylock(&ufile->hw_destroy_rwsem))
>> + if (!mutex_trylock(&ufile->disassociation_lock))
>> goto out_zap;
>
> Nonon, don't touch this stuff! It is fine as is
>
> The extra lock should be in the mmap zap functions only
>
> Jason

Okay, I got it. Thanks

Junxian