Arjan van de Ven wrote:
> as first general comment, I think that some of the memory allocation
> GFP_ flags are possibly incorrect; I would expect several places to use
> GFP_NOIO rather than GFP_KERNEL, to avoid recursion/deadlocks
>
>
>> +static void blkif_recover(struct blkfront_info *info)
>>
>> + /* Stage 1: Make a safe copy of the shadow state. */
>> + copy = kmalloc(sizeof(info->shadow), GFP_KERNEL | __GFP_NOFAIL);
>>
>
> like here..
>
>
> and __GFP_NOFAIL is usually horrid; is this because error recovery was
> an afterthought, or because it's physically impossible? In addition
> __GFP_NOFAIL in a block device driver is... an interesting way to add
> OOM deadlocks... have the VM guys looked into this yet?
>
In this particular case, it's only used on the resume path, which I'm
guessing would not lead to IO recursion. There doesn't seem to be any
particular reason for this to be NOFAIL though (but I haven't really
analyzed it).
There don't appear to be any memory allocations on the IO path; they're
all in setup code.
J
