[openib-general] Re: [Andrew Morton] inappropriate use of in_atomic()

[Michael S. Tsirkin]

Quoting r. Hal Rosenstock <halr at voltaire.com>:
> Subject: Re: [Andrew Morton] inappropriate use of in_atomic()
> 
> On Thu, 2005-03-10 at 23:50, Roland Dreier wrote:
> > drivers/infiniband/core/mad.c is in Andrew's list...
> > 
> > >From looking at the code, the best fix I can come up with is just to
> > always use GFP_ATOMIC ... worst case we drop a MAD under memory
> > pressure.  
> 
> That could be bad if this persists but I suppose there are other ill
> effects of this.
> 
> > The other option is to change ib_post_send_mad() to take a
> > GFP_ mask as a parameter, but that doesn't seem worth doing...
> 
> There aren't that many places this is called. Also, it appears to me
> that sa_query.c is already doing this for some of it's memory allocation
> and this could be passed down to ib_post_send_mad.
> 
> int ib_sa_path_rec_get(struct ib_device *device, u8 port_num,
>                        struct ib_sa_path_rec *rec,
>                        ib_sa_comp_mask comp_mask,
>                        int timeout_ms, int gfp_mask,
> ...
> 
> This approach seems better to me from a robustness standpoint. 
> 
> Is the difficulty determing what to set the mask to for each call ? If
> they all end up being GFP_ATOMIC, this reduces to your preferred
> solution.
> 
> The biggest impact appears to be on CM (at least currently).

As far as I remember most CM code is thread level anyway, isnt it?

-- 
MST - Michael S. Tsirkin