Messages in this thread | | | Date | Mon, 3 Dec 2012 11:07:15 -0800 | From | Zach Brown <> | Subject | Re: [patch] bdi: add a user-tunable cpu_list for the bdi flusher threads |
| |
On Mon, Dec 03, 2012 at 11:22:31AM -0500, Jeff Moyer wrote: > Jeff Moyer <jmoyer@redhat.com> writes: > > >>> + bdi->flusher_cpumask = kmalloc(sizeof(cpumask_t), GFP_KERNEL); > >>> + if (!bdi->flusher_cpumask) > >>> + return -ENOMEM; > >> > >> The bare GFP_KERNEL raises an eyebrow. Some bdi_init() callers like > >> blk_alloc_queue_node() look like they'll want to pass in a gfp_t for the > >> allocation. > > > > I'd be surprised if that was necessary, seeing how every single caller > > of blk_alloc_queue_node passes in GFP_KERNEL. I'll make the change, > > though, there aren't too many callers of bdi_init out there. > > No other callers of bdi_init want anything but GFP_KERNEL. In the case > of blk_alloc_queue_node, even *it* doesn't honor the gfp_t passed in! > Have a look at blkcg_init_queue (called from blk_alloc_queue_node) to > see what I mean. Maybe that's a bug?
Heh, indeed.
> I've written the patch to modify bdi_init to take a gfp_t, but I'm > actually not in favor of this change, so I'm not going to post it > (unless, of course, you can provide a compelling argument). :-)
No argument here, it just jumped out at me in the code. I didn't check out the callers or history of why it was that way :).
- z
| |