On Fri, 17 Dec 2010, KOSAKI Motohiro wrote: > > (Patch based on 2.6.36 tag). > > > > These GFP_KERNEL allocations could happen even though the caller of __vmalloc() > > requested a stricter gfp mask (such as GFP_NOFS or GFP_ATOMIC). > > > > This was first noticed in Lustre, where it led to deadlocks due to a filesystem > > thread which requested a GFP_NOFS __vmalloc() allocation ended up calling down > > to Lustre itself to free memory, despite this not being allowed by GFP_NOFS. > > > > Further analysis showed that some in-tree filesystems (namely GFS, Ceph and XFS) > > were vulnerable to the same bug due to calling __vmalloc() or vm_map_ram() in > > contexts where __GFP_FS allocations are not allowed. > > > > Fixing this bug required changing a few mm interfaces to accept gfp flags. > > This needed to be done in all architectures, thus the large number of changes. > > I like this patch. but please separate it two patches. > > 1) add gfp_mask argument to some function > 2) vmalloc use flexible mask instead GFP_KERNEL always. > > I mean please consider to make reviewers friendly patch. > IOW, please see your diffstat. ;) > I agree, I'm also wondering if it would be easier to introduce seperate, lower-level versions of the functions that the current interfaces would then use instead of converting all of their current use cases. Using pmd_alloc_one() as an example: convert existing pmd_alloc_one() to __pmd_alloc_one() for each arch and add the gfp_t formal), then introduce a new pmd_alloc_one() that does __pmd_alloc_one(..., GFP_KERNEL). -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
