On Thu, 8 Mar 2012 13:40:50 -0800, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote: > On Thu, 8 Mar 2012 15:19:27 -0600 > Tyler Hicks <tyhicks@xxxxxxxxxxxxx> wrote: > > > > > > > > > > Sigh. Was lockdep_annotate_inode_mutex_key() sufficiently > > > self-explanatory to justify leaving it undocumented? > > > > > > <goes off and reads e096d0c7e2e> > > > > > > OK, the patch looks correct given the explanation in e096d0c7e2e, but > > > I'd like to understand why it becomes necessary only now. > > > > > > > NOTE: This patch also require > > > > http://thread.gmane.org/gmane.linux.file-systems/58795/focus=59565 > > > > to remove the lockdep warning > > > > > > And that patch has been basically ignored. > > > > Al commented on it here: > > > > https://lkml.org/lkml/2012/2/16/518 > > > > He said that while my patch is correct, taking i_mutex inside mmap_sem > > is still wrong. > > OK, thanks, yup. Taking i_mutex in file_operations.mmap() is wrong. > > Is hugetlbfs actually deadlockable because of this, or is it the case > that the i_mutex->mmap_sem ordering happens to never happen for this > filesystem? Although we shouldn't go and create incompatible lock > ranking rules for different filesystems! > > So we need to pull the i_mutex out of hugetlbfs_file_mmap(). What's it > actually trying to do in there? If we switch to > i_size_read()/i_size_write() then AFAICT the problem comes down to > hugetlb_reserve_pages(). > > hugetlb_reserve_pages() fiddles with i_mapping->private_list and the fs > owns private_list and is free to use a lock other than i_mutex to > protect it. (In fact i_mapping.private_lock is the usual lock for > private_list). > > > > So from a quick scan here I'm thinking that a decent fix is to remove > the i_mutex locking from hugetlbfs_file_mmap(), switch > hugetlbfs_file_mmap() to i_size_read/write then use a hugetlb-private > lock to protect i_mapping->private_list. region_chg() will do > GFP_KERNEL allocations under that lock, so some care is needed. > But as per 7762f5a0b709b415fda132258ad37b9f2a1db994 i_size_write should always happen with i_mutex held -aneesh -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>