On Sun, 5 Jun 2011, Al Viro wrote: > When alloc_huge_page() runs afoul of quota, it returns ERR_PTR(-ENOSPC). > Callers do not expect that - hugetlb_cow() returns ENOSPC if it gets that > and so does hugetlb_no_page(). Eventually the thing propagates back to > hugetlb_fault() and is returned by it. > > Callers of hugetlb_fault() clearly expect a bitmap of VM_... and > not something from errno.h: one place is > ret = hugetlb_fault(mm, vma, vaddr, > (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); > spin_lock(&mm->page_table_lock); > if (!(ret & VM_FAULT_ERROR)) > continue; > and another is handle_mm_fault(), which ends up returning ENOSPC and *its* > callers are definitely not ready to deal with that. > > ENOSPC is 28, i.e. VM_FAULT_MAJOR | VM_FAULT_WRITE | VM_FAULT_HWPOISON; > it's also theoretically possible to get ENOMEM if region_chg() ends up > hitting > nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); > if (!nrg) > return -ENOMEM; > region_chg() <- vma_needs_reservation() <- alloc_huge_page() and from that > point as with ENOSPC. ENOMEM is 12, i.e. VM_FAULT_MAJOR | VM_FAULT_WRITE... Good find, news to me. Interesting uses of -PTR_ERR()! Looks like we'd better not have more than 12 VM_FAULT_ flags. > > Am I right assuming that we want VM_FAULT_OOM in both cases? No, where hugetlb_get_quota() fails it should be VM_FAULT_SIGBUS: there's no excuse to go on an OOM-killing spree just because hugetlb quota is exhausted. VM_FAULT_OOM is appropriate where vma_needs_reservation() fails, because region_chg() couldn't kmalloc a structure, as you point out. (Though that doesn't matter much, since the only way the kmalloc can fail is when this task is already selected for OOM-kill - I think.) Hugh -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx 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>