On Wed, 10 May 2023, Matthew Wilcox wrote: > On Tue, May 09, 2023 at 09:42:44PM -0700, Hugh Dickins wrote: > > diff --git a/arch/arm/lib/uaccess_with_memcpy.c b/arch/arm/lib/uaccess_with_memcpy.c > > index e4c2677cc1e9..2f6163f05e93 100644 > > --- a/arch/arm/lib/uaccess_with_memcpy.c > > +++ b/arch/arm/lib/uaccess_with_memcpy.c > > @@ -74,6 +74,9 @@ pin_page_for_write(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp) > > return 0; > > > > pte = pte_offset_map_lock(current->mm, pmd, addr, &ptl); > > + if (unlikely(!pte)) > > + return 0; > > Failing seems like the wrong thig to do if we transitioned from a PTE > to PMD here? Looks to me like we should goto a new label right after > the 'pmd = pmd_offset(pud, addr);', no? I'm pretty sure it's right as is; but probably more by luck than care - I do not think I studied this code as closely as you have now made me do; and it's clear that this is a piece of code where rare transient issues could come up, and must be handled correctly. Thank you for making me look again. The key is in the callers of pin_page_for_write(): __copy_to_user_memcpy() and __clear_user_memset(). They're doing "while (!pin_page_for_write())" loops - they hope for the fast path of getting pte_lock or pmd_lock on the page, and doing a __memcpy() or __memset() to the user address; but if anything goes "wrong", a __put_user() to fault in the page (or fail) then pin_page_for_write() again. "if (unlikely(!pte)) return 0" says that the expected fast path did not succeed, so please __put_user() and have another go. It is somewhere I could have done a "goto again", but that would be superfluous when it's already designed that way at the outer level. Hugh
