On Fri, Oct 11, 2024 at 11:12 AM Jann Horn <jannh@xxxxxxxxxx> wrote: > > On Fri, Sep 27, 2024 at 2:51 PM Lorenzo Stoakes > <lorenzo.stoakes@xxxxxxxxxx> wrote: > > Implement a new lightweight guard page feature, that is regions of userland > > virtual memory that, when accessed, cause a fatal signal to arise. > [...] > > --- > > arch/alpha/include/uapi/asm/mman.h | 3 + > > arch/mips/include/uapi/asm/mman.h | 3 + > > arch/parisc/include/uapi/asm/mman.h | 3 + > > arch/xtensa/include/uapi/asm/mman.h | 3 + > > include/uapi/asm-generic/mman-common.h | 3 + > > I kinda wonder if we could start moving the parts of those headers > that are the same for all architectures to include/uapi/linux/mman.h > instead... but that's maybe out of scope for this series. > > [...] > > diff --git a/mm/madvise.c b/mm/madvise.c > > index e871a72a6c32..7216e10723ae 100644 > > --- a/mm/madvise.c > > +++ b/mm/madvise.c > > @@ -60,6 +60,7 @@ static int madvise_need_mmap_write(int behavior) > > case MADV_POPULATE_READ: > > case MADV_POPULATE_WRITE: > > case MADV_COLLAPSE: > > + case MADV_GUARD_UNPOISON: /* Only poisoning needs a write lock. */ > > What does poisoning need a write lock for? anon_vma_prepare() doesn't > need it (it only needs mmap_lock held for reading), > zap_page_range_single() doesn't need it, and pagewalk also doesn't > need it as long as the range being walked is covered by a VMA, which > it is... > > I see you set PGWALK_WRLOCK in guard_poison_walk_ops with a comment > saying "We might need to install an anon_vma" - is that referring to > an older version of the patch where the anon_vma_prepare() call was > inside the pagewalk callback or something like that? Either way, > anon_vma_prepare() doesn't need write locks (it can't, it has to work > from the page fault handling path). I was wondering about that too and I can't find any reason for write-locking the mm for this operation. PGWALK_WRLOCK should also be changed to PGWALK_RDLOCK as we are not modifying the VMA. BTW, I'm testing your patchset on Android and so far it is stable! > > > return 0; > > default: > > /* be safe, default to 1. list exceptions explicitly */ > [...] > > +static long madvise_guard_poison(struct vm_area_struct *vma, > > + struct vm_area_struct **prev, > > + unsigned long start, unsigned long end) > > +{ > > + long err; > > + bool retried = false; > > + > > + *prev = vma; > > + if (!is_valid_guard_vma(vma, /* allow_locked = */false)) > > + return -EINVAL; > > + > > + /* > > + * Optimistically try to install the guard poison pages first. If any > > + * non-guard pages are encountered, give up and zap the range before > > + * trying again. > > + */ > > + while (true) { > > + unsigned long num_installed = 0; > > + > > + /* Returns < 0 on error, == 0 if success, > 0 if zap needed. */ > > + err = walk_page_range_mm(vma->vm_mm, start, end, > > + &guard_poison_walk_ops, > > + &num_installed); > > + /* > > + * If we install poison markers, then the range is no longer > > + * empty from a page table perspective and therefore it's > > + * appropriate to have an anon_vma. > > + * > > + * This ensures that on fork, we copy page tables correctly. > > + */ > > + if (err >= 0 && num_installed > 0) { > > + int err_anon = anon_vma_prepare(vma); > > I'd move this up, to before we create poison PTEs. There's no harm in > attaching an anon_vma to the VMA even if the rest of the operation > fails; and I think it would be weird to have error paths that don't > attach an anon_vma even though they . > > > + if (err_anon) > > + err = err_anon; > > + } > > + > > + if (err <= 0) > > + return err; > > + > > + if (!retried) > > + /* > > + * OK some of the range have non-guard pages mapped, zap > > + * them. This leaves existing guard pages in place. > > + */ > > + zap_page_range_single(vma, start, end - start, NULL); > > + else > > + /* > > + * If we reach here, then there is a racing fault that > > + * has populated the PTE after we zapped. Give up and > > + * let the user know to try again. > > + */ > > + return -EAGAIN; > > Hmm, yeah, it would be nice if we could avoid telling userspace to > loop on -EAGAIN but I guess we don't have any particularly good > options here? Well, we could bail out with -EINTR if a (fatal?) signal > is pending and otherwise keep looping... if we'd tell userspace "try > again on -EAGAIN", we might as well do that in the kernel... > > (Personally I would put curly braces around these branches because > they occupy multiple lines, though the coding style doesn't explicitly > say that, so I guess maybe it's a matter of personal preference... > adding curly braces here would match what is done, for example, in > relocate_vma_down().) > > > + retried = true; > > + } > > +} > > + > > +static int guard_unpoison_pte_entry(pte_t *pte, unsigned long addr, > > + unsigned long next, struct mm_walk *walk) > > +{ > > + pte_t ptent = ptep_get(pte); > > + > > + if (is_guard_pte_marker(ptent)) { > > + /* Simply clear the PTE marker. */ > > + pte_clear_not_present_full(walk->mm, addr, pte, true); > > I think that last parameter probably should be "false"? The sparc code > calls it "fullmm", which is a term the MM code uses when talking about > operations that remove all mappings in the entire mm_struct because > the process has died, which allows using some faster special-case > version of TLB shootdown or something along those lines. > > > + update_mmu_cache(walk->vma, addr, pte); > > + } > > + > > + return 0; > > +} > > + > > +static const struct mm_walk_ops guard_unpoison_walk_ops = { > > + .pte_entry = guard_unpoison_pte_entry, > > + .walk_lock = PGWALK_RDLOCK, > > +}; > > It is a _little_ weird that unpoisoning creates page tables when they > don't already exist, which will also prevent creating THP entries on > fault in such areas afterwards... but I guess it doesn't really matter > given that poisoning has that effect, too, and you probably usually > won't call MADV_GUARD_UNPOISON on an area that hasn't been poisoned > before... so I guess this is not an actionable comment.