| Message ID | 20240412142120.220087-4-david@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | s390: PG_arch_1+folio cleanups for uv+hugetlb | expand |
On Fri, 12 Apr 2024 16:21:13 +0200 David Hildenbrand <david@redhat.com> wrote: > While s390x makes sure to never have PMD-mapped THP in processes that use > KVM -- by remapping them using PTEs in > thp_split_walk_pmd_entry()->split_huge_pmd() -- there is still the > possibility of having PTE-mapped THPs (large folios) mapped into guest > memory. > > This would happen if user space allocates memory before calling > KVM_CREATE_VM (which would call s390_enable_sie()). With upstream QEMU, > this currently doesn't happen, because guest memory is setup and > condiitonally preallocated after KVM_CREATE_VM. *conditionally > > Could it happen with shmem/file-backed memory when another process > allocated memory in the pagecache? Likely, although currently not a > common setup. > > Trying to split any PTE-mapped large folios sounds like the right and > future-proof thing to do here. So let's call split_folio() and handle the > return values accordingly. Reviewed-by: Claudio Imbrenda <imbrenda@linux.ibm.com> > > Signed-off-by: David Hildenbrand <david@redhat.com> > --- > arch/s390/kernel/uv.c | 31 +++++++++++++++++++++++++------ > 1 file changed, 25 insertions(+), 6 deletions(-) > > diff --git a/arch/s390/kernel/uv.c b/arch/s390/kernel/uv.c > index 25fe28d189df..3c6d86e3e828 100644 > --- a/arch/s390/kernel/uv.c > +++ b/arch/s390/kernel/uv.c > @@ -338,11 +338,10 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb) > goto out; > if (pte_present(*ptep) && !(pte_val(*ptep) & _PAGE_INVALID) && pte_write(*ptep)) { > folio = page_folio(pte_page(*ptep)); > - rc = -EINVAL; > - if (folio_test_large(folio)) > - goto unlock; > rc = -EAGAIN; > - if (folio_trylock(folio)) { > + if (folio_test_large(folio)) { > + rc = -E2BIG; > + } else if (folio_trylock(folio)) { > if (should_export_before_import(uvcb, gmap->mm)) > uv_convert_from_secure(PFN_PHYS(folio_pfn(folio))); > rc = make_folio_secure(folio, uvcb); > @@ -353,15 +352,35 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb) > * Once we drop the PTL, the folio may get unmapped and > * freed immediately. We need a temporary reference. > */ > - if (rc == -EAGAIN) > + if (rc == -EAGAIN || rc == -E2BIG) > folio_get(folio); > } > -unlock: > pte_unmap_unlock(ptep, ptelock); > out: > mmap_read_unlock(gmap->mm); > > switch (rc) { > + case -E2BIG: > + folio_lock(folio); > + rc = split_folio(folio); > + folio_unlock(folio); > + folio_put(folio); > + > + switch (rc) { > + case 0: > + /* Splitting succeeded, try again immediately. */ > + goto again; > + case -EAGAIN: > + /* Additional folio references. */ > + if (drain_lru(&drain_lru_called)) > + goto again; > + return -EAGAIN; > + case -EBUSY: > + /* Unexpected race. */ > + return -EAGAIN; > + } > + WARN_ON_ONCE(1); > + return -ENXIO; > case -EAGAIN: > /* > * If we are here because the UVC returned busy or partial
diff --git a/arch/s390/kernel/uv.c b/arch/s390/kernel/uv.c index 25fe28d189df..3c6d86e3e828 100644 --- a/arch/s390/kernel/uv.c +++ b/arch/s390/kernel/uv.c @@ -338,11 +338,10 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb) goto out; if (pte_present(*ptep) && !(pte_val(*ptep) & _PAGE_INVALID) && pte_write(*ptep)) { folio = page_folio(pte_page(*ptep)); - rc = -EINVAL; - if (folio_test_large(folio)) - goto unlock; rc = -EAGAIN; - if (folio_trylock(folio)) { + if (folio_test_large(folio)) { + rc = -E2BIG; + } else if (folio_trylock(folio)) { if (should_export_before_import(uvcb, gmap->mm)) uv_convert_from_secure(PFN_PHYS(folio_pfn(folio))); rc = make_folio_secure(folio, uvcb); @@ -353,15 +352,35 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb) * Once we drop the PTL, the folio may get unmapped and * freed immediately. We need a temporary reference. */ - if (rc == -EAGAIN) + if (rc == -EAGAIN || rc == -E2BIG) folio_get(folio); } -unlock: pte_unmap_unlock(ptep, ptelock); out: mmap_read_unlock(gmap->mm); switch (rc) { + case -E2BIG: + folio_lock(folio); + rc = split_folio(folio); + folio_unlock(folio); + folio_put(folio); + + switch (rc) { + case 0: + /* Splitting succeeded, try again immediately. */ + goto again; + case -EAGAIN: + /* Additional folio references. */ + if (drain_lru(&drain_lru_called)) + goto again; + return -EAGAIN; + case -EBUSY: + /* Unexpected race. */ + return -EAGAIN; + } + WARN_ON_ONCE(1); + return -ENXIO; case -EAGAIN: /* * If we are here because the UVC returned busy or partial
While s390x makes sure to never have PMD-mapped THP in processes that use KVM -- by remapping them using PTEs in thp_split_walk_pmd_entry()->split_huge_pmd() -- there is still the possibility of having PTE-mapped THPs (large folios) mapped into guest memory. This would happen if user space allocates memory before calling KVM_CREATE_VM (which would call s390_enable_sie()). With upstream QEMU, this currently doesn't happen, because guest memory is setup and condiitonally preallocated after KVM_CREATE_VM. Could it happen with shmem/file-backed memory when another process allocated memory in the pagecache? Likely, although currently not a common setup. Trying to split any PTE-mapped large folios sounds like the right and future-proof thing to do here. So let's call split_folio() and handle the return values accordingly. Signed-off-by: David Hildenbrand <david@redhat.com> --- arch/s390/kernel/uv.c | 31 +++++++++++++++++++++++++------ 1 file changed, 25 insertions(+), 6 deletions(-)