Message ID | 20191203104853.4378-3-thomas_os@shipmail.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [v2,1/2] mm: Add and export vmf_insert_mixed_prot() | expand |
On Tue 03-12-19 11:48:53, Thomas Hellström (VMware) wrote: > From: Thomas Hellstrom <thellstrom@vmware.com> > > TTM graphics buffer objects may, transparently to user-space, move > between IO and system memory. When that happens, all PTEs pointing to the > old location are zapped before the move and then faulted in again if > needed. When that happens, the page protection caching mode- and > encryption bits may change and be different from those of > struct vm_area_struct::vm_page_prot. > > We were using an ugly hack to set the page protection correctly. > Fix that and instead use vmf_insert_mixed_prot() and / or > vmf_insert_pfn_prot(). > Also get the default page protection from > struct vm_area_struct::vm_page_prot rather than using vm_get_page_prot(). > This way we catch modifications done by the vm system for drivers that > want write-notification. So essentially this should have any new side effect on functionality it is just making a hacky/ugly code less so? In other words what are the consequences of having page protection inconsistent from vma's? > Cc: Andrew Morton <akpm@linux-foundation.org> > Cc: Michal Hocko <mhocko@suse.com> > Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org> > Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> > Cc: Ralph Campbell <rcampbell@nvidia.com> > Cc: "Jérôme Glisse" <jglisse@redhat.com> > Cc: "Christian König" <christian.koenig@amd.com> > Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com> > Reviewed-by: Christian König <christian.koenig@amd.com> > --- > drivers/gpu/drm/ttm/ttm_bo_vm.c | 14 +++++++------- > 1 file changed, 7 insertions(+), 7 deletions(-) > > diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c > index e6495ca2630b..2098f8d4dfc5 100644 > --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c > +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c > @@ -173,7 +173,6 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, > pgoff_t num_prefault) > { > struct vm_area_struct *vma = vmf->vma; > - struct vm_area_struct cvma = *vma; > struct ttm_buffer_object *bo = vma->vm_private_data; > struct ttm_bo_device *bdev = bo->bdev; > unsigned long page_offset; > @@ -244,7 +243,7 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, > goto out_io_unlock; > } > > - cvma.vm_page_prot = ttm_io_prot(bo->mem.placement, prot); > + prot = ttm_io_prot(bo->mem.placement, prot); > if (!bo->mem.bus.is_iomem) { > struct ttm_operation_ctx ctx = { > .interruptible = false, > @@ -260,7 +259,7 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, > } > } else { > /* Iomem should not be marked encrypted */ > - cvma.vm_page_prot = pgprot_decrypted(cvma.vm_page_prot); > + prot = pgprot_decrypted(prot); > } > > /* > @@ -284,10 +283,11 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, > } > > if (vma->vm_flags & VM_MIXEDMAP) > - ret = vmf_insert_mixed(&cvma, address, > - __pfn_to_pfn_t(pfn, PFN_DEV)); > + ret = vmf_insert_mixed_prot(vma, address, > + __pfn_to_pfn_t(pfn, PFN_DEV), > + prot); > else > - ret = vmf_insert_pfn(&cvma, address, pfn); > + ret = vmf_insert_pfn_prot(vma, address, pfn, prot); > > /* Never error on prefaulted PTEs */ > if (unlikely((ret & VM_FAULT_ERROR))) { > @@ -319,7 +319,7 @@ vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf) > if (ret) > return ret; > > - prot = vm_get_page_prot(vma->vm_flags); > + prot = vma->vm_page_prot; > ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT); > if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) > return ret; > -- > 2.21.0
On 12/4/19 2:52 PM, Michal Hocko wrote: > On Tue 03-12-19 11:48:53, Thomas Hellström (VMware) wrote: >> From: Thomas Hellstrom <thellstrom@vmware.com> >> >> TTM graphics buffer objects may, transparently to user-space, move >> between IO and system memory. When that happens, all PTEs pointing to the >> old location are zapped before the move and then faulted in again if >> needed. When that happens, the page protection caching mode- and >> encryption bits may change and be different from those of >> struct vm_area_struct::vm_page_prot. >> >> We were using an ugly hack to set the page protection correctly. >> Fix that and instead use vmf_insert_mixed_prot() and / or >> vmf_insert_pfn_prot(). >> Also get the default page protection from >> struct vm_area_struct::vm_page_prot rather than using vm_get_page_prot(). >> This way we catch modifications done by the vm system for drivers that >> want write-notification. > So essentially this should have any new side effect on functionality it > is just making a hacky/ugly code less so? Functionality is unchanged. The use of a on-stack vma copy was severely frowned upon in an earlier thread, which also points to another similar example using vmf_insert_pfn_prot(). https://lore.kernel.org/lkml/20190905103541.4161-2-thomas_os@shipmail.org/ > In other words what are the > consequences of having page protection inconsistent from vma's? During the years, it looks like the caching- and encryption flags of vma::vm_page_prot have been largely removed from usage. From what I can tell, there are no more places left that can affect TTM. We discussed __split_huge_pmd_locked() towards the end of that thread, but that doesn't affect TTM even with huge page-table entries. /Thomas
On Wed 04-12-19 15:16:09, Thomas Hellström (VMware) wrote: > On 12/4/19 2:52 PM, Michal Hocko wrote: > > On Tue 03-12-19 11:48:53, Thomas Hellström (VMware) wrote: > > > From: Thomas Hellstrom <thellstrom@vmware.com> > > > > > > TTM graphics buffer objects may, transparently to user-space, move > > > between IO and system memory. When that happens, all PTEs pointing to the > > > old location are zapped before the move and then faulted in again if > > > needed. When that happens, the page protection caching mode- and > > > encryption bits may change and be different from those of > > > struct vm_area_struct::vm_page_prot. > > > > > > We were using an ugly hack to set the page protection correctly. > > > Fix that and instead use vmf_insert_mixed_prot() and / or > > > vmf_insert_pfn_prot(). > > > Also get the default page protection from > > > struct vm_area_struct::vm_page_prot rather than using vm_get_page_prot(). > > > This way we catch modifications done by the vm system for drivers that > > > want write-notification. > > So essentially this should have any new side effect on functionality it > > is just making a hacky/ugly code less so? > > Functionality is unchanged. The use of a on-stack vma copy was severely > frowned upon in an earlier thread, which also points to another similar > example using vmf_insert_pfn_prot(). > > https://lore.kernel.org/lkml/20190905103541.4161-2-thomas_os@shipmail.org/ > > > In other words what are the > > consequences of having page protection inconsistent from vma's? > > During the years, it looks like the caching- and encryption flags of > vma::vm_page_prot have been largely removed from usage. From what I can > tell, there are no more places left that can affect TTM. We discussed > __split_huge_pmd_locked() towards the end of that thread, but that doesn't > affect TTM even with huge page-table entries. Please state all those details/assumptions you are operating on in the changelog.
On 12/4/19 3:35 PM, Michal Hocko wrote: > On Wed 04-12-19 15:16:09, Thomas Hellström (VMware) wrote: >> On 12/4/19 2:52 PM, Michal Hocko wrote: >>> On Tue 03-12-19 11:48:53, Thomas Hellström (VMware) wrote: >>>> From: Thomas Hellstrom <thellstrom@vmware.com> >>>> >>>> TTM graphics buffer objects may, transparently to user-space, move >>>> between IO and system memory. When that happens, all PTEs pointing to the >>>> old location are zapped before the move and then faulted in again if >>>> needed. When that happens, the page protection caching mode- and >>>> encryption bits may change and be different from those of >>>> struct vm_area_struct::vm_page_prot. >>>> >>>> We were using an ugly hack to set the page protection correctly. >>>> Fix that and instead use vmf_insert_mixed_prot() and / or >>>> vmf_insert_pfn_prot(). >>>> Also get the default page protection from >>>> struct vm_area_struct::vm_page_prot rather than using vm_get_page_prot(). >>>> This way we catch modifications done by the vm system for drivers that >>>> want write-notification. >>> So essentially this should have any new side effect on functionality it >>> is just making a hacky/ugly code less so? >> Functionality is unchanged. The use of a on-stack vma copy was severely >> frowned upon in an earlier thread, which also points to another similar >> example using vmf_insert_pfn_prot(). >> >> https://lore.kernel.org/lkml/20190905103541.4161-2-thomas_os@shipmail.org/ >> >>> In other words what are the >>> consequences of having page protection inconsistent from vma's? >> During the years, it looks like the caching- and encryption flags of >> vma::vm_page_prot have been largely removed from usage. From what I can >> tell, there are no more places left that can affect TTM. We discussed >> __split_huge_pmd_locked() towards the end of that thread, but that doesn't >> affect TTM even with huge page-table entries. > Please state all those details/assumptions you are operating on in the > changelog. Thanks. I'll update the patchset and add that. /Thomas
On Wed 04-12-19 15:36:58, Thomas Hellström (VMware) wrote: > On 12/4/19 3:35 PM, Michal Hocko wrote: > > On Wed 04-12-19 15:16:09, Thomas Hellström (VMware) wrote: > > > On 12/4/19 2:52 PM, Michal Hocko wrote: > > > > On Tue 03-12-19 11:48:53, Thomas Hellström (VMware) wrote: > > > > > From: Thomas Hellstrom <thellstrom@vmware.com> > > > > > > > > > > TTM graphics buffer objects may, transparently to user-space, move > > > > > between IO and system memory. When that happens, all PTEs pointing to the > > > > > old location are zapped before the move and then faulted in again if > > > > > needed. When that happens, the page protection caching mode- and > > > > > encryption bits may change and be different from those of > > > > > struct vm_area_struct::vm_page_prot. > > > > > > > > > > We were using an ugly hack to set the page protection correctly. > > > > > Fix that and instead use vmf_insert_mixed_prot() and / or > > > > > vmf_insert_pfn_prot(). > > > > > Also get the default page protection from > > > > > struct vm_area_struct::vm_page_prot rather than using vm_get_page_prot(). > > > > > This way we catch modifications done by the vm system for drivers that > > > > > want write-notification. > > > > So essentially this should have any new side effect on functionality it > > > > is just making a hacky/ugly code less so? > > > Functionality is unchanged. The use of a on-stack vma copy was severely > > > frowned upon in an earlier thread, which also points to another similar > > > example using vmf_insert_pfn_prot(). > > > > > > https://lore.kernel.org/lkml/20190905103541.4161-2-thomas_os@shipmail.org/ > > > > > > > In other words what are the > > > > consequences of having page protection inconsistent from vma's? > > > During the years, it looks like the caching- and encryption flags of > > > vma::vm_page_prot have been largely removed from usage. From what I can > > > tell, there are no more places left that can affect TTM. We discussed > > > __split_huge_pmd_locked() towards the end of that thread, but that doesn't > > > affect TTM even with huge page-table entries. > > Please state all those details/assumptions you are operating on in the > > changelog. > > Thanks. I'll update the patchset and add that. And thinking about that this also begs for a comment in the code to explain that some (which?) mappings might have a mismatch and the generic code have to be careful. Because as things stand now this seems to be really subtle and happen to work _now_ and might break in the future. Or what does prevent a generic code to stumble over this discrepancy?
On 12/4/19 3:42 PM, Michal Hocko wrote: > On Wed 04-12-19 15:36:58, Thomas Hellström (VMware) wrote: >> On 12/4/19 3:35 PM, Michal Hocko wrote: >>> On Wed 04-12-19 15:16:09, Thomas Hellström (VMware) wrote: >>>> On 12/4/19 2:52 PM, Michal Hocko wrote: >>>>> On Tue 03-12-19 11:48:53, Thomas Hellström (VMware) wrote: >>>>>> From: Thomas Hellstrom <thellstrom@vmware.com> >>>>>> >>>>>> TTM graphics buffer objects may, transparently to user-space, move >>>>>> between IO and system memory. When that happens, all PTEs pointing to the >>>>>> old location are zapped before the move and then faulted in again if >>>>>> needed. When that happens, the page protection caching mode- and >>>>>> encryption bits may change and be different from those of >>>>>> struct vm_area_struct::vm_page_prot. >>>>>> >>>>>> We were using an ugly hack to set the page protection correctly. >>>>>> Fix that and instead use vmf_insert_mixed_prot() and / or >>>>>> vmf_insert_pfn_prot(). >>>>>> Also get the default page protection from >>>>>> struct vm_area_struct::vm_page_prot rather than using vm_get_page_prot(). >>>>>> This way we catch modifications done by the vm system for drivers that >>>>>> want write-notification. >>>>> So essentially this should have any new side effect on functionality it >>>>> is just making a hacky/ugly code less so? >>>> Functionality is unchanged. The use of a on-stack vma copy was severely >>>> frowned upon in an earlier thread, which also points to another similar >>>> example using vmf_insert_pfn_prot(). >>>> >>>> https://lore.kernel.org/lkml/20190905103541.4161-2-thomas_os@shipmail.org/ >>>> >>>>> In other words what are the >>>>> consequences of having page protection inconsistent from vma's? >>>> During the years, it looks like the caching- and encryption flags of >>>> vma::vm_page_prot have been largely removed from usage. From what I can >>>> tell, there are no more places left that can affect TTM. We discussed >>>> __split_huge_pmd_locked() towards the end of that thread, but that doesn't >>>> affect TTM even with huge page-table entries. >>> Please state all those details/assumptions you are operating on in the >>> changelog. >> Thanks. I'll update the patchset and add that. > And thinking about that this also begs for a comment in the code to > explain that some (which?) mappings might have a mismatch and the > generic code have to be careful. Because as things stand now this seems > to be really subtle and happen to work _now_ and might break in the future. > Or what does prevent a generic code to stumble over this discrepancy? Yes we had that discussion in the thread I pointed to. I initially suggested and argued for updating the vma::vm_page_prot using a WRITE_ONCE() (we only have the mmap_sem in read mode), there seems to be other places in generic code that does the same. But I was convinced by Andy that this was the right way and also was used elsewhere. (See also https://elixir.bootlin.com/linux/latest/source/arch/x86/entry/vdso/vma.c#L116) I guess to have this properly formulated, what's required is that generic code doesn't build page-table entries using vma::vm_page_prot for VM_PFNMAP and VM_MIXEDMAP outside of driver control. /Thomas
On Wed 04-12-19 16:19:27, Thomas Hellström (VMware) wrote: > On 12/4/19 3:42 PM, Michal Hocko wrote: > > On Wed 04-12-19 15:36:58, Thomas Hellström (VMware) wrote: > > > On 12/4/19 3:35 PM, Michal Hocko wrote: > > > > On Wed 04-12-19 15:16:09, Thomas Hellström (VMware) wrote: > > > > > On 12/4/19 2:52 PM, Michal Hocko wrote: > > > > > > On Tue 03-12-19 11:48:53, Thomas Hellström (VMware) wrote: > > > > > > > From: Thomas Hellstrom <thellstrom@vmware.com> > > > > > > > > > > > > > > TTM graphics buffer objects may, transparently to user-space, move > > > > > > > between IO and system memory. When that happens, all PTEs pointing to the > > > > > > > old location are zapped before the move and then faulted in again if > > > > > > > needed. When that happens, the page protection caching mode- and > > > > > > > encryption bits may change and be different from those of > > > > > > > struct vm_area_struct::vm_page_prot. > > > > > > > > > > > > > > We were using an ugly hack to set the page protection correctly. > > > > > > > Fix that and instead use vmf_insert_mixed_prot() and / or > > > > > > > vmf_insert_pfn_prot(). > > > > > > > Also get the default page protection from > > > > > > > struct vm_area_struct::vm_page_prot rather than using vm_get_page_prot(). > > > > > > > This way we catch modifications done by the vm system for drivers that > > > > > > > want write-notification. > > > > > > So essentially this should have any new side effect on functionality it > > > > > > is just making a hacky/ugly code less so? > > > > > Functionality is unchanged. The use of a on-stack vma copy was severely > > > > > frowned upon in an earlier thread, which also points to another similar > > > > > example using vmf_insert_pfn_prot(). > > > > > > > > > > https://lore.kernel.org/lkml/20190905103541.4161-2-thomas_os@shipmail.org/ > > > > > > > > > > > In other words what are the > > > > > > consequences of having page protection inconsistent from vma's? > > > > > During the years, it looks like the caching- and encryption flags of > > > > > vma::vm_page_prot have been largely removed from usage. From what I can > > > > > tell, there are no more places left that can affect TTM. We discussed > > > > > __split_huge_pmd_locked() towards the end of that thread, but that doesn't > > > > > affect TTM even with huge page-table entries. > > > > Please state all those details/assumptions you are operating on in the > > > > changelog. > > > Thanks. I'll update the patchset and add that. > > And thinking about that this also begs for a comment in the code to > > explain that some (which?) mappings might have a mismatch and the > > generic code have to be careful. Because as things stand now this seems > > to be really subtle and happen to work _now_ and might break in the future. > > Or what does prevent a generic code to stumble over this discrepancy? > > Yes we had that discussion in the thread I pointed to. I initially suggested > and argued for updating the vma::vm_page_prot using a WRITE_ONCE() (we only > have the mmap_sem in read mode), there seems to be other places in generic > code that does the same. > > But I was convinced by Andy that this was the right way and also was used > elsewhere. > > (See also https://elixir.bootlin.com/linux/latest/source/arch/x86/entry/vdso/vma.c#L116) > > I guess to have this properly formulated, what's required is that generic > code doesn't build page-table entries using vma::vm_page_prot for VM_PFNMAP > and VM_MIXEDMAP outside of driver control. Let me repeat that this belongs to a code somewhere everybody can see it rather than a "random" discussion at mailing list. Thanks!
On 12/4/19 4:26 PM, Michal Hocko wrote: > On Wed 04-12-19 16:19:27, Thomas Hellström (VMware) wrote: >> On 12/4/19 3:42 PM, Michal Hocko wrote: >>> On Wed 04-12-19 15:36:58, Thomas Hellström (VMware) wrote: >>>> On 12/4/19 3:35 PM, Michal Hocko wrote: >>>>> On Wed 04-12-19 15:16:09, Thomas Hellström (VMware) wrote: >>>>>> On 12/4/19 2:52 PM, Michal Hocko wrote: >>>>>>> On Tue 03-12-19 11:48:53, Thomas Hellström (VMware) wrote: >>>>>>>> From: Thomas Hellstrom <thellstrom@vmware.com> >>>>>>>> >>>>>>>> TTM graphics buffer objects may, transparently to user-space, move >>>>>>>> between IO and system memory. When that happens, all PTEs pointing to the >>>>>>>> old location are zapped before the move and then faulted in again if >>>>>>>> needed. When that happens, the page protection caching mode- and >>>>>>>> encryption bits may change and be different from those of >>>>>>>> struct vm_area_struct::vm_page_prot. >>>>>>>> >>>>>>>> We were using an ugly hack to set the page protection correctly. >>>>>>>> Fix that and instead use vmf_insert_mixed_prot() and / or >>>>>>>> vmf_insert_pfn_prot(). >>>>>>>> Also get the default page protection from >>>>>>>> struct vm_area_struct::vm_page_prot rather than using vm_get_page_prot(). >>>>>>>> This way we catch modifications done by the vm system for drivers that >>>>>>>> want write-notification. >>>>>>> So essentially this should have any new side effect on functionality it >>>>>>> is just making a hacky/ugly code less so? >>>>>> Functionality is unchanged. The use of a on-stack vma copy was severely >>>>>> frowned upon in an earlier thread, which also points to another similar >>>>>> example using vmf_insert_pfn_prot(). >>>>>> >>>>>> https://lore.kernel.org/lkml/20190905103541.4161-2-thomas_os@shipmail.org/ >>>>>> >>>>>>> In other words what are the >>>>>>> consequences of having page protection inconsistent from vma's? >>>>>> During the years, it looks like the caching- and encryption flags of >>>>>> vma::vm_page_prot have been largely removed from usage. From what I can >>>>>> tell, there are no more places left that can affect TTM. We discussed >>>>>> __split_huge_pmd_locked() towards the end of that thread, but that doesn't >>>>>> affect TTM even with huge page-table entries. >>>>> Please state all those details/assumptions you are operating on in the >>>>> changelog. >>>> Thanks. I'll update the patchset and add that. >>> And thinking about that this also begs for a comment in the code to >>> explain that some (which?) mappings might have a mismatch and the >>> generic code have to be careful. Because as things stand now this seems >>> to be really subtle and happen to work _now_ and might break in the future. >>> Or what does prevent a generic code to stumble over this discrepancy? >> Yes we had that discussion in the thread I pointed to. I initially suggested >> and argued for updating the vma::vm_page_prot using a WRITE_ONCE() (we only >> have the mmap_sem in read mode), there seems to be other places in generic >> code that does the same. >> >> But I was convinced by Andy that this was the right way and also was used >> elsewhere. >> >> (See also https://elixir.bootlin.com/linux/latest/source/arch/x86/entry/vdso/vma.c#L116) >> >> I guess to have this properly formulated, what's required is that generic >> code doesn't build page-table entries using vma::vm_page_prot for VM_PFNMAP >> and VM_MIXEDMAP outside of driver control. > Let me repeat that this belongs to a code somewhere everybody can see it > rather than a "random" discussion at mailing list. > > Thanks! Yes, I agree. I'll of course follow up with the comments added to the code. Thomas
diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c index e6495ca2630b..2098f8d4dfc5 100644 --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c @@ -173,7 +173,6 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, pgoff_t num_prefault) { struct vm_area_struct *vma = vmf->vma; - struct vm_area_struct cvma = *vma; struct ttm_buffer_object *bo = vma->vm_private_data; struct ttm_bo_device *bdev = bo->bdev; unsigned long page_offset; @@ -244,7 +243,7 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, goto out_io_unlock; } - cvma.vm_page_prot = ttm_io_prot(bo->mem.placement, prot); + prot = ttm_io_prot(bo->mem.placement, prot); if (!bo->mem.bus.is_iomem) { struct ttm_operation_ctx ctx = { .interruptible = false, @@ -260,7 +259,7 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, } } else { /* Iomem should not be marked encrypted */ - cvma.vm_page_prot = pgprot_decrypted(cvma.vm_page_prot); + prot = pgprot_decrypted(prot); } /* @@ -284,10 +283,11 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf, } if (vma->vm_flags & VM_MIXEDMAP) - ret = vmf_insert_mixed(&cvma, address, - __pfn_to_pfn_t(pfn, PFN_DEV)); + ret = vmf_insert_mixed_prot(vma, address, + __pfn_to_pfn_t(pfn, PFN_DEV), + prot); else - ret = vmf_insert_pfn(&cvma, address, pfn); + ret = vmf_insert_pfn_prot(vma, address, pfn, prot); /* Never error on prefaulted PTEs */ if (unlikely((ret & VM_FAULT_ERROR))) { @@ -319,7 +319,7 @@ vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf) if (ret) return ret; - prot = vm_get_page_prot(vma->vm_flags); + prot = vma->vm_page_prot; ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT); if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) return ret;