On Fri, Sep 6, 2019 at 12:29 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Fri, Sep 06, 2019 at 12:13:00AM +0530, Souptick Joarder wrote:
> > +++ b/mm/memory.c
> > @@ -1750,13 +1750,10 @@ static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
> > } else {
> > return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
> > }
> > -
> > - if (err == -ENOMEM)
> > - return VM_FAULT_OOM;
> > - if (err < 0 && err != -EBUSY)
> > - return VM_FAULT_SIGBUS;
> > -
> > - return VM_FAULT_NOPAGE;
> > + if (!err || err == -EBUSY)
> > + return VM_FAULT_NOPAGE;
> > + else
> > + return vmf_error(err);
> > }
>
> My plan is to convert insert_page() to return a VM_FAULT error code like
> insert_pfn() does. Need to finish off the vm_insert_page() conversions
> first ;-)
Previously we have problem while converting vm_insert_page() to return
vm_fault_t.
vm_insert_page() is called from different drivers. Some of them are
already converted
to use vm_map_pages()/ vm_map_pages_zero(). But still we left with few users.
drivers/media/usb/usbvision/usbvision-video.c#L1045
mm/vmalloc.c#L2969
These 2 can be converted with something like vm_map_vmalloc_pages().
I am working on it. Will post it in sometime.
drivers/android/binder_alloc.c#L259 (have objection)
drivers/infiniband/hw/efa/efa_verbs.c#L1701
drivers/infiniband/hw/mlx5/main.c#L2085 (have objection as using
vm_map_pages_zero() doesn't make sense)
drivers/xen/gntalloc.c#L548 (have limitation)
kernel/kcov.c#L297 (have objection)
net/ipv4/tcp.c#L1799
net/packet/af_packet.c#L4453
But these are the places where replacing vm_insert_page() is bit
difficult/ have objection.
In some cases, maintainers/ reviewers will not agree to replace
vm_insert_page().
In other scenario, if we change return type of vm_insert_page() to vm_fault_t,
we end up with adding few lines of conversion code from vm_fault_t to errno
in drivers which is not a correct way to go with.
Any suggestion, how to solve this ?
