On Mon, Apr 23, 2018 at 7:29 PM, Jan Kara <jack@xxxxxxx> wrote:
> On Sun 22-04-18 19:25:05, Matthew Wilcox wrote:
>> On Mon, Apr 23, 2018 at 01:09:48AM +0200, Jan Kara wrote:
>> > > -int vm_insert_mixed_mkwrite(struct vm_area_struct *vma, unsigned long addr,
>> > > - pfn_t pfn)
>> > > +vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
>> > > + unsigned long addr, pfn_t pfn)
>> > > {
>> > > - return __vm_insert_mixed(vma, addr, pfn, true);
>> > > + int err;
>> > > +
>> > > + err = __vm_insert_mixed(vma, addr, pfn, true);
>> > > + if (err == -ENOMEM)
>> > > + return VM_FAULT_OOM;
>> > > + if (err < 0 && err != -EBUSY)
>> > > + return VM_FAULT_SIGBUS;
>> > > + return VM_FAULT_NOPAGE;
>> > > }
>> > > -EXPORT_SYMBOL(vm_insert_mixed_mkwrite);
>> > > +EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
>> >
>> > So are we sure that all the callers of this function (and also of
>> > vmf_insert_mixed()) are OK with EBUSY? Because especially in the
>> > vmf_insert_mixed() case other page than the caller provided is in page
>> > tables and thus possibly the caller needs to do some error recovery (such
>> > as drop page refcount) in such case...
>>
>> I went through all the users and didn't find any that did anything
>> with -EBUSY other than turn it into VM_FAULT_NOPAGE. I agree that it's
>> possible that there might have been someone who wanted to do that, but
>> we tend to rely on mapcount (through rmap) rather than refcount (ie we
>> use refcount to mean the number of kernel references to the page and then
>> use mapcount for the number of times it's mapped into a process' address
>> space). All the drivers I audited would allocagte the page first, store
>> it in their own data structures, then try to insert it into the virtual
>> address space. So an EBUSY always meant "the same page was inserted".
>>
>> If we did want to support "This happened already" in the future, we
>> could define a VM_FAULT flag for that.
>
> OK, fair enough and thanks for doing an audit! So possibly just add a
> comment above vmf_insert_mixed() and vmf_insert_mixed_mkwrite() like:
>
> /*
> * If the insertion of PTE failed because someone else already added a
> * different entry in the mean time, we treat that as success as we assume
> * the same entry was actually inserted.
> */
>
> After that feel free to add:
>
> Reviewed-by: Jan Kara <jack@xxxxxxx>
>
> to the patch.
>
Thanks , will add this in change log and send v4.
