[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH v3] altp2m: Allow the hostp2m to be shared



On Wed, May 25, 2016 at 12:25 PM, George Dunlap
<george.dunlap@xxxxxxxxxx> wrote:
> On Wed, May 25, 2016 at 5:58 PM, Tamas K Lengyel <tamas@xxxxxxxxxxxxx> wrote:
>> On Wed, May 25, 2016 at 10:08 AM, George Dunlap
>> <george.dunlap@xxxxxxxxxx> wrote:
>>> On Wed, May 25, 2016 at 4:31 PM, Tamas K Lengyel <tamas@xxxxxxxxxxxxx> 
>>> wrote:
>>>>
>>>> On May 25, 2016 05:27, "George Dunlap" <george.dunlap@xxxxxxxxxx> wrote:
>>>>>
>>>>> On Fri, Apr 29, 2016 at 6:42 PM, Tamas K Lengyel <tamas@xxxxxxxxxxxxx>
>>>>> wrote:
>>>>> > Don't propagate altp2m changes from ept_set_entry for memshare as
>>>>> > memshare
>>>>> > already has the lock. We call altp2m propagate changes once memshare
>>>>> > successfully finishes. Allow the hostp2m entries to be of type
>>>>> > p2m_ram_shared when applying mem_access. Also, do not trigger PoD for
>>>>> > hostp2m
>>>>> > when setting altp2m mem_access to be in-line with non-altp2m mem_access
>>>>> > path.
>>>>>
>>>>> Hey Tamas,
>>>>>
>>>>> Sorry for the long delay in getting back to you on this.
>>>>
>>>> No problem, thanks for taking a closer look!
>>>>
>>>>>
>>>>> So the main issue here (correct me if I'm wrong) is the locking
>>>>> discipline: namely, men_sharing_share_pages():
>>>>> - Grabs the hostp2m lock
>>>>> - Grabs the appropriate domain memsharing locks
>>>>> - Calls set_shared_p2m_entry(), which ends up calling ept_set_entry(),
>>>>> which (when altp2m is active) grabs the altp2mlist and altp2m locks.
>>>>>
>>>>> This causes an ASSERT(), since the altp2mlist lock is ahead of the
>>>>> memsharing locks in the list.
>>>>>
>>>>> But having taken a closer look at the code, I'm not sure the change is
>>>>> quite correct.  Please correct me if I've misread something:
>>>>>
>>>>> mem_sharing_share_pages() is passed two <domain,gfn> pairs -- the
>>>>> <sd,sgfn> (which I assume stands for "shared gfn") and <cd,cgfn>
>>>>> (which I assume stands for "copy"); and it
>>>>
>>>> Here s/c stands for source/client.
>>>>
>>>>> 1) Looks up smfn and cmfn, which back sgfn and cmfn respectively
>>>>> 2) Looks up cmfn, which backs cgfn then replaces all gfn entries which
>>>>> point to cmfn with smfn (updating accounting as appropriate)
>>>>
>>>> Hm, I might have missed that. Where does it do the lookup for all other
>>>> cgfns backed by this cmfn?
>>>
>>> I was looking at the loop in the middle of the function:
>>>
>>> while ( (gfn = rmap_iterate(cpage, &ri)) != NULL) {
>>>  ...
>>> }
>>>
>>> I haven't chased it down, but it looks like this walks the reverse map
>>> of all gfns which map cpage; and for each such gfn it finds it:
>>> * removes the cpage -> gfn rmap
>>> * Adds an spage -> gfn map
>>> * Reduces the type count of cpage
>>> * Sets the p2m entry for that gfn to the smfn (rather than cmfn).
>>>
>>> Obviously the common case is that the number of mappings is exactly 1;
>>> but we need to either ensure that this is always true, or we need to
>>> handle the case where it's not true. :-)
>>>
>>>>> But this change will only call p2m_altp2m_propagate_change() for the
>>>>> original cgfn -- any other gfns which are backed by cmfn will not have
>>>>> the corresponding altp2m entries propagated properly.
>>>>
>>>> Right, if there is some other place where it does sharing in the back we
>>>> would have to propagate that change.
>>>>
>>>>> This sort of mistake is easy to make, which is why I think we should
>>>>> try to always update the altp2ms in ept_set_entry() if we can, to
>>>>> minimize the opportunity for making this sort of mistake.
>>>>>
>>>>> Is there ever a reason to grab the altp2m lock and *then* grab the
>>>>> sharing lock?  Could we just move the sharing lock up between the p2m
>>>>> lock and the altp2mlist lock instead?
>>>>>
>>>>
>>>> I can't think of a scenario where we would get to sharing from altp2m with
>>>> altp2m locking first. Not sure what you mean by moving the sharing lock up
>>>> though. The problem is that sharing already has the lock by the time altp2m
>>>> tries to lock, so we could pass that info down to make altp2m aware it 
>>>> needs
>>>> no locking. It would require extending a bunch of functions though with an
>>>> extra input that is barely ever used..
>>>
>>> If you have altp2m there are three locks.  There's one p2m lock for
>>> the "host" p2m (that is, Xen's idea of what the mapping should look
>>> like).  Then there's the altp2mlist lock, which protects the *list* of
>>> altp2ms; then each altp2m itself has its own lock.  These are defined
>>> in mm-lock.h and  must be grabbed in that order: p2m before
>>> altp2mlist, altp2mlist before altp2m.
>>>
>>> I assume that the memsharing code is grabbing the hostp2m lock (it
>>> should be anyway), then grabbing the memsharing locks. This is allowed
>>> because the memsharing locks are defined after the p2m lock in
>>> mm-lock.h.  But then when updating the p2m entry, if you have an
>>> altp2m active, it then tries to grab the altp2mlist lock so it can
>>> iterate over the altp2ms.  Since the altp2mlist lock is *before* the
>>> sharing lock in mm-lock.h, this triggers an assert.
>>>
>>> Is that not what your issue is?
>>
>> Ahh, I see! Let me give that a try - TBH this locking order
>> enforcement based on position in mm-lock.h was not entirely clear to
>> me =)
>
> Indeed, it is a bit strange, but if you see the number of locks that
> must be ordered properly to avoid deadlock (what, 8 or so?) it's
> really only the sane way to make sure things are kept straight.
>
> The original implementation actually uses line numbers from mm-lock.h
> to declare the order, but I *think* there recently went in a patch to
> change those to explicit enumerations, to make xsplice patches easier.
>

Alright, moving up the locks in mm-locks.h does resolve the problem
and we can keep the altp2m propagate in ept_set_entry as well! The
xen-access altp2m tests pass as well on a fully memshared domain. Will
send the new patch shortly.

Thanks,
Tamas

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.