Re: [Xen-devel] [PATCH v3 1/4] x86/mem_sharing: reorder when pages are unlocked and released

[George Dunlap <george.dunlap@xxxxxxxxxx>]

On 4/29/19 3:49 PM, Andrew Cooper wrote:
> On 29/04/2019 15:46, George Dunlap wrote:
>> On 4/29/19 3:32 PM, George Dunlap wrote:
>>> On 4/26/19 6:21 PM, Tamas K Lengyel wrote:
>>>> Calling _put_page_type while also holding the page_lock
>>>> for that page can cause a deadlock.
>>>>
>>>> Signed-off-by: Tamas K Lengyel <tamas@xxxxxxxxxxxxx>
>>>> Cc: Jan Beulich <jbeulich@xxxxxxxx>
>>>> Cc: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
>>>> Cc: George Dunlap <george.dunlap@xxxxxxxxxxxxx>
>>>> Cc: Wei Liu <wei.liu2@xxxxxxxxxx>
>>>> Cc: Roger Pau Monne <roger.pau@xxxxxxxxxx>
>>>> ---
>>>> v3: simplified patch by keeping the additional references already in-place
>>>> ---
>>>>  xen/arch/x86/mm/mem_sharing.c | 19 ++++++++++---------
>>>>  1 file changed, 10 insertions(+), 9 deletions(-)
>>>>
>>>> diff --git a/xen/arch/x86/mm/mem_sharing.c b/xen/arch/x86/mm/mem_sharing.c
>>>> index dfc279d371..e2f74ac770 100644
>>>> --- a/xen/arch/x86/mm/mem_sharing.c
>>>> +++ b/xen/arch/x86/mm/mem_sharing.c
>>>> @@ -648,10 +648,6 @@ static int page_make_private(struct domain *d, struct 
>>>> page_info *page)
>>>>          return -EBUSY;
>>>>      }
>>>>  
>>>> -    /* We can only change the type if count is one */
>>>> -    /* Because we are locking pages individually, we need to drop
>>>> -     * the lock here, while the page is typed. We cannot risk the 
>>>> -     * race of page_unlock and then put_page_type. */
>>>>      expected_type = (PGT_shared_page | PGT_validated | PGT_locked | 2);
>>>>      if ( page->u.inuse.type_info != expected_type )
>>>>      {
>>>> @@ -660,12 +656,12 @@ static int page_make_private(struct domain *d, 
>>>> struct page_info *page)
>>>>          return -EEXIST;
>>>>      }
>>>>  
>>>> -    /* Drop the final typecount */
>>>> -    put_page_and_type(page);
>>>> -
>>>>      /* Now that we've dropped the type, we can unlock */
>>>>      mem_sharing_page_unlock(page);
>>>>  
>>>> +    /* Drop the final typecount */
>>>> +    put_page_and_type(page);
>>>> +
>>>>      /* Change the owner */
>>>>      ASSERT(page_get_owner(page) == dom_cow);
>>>>      page_set_owner(page, d);
>>>> @@ -900,6 +896,7 @@ static int share_pages(struct domain *sd, gfn_t sgfn, 
>>>> shr_handle_t sh,
>>>>      p2m_type_t smfn_type, cmfn_type;
>>>>      struct two_gfns tg;
>>>>      struct rmap_iterator ri;
>>>> +    unsigned long put_count = 0;
>>>>  
>>>>      get_two_gfns(sd, sgfn, &smfn_type, NULL, &smfn,
>>>>                   cd, cgfn, &cmfn_type, NULL, &cmfn, 0, &tg);
>>>> @@ -984,7 +981,7 @@ static int share_pages(struct domain *sd, gfn_t sgfn, 
>>>> shr_handle_t sh,
>>>>           * Don't change the type of rmap for the client page. */
>>>>          rmap_del(gfn, cpage, 0);
>>>>          rmap_add(gfn, spage);
>>>> -        put_page_and_type(cpage);
>>>> +        put_count++;
>>>>          d = get_domain_by_id(gfn->domain);
>>>>          BUG_ON(!d);
>>>>          BUG_ON(set_shared_p2m_entry(d, gfn->gfn, smfn));
>>>> @@ -999,6 +996,10 @@ static int share_pages(struct domain *sd, gfn_t sgfn, 
>>>> shr_handle_t sh,
>>>>      mem_sharing_page_unlock(secondpg);
>>>>      mem_sharing_page_unlock(firstpg);
>>>>  
>>>> +    BUG_ON(!put_count);
>>>> +    while ( put_count-- )
>>>> +        put_page_and_type(cpage);
>>>> +
>>>>      /* Free the client page */
>>>>      if(test_and_clear_bit(_PGC_allocated, &cpage->count_info))
>>>>          put_page(cpage);
>>>> @@ -1167,8 +1168,8 @@ int __mem_sharing_unshare_page(struct domain *d,
>>>>      {
>>>>          if ( !last_gfn )
>>>>              mem_sharing_gfn_destroy(page, d, gfn_info);
>>>> -        put_page_and_type(page);
>>>>          mem_sharing_page_unlock(page);
>>>> +        put_page_and_type(page);
>>>>          if ( last_gfn )
>>>>          {
>>>>              if ( !get_page(page, dom_cow) )
>>> ...Probably should have mentioned that this needs to be applied after
>>> your other patch. :-)
>> Hmm -- actually, the base appears to be a non-publicly-available tree
>> (Andy's private x86-next).
> 
> Its perfectly public
> 
> http://xenbits.xen.org/gitweb/?p=people/andrewcoop/xen.git;a=shortlog;h=refs/heads/x86-next

OK, didn't realize this.

> 
> and has been in effect during commit freezes for the past several years
> worth of releases.
> 
> It is fine to post a tree based on something other than xen.git/staging
> so long as the cover letter identifies where the series is based.  A lot
> of development already occurs in this way on xen-devel.

I agree it's fine when a URL is provided.

 -George

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel