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

Re: [Xen-devel] [PATCH v4 5/7] mm: make MEMF_no_refcount pages safe to assign



> -----Original Message-----
> From: Jan Beulich <jbeulich@xxxxxxxx>
> Sent: 28 January 2020 15:23
> To: Durrant, Paul <pdurrant@xxxxxxxxxxxx>
> Cc: xen-devel@xxxxxxxxxxxxxxxxxxxx; Andrew Cooper
> <andrew.cooper3@xxxxxxxxxx>; George Dunlap <George.Dunlap@xxxxxxxxxxxxx>;
> Ian Jackson <ian.jackson@xxxxxxxxxxxxx>; Julien Grall <julien@xxxxxxx>;
> Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>; Stefano Stabellini
> <sstabellini@xxxxxxxxxx>; Wei Liu <wl@xxxxxxx>; Volodymyr Babchuk
> <Volodymyr_Babchuk@xxxxxxxx>; Roger Pau Monné <roger.pau@xxxxxxxxxx>
> Subject: Re: [PATCH v4 5/7] mm: make MEMF_no_refcount pages safe to assign
> 
> On 24.01.2020 16:31, Paul Durrant wrote:
> > Currently it is unsafe to assign a domheap page allocated with
> > MEMF_no_refcount to a domain because the domain't 'tot_pages' will not
> > be incremented, but will be decrement when the page is freed (since
> > free_domheap_pages() has no way of telling that the increment was
> skipped).
> >
> > This patch allocates a new 'count_info' bit for a PGC_no_refcount flag
> > which is then used to mark domheap pages allocated with
> MEMF_no_refcount.
> > This then allows free_domheap_pages() to skip decrementing tot_pages
> when
> > appropriate and hence makes the pages safe to assign.
> >
> > NOTE: The patch sets MEMF_no_refcount directly in alloc_domheap_pages()
> >       rather than in assign_pages() because the latter is called with
> >       MEMF_no_refcount by memory_exchange() as an optimization, to avoid
> >       too many calls to domain_adjust_tot_pages() (which acquires and
> >       releases the global 'heap_lock').
> 
> I don't think there were any optimization thoughts with this. The
> MEMF_no_refcount use is because otherwise for a domain with
> tot_pages == max_pages the assignment would fail.
> 

That would not be the case if the calls to steal_page() further up didn't pass 
MEMF_no_refcount (which would be the correct thing to do if not passing it to 
assign_pages(). I had originally considered doing that because I think it 
allows the somewhat complex error path after assign_pages() to be dropped. But 
avoiding thrashing the global lock seemed a good reason to leave 
memory_exchange() the way it is.

> > --- a/xen/common/page_alloc.c
> > +++ b/xen/common/page_alloc.c
> > @@ -460,6 +460,9 @@ unsigned long domain_adjust_tot_pages(struct domain
> *d, long pages)
> >  {
> >      long dom_before, dom_after, dom_claimed, sys_before, sys_after;
> >
> > +    if ( !pages )
> > +        goto out;
> 
> Unrelated change? Are there, in fact, any callers passing in 0?
> Oh, further down you add one which may do so, but then perhaps
> better to make the caller not call here (as is done e.g. in
> memory_exchange())?

I think it's preferable for domain_adjust_tot_pages() to handle zero gracefully.

> 
> > @@ -2331,11 +2331,20 @@ struct page_info *alloc_domheap_pages(
> >                                    memflags, d)) == NULL)) )
> >           return NULL;
> >
> > -    if ( d && !(memflags & MEMF_no_owner) &&
> > -         assign_pages(d, pg, order, memflags) )
> > +    if ( d && !(memflags & MEMF_no_owner) )
> >      {
> > -        free_heap_pages(pg, order, memflags & MEMF_no_scrub);
> > -        return NULL;
> > +        if ( assign_pages(d, pg, order, memflags) )
> > +        {
> > +            free_heap_pages(pg, order, memflags & MEMF_no_scrub);
> > +            return NULL;
> > +        }
> > +        if ( memflags & MEMF_no_refcount )
> > +        {
> > +            unsigned long i;
> > +
> > +            for ( i = 0; i < (1 << order); i++ )
> > +                pg[i].count_info |= PGC_no_refcount;
> > +        }
> 
> I would seem to me that this needs doing the other way around:
> First set PGC_no_refcount, then assign_pages(). After all, the
> moment assign_pages() drops its lock, the domain could also
> decide to get rid of (some of) the pages again.

True. Yes, this needs to be swapped.

> For this (and
> also to slightly simplify things in free_domheap_pages())
> perhaps it would be better not to add that ASSERT() to
> free_heap_pages(). The function shouldn't really be concerned
> of any refcounting, and hence could as well be ignorant to
> PGC_no_refcount being set on a page.
> 

Not sure I understand here. What would you like to see free_heap_pages() assert?

> > @@ -2368,24 +2377,32 @@ void free_domheap_pages(struct page_info *pg,
> unsigned int order)
> >
> >          if ( likely(d) && likely(d != dom_cow) )
> >          {
> > +            long pages = 0;
> > +
> >              /* NB. May recursively lock from relinquish_memory(). */
> >              spin_lock_recursive(&d->page_alloc_lock);
> >
> >              for ( i = 0; i < (1 << order); i++ )
> >              {
> > +                unsigned long count_info = pg[i].count_info;
> > +
> >                  if ( pg[i].u.inuse.type_info & PGT_count_mask )
> >                  {
> >                      printk(XENLOG_ERR
> >                             "pg[%u] MFN %"PRI_mfn" c=%#lx o=%u v=%#lx
> t=%#x\n",
> >                             i, mfn_x(page_to_mfn(pg + i)),
> > -                           pg[i].count_info, pg[i].v.free.order,
> > +                           count_info, pg[i].v.free.order,
> >                             pg[i].u.free.val, pg[i].tlbflush_timestamp);
> >                      BUG();
> >                  }
> >                  arch_free_heap_page(d, &pg[i]);
> > +                if ( count_info & PGC_no_refcount )
> > +                    pg[i].count_info &= ~PGC_no_refcount;
> > +                else
> > +                    pages--;
> 
> Not only to reduce code churn, may I recommend to avoid introducing
> the local variable? There's no strict rule preventing
> arch_free_heap_page() from possibly playing with the field you
> latch up front.

Ok.

  Paul

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

 


Rackspace

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