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

[Xen-devel] Re: Some problems about xenpaging



On Fri, Sep 23, Hongkaixing wrote:

> Hi, Olaf

Hello Hongkaixing,

thanks for the feedback and the patches!

> we have tested the xenpaging feature and found some problems.
> (1) the test case like this : when we start a VM with POD enable, the 
> xenpaging is started at the same time.
>     this case will cause many problems ,finally, we fixed the BUG, the patch 
> is attached below.

The changes for p2m_pod_decrease_reservation() and
p2m_mem_paging_nominate() look ok to me, but I will leave the final word
to Tim and George.

If I understand the current logic in p2m_mem_paging_nominate()
correctly, if a gfn is p2m_populate_on_demand, then gfn_to_mfn() will
turn the gfn into p2m_ram_rw and let xenpaging write it to disk. That
does not look correct. With your change the gfn will not be nominated
because its not backed by RAM anyway.

The locking in the p2m_mem_paging_nominate(), p2m_mem_paging_evict() and
p2m_mem_paging_populate() can be improved. Using gfn_to_mfn_query()
inside the p2m_lock looks correct to me. Even p2m_mem_paging_resume()
could use the query variant and forward the type to p2m_ram_rw only if
the p2mt was p2m_ram_paging_in_start.

Please send proper patches for xen-unstable as described in the
MAINTAINERS file.

> (2) there is a very serious problem. we have observed many VM crash examples, 
> the error code is not always the same. 
>      we guess there exists conflict between xenpaging and memory mapping. 
>      for instance, if the Dom0 map the DomU's memory to its own space , then 
> the memory of DomainU is paged out, when the Domain0 access this memory area, 
> a panic is caused.
>      And I really do not know whether the qemu device can perceive the memory 
> modified by xenpaging? 

The issue is that gfn_to_mfn() returns the p2m_ram_paging* types and
expects the caller to retry until the gfn is accessible, instead of
calling p2m_mem_paging_populate() and going to sleep until the page is
available.
Thats still on my list of things to do.


Olaf


> here is the patch to solve the pod problem
> 1) fix the p2m_pod_decrease_reservation() function, take care of the paging 
> type
> --- ./origin/xen/arch/x86/mm/p2m.c      2011-09-05 20:39:30.000000000 +0800
> +++ ./b/xen/arch/x86/mm/p2m.c   2011-09-23 23:46:19.000000000 +0800
> @@ -675,6 +675,23 @@
>              BUG_ON(p2md->pod.entry_count < 0);
>              pod--;
>          }
> +        else if ( steal_for_cache && p2m_is_paging(t) )
> +        {
> +            struct page_info *page;
> +            /* alloc a new page to compensate the pod list */
> +            page = alloc_domheap_page(d, 0);
> +            if ( unlikely(page == NULL) )
> +            {
> +                goto out_entry_check;
> +            }
> +            set_p2m_entry(d, gpfn + i, _mfn(INVALID_MFN), 0, p2m_invalid);
> +            p2m_mem_paging_drop_page(d, gpfn+i);
> +            p2m_pod_cache_add(d, page, 0);
> +            steal_for_cache =  ( p2md->pod.entry_count > p2md->pod.count );
> +            nonpod--;
> +            ram--;
> +        }
> +        /* for other ram types */
>          else if ( steal_for_cache && p2m_is_ram(t) )
>          {
>              struct page_info *page;
> 
> 2) fix the race between POD and xenpaging
>    situation can be described as the follow
>    
>    xenpaging                  POD
>   
>  mfn = gfn_to_mfn()         mfn = gfn_to_mfn()
>    check p2m type           check p2mt
>                             p2m_lock
>                             change p2m type
>                             p2m_unlock
>                             add to pod list
>     p2m_lock()
>     change p2m type
>     p2m_unlock
>     put_page()
> 
> the result is a page is added to the pod list,and then it's removed from the 
> list by put_page.
> my suggestion is to extend the range of the p2m lock to contain the p2m check.
> 
> in  p2m_mem_paging_nominate() function
> @@ -2532,7 +2561,8 @@
>      mfn_t mfn;
>      int ret;
>  
> -    mfn = gfn_to_mfn(d, gfn, &p2mt);
> +    p2m_lock(d->arch.p2m);
> +    mfn = gfn_to_mfn_query(d, gfn, &p2mt);
>  
>      /* Check if mfn is valid */
>      ret = -EINVAL;
> @@ -2580,13 +2610,12 @@
>          goto out;
>  
>      /* Fix p2m entry */
> -    p2m_lock(d->arch.p2m);
>      set_p2m_entry(d, gfn, mfn, 0, p2m_ram_paging_out);
> -    p2m_unlock(d->arch.p2m);
>  
>      ret = 0;
>  
>   out:
> +    p2m_unlock(d->arch.p2m);
>      return ret;
>  }
> in 
> @@ -2595,34 +2624,39 @@
>      struct page_info *page;
>      p2m_type_t p2mt;
>      mfn_t mfn;
> -
> +    int ret;
> +    p2m_lock(d->arch.p2m);
>      /* Get mfn */
> -    mfn = gfn_to_mfn(d, gfn, &p2mt);
> +    mfn = gfn_to_mfn_query(d, gfn, &p2mt);
> +    
> +    ret = -EINVAL;
>      if ( unlikely(!mfn_valid(mfn)) )
> -        return -EINVAL;
> +        goto out;
>  
>      if (p2mt != p2m_ram_paging_out)
>       {
>           printk("p2m_mem_paging_evict type %d\n", p2mt);
> -               return -EINVAL;
> +         goto out;
>       }
>      /* Get the page so it doesn't get modified under Xen's feet */
>      page = mfn_to_page(mfn);
>      if ( unlikely(!get_page(page, d)) )
> -        return -EINVAL;
> +        goto out;
>  
>      /* Decrement guest domain's ref count of the page */
>      if ( test_and_clear_bit(_PGC_allocated, &page->count_info) )
>          put_page(page);
>  
>      /* Remove mapping from p2m table */
> -    p2m_lock(d->arch.p2m);
>      set_p2m_entry(d, gfn, _mfn(PAGING_MFN), 0, p2m_ram_paged);
> -    p2m_unlock(d->arch.p2m);
>  
>      /* Put the page back so it gets freed */
>      put_page(page);
> +    
> +    ret = 0;
>  
> +out:
> +    p2m_unlock(d->arch.p2m);
>      return 0;
>  }
> 
> 3) fix the vmx_load_pdptrs() function in  vmx.c
> in this situation the page directory table is paged out.
> Although using mdelay() is a bad idea, it's better than making the xen crash  
> 
> void vmx_load_pdptrs(struct vcpu *v)
> {
>     unsigned long cr3 = v->arch.hvm_vcpu.guest_cr[3], mfn;
>     uint64_t *guest_pdptrs;
>     p2m_type_t p2mt;
>     char *p;
>     unsigned int try_count = 0;
>     /* EPT needs to load PDPTRS into VMCS for PAE. */
>     if ( !hvm_pae_enabled(v) || (v->arch.hvm_vcpu.guest_efer & EFER_LMA) )
>         return;
>     if ( cr3 & 0x1fUL )
>         goto crash;
>     mfn = mfn_x(gfn_to_mfn(v->domain, cr3 >> PAGE_SHIFT, &p2mt));
>     if ( !p2m_is_ram(p2mt) )
>         goto crash;
> +  if( p2m_is_paging(p2mt))
> +  {
> +        p2m_mem_paging_populate(v->domain, cr3 >> PAGE_SHIFT);
> +        do
> +        {
> +             mdelay(1);
> +             try_count ++;
> +             if ( try_count > 65535 )
> +             {
> +                 goto crash;
> +             }
> +             mfn = mfn_x(gfn_to_mfn(v->domain, cr3 >> PAGE_SHIFT, &p2mt));
> +        }while( !mfn_valid(mfn));
> +    }
> +
>     p = map_domain_page(mfn);

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


 


Rackspace

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