Re: [PATCH] xen/tmem: Don't use map_domain_page for long-life-time pages.

[George Dunlap <george.dunlap@eu.citrix.com>]

On 13/06/13 13:50, Konrad Rzeszutek Wilk wrote:
> When using tmem with Xen 4.3 (and debug build) we end up with:
>
> (XEN) Xen BUG at domain_page.c:143
> (XEN) ----[ Xen-4.3-unstable  x86_64  debug=y  Not tainted ]----
> (XEN) CPU:    3
> (XEN) RIP:    e008:[<ffff82c4c01606a7>] map_domain_page+0x61d/0x6e1
> ..
> (XEN) Xen call trace:
> (XEN)    [<ffff82c4c01606a7>] map_domain_page+0x61d/0x6e1
> (XEN)    [<ffff82c4c01373de>] cli_get_page+0x15e/0x17b
> (XEN)    [<ffff82c4c01377c4>] tmh_copy_from_client+0x150/0x284
> (XEN)    [<ffff82c4c0135929>] do_tmem_put+0x323/0x5c4
> (XEN)    [<ffff82c4c0136510>] do_tmem_op+0x5a0/0xbd0
> (XEN)    [<ffff82c4c022391b>] syscall_enter+0xeb/0x145
> (XEN)
>
> A bit of debugging revealed that the map_domain_page and unmap_domain_page
> are meant for short life-time mappings. And that those mappings are finite.
> In the 2 VCPU guest we only have 32 entries and once we have exhausted those
> we trigger the BUG_ON condition.
>
> The two functions - tmh_persistent_pool_page_[get,put] are used by the xmem_pool
> when xmem_pool_[alloc,free] are called. These xmem_pool_* function are wrapped
> in macro and functions - the entry points are via: tmem_malloc
> and tmem_page_alloc. In both cases the users are in the hypervisor and they
> do not seem to suffer from using the hypervisor virtual addresses.
>
> CC: Bob Liu <bob.liu@oracle.com>
> CC: Dan Magenheimer <dan.magenheimer@oracle.com>
> Suggested-by: Jan Beulich <jbeulich@suse.com>
> Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> ---
>   xen/common/tmem_xen.c |    5 ++---
>   1 files changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/xen/common/tmem_xen.c b/xen/common/tmem_xen.c
> index 3a1f3c9..736a8c3 100644
> --- a/xen/common/tmem_xen.c
> +++ b/xen/common/tmem_xen.c
> @@ -385,7 +385,7 @@ static void *tmh_persistent_pool_page_get(unsigned long size)
>       if ( (pi = _tmh_alloc_page_thispool(d)) == NULL )
>           return NULL;
>       ASSERT(IS_VALID_PAGE(pi));
> -    return __map_domain_page(pi);
> +    return page_to_virt(pi);

Did I understand correctly that the map_domain_page() was required on 
 >5TiB systems, presumably because of limited virtual address space?  In 
which case this code will fail on those systems?

If that is the case, then we need to have a way to make sure tmem cannot 
be enabled on such systems.

  -George