Re: [PATCH] xen-mapcache: avoid a race on memory map while using MAP_FIXED

["Roger Pau Monné via" <qemu-devel@nongnu.org>]

On Tue, Apr 20, 2021 at 04:35:02AM +0100, Igor Druzhinin wrote:
> When we're replacing the existing mapping there is possibility of a race
> on memory map with other threads doing mmap operations - the address being
> unmapped/re-mapped could be occupied by another thread in between.
> 
> Linux mmap man page recommends keeping the existing mappings in place to
> reserve the place and instead utilize the fact that the next mmap operation
> with MAP_FIXED flag passed will implicitly destroy the existing mappings
> behind the chosen address. This behavior is guaranteed by POSIX / BSD and
> therefore is portable.
> 
> Note that it wouldn't make the replacement atomic for parallel accesses to
> the replaced region - those might still fail with SIGBUS due to
> xenforeignmemory_map not being atomic. So we're still not expecting those.
> 
> Tested-by: Anthony PERARD <anthony.perard@citrix.com>
> Signed-off-by: Igor Druzhinin <igor.druzhinin@citrix.com>

Should we add a 'Fixes: 759235653de ...' or similar tag here?

> ---
>  hw/i386/xen/xen-mapcache.c | 15 ++++++++++++++-
>  1 file changed, 14 insertions(+), 1 deletion(-)
> 
> diff --git a/hw/i386/xen/xen-mapcache.c b/hw/i386/xen/xen-mapcache.c
> index 5b120ed..e82b7dc 100644
> --- a/hw/i386/xen/xen-mapcache.c
> +++ b/hw/i386/xen/xen-mapcache.c
> @@ -171,7 +171,20 @@ static void xen_remap_bucket(MapCacheEntry *entry,
>          if (!(entry->flags & XEN_MAPCACHE_ENTRY_DUMMY)) {
>              ram_block_notify_remove(entry->vaddr_base, entry->size);
>          }
> -        if (munmap(entry->vaddr_base, entry->size) != 0) {
> +
> +        /*
> +         * If an entry is being replaced by another mapping and we're using
> +         * MAP_FIXED flag for it - there is possibility of a race for vaddr
> +         * address with another thread doing an mmap call itself
> +         * (see man 2 mmap). To avoid that we skip explicit unmapping here
> +         * and allow the kernel to destroy the previous mappings by replacing
> +         * them in mmap call later.
> +         *
> +         * Non-identical replacements are not allowed therefore.
> +         */
> +        assert(!vaddr || (entry->vaddr_base == vaddr && entry->size == size));
> +
> +        if (!vaddr && munmap(entry->vaddr_base, entry->size) != 0) {

Oh, so remappings of foreign addresses must always use the same
virtual address space, I somehow assumed that you could remap an
existing foreign map (or dummy mapping) into a different virtual
address if you so wanted.

Thanks, Roger.