Re: [PATCH] mmap.2: document new MAP_FIXED_SAFE flag

From: John Hubbard
Date: Wed Nov 29 2017 - 22:16:49 EST


On 11/29/2017 06:45 AM, Michal Hocko wrote:
> From: Michal Hocko <mhocko@xxxxxxxx>
>
> 4.16+ kernels offer a new MAP_FIXED_SAFE flag which allows to atomicaly

"allows the caller to atomically"

, if you care about polishing the commit message...see the real review,
below. :)

> probe for a given address range.
>
> Signed-off-by: Michal Hocko <mhocko@xxxxxxxx>
> ---
> man2/mmap.2 | 18 ++++++++++++++++++
> 1 file changed, 18 insertions(+)
>
> diff --git a/man2/mmap.2 b/man2/mmap.2
> index 385f3bfd5393..622a7000de83 100644
> --- a/man2/mmap.2
> +++ b/man2/mmap.2
> @@ -225,6 +225,18 @@ will fail.
> Because requiring a fixed address for a mapping is less portable,
> the use of this option is discouraged.
> .TP
> +.B MAP_FIXED_SAFE (since 4.16)
> +Similar to MAP_FIXED wrt. to the
> +.I
> +addr
> +enforcement except it never clobbers a colliding mapped range and rather fail with
> +.B EEXIST
> +in such a case. This flag can therefore be used as a safe and atomic probe for the
> +the specific address range. Please note that older kernels which do not recognize
> +this flag can fallback to the hint based implementation and map to a different
> +location. Any backward compatible software should therefore check the returning
> +address with the given one.
> +.TP
> .B MAP_GROWSDOWN
> This flag is used for stacks.
> It indicates to the kernel virtual memory system that the mapping

Hi Michal,

I've taken the liberty of mostly rewriting this part, in order to more closely
match the existing paragraphs; to fix minor typos; and to attempt to slightly
clarify the paragraph.

+.BR MAP_FIXED_SAFE " (since Linux 4.16)"
+Similar to MAP_FIXED with respect to the
+.I
+addr
+enforcement, but different in that MAP_FIXED_SAFE never clobbers a pre-existing
+mapped range. If the requested range would collide with an existing
+mapping, then this call fails with
+.B EEXIST.
+This flag can therefore be used as a way to atomically (with respect to other
+threads) attempt to map an address range: one thread will succeed; all others
+will report failure. Please note that older kernels which do not recognize this
+flag will typically (upon detecting a collision with a pre-existing mapping)
+fall back a "non-MAP_FIXED" type of behavior: they will return an address that
+is different than the requested one. Therefore, backward-compatible software
+should check the returned address against the requested address.
+.TP

(I'm ignoring the naming, because there is another thread about that,
so please just the above as "MAP_FIXED_whatever-is-chosen".)

> @@ -449,6 +461,12 @@ is not a valid file descriptor (and
> .B MAP_ANONYMOUS
> was not set).
> .TP
> +.B EEXIST
> +range covered by
> +.IR addr ,

nit: trailing space on the above line.

> +.IR length
> +is clashing with an existing mapping.
> +.TP
> .B EINVAL
> We don't like
> .IR addr ,
>

One other thing: reading through mmap.2, I now want to add this as well:

diff --git a/man2/mmap.2 b/man2/mmap.2
index 622a7000d..780cad6d9 100644
--- a/man2/mmap.2
+++ b/man2/mmap.2
@@ -222,20 +222,25 @@ part of the existing mapping(s) will be discarded.
If the specified address cannot be used,
.BR mmap ()
will fail.
-Because requiring a fixed address for a mapping is less portable,
-the use of this option is discouraged.
+Software that aspires to be as portable as possible should use this option with
+care, keeping in mind that different kernels and C libraries may set up quite
+different mapping ranges.


...because that advice is just wrong (it presumes that "less portable" ==
"must be discouraged").

Should I send out a separate patch for that, or is it better to glom it together
with this one?

thanks,
John Hubbard
NVIDIA