Re: [PATCH bpf-next v5 11/11] bpf: Document new atomic instructions

From: Yonghong Song
Date: Wed Dec 16 2020 - 10:27:10 EST




On 12/16/20 3:44 AM, Brendan Jackman wrote:
On Wed, 16 Dec 2020 at 08:08, Yonghong Song <yhs@xxxxxx> wrote:



On 12/15/20 4:18 AM, Brendan Jackman wrote:
Document new atomic instructions.

Signed-off-by: Brendan Jackman <jackmanb@xxxxxxxxxx>

Ack with minor comments below.

Acked-by: Yonghong Song <yhs@xxxxxx>

---
Documentation/networking/filter.rst | 26 ++++++++++++++++++++++++++
1 file changed, 26 insertions(+)

diff --git a/Documentation/networking/filter.rst b/Documentation/networking/filter.rst
index 1583d59d806d..26d508a5e038 100644
--- a/Documentation/networking/filter.rst
+++ b/Documentation/networking/filter.rst
@@ -1053,6 +1053,32 @@ encoding.
.imm = BPF_ADD, .code = BPF_ATOMIC | BPF_W | BPF_STX: lock xadd *(u32 *)(dst_reg + off16) += src_reg
.imm = BPF_ADD, .code = BPF_ATOMIC | BPF_DW | BPF_STX: lock xadd *(u64 *)(dst_reg + off16) += src_reg

+The basic atomic operations supported (from architecture v4 onwards) are:

Remove "(from architecture v4 onwards)".

Oops, thanks.

+
+ BPF_ADD
+ BPF_AND
+ BPF_OR
+ BPF_XOR
+
+Each having equivalent semantics with the ``BPF_ADD`` example, that is: the
+memory location addresed by ``dst_reg + off`` is atomically modified, with
+``src_reg`` as the other operand. If the ``BPF_FETCH`` flag is set in the
+immediate, then these operations also overwrite ``src_reg`` with the
+value that was in memory before it was modified.
+
+The more special operations are:
+
+ BPF_XCHG
+
+This atomically exchanges ``src_reg`` with the value addressed by ``dst_reg +
+off``.
+
+ BPF_CMPXCHG
+
+This atomically compares the value addressed by ``dst_reg + off`` with
+``R0``. If they match it is replaced with ``src_reg``, The value that was there
+before is loaded back to ``R0``.
+
Note that 1 and 2 byte atomic operations are not supported.

Adding something like below.

Except xadd for legacy reason, all other 4 byte atomic operations
require alu32 mode.
The alu32 mode can be enabled with clang flags "-Xclang -target-feature
-Xclang +alu32" or "-mcpu=v3". The cpu version 3 has alu32 mode on by
default.

Thanks, I've written it as:

Except ``BPF_ADD`` _without_ ``BPF_FETCH`` (for legacy reasons), all 4
byte atomic operations require alu32 mode. Clang enables this mode by
default in architecture v3 (``-mcpu=v3``). For older versions it can
be enabled with ``-Xclang -target-feature -Xclang +alu32``.

Sounds good. thanks!



You may encounter BPF_XADD - this is a legacy name for BPF_ATOMIC, referring to