[RFC] Introduce verbs API for multi-packet work request

This RFC patch introduces a libibverbs API for receiving multiple packets
on a single work request (aka: "MP WR").

Traditional verbs work request maps a single WR to a single received
message. The entire WR buffer is consumed, regardless of the ingress
message size. The WR has a single completion, which reports the message
length and some additional flags and/or values.
Some limitations of the traditional WR include:
1. If the ingress message is much smaller than the WR buffer, the buffer
   memory is not well utilized.
2. If the ingress message is larger than the WR buffer, the QP might
   transition to error or the message might be dropped.

The motivation for a MP WR is to enable:
1. High efficiency of receive buffer memory utilization by:
 a. Allowing multiple ingress packets to be written in a single WR
    buffer, into different memory parts of the entire WR buffer.
    Each packet start offset in the buffer will be according to a packet
    alignment size defined by user. The packet alignment size can be equal
    to cache line size, a page size, or other desired application logic
    values. A packet can be delivered while consuming multiple aligned
    memory segments. This allows multiple different packet sizes to be
    received within a single MP WR buffer.
    Work completions are generated for the received data similar to when
    completions are generated for a traditional work request.
 b. Allowing a FIRST and MIDDLE packets to be writen to WR memory in a
    dis-contiguous fashion. This allows very large transfers to MP WRs QPs
    without having to increase the WR buffer size to the largest possible
    message length.
    After any FIRST or MIDDLE packet the hardware can write a CQE with the
    'MORE_IN_MSG' flag to indicate it is not the end of the logical transfer.
    The entire message, built from the multi-packet completions, can span
    over multiple work request.
2. Improved device PCI utilization: device PCI fetch of a single WR entry
   can handle multiple packets, rather than having to fetch WR entires for
   each received packet as is traditionally required.

Definitions of verbs MP WR:
- MP WR capability is supported by a device when struct ibv_mp_wr_caps
  values are greater than zero bytes, both max_wr_buffer_sz and
  max_packet_align_sz.
- MP WR receive queue can be defined for a QP, SRQ, or WQ (of type RQ).
- A MP WR is defined with struct ibv_mp_wr_caps, by its WR buffer size
  and the packet alignment size (both in bytes). User sets the requested
  values during object creation, and the returned values are the actual
  values used by provider library (equal or greater from user
  requested). All post receive must have same size WR buffers, matching
  the buffer size specified during creation.
- A MP WR requires additional completion flags. For this, the QP, SRQ or
  WQ must be created with an extended CQ using ibv_create_cq_ex() with the
  IBV_WC_EX_WITH_MP_WR flag.
- The reported MP WR completion flags include:
  a. IBV_WC_MP_WR_MORE_IN_MSG: is reported by a multi-packet work
     request that has more packet completions expected in this message for
     this qp_num. This is set after receiving a FIRST or MIDDLE packet
     into the WR.
  b. IBV_WC_MP_WR_CONSUMED: is reported once an entire multi-packet
     work request buffer is consumed, so that user knows the device releases
     ownership of that wr_id and buffer. IBV_WC_RECV_NOP opcode is reported
     in WC for a 'consumed' WR that is without a received message.
- The byte offset in the work request buffer for the start of a specific logical
  transfer is report by ibv_wc_read_mp_wr_offset(). This may be the start of a
  complete full packet, or the start of a FIRST, MIDDLE or LAST segment.

Application Notes:
- When using the MP WR, multiple packets can be reported for each wr_id. In
this case the wr_id reflects the MP WR buffer submitted to the hardware by
the application can be repeated for multiple completions. Application's will
need to use different logic around with wr_id to support MP WR.
- It's the user's responsibility to reconstruct the full packet if it was
segmented across multiple WC buffers, and across multiple WR buffers.

Example A:
1. Create MP WR QP with:
   - wr_buffer_sz = 64 KB
   - packet_align_sz = 512 bytes
2. Lets assume MTU is 4 KB
3. In which case each wr can receive
   a. 128 RDMA messages of 512 bytes each until WR is entirely consumed.
   b. A 12,000 bytes RDMA message will report up to 3 WCs. FIRST and
      MIDDLE packets have 2 WC's with MP_WR_MORE_IN_MSG of length 4,096 bytes,
      ending in a WC with 3808 bytes. WR will still have 52 KB left for
      ingress packets handling before reporting MP_WR_CONSUMED.

Example B:
1. Create MP WR QP with:
   - wr_buffer_sz = 1 MB
   - packet_align_sz = 4 KB
2. Lets assume MTU is 4 KB
3. In which case each WR can receive up to 256 packets. We cut the post_recv
and PCI WR fetch by a factor of 1:250.
Packets are received in page (4K) alignment.

Issue: 1215816
Change-Id: I8f9cca81c7c70d79f2bbf25401f62b06e4f61b27
Signed-off-by: Alex Rosenbaum <alexr@mellanox.com>
---
 libibverbs/man/ibv_create_cq_ex.3    | 22 +++++++++++++++++++-
 libibverbs/man/ibv_create_qp_ex.3    | 40 ++++++++++++++++++++++++++++++++----
 libibverbs/man/ibv_create_srq_ex.3   | 33 ++++++++++++++++++++++++++++-
 libibverbs/man/ibv_create_wq.3       | 34 +++++++++++++++++++++++++++++-
 libibverbs/man/ibv_query_device_ex.3 |  9 ++++++++
 libibverbs/verbs.h                   | 35 ++++++++++++++++++++++++++++---
 6 files changed, 163 insertions(+), 10 deletions(-)

 enum ibv_qp_open_attr_mask {
@@ -1209,6 +1232,7 @@ struct ibv_cq_ex {
  uint32_t (*read_flow_tag)(struct ibv_cq_ex *current);
  void (*read_tm_info)(struct ibv_cq_ex *current,
       struct ibv_wc_tm_info *tm_info);
+ size_t  (*read_mp_wr_offset)(struct ibv_cq_ex *cq);
 };

 static inline struct ibv_cq *ibv_cq_ex_to_cq(struct ibv_cq_ex *cq)
@@ -1327,6 +1351,11 @@ static inline void ibv_wc_read_tm_info(struct
ibv_cq_ex *cq,
  cq->read_tm_info(cq, tm_info);
 }

+static inline size_t ibv_wc_read_mp_wr_offset(struct ibv_cq_ex *cq)
+{
+ return cq->read_mp_wr_offset(cq);
+}
+
 static inline int ibv_post_wq_recv(struct ibv_wq *wq,
     struct ibv_recv_wr *recv_wr,
     struct ibv_recv_wr **bad_recv_wr)


Message ID	CAFgAxU-rJb6mvH_XRt0S5_-wKwv3rOnweiLJ8WSuc=9F+P0BrQ@mail.gmail.com (mailing list archive)
State	RFC
Delegated to:	Leon Romanovsky
Headers	show Return-Path: <linux-rdma-owner@kernel.org> MIME-Version: 1.0 From: Alex Rosenbaum <rosenbaumalex@gmail.com> Date: Tue, 12 Dec 2017 23:15:36 +0200 Message-ID: <CAFgAxU-rJb6mvH_XRt0S5_-wKwv3rOnweiLJ8WSuc=9F+P0BrQ@mail.gmail.com> Subject: [PATCH RFC] Introduce verbs API for multi-packet work request To: linux-rdma@vger.kernel.org Cc: Yishai Hadas <yishaih@mellanox.com>, "Guy Levi(SW)" <guyle@mellanox.com>, Leon Romanovsky <leonro@mellanox.com>, "Alex @ Mellanox" <alexr@mellanox.com> Content-Type: text/plain; charset="UTF-8" Sender: linux-rdma-owner@vger.kernel.org Precedence: bulk

[RFC] Introduce verbs API for multi-packet work request

Commit Message

Comments

Patch