libxscale: Introduce xscale provider - #1773
Open
tianx666 wants to merge 13 commits into
Open
Conversation
tianx666
force-pushed
the
master
branch
5 times, most recently
from
July 20, 2026 03:14
460486f to
d013e80
Compare
This patch adds a new userspace RDMA provider driver for the xscale
family of RDMA devices. The provider implements the standard
libibverbs interface, allowing applications to
use RDMA operations via the verbs API.
The initial implementation includes:
- Basic CMake build infrastructure and provider registration.
- Device probing and context allocation via the verbs framework.
- Query operations: query_port and query_device_ex with firmware
version formatting.
- Import of necessary kernel ABI headers (xsc-abi.h) and integration
with the kernel driver.
This is the foundational patch; subsequent patches will add full
functionality for PD, MR, CQ, QP, and advanced features.
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
tianx666
force-pushed
the
master
branch
4 times, most recently
from
July 20, 2026 05:12
25970ee to
80a87fe
Compare
Define the hardware-specific structures, constants, and inline helper
functions required to interact with the XScale RDMA engine.
This patch introduces:
- xsc_hsi.h: Host-software interface definitions, including message
opcodes, completion queue entries (CQE), work queue elements (WQE),
doorbell formats, and register layouts.
- xsc_hw.h: Low-level hardware access routines for Andes and Diamond
device families, such as setting data segments, ringing doorbells,
updating CQ consumer indices, and translating error codes to verbs
statuses.
These definitions are used by all subsequent patches to construct and
post work requests, poll completions, and manage hardware resources.
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
Implement the basic memory management verbs required for RDMA operations:
- alloc_pd / dealloc_pd: Manage protection domains, with reference
counting to prevent premature deallocation.
- reg_mr / dereg_mr: Register and deregister memory regions, supporting
standard access flags. The implementation uses the common ibv_cmd_*
wrappers and stores local keys (lkey/rkey) for later use.
These verbs are prerequisites for creating QPs and posting work requests.
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
This patch adds complete CQ management:
- create_cq / create_cq_ex: Allocate CQ buffers using either huge
pages, contiguous memory, or regular anonymous memory (controlled
by environment variables). Support for GPU‑backed buffers via dmabuf
is also provided.
- resize_cq: Resize an existing CQ.
- destroy_cq: Free CQ resources and clean up associated QP error lists.
This lays the groundwork for polling completions and event handling.
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
Implement the core QP lifecycle operations:
- create_qp / create_qp_ex: Allocate QP buffers (SQ and RQ), calculate
WQE sizes based on device capabilities, and initialize QP indices.
Supports raw packet QPs and RC QPs.
- modify_qp: Transition QP states (RTR, RTS, etc.) and handle error
state transitions, including flushing pending work requests.
- query_qp: Retrieve current QP attributes.
- destroy_qp: Clean up QP resources, remove from the QP table, and
flush any remaining CQEs.
A global QP table is maintained to map QP numbers to internal structures
for fast lookups during completion processing. Error state handling
ensures that CQs are notified when a QP enters the error state.
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
Add the ability to post work requests to the hardware:
- post_send: Construct send WQEs (including RDMA write/read, send,
atomic operations) and ring the send doorbell. Inline data and
scatter‑gather lists are supported.
- post_recv: Build receive WQEs and ring the receive doorbell.
The implementation carefully manages WQE rings, tracks work request IDs
and completion flags, and handles overflow conditions. Doorbell ringing
is done via MMIO writes with proper barriers.
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
tianx666
force-pushed
the
master
branch
6 times, most recently
from
July 24, 2026 10:09
6c74a18 to
cd5a002
Compare
Implement the completion retrieval path:
- poll_cq: Scan the CQ ring for new CQEs, parse them (both good and
error completions), and fill ibv_wc structures.
- arm_cq / cq_event: Arm the CQ for notification and handle event
callbacks.
Error CQEs are translated to appropriate verbs status codes, and QPs
are moved to error state when fatal errors occur. Flush error CQEs are
generated for QPs that have been moved to error state to ensure all
pending WRs are completed.
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
Extend the MR registration to handle dmabuf file descriptors, allowing external (e.g., GPU) memory to be used as RDMA buffers. The new function reg_dmabuf_mr wraps the common ibv_cmd_reg_dmabuf_mr and stores the access flags for later use. Signed-off-by: Tian Xin <tianx@yunsilicon.com> Signed-off-by: Wei Honggang <weihg@yunsilicon.com> Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com> Signed-off-by: Li Qiang <liq@yunsilicon.com> Signed-off-by: Yan Lei <jacky@yunsilicon.com>
Implement the ibv_wr_api (start, set SGE/inline data, complete).
The following operations are supported for RC QPs:
- wr_send / wr_send_imm
- wr_rdma_write / wr_rdma_write_imm / wr_rdma_read
- wr_atomic_cmp_swp / wr_atomic_fetch_add
- wr_set_sge / wr_set_sge_list / wr_set_inline_data /
wr_set_inline_data_list
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
Add the cq_ex interface for efficient completion polling: - start_poll: Begin polling a CQ, lock it, and retrieve the first CQE. - next_poll: Fetch subsequent CQEs without releasing the lock. - end_poll: Release the CQ lock and update the consumer index. Signed-off-by: Tian Xin <tianx@yunsilicon.com> Signed-off-by: Wei Honggang <weihg@yunsilicon.com> Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com> Signed-off-by: Li Qiang <liq@yunsilicon.com> Signed-off-by: Yan Lei <jacky@yunsilicon.com>
Introduce a set of xscale‑specific direct verbs that provide advanced
functionality not covered by the standard verbs API:
- devx_create_qp / devx_destroy_qp: Create/destroy QPs with user‑supplied
memory (e.g., GPU buffers) via devx_umem_in.
- devx_create_cq / devx_destroy_cq: Similar for CQs.
- devx_alloc_uar / devx_free_uar: Allocate/free User Access Regions
(UAR) for direct doorbell ringing.
- devx_alloc_sq_uar: Allocate dedicated SQ UAR for multi‑QP performance.
- devx_modify_qp: Wrapper for modify_qp with DV context.
- devx_get_device_id: Retrieve the hardware device ID.
- devx_exp_post_send: Extended post_send that supports masked atomic
operations (8‑byte and 4‑byte masked compare‑and‑swap and fetch‑and‑add).
Also add ECE (Extended Connection Establishment) support via set_ece and
query_ece, enabling negotiation of protocol profiles (e.g., Veroce).
These extensions enable high‑performance and custom use cases for
applications using the xscale provider.
Signed-off-by: Tian Xin <tianx@yunsilicon.com>
Signed-off-by: Wei Honggang <weihg@yunsilicon.com>
Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com>
Signed-off-by: Li Qiang <liq@yunsilicon.com>
Signed-off-by: Yan Lei <jacky@yunsilicon.com>
This patch integrates the xscale RDMA provider into the rdma-core build and packaging system for both Debian and RPM-based distributions. Signed-off-by: Tian Xin <tianx@yunsilicon.com> Signed-off-by: Wei Honggang <weihg@yunsilicon.com> Signed-off-by: Zhao Qianwei <zhaoqw@yunsilicon.com> Signed-off-by: Li Qiang <liq@yunsilicon.com> Signed-off-by: Yan Lei <jacky@yunsilicon.com>
fix pipeline issues Signed-off-by: Tian Xin <tianx@yunsilicon.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Introduce a new RDMA provider for Yunsilicon nics