- GPU Direct RDMA (P2P DMA) for Device Private Pages
This patch series aims to enable Peer-to-Peer (P2P) DMA access in GPU-centric applications that use RDMA and device private memory. By allowing the GPU to directly expose its private pages to devices like NICs, this enhancement avoids routing data through system RAM, which is typically required to access device private memory.
本补丁系列旨在为依赖 RDMA 和设备私有内存的 GPU 中心应用程序启用点对点（P2P）DMA 访问。通过允许 GPU 直接将其私有页暴露给诸如 NIC 等设备，该改进避免了必须通过系统内存访问设备私有页的传统方式，从而减少了数据传输开销。
To fully support P2P access for device private memory, the following enhancements are proposed:
为了全面支持设备私有内存的 P2P 访问，提出了以下改进：
**Memory Management (MM)**
* Extend `struct pagemap_ops` to support P2P page operations: This change allows GPUs to directly map device private pages for P2P DMA.
* Use `hmm_range_fault` to handle P2P access for device private pages instead of relying on standard page faults.
**内存管理（MM）**
* 扩展 `struct pagemap_ops` 以支持 P2P 页操作：此更改使 GPU 能够直接映射设备私有页以支持 P2P DMA。
* 使用 `hmm_range_fault` 来处理设备私有页的 P2P 访问，而非依赖传统的页错误机制。
**IB Drivers**
* Add a `TRY_P2P_REQ` flag to the `hmm_range_fault` call: This flag signals the need for P2P mapping, enabling InfiniB