[{"data":1,"prerenderedAt":382},["ShallowReactive",2],{"\u002Fen\u002Fblogs\u002F20261116UBMC-bootraid-b310n":3,"content-doc-\u002Fzh\u002Fblogs\u002F20261116UBMC-bootraid-b310n":4,"surround-\u002Fzh\u002Fblogs\u002F20261116UBMC-bootraid-b310n":375},null,{"id":5,"title":6,"author":7,"body":11,"category":364,"city":3,"company":3,"coverImage":365,"date":366,"description":367,"extension":368,"footer":3,"header":3,"location":3,"logo":3,"logoDark":3,"meta":369,"name":3,"navigation":370,"path":371,"seo":372,"stem":373,"version":3,"__hash__":374},"content\u002Fzh\u002Fblogs\u002F20261116UBMC-bootraid-b310n.md","打破存储管理边界：百信信息基于openUBMC的BootRAID卡适配实践",[8],{"name":9,"description":10},"罗昊","百信信息BMC高级工程师，openUBMC社区AI SIG和CICD SIG Committer。",{"type":12,"value":13,"toc":344},"minimark",[14,18,22,25,28,32,35,64,68,71,103,110,114,119,122,142,145,149,152,155,161,164,168,171,179,185,189,192,195,251,255,258,262,282,286,312,315,318,321,341],[15,16,17],"h2",{"id":17},"概要",[19,20,21],"p",{},"在AI训练、HCI超融合和分布式存储场景中，系统盘承担操作系统启动、配置与日志记录，其可用性直接关系服务器节点能否持续运行。传统“一卡两用”方案由同一张RAID卡管理系统盘和数据盘：控制器成为共同单点故障点，两类磁盘还会争用通道与缓存。对于数据盘仅需直通访问的场景，通用RAID功能超配还会增加建设成本。",[19,23,24],{},"为此，百信信息采用TrustRAID B310n国产NVMe BootRAID卡与HBA卡构建“专卡专用”方案，以独立控制器隔离系统盘与数据盘故障域，并通过硬件RAID1保护系统盘。在此基础上，依托openUBMC完成盘卡一体形态的带外管理适配，将B310n纳入统一存储管理，为“专卡专用”方案补齐监控与管理能力。",[19,26,27],{},"本文围绕这项适配展开：先说明新厂商SDK、无背板形态及数据模型带来的挑战，再介绍基于openUBMC的总体方案、关键实现和分层验证方法，说明如何在复用既有管理框架的同时接入新的存储部件。",[15,29,31],{"id":30},"_01-bootraid接入bmc的适配挑战","01 BootRAID接入BMC的适配挑战",[19,33,34],{},"B310n将两块M.2 NVMe SSD集成在PCIe卡上，接入openUBMC存储管理框架时，需要处理以下差异：",[36,37,38,46,52,58],"ul",{},[39,40,41,45],"li",{},[42,43,44],"strong",{},"新厂商SDK接入："," Marvell SDK的初始化、数据结构和通信回调需要适配SML统一接口。",[39,47,48,51],{},[42,49,50],{},"无背板槽位映射："," 盘卡一体形态没有独立硬盘背板，需要将卡侧物理盘与CSR中的Drive对象关联。",[39,53,54,57],{},[42,55,56],{},"存储模型兼容："," SDK未直接提供SML所需的Array信息，需根据逻辑盘及其成员物理盘构造兼容的阵列对象。",[39,59,60,63],{},[42,61,62],{},"链路负载优化："," 需要兼顾动态状态更新与I2C总线负载，减少短时间内的重复读取。",[15,65,67],{"id":66},"_02-基于openubmc的总体方案","02 基于openUBMC的总体方案",[19,69,70],{},"本次适配从板卡南向接入到北向接口展示，复用openUBMC存储管理框架，各层职责如下：",[36,72,73,79,85,91,97],{},[39,74,75,78],{},[42,76,77],{},"盘卡一体："," B310n的Marvell控制器管理卡上两块M.2 NVMe SSD，通过I2C带外链路提供控制器、逻辑盘和物理盘的原始数据。",[39,80,81,84],{},[42,82,83],{},"CSR配置："," CSR描述拓扑结构、器件对象、告警事件、传感器、调速策略等。",[39,86,87,90],{},[42,88,89],{},"SML适配："," 在存储设备管理库SML（Storage Management Library）中新增sml_hrdt模块，封装Marvell SDK调用、数据转换与缓存，由l_sml提供统一Lua接口。",[39,92,93,96],{},[42,94,95],{},"存储管理："," storage维护控制器、物理盘、逻辑盘与阵列对象，完成槽位映射、周期刷新与数据整合，并将管理信息推送到资源树上。",[39,98,99,102],{},[42,100,101],{},"北向接口："," 通过资源树及rackmount接口映射配置，向Web、Redfish、CLI等提供统一存储管理能力。",[19,104,105],{},[106,107],"img",{"alt":108,"src":109,"title":108},"B310n接入openUBMC存储管理的总体架构","\u002Fcategory\u002Fblog\u002F20261116UBMC-bootraid-b310n\u002F1.png",[15,111,113],{"id":112},"_03-关键设计与实现","03 关键设计与实现",[115,116,118],"h3",{"id":117},"_31-marvell-sdk接入-扩展控制器支持厂商","3.1 Marvell SDK接入 扩展控制器支持厂商",[19,120,121],{},"本次在SML厂商适配层完成Marvell SDK封装与数据转换，通信接入包括以下环节：",[36,123,124,130,136],{},[39,125,126,129],{},[42,127,128],{},"回调注册："," 针对SDK分别调用读、写接口的约定，补充并注册独立的I2C读写回调。",[39,131,132,135],{},[42,133,134],{},"语言绑定："," l_sml将SML的C接口封装为Lua接口，并转发硬件访问请求。",[39,137,138,141],{},[42,139,140],{},"硬件访问："," storage提供SML扩展插件，在hwproxy侧执行SDK所需的I2C读写。",[19,143,144],{},"在既有博通、华为和PMC RAID卡适配基础上，storage新增基于Marvell控制器的B310n支持，上层继续使用统一存储对象开展管理。",[115,146,148],{"id":147},"_32-无背板槽位映射-适配盘卡一体形态","3.2 无背板槽位映射 适配盘卡一体形态",[19,150,151],{},"PD（Physical Drive）是RAID卡侧的物理盘，Drive是CSR中的硬盘对象。槽位映射将两者关联，确定物理盘所在槽位。",[19,153,154],{},"依赖背板反馈的传统映射方案，通过RAID卡点定位灯操作与背板SMC上报的点灯状态确定槽位。B310n的M.2盘直接安装在卡上，需要使用新的映射依据。",[19,156,157],{},[106,158],{"alt":159,"src":160,"title":159},"传统背板反馈槽位映射与B310n无背板槽位映射对比","\u002Fcategory\u002Fblog\u002F20261116UBMC-bootraid-b310n\u002F2.png",[19,162,163],{},"本次适配根据B310n卡侧PD ID与CSR中Drive名称的对应关系，在Drive在位且PD有效时建立槽位映射，无需硬盘背板SMC反馈。拔盘时解除关联并保留Drive对象，重新插盘后恢复关联。",[115,165,167],{"id":166},"_33-存储模型兼容-衔接sdk数据与sml对象","3.3 存储模型兼容 衔接SDK数据与SML对象",[19,169,170],{},"B310n SDK提供逻辑盘VD及成员物理盘信息，未直接提供SML所需的Array信息。适配层通过以下转换补齐阵列模型，供storage维护Volume、DiskArray与Drive等资源对象：",[36,172,173,176],{},[39,174,175],{},"将SDK逻辑盘VD映射为SML逻辑盘LD，以VD ID生成对应的Array引用。",[39,177,178],{},"本次适配中LD与Array一一对应，Array成员列表由SDK返回的物理盘ID填充。",[19,180,181],{},[106,182],{"alt":183,"src":184,"title":183},"B310n SDK对象与SML对象的适配关系","\u002Fcategory\u002Fblog\u002F20261116UBMC-bootraid-b310n\u002F3.png",[115,186,188],{"id":187},"_34-链路负载优化-分类执行缓存策略","3.4 链路负载优化 分类执行缓存策略",[19,190,191],{},"storage需要周期性获取管理信息，不同业务逻辑可能重复查询相同数据。频繁读取会增加I2C总线负载，而物理盘在位、逻辑盘变化与定位灯状态等动态信息又需要及时更新。",[19,193,194],{},"B310n适配模块按控制器、数据类型与对象ID建立哈希索引。缓存有效期内复用查询结果，减少重复读取；动态信息按查询更新，分类执行以下策略：",[196,197,198,214],"table",{},[199,200,201],"thead",{},[202,203,204,208,211],"tr",{},[205,206,207],"th",{},"信息类别",[205,209,210],{},"当前实现策略",[205,212,213],{},"设计目的",[215,216,217,229,240],"tbody",{},[202,218,219,223,226],{},[220,221,222],"td",{},"控制器信息、部分SMART与健康信息",[220,224,225],{},"缓存有效期为10秒",[220,227,228],{},"复用短时间内的查询结果，减少I2C访问",[202,230,231,234,237],{},[220,232,233],{},"物理盘基本信息、逻辑盘及列表、定位灯状态等",[220,235,236],{},"每次查询重新读取板卡",[220,238,239],{},"及时更新对象关系和设备状态",[202,241,242,245,248],{},[220,243,244],{},"本次读取失败时",[220,246,247],{},"已有缓存时返回上次成功读取的结果",[220,249,250],{},"保留已有管理信息，供异常处理使用",[15,252,254],{"id":253},"_04-分层调试与工程验证","04 分层调试与工程验证",[19,256,257],{},"团队先验证SDK与带外链路，再联调SML和storage，分别检查通信、接口返回值和对象关联，缩小问题定位范围。",[115,259,261],{"id":260},"_41-独立验证sdk与通信链路","4.1 独立验证SDK与通信链路",[36,263,264,270,276],{},[39,265,266,269],{},[42,267,268],{},"调试解耦："," 实现并注册I2C读写和日志接口，构建独立SDK调试程序。",[39,271,272,275],{},[42,273,274],{},"数据解析："," 分别调用控制器、逻辑盘和物理盘API，将返回结构体解析为可读信息。",[39,277,278,281],{},[42,279,280],{},"交叉核对："," 与OS、BIOS侧信息对照，确认通信正常、解析结果一致后再接入框架。",[115,283,285],{"id":284},"_42-围绕对象关系开展框架联调","4.2 围绕对象关系开展框架联调",[36,287,288,294,300,306],{},[39,289,290,293],{},[42,291,292],{},"槽位映射："," 覆盖首次识别、拔盘解除关联和重新插入后的配对。",[39,295,296,299],{},[42,297,298],{},"存储模型："," 核对SML对象及其对应资源的ID、成员盘与关联关系。",[39,301,302,305],{},[42,303,304],{},"刷新机制："," 检查动态状态更新、重复查询与通信异常后的恢复。",[39,307,308,311],{},[42,309,310],{},"状态呈现："," 将RAID1降级、重构等界面状态与板卡实际状态对照。",[15,313,314],{"id":314},"总结",[19,316,317],{},"百信信息通过BootRAID卡与HBA卡分工管理系统盘和数据盘，并依托openUBMC完成B310n统一带外管理适配，形成从硬件故障隔离到运维状态呈现的完整方案。硬件架构解决两类磁盘共用控制器的问题，BMC适配则让新形态设备接入既有管理入口。",[19,319,320],{},"面向用户，这套方案的价值体现在三个方面：",[36,322,323,329,335],{},[39,324,325,328],{},[42,326,327],{},"业务连续性："," 故障域隔离与硬件RAID1保护系统盘，统一监控支撑故障处置。",[39,330,331,334],{},[42,332,333],{},"部署与维护："," 盘卡一体、预配置RAID1及热插拔简化部署维护，openUBMC提供统一管理。",[39,336,337,340],{},[42,338,339],{},"成本优化："," 在HCI和分布式存储的数据盘直通场景中，专卡分工与硬件精简带来20%以上的整体成本优化。",[19,342,343],{},"从技术实现看，本次实践通过Marvell SDK封装、无背板槽位映射、存储模型兼容与分类缓存策略，扩展了openUBMC对新厂商和盘卡一体设备的支持。厂商差异在适配层处理，上层继续复用统一存储资源模型，为后续同类存储部件接入提供了可参考的实现路径。",{"title":345,"searchDepth":346,"depth":346,"links":347},"",4,[348,350,351,352,359,363],{"id":17,"depth":349,"text":17},2,{"id":30,"depth":349,"text":31},{"id":66,"depth":349,"text":67},{"id":112,"depth":349,"text":113,"children":353},[354,356,357,358],{"id":117,"depth":355,"text":118},3,{"id":147,"depth":355,"text":148},{"id":166,"depth":355,"text":167},{"id":187,"depth":355,"text":188},{"id":253,"depth":349,"text":254,"children":360},[361,362],{"id":260,"depth":355,"text":261},{"id":284,"depth":355,"text":285},{"id":314,"depth":349,"text":314},"essentials","\u002Fcategory\u002Fblog\u002F20261116UBMC-bootraid-b310n\u002Fcover.webp","2026\u002F11\u002F16","AI训练、HCI超融合与分布式存储场景对系统盘可用性提出更高要求。百信信息基于 openUBMC 完成 TrustRAID B310n 国产 NVMe BootRAID 卡的带外管理适配，通过专卡专用、无背板槽位映射、存储模型兼容与分类缓存策略，实现系统盘与数据盘的故障域隔离与统一管理。","md",{},true,"\u002Fzh\u002Fblogs\u002F20261116UBMC-bootraid-b310n",{"title":6,"description":367},"zh\u002Fblogs\u002F20261116UBMC-bootraid-b310n","uyedwbT0FVm6Qgg146tM5xaIb4Rq6Nza4E4RvkivnU8",[376,379],{"path":377,"title":378},"\u002Fzh\u002Fblogs\u002F20261112UBMC-component-score","一条命令，量化你的部件管理能力 — openUBMC 部件兼容性评分工具实战指南",{"path":380,"title":381},"\u002Fzh\u002Fblogs\u002F2026UBMC-Changsha","开源BMC筑基AI算力——openUBMC 走进长沙高校",1795081488775]