{"id":5483,"date":"2021-08-17T16:44:52","date_gmt":"2021-08-17T08:44:52","guid":{"rendered":"https:\/\/damogame.cn\/wordpress\/?p=5483"},"modified":"2021-08-17T16:44:52","modified_gmt":"2021-08-17T08:44:52","slug":"aio-%e7%9a%84%e6%96%b0%e5%bd%92%e5%ae%bf%ef%bc%9aio_uring","status":"publish","type":"post","link":"https:\/\/i007.cc\/wordpress\/archives\/5483","title":{"rendered":"AIO \u7684\u65b0\u5f52\u5bbf\uff1aio_uring"},"content":{"rendered":"<header class=\"Post-Header\">\n<div class=\"Post-Author\"><a href=\"https:\/\/zhuanlan.zhihu.com\/p\/62682475\">\u539f\u6587\u5730\u5740<\/a><\/div>\n<div><\/div>\n<div class=\"LabelContainer-wrapper\"><\/div>\n<div tabindex=\"0\" role=\"button\"><span class=\"Voters\"><button class=\"Button Button--plain\" type=\"button\">503 \u4eba\u8d5e\u540c\u4e86\u8be5\u6587\u7ae0<\/button><\/span><\/div>\n<\/header>\n<div class=\"Post-RichTextContainer\">\n<div class=\"RichText ztext Post-RichText css-hnrfcf\">\n<p>\u5728\u6b63\u6587\u5f00\u59cb\u4e4b\u524d\uff0c\u6211\u4eec\u5148\u6765\u770b\u4e00\u6bb5\u5bf9\u8bdd\uff08\u6f14\u7ece\u7248\u672c\uff09<\/p>\n<blockquote><p>Jens Axboe \uff1aLinus\uff0c\u6211\u6709\u4e2a\u597d\u4e1c\u897f\uff0c\u4f60\u7785\u7785\uff1f<br \/>\nLinus\uff1a\u5565\u73a9\u610f\u513f\uff0c\u4e0d\u662f\u5df2\u7ecf\u6709 aio \u4e86\u4e48\uff0c\u4e3a\u5565\u53c8\u6765\u4e00\u5957\uff0c\u4f60\u548b\u4e0d\u53bb\u597d\u597d\u4fee aio \u7684\u95ee\u9898\u3002aio \u8fd8\u6709 balabala \u95ee\u9898\u6ca1\u6709\u4fee\u5462\u3002<br \/>\nJens Axboe \uff1a\u6211\u8fd9\u4e2a DIAO \u554a\uff0cbalabala<br \/>\nLinus \uff1a\u770b\u4f60\u8bda\u610f\u8fd9\u4e48\u8db3\uff0c\u90a3\u884c\u5427\uff0c\u6211\u5148\u6536\u5230\u6211\u7684 tree \u4e0b\uff0c\u4e0d push out \u51fa\u53bb\uff0c\u8ba9\u6211\u6d4b\u6d4b\u5148\u3002<br \/>\n\u3010\u4e0d\u4e00\u4f1a\u513f\u3011<br \/>\nLinus \uff1a\u4f60\u8fd9 IO \u5f15\u7528\u8ba1\u6570\u5199\u7684\u8fa3\u9e21\uff0c\u4e00\u770b\u5c31\u6709\u95ee\u9898\uff0c\u53bb\u6539\u5427\u3002<br \/>\n\u2026\u2026<\/p><\/blockquote>\n<h2>\u80cc\u666f<\/h2>\n<p>Linus \u548c Jens \u5728\u8ba8\u8bba\u7684\uff0c\u5c31\u662f Linux Kernel \u5373\u5c06\u5728 5.1 \u7248\u672c\u52a0\u5165\u4e00\u4e2a\u91cd\u5927 feature\uff1aio_uring\u3002<\/p>\n<p>\u5bf9\u505a\u5b58\u50a8\u7684\u6765\u8bf4\uff0c\u8fd9\u662f\u4e00\u4e2a\u5927\u4e8b\u60c5\uff0c\u503c\u5f97\u666e\u5927\u559c\u5954\uff0c\u5e7f\u800c\u544a\u4e4b\u3002libaio \u5373\u5c06\u57cb\u5165\u9ec4\u571f\uff0cio_uring \u62d4\u5730\u800c\u8d77\u3002<\/p>\n<p>\u4e00\u53e5\u8bdd\u603b\u7ed3 io_uring \u5c31\u662f\uff1a\u4e00\u5957\u5168\u65b0\u7684 syscall\uff0c\u4e00\u5957\u5168\u65b0\u7684 async API\uff0c\u66f4\u9ad8\u7684\u6027\u80fd\uff0c\u66f4\u597d\u7684\u517c\u5bb9\u6027\uff0c\u6765\u8fce\u63a5\u9ad8 IOPS\uff0c\u9ad8\u541e\u5410\u91cf\u7684\u672a\u6765\u3002<\/p>\n<p>\u5148\u770b\u4e00\u4e0b\u6027\u80fd\u6570\u636e\uff08\u6570\u636e\u6765\u81ea Jens Axboe\uff09\u3002<\/p>\n<p>4k randread\uff0c3D Xpoint \u76d8\uff1a<\/p>\n<div class=\"highlight\">\n<pre><code class=\"language-text\">Interface       QD      Polled          Latency         IOPS\r\n--------------------------------------------------------------------------\r\nio_uring        1       0                9.5usec         77K\r\nio_uring        2       0                8.2usec        183K\r\nio_uring        4       0                8.4usec        383K\r\nio_uring        8       0               13.3usec        449K\r\n\r\nlibaio          1       0                9.7usec         74K\r\nlibaio          2       0                8.5usec        181K\r\nlibaio          4       0                8.5usec        373K\r\nlibaio          8       0               15.4usec        402K\r\n\r\nio_uring        1       1                6.1usec        139K\r\nio_uring        2       1                6.1usec        272K    \r\nio_uring        4       1                6.3usec        519K\r\nio_uring        8       1               11.5usec        592K\r\n\r\nspdk            1       1                6.1usec        151K\r\nspdk            2       1                6.2usec        293K\r\nspdk            4       1                6.7usec        536K\r\nspdk            8       1               12.6usec        586K<\/code><\/pre>\n<\/div>\n<p>io_uring vs libaio\uff0c\u5728\u975e polling \u6a21\u5f0f\u4e0b\uff0cio_uring \u6027\u80fd\u63d0\u5347\u4e0d\u5230 10%\uff0c\u597d\u50cf\u5e76\u6ca1\u6709\u4ec0\u4e48\u4e86\u4e0d\u8d77\u7684\u5730\u65b9\u3002<\/p>\n<p>\u7136\u800c io_uring \u63d0\u4f9b\u4e86 polling \u6a21\u5f0f\u3002\u5728 polling \u6a21\u5f0f\u4e0b\uff0cio_uring \u548c SPDK \u7684\u6027\u80fd\u975e\u5e38\u63a5\u8fd1\uff0c\u7279\u522b\u662f\u9ad8 QueueDepth \u4e0b\uff0cio_uring \u6709\u8d76\u8d85\u7684\u67b6\u52bf\uff0c\u540c\u65f6\u5b8c\u7206 libaio\u3002<\/p>\n<p>\u6d4b\u8bd5 per-core\uff0c4k randread \u591a\u8bbe\u5907\u4e0b\u7684\u6700\u9ad8 IOPS \u80fd\u529b\uff1a<\/p>\n<div class=\"highlight\">\n<pre><code class=\"language-text\">Interface       QD      Polled          IOPS\r\n--------------------------------------------------------------------------\r\nio_uring        128     1               1620K\r\nlibaio          128     0                608K\r\nspdk            128     1               1739K<\/code><\/pre>\n<\/div>\n<p>\u6700\u8fd1\u51e0\u5e74\u4e00\u76f4\u6d41\u884c kernel bypass\uff0c\u4ece\u7f51\u7edc\u5230\u5b58\u50a8\uff0c\u5404\u4e2a\u9886\u57df\u5f00\u82b1\uff0c\u5185\u6838\u5728\u6027\u80fd\u65b9\u9762\u88ab\u5404\u79cd\u8bdf\u75c5\u3002io_uring \u51fa\u73b0\u4ee5\u540e\uff0c\u7b97\u662f\u6273\u56de\u4e00\u5c40\u3002<\/p>\n<p>io_uring \u6709\u5982\u6b64\u51fa\u4f17\u7684\u6027\u80fd\uff0c\u4e3b\u8981\u6765\u6e90\u4e8e\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\uff1a<\/p>\n<ul>\n<li>\u7528\u6237\u6001\u548c\u5185\u6838\u6001\u5171\u4eab\u63d0\u4ea4\u961f\u5217\uff08submission queue\uff09\u548c\u5b8c\u6210\u961f\u5217\uff08completion queue\uff09<\/li>\n<li>IO \u63d0\u4ea4\u548c\u6536\u5272\u53ef\u4ee5 offload \u7ed9 Kernel\uff0c\u4e14\u63d0\u4ea4\u548c\u5b8c\u6210\u4e0d\u9700\u8981\u7ecf\u8fc7\u7cfb\u7edf\u8c03\u7528\uff08system call\uff09<\/li>\n<li>\u652f\u6301 Block \u5c42\u7684 Polling \u6a21\u5f0f<\/li>\n<li>\u901a\u8fc7\u63d0\u524d\u6ce8\u518c\u7528\u6237\u6001\u5185\u5b58\u5730\u5740\uff0c\u51cf\u5c11\u5730\u5740\u6620\u5c04\u7684\u5f00\u9500<\/li>\n<\/ul>\n<p>\u4e0d\u4ec5\u5982\u6b64\uff0cio_uring \u8fd8\u53ef\u4ee5\u5b8c\u7f8e\u652f\u6301 buffered IO\uff0c\u800c libaio \u5bf9\u4e8e buffered IO \u7684\u652f\u6301\u5219\u4e00\u76f4\u662f\u88ab\u8bdf\u75c5\u7684\u5730\u65b9\u3002<\/p>\n<h2>io_uring<\/h2>\n<p>io_uring \u63d0\u4f9b\u4e86\u4e00\u5957\u65b0\u7684\u7cfb\u7edf\u8c03\u7528\uff0c\u5e94\u7528\u7a0b\u5e8f\u53ef\u4ee5\u4f7f\u7528\u4e24\u4e2a\u961f\u5217\uff0cSubmission Queue\uff08SQ) \u548c Completion Queue\uff08CQ) \u6765\u548c Kernel \u8fdb\u884c\u901a\u4fe1\u3002\u8fd9\u79cd\u65b9\u5f0f\u7c7b\u4f3c RDMA \u6216\u8005 NVMe \u7684\u65b9\u5f0f\uff0c\u53ef\u4ee5\u9ad8\u6548\u5904\u7406 IO\u3002<\/p>\n<div class=\"highlight\">\n<pre><code class=\"language-text\">syscall \r\n425        io_uring_setup\r\n426        io_uring_enter\r\n427        io_uring_register<\/code><\/pre>\n<\/div>\n<h2>io_uring \u51c6\u5907\u9636\u6bb5<\/h2>\n<p>io_uring_setup \u9700\u8981\u4e24\u4e2a\u53c2\u6570\uff0centries \u548c io_uring_params\u3002<\/p>\n<p>\u5176\u4e2d entries\uff0c\u4ee3\u8868 queue depth\u3002<\/p>\n<p>io_uring_params \u7684\u5b9a\u4e49\u5982\u4e0b\u3002<\/p>\n<div class=\"highlight\">\n<pre><code class=\"language-text\">struct io_uring_params {\r\n\t__u32 sq_entries;\r\n\t__u32 cq_entries;\r\n\t__u32 flags;\r\n\t__u32 sq_thread_cpu;\r\n\t__u32 sq_thread_idle;\r\n\t__u32 resv[5];\r\n\tstruct io_sqring_offsets sq_off;\r\n\tstruct io_cqring_offsets cq_off;\r\n};\r\n\r\nstruct io_sqring_offsets {\r\n\t__u32 head;\r\n\t__u32 tail;\r\n\t__u32 ring_mask;\r\n\t__u32 ring_entries;\r\n\t__u32 flags;\r\n\t__u32 dropped;\r\n\t__u32 array;\r\n\t__u32 resv1;\r\n\t__u64 resv2;\r\n};\r\n\r\nstruct io_cqring_offsets {\r\n\t__u32 head;\r\n\t__u32 tail;\r\n\t__u32 ring_mask;\r\n\t__u32 ring_entries;\r\n\t__u32 overflow;\r\n\t__u32 cqes;\r\n\t__u64 resv[2];\r\n};<\/code><\/pre>\n<\/div>\n<p>\u5176\u4e2d\uff0cflags\u3001sq_thread_cpu\u3001sq_thread_idle \u5c5e\u4e8e\u8f93\u5165\u53c2\u6570\uff0c\u7528\u4e8e\u5b9a\u4e49 io_uring \u5728\u5185\u6838\u4e2d\u7684\u884c\u4e3a\u3002\u5176\u4ed6\u53c2\u6570\u5c5e\u4e8e\u8f93\u51fa\u53c2\u6570\uff0c\u7531\u5185\u6838\u8d1f\u8d23\u8bbe\u7f6e\u3002<\/p>\n<p>\u5728 io_setup \u8fd4\u56de\u7684\u65f6\u5019\uff0c\u5185\u6838\u5df2\u7ecf\u521d\u59cb\u5316\u597d\u4e86 SQ \u548c CQ\uff0c\u6b64\u5916\uff0c\u8fd8\u6709\u5185\u6838\u8fd8\u63d0\u4f9b\u4e86\u4e00\u4e2a Submission Queue Entries\uff08SQEs\uff09\u6570\u7ec4\u3002<\/p>\n<figure data-size=\"normal\"><img decoding=\"async\" class=\"origin_image zh-lightbox-thumb lazy\" src=\"https:\/\/pic2.zhimg.com\/80\/v2-212278cc2fe40d4ce708a93b119b9031_720w.png\" width=\"888\" data-caption=\"\" data-size=\"normal\" data-rawwidth=\"888\" data-rawheight=\"390\" data-original=\"https:\/\/pic2.zhimg.com\/v2-212278cc2fe40d4ce708a93b119b9031_r.jpg\" data-actualsrc=\"https:\/\/pic2.zhimg.com\/v2-212278cc2fe40d4ce708a93b119b9031_b.png\" data-lazy-status=\"ok\" \/><\/figure>\n<p>\u4e4b\u6240\u4ee5\u989d\u5916\u91c7\u7528\u4e86\u4e00\u4e2a\u6570\u7ec4\u4fdd\u5b58 SQEs\uff0c\u662f\u4e3a\u4e86\u65b9\u4fbf\u901a\u8fc7 RingBuffer \u63d0\u4ea4\u5185\u5b58\u4e0a\u4e0d\u8fde\u7eed\u7684\u8bf7\u6c42\u3002SQ \u548c CQ \u4e2d\u6bcf\u4e2a\u8282\u70b9\u4fdd\u5b58\u7684\u90fd\u662f SQEs \u6570\u7ec4\u7684\u504f\u79fb\u91cf\uff0c\u800c\u4e0d\u662f\u5b9e\u9645\u7684\u8bf7\u6c42\uff0c\u5b9e\u9645\u7684\u8bf7\u6c42\u53ea\u4fdd\u5b58\u5728 SQEs \u6570\u7ec4\u4e2d\u3002\u8fd9\u6837\u5728\u63d0\u4ea4\u8bf7\u6c42\u65f6\uff0c\u5c31\u53ef\u4ee5\u6279\u91cf\u63d0\u4ea4\u4e00\u7ec4 SQEs \u4e0a\u4e0d\u8fde\u7eed\u7684\u8bf7\u6c42\u3002<\/p>\n<p>\u4f46\u7531\u4e8e SQ\uff0cCQ\uff0cSQEs \u662f\u5728\u5185\u6838\u4e2d\u5206\u914d\u7684\uff0c\u6240\u4ee5\u7528\u6237\u6001\u7a0b\u5e8f\u5e76\u4e0d\u80fd\u76f4\u63a5\u8bbf\u95ee\u3002io_setup \u7684\u8fd4\u56de\u503c\u662f\u4e00\u4e2a fd\uff0c\u5e94\u7528\u7a0b\u5e8f\u4f7f\u7528\u8fd9\u4e2a fd \u8fdb\u884c mmap\uff0c\u548c kernel \u5171\u4eab\u4e00\u5757\u5185\u5b58\u3002<\/p>\n<p>\u8fd9\u5757\u5185\u5b58\u5171\u5206\u4e3a\u4e09\u4e2a\u533a\u57df\uff0c\u5206\u522b\u662f SQ\uff0cCQ\uff0cSQEs\u3002kernel \u8fd4\u56de\u7684 io_sqring_offset \u548c io_cqring_offset \u5206\u522b\u63cf\u8ff0\u4e86 SQ \u548c CQ \u7684\u6307\u9488\u5728 mmap \u4e2d\u7684 offset\u3002\u800c SQEs \u5219\u76f4\u63a5\u5bf9\u5e94\u4e86 mmap \u4e2d\u7684 SQEs \u533a\u57df\u3002<\/p>\n<p>mmap \u7684\u65f6\u5019\u9700\u8981\u4f20\u5165 MAP_POPULATE \u53c2\u6570\uff0c\u4ee5\u9632\u6b62\u5185\u5b58\u88ab page fault\u3002<\/p>\n<h2>IO \u63d0\u4ea4<\/h2>\n<p>IO \u63d0\u4ea4\u7684\u505a\u6cd5\u662f\u627e\u5230\u4e00\u4e2a\u7a7a\u95f2\u7684 SQE\uff0c\u6839\u636e\u8bf7\u6c42\u8bbe\u7f6e SQE\uff0c\u5e76\u5c06\u8fd9\u4e2a SQE \u7684\u7d22\u5f15\u653e\u5230 SQ \u4e2d\u3002SQ \u662f\u4e00\u4e2a\u5178\u578b\u7684 RingBuffer\uff0c\u6709 head\uff0ctail \u4e24\u4e2a\u6210\u5458\uff0c\u5982\u679c head == tail\uff0c\u610f\u5473\u7740\u961f\u5217\u4e3a\u7a7a\u3002SQE \u8bbe\u7f6e\u5b8c\u6210\u540e\uff0c\u9700\u8981\u4fee\u6539 SQ \u7684 tail\uff0c\u4ee5\u8868\u793a\u5411 RingBuffer \u4e2d\u63d2\u5165\u4e00\u4e2a\u8bf7\u6c42\u3002<\/p>\n<p>\u5f53\u6240\u6709\u8bf7\u6c42\u90fd\u52a0\u5165 SQ \u540e\uff0c\u5c31\u53ef\u4ee5\u4f7f\u7528 \uff1a<\/p>\n<div class=\"highlight\">\n<pre><code class=\"language-text\">int io_uring_enter(unsigned int fd, u32 to_submit, u32 min_complete, u32 flags);<\/code><\/pre>\n<\/div>\n<p>\u6765\u63d0\u4ea4 IO \u8bf7\u6c42\u3002<\/p>\n<p>io_uring_enter \u88ab\u8c03\u7528\u540e\u4f1a\u9677\u5165\u5230\u5185\u6838\uff0c\u5185\u6838\u5c06 SQ \u4e2d\u7684\u8bf7\u6c42\u63d0\u4ea4\u7ed9 Block \u5c42\u3002to_submit \u8868\u793a\u4e00\u6b21\u63d0\u4ea4\u591a\u5c11\u4e2a IO\u3002<\/p>\n<p>\u5982\u679c flags \u8bbe\u7f6e\u4e86 IORING_ENTER_GETEVENTS\uff0c\u5e76\u4e14 min_complete &gt; 0\uff0c\u90a3\u4e48\u8fd9\u4e2a\u7cfb\u7edf\u8c03\u7528\u4f1a\u540c\u65f6\u5904\u7406 IO \u6536\u5272\u3002\u8fd9\u4e2a\u7cfb\u7edf\u8c03\u7528\u4f1a\u4e00\u76f4 block\uff0c\u76f4\u5230 min_complete \u4e2a IO \u5df2\u7ecf\u5b8c\u6210\u3002<\/p>\n<p>\u8fd9\u4e2a\u6d41\u7a0b\u8c8c\u4f3c\u548c libaio \u6ca1\u6709\u4ec0\u4e48\u533a\u522b\uff0cIO \u63d0\u4ea4\u7684\u8fc7\u7a0b\u4e2d\u4f9d\u7136\u4f1a\u4ea7\u751f\u7cfb\u7edf\u8c03\u7528\u3002<\/p>\n<p><b>\u4f46 io_uring \u7684\u7cbe\u9ad3\u5728\u4e8e\uff0c\u63d0\u4f9b\u4e86 submission offload \u6a21\u5f0f\uff0c\u4f7f\u5f97\u63d0\u4ea4\u8fc7\u7a0b\u5b8c\u5168\u4e0d\u9700\u8981\u8fdb\u884c\u7cfb\u7edf\u8c03\u7528\u3002<\/b><\/p>\n<p>\u5982\u679c\u5728\u8c03\u7528 io_uring_setup \u65f6\u8bbe\u7f6e\u4e86 IORING_SETUP_SQPOLL \u7684 flag\uff0c\u5185\u6838\u4f1a\u989d\u5916\u542f\u52a8\u4e00\u4e2a\u5185\u6838\u7ebf\u7a0b\uff0c\u6211\u4eec\u79f0\u4f5c SQ \u7ebf\u7a0b\u3002\u8fd9\u4e2a\u5185\u6838\u7ebf\u7a0b\u53ef\u4ee5\u8fd0\u884c\u5728\u67d0\u4e2a\u6307\u5b9a\u7684 core \u4e0a\uff08\u901a\u8fc7 sq_thread_cpu \u914d\u7f6e\uff09\u3002\u8fd9\u4e2a\u5185\u6838\u7ebf\u7a0b\u4f1a\u4e0d\u505c\u7684 Poll SQ\uff0c\u9664\u975e\u5728\u4e00\u6bb5\u65f6\u95f4\u5185\u6ca1\u6709 Poll \u5230\u4efb\u4f55\u8bf7\u6c42\uff08\u901a\u8fc7 sq_thread_idle \u914d\u7f6e\uff09\uff0c\u624d\u4f1a\u88ab\u6302\u8d77\u3002<\/p>\n<figure data-size=\"normal\"><img decoding=\"async\" class=\"origin_image zh-lightbox-thumb lazy\" src=\"https:\/\/pic3.zhimg.com\/80\/v2-549f6af4518938aed8f7065d2c5d1022_720w.png\" width=\"602\" data-caption=\"\" data-size=\"normal\" data-rawwidth=\"602\" data-rawheight=\"531\" data-original=\"https:\/\/pic3.zhimg.com\/v2-549f6af4518938aed8f7065d2c5d1022_r.jpg\" data-actualsrc=\"https:\/\/pic3.zhimg.com\/v2-549f6af4518938aed8f7065d2c5d1022_b.png\" data-lazy-status=\"ok\" \/><\/figure>\n<p>\u5f53\u7a0b\u5e8f\u5728\u7528\u6237\u6001\u8bbe\u7f6e\u5b8c SQE\uff0c\u5e76\u901a\u8fc7\u4fee\u6539 SQ \u7684 tail \u5b8c\u6210\u4e00\u6b21\u63d2\u5165\u65f6\uff0c\u5982\u679c\u6b64\u65f6 SQ \u7ebf\u7a0b\u5904\u4e8e\u5524\u9192\u72b6\u6001\uff0c\u90a3\u4e48\u53ef\u4ee5\u7acb\u523b\u6355\u83b7\u5230\u8fd9\u6b21\u63d0\u4ea4\uff0c\u8fd9\u6837\u5c31\u907f\u514d\u4e86\u7528\u6237\u7a0b\u5e8f\u8c03\u7528 io_uring_enter \u8fd9\u4e2a\u7cfb\u7edf\u8c03\u7528\u3002\u5982\u679c SQ \u7ebf\u7a0b\u5904\u4e8e\u4f11\u7720\u72b6\u6001\uff0c\u5219\u9700\u8981\u901a\u8fc7\u8c03\u7528 io_uring_enter\uff0c\u5e76\u4f7f\u7528 IORING_SQ_NEED_WAKEUP \u53c2\u6570\uff0c\u6765\u5524\u9192 SQ \u7ebf\u7a0b\u3002\u7528\u6237\u6001\u53ef\u4ee5\u901a\u8fc7 sqring \u7684 flags \u53d8\u91cf\u83b7\u53d6 SQ \u7ebf\u7a0b\u7684\u72b6\u6001\u3002<\/p>\n<h2>IO \u6536\u5272<\/h2>\n<p>\u5f53 IO \u5b8c\u6210\u65f6\uff0c\u5185\u6838\u8d1f\u8d23\u5c06\u5b8c\u6210 IO \u5728 SQEs \u4e2d\u7684 index \u653e\u5230 CQ \u4e2d\u3002\u7531\u4e8e IO \u5728\u63d0\u4ea4\u7684\u65f6\u5019\u53ef\u4ee5\u987a\u4fbf\u8fd4\u56de\u5b8c\u6210\u7684 IO\uff0c\u6240\u4ee5\u6536\u5272 IO \u4e0d\u9700\u8981\u989d\u5916\u7cfb\u7edf\u8c03\u7528\u3002\u8fd9\u662f\u8ddf libaio \u6bd4\u8f83\u5927\u7684\u4e0d\u540c\uff0c\u7701\u53bb\u4e86\u4e00\u6b21\u7cfb\u7edf\u8c03\u7528\u3002<\/p>\n<p>\u5982\u679c\u4f7f\u7528\u4e86 IORING_SETUP_SQPOLL \u53c2\u6570\uff0cIO \u6536\u5272\u4e5f\u4e0d\u9700\u8981\u7cfb\u7edf\u8c03\u7528\u7684\u53c2\u4e0e\u3002\u7531\u4e8e\u5185\u6838\u548c\u7528\u6237\u6001\u5171\u4eab\u5185\u5b58\uff0c\u6240\u4ee5\u6536\u5272\u7684\u65f6\u5019\uff0c\u7528\u6237\u6001\u904d\u5386 [cring-&gt;head, cring-&gt;tail) \u533a\u95f4\uff0c\u8fd9\u662f\u5df2\u7ecf\u5b8c\u6210\u7684 IO \u961f\u5217\uff0c\u7136\u540e\u627e\u5230\u76f8\u5e94\u7684 CQE \u5e76\u8fdb\u884c\u5904\u7406\uff0c\u6700\u540e\u79fb\u52a8 head \u6307\u9488\u5230 tail\uff0cIO \u6536\u5272\u5c31\u5230\u6b64\u7ed3\u675f\u4e86\u3002<\/p>\n<p>\u7531\u4e8e\u63d0\u4ea4\u548c\u6536\u5272\u7684\u65f6\u5019\u9700\u8981\u8bbf\u95ee\u5171\u4eab\u5185\u5b58\u7684 head\uff0ctail \u6307\u9488\uff0c\u6240\u4ee5\u9700\u8981\u4f7f\u7528 rmb\/wmb \u5185\u5b58\u5c4f\u969c\u64cd\u4f5c\u786e\u4fdd\u65f6\u5e8f\u3002<\/p>\n<p>\u6240\u4ee5\u5728\u6700\u7406\u60f3\u7684\u60c5\u51b5\u4e0b\uff0cIO \u63d0\u4ea4\u548c\u6536\u5272\u90fd\u4e0d\u9700\u8981\u4f7f\u7528\u7cfb\u7edf\u8c03\u7528\u3002<\/p>\n<h2>\u5176\u5b83\u9ad8\u7ea7\u7279\u6027<\/h2>\n<p>io_uring \u652f\u6301\u8fd8\u652f\u6301\u4ee5\u4e0b\u7279\u6027\u3002<\/p>\n<h3>IORING_REGISTER_FILES<\/h3>\n<p>\u8fd9\u4e2a\u7684\u7528\u9014\u662f\u907f\u514d\u6bcf\u6b21 IO \u5bf9\u6587\u4ef6\u505a fget\/fput \u64cd\u4f5c\uff0c\u5f53\u6279\u91cf IO \u7684\u65f6\u5019\uff0c\u8fd9\u7ec4\u539f\u5b50\u64cd\u4f5c\u53ef\u4ee5\u907f\u514d\u6389\u3002<\/p>\n<h3>IORING_SETUP_IOPOLL<\/h3>\n<p>\u8fd9\u4e2a\u529f\u80fd\u8ba9\u5185\u6838\u91c7\u7528 Polling \u7684\u6a21\u5f0f\u6536\u5272 Block \u5c42\u7684\u8bf7\u6c42\u3002\u5f53\u6ca1\u6709\u4f7f\u7528 SQ \u7ebf\u7a0b\u65f6\uff0cio_uring_enter \u51fd\u6570\u4f1a\u4e3b\u52a8\u7684 Poll\uff0c\u4ee5\u68c0\u67e5\u63d0\u4ea4\u7ed9 Block \u5c42\u7684\u8bf7\u6c42\u662f\u5426\u5df2\u7ecf\u5b8c\u6210\uff0c\u800c\u4e0d\u662f\u6302\u8d77\uff0c\u5e76\u7b49\u5f85 Block \u5c42\u5b8c\u6210\u540e\u518d\u88ab\u5524\u9192\u3002\u4f7f\u7528 SQ \u7ebf\u7a0b\u65f6\u4e5f\u662f\u540c\u7406\u3002<\/p>\n<p>\u901a\u8fc7 perf \u53ef\u4ee5\u770b\u5230\uff0c\u5f53\u4f7f\u7528 IOPOLL \u65f6\uff0c88% \u7684 CPU \u65f6\u95f4\u82b1\u8d39\u5728\u8c03\u7528 blkdev_iopoll \u548c blk_poll \u4e0a\u3002<\/p>\n<figure data-size=\"normal\"><img decoding=\"async\" class=\"origin_image zh-lightbox-thumb lazy\" src=\"https:\/\/pic1.zhimg.com\/80\/v2-01b21e6d63e42d0d37a81f67a2c3bf64_720w.png\" width=\"698\" data-caption=\"\" data-size=\"normal\" data-rawwidth=\"698\" data-rawheight=\"240\" data-original=\"https:\/\/pic1.zhimg.com\/v2-01b21e6d63e42d0d37a81f67a2c3bf64_r.jpg\" data-actualsrc=\"https:\/\/pic1.zhimg.com\/v2-01b21e6d63e42d0d37a81f67a2c3bf64_b.png\" data-lazy-status=\"ok\" \/><\/figure>\n<h3>IORING_REGISTER_BUFFERS<\/h3>\n<p>\u5982\u679c\u5e94\u7528\u63d0\u4ea4\u5230\u5185\u6838\u7684\u865a\u62df\u5185\u5b58\u5730\u5740\u662f\u56fa\u5b9a\u7684\uff0c\u90a3\u4e48\u53ef\u4ee5\u63d0\u524d\u5b8c\u6210\u865a\u62df\u5730\u5740\u5230\u7269\u7406 pages \u7684\u6620\u5c04\uff0c\u907f\u514d\u5728 IO \u8def\u5f84\u4e0a\u8fdb\u884c\u8f6c\u6362\uff0c\u4ece\u800c\u4f18\u5316\u6027\u80fd\u3002\u7528\u6cd5\u662f\uff0c\u5728 setup io_uring \u4e4b\u540e\uff0c\u8c03\u7528 io_uring_register\uff0c\u4f20\u9012 IORING_REGISTER_BUFFERS \u4f5c\u4e3a opcode\uff0c\u53c2\u6570\u662f\u4e00\u4e2a\u6307\u5411 iovec \u7684\u6570\u7ec4\uff0c\u8868\u793a\u8fd9\u4e9b\u5730\u5740\u9700\u8981 map \u5230\u5185\u6838\u3002\u5728\u505a IO \u7684\u65f6\u5019\uff0c\u4f7f\u7528\u5e26 FIXED \u7248\u672c\u7684opcode\uff08IORING_OP_READ_FIXED \/IORING_OP_WRITE_FIXED\uff09\u6765\u64cd\u4f5c IO \u5373\u53ef\u3002<\/p>\n<p>\u5185\u6838\u5728\u5904\u7406 IORING_REGISTER_BUFFERS \u65f6\uff0c\u63d0\u524d\u4f7f\u7528 get_user_pages \u6765\u83b7\u5f97 userspace \u865a\u62df\u5730\u5740\u5bf9\u5e94\u7684\u7269\u7406 pages\u3002\u5728\u505a IO \u7684\u65f6\u5019\uff0c\u5982\u679c\u63d0\u4ea4\u7684\u865a\u62df\u5730\u5740\u66fe\u7ecf\u88ab\u6ce8\u518c\u8fc7\uff0c\u90a3\u4e48\u5c31\u514d\u53bb\u4e86\u865a\u62df\u5730\u5740\u5230 pages \u7684\u8f6c\u6362\u3002<\/p>\n<p>\u4e0b\u9762\u662f\u4e24\u4e2a\u7248\u672c\u7684 perf \u6570\u636e\u3002<\/p>\n<p>\u5e26 fixed buffer\uff1a<\/p>\n<figure data-size=\"normal\"><img decoding=\"async\" class=\"origin_image zh-lightbox-thumb lazy\" src=\"https:\/\/pic2.zhimg.com\/80\/v2-b792ddde25c6cb6463a390a7425fab09_720w.png\" width=\"558\" data-caption=\"\" data-size=\"normal\" data-rawwidth=\"558\" data-rawheight=\"540\" data-original=\"https:\/\/pic2.zhimg.com\/v2-b792ddde25c6cb6463a390a7425fab09_r.jpg\" data-actualsrc=\"https:\/\/pic2.zhimg.com\/v2-b792ddde25c6cb6463a390a7425fab09_b.png\" data-lazy-status=\"ok\" \/><\/figure>\n<p>\u4e0d\u5e26 fixed buffer\uff1a<\/p>\n<figure data-size=\"normal\"><img decoding=\"async\" class=\"origin_image zh-lightbox-thumb lazy\" src=\"https:\/\/pic1.zhimg.com\/80\/v2-7d6017d82bd6b06c5a7b4142bcc72e24_720w.png\" width=\"558\" data-caption=\"\" data-size=\"normal\" data-rawwidth=\"558\" data-rawheight=\"540\" data-original=\"https:\/\/pic1.zhimg.com\/v2-7d6017d82bd6b06c5a7b4142bcc72e24_r.jpg\" data-actualsrc=\"https:\/\/pic1.zhimg.com\/v2-7d6017d82bd6b06c5a7b4142bcc72e24_b.png\" data-lazy-status=\"ok\" \/><\/figure>\n<p>\u53ef\u4ee5\u660e\u663e\u770b\u5230\uff0c\u63d0\u524d map pages\uff0c\u53ef\u4ee5\u51cf\u5c11 iov_iter_get_pages 7% \u7684 CPU \u65f6\u95f4\u6d88\u8017\u3002<\/p>\n<h2>\u5173\u4e8e\u540d\u5b57<\/h2>\n<p>\u53d6\u540d\u4e00\u76f4\u662f\u4e00\u4e2a\u8001\u5927\u96be\u7684\u95ee\u9898\uff0cio_uring \u8fd9\u4e2a\u540d\u5b57\uff0c\u6709\u70b9\u610f\u601d\uff0c\u793e\u533a\u6709\u4eba\u5410\u69fd\uff0c\u770b\u8d77\u6765\u50cf io urine\uff08*\u4e0d\u662f\u4e00\u4e2a\u5f88\u597d\u7684\u8bcd*\uff09\uff0c\u592a\u6b22\u4e50\u4e86\u3002<\/p>\n<p>\u6709\u4eba\u8bf4\u53ef\u4ee5\u53eb\u505a aio_ring\uff0cio_ring\uff0cring_io\u3002<\/p>\n<h2>\u603b\u7ed3<\/h2>\n<p>io_uring \u7684\u63a5\u53e3\u867d\u7136\u7b80\u5355\uff0c\u4f46\u64cd\u4f5c\u8d77\u6765\u6709\u4e9b\u590d\u6742\uff0c\u9700\u8981\u624b\u52a8 mmap \u6765\u6620\u5c04\u5185\u5b58\u3002\u53ef\u4ee5\u770b\u5230\uff0cio_uring \u662f\u5b8c\u5168\u4e3a\u6027\u80fd\u800c\u751f\u7684\u65b0\u4e00\u4ee3 native async IO \u6a21\u578b\uff0c\u6bd4 libaio \u9ad8\u7ea7\u4e0d\u5c11\u3002\u901a\u8fc7\u5168\u65b0\u7684\u8bbe\u8ba1\uff0c\u5171\u4eab\u5185\u5b58\uff0cIO \u8fc7\u7a0b\u4e0d\u9700\u8981\u7cfb\u7edf\u8c03\u7528\uff0c\u7531\u5185\u6838\u5b8c\u6210 IO \u7684\u63d0\u4ea4\uff0c \u4ee5\u53ca IO completion polling \u673a\u5236\uff0c\u5b9e\u73b0\u4e86\u9ad8IOPS\uff0c\u9ad8 Bandwidth\u3002\u76f8\u6bd4 kernel bypass\uff0c\u8fd9\u79cd native \u7684\u65b9\u5f0f\u663e\u5f97\u53cb\u597d\u4e00\u4e9b\u3002<\/p>\n<p>\u5f53\u7136\uff0c\u4e0d\u53ef\u5426\u8ba4\uff0caio \u4e5f\u5728\u4e0e\u65f6\u4ff1\u8fdb\u3002\u81ea\u4ece kernel 2.5 \u8fdb\u5165 upstream \u4ee5\u6765\uff0caio \u4e00\u76f4\u90fd\u6ca1\u6709\u5b9e\u73b0\u5b8c\u6574\u3002 aio \u5bf9 Direct IO \u652f\u6301\u7684\u5f88\u597d\uff0c\u4f46\u662f\u5176\u4ed6\u7684 IO \u7c7b\u578b\u652f\u6301\u7684\u4e0d\u5b8c\u5584\u3002\u5c1d\u8bd5\u4f7f\u7528\u5176\u4ed6\u7c7b\u578b\u7684 IO\uff0c\u4f8b\u5982 buffered IO\uff0c\u53ef\u80fd\u5bfc\u81f4\u540c\u6b65\u7684\u884c\u4e3a\u3002polling \u4e5f\u662f\u4e00\u4e2a\u65b9\u5411\uff0c\u6700\u8fd1 aio \u7684 polling \u673a\u5236\u5df2\u7ecf\u5b9e\u73b0\uff0c\u611f\u5174\u8da3\u7684\u53ef\u4ee5\u5c1d\u8bd5\u4e00\u4e0b\u3002<\/p>\n<h2>\u53c2\u8003<\/h2>\n<ol>\n<li><u><a class=\" external\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/lore.kernel.org\/linux-block\/20190116175003.17880-1-axboe%40kernel.dk\/\" target=\"_blank\" rel=\"nofollow noopener noreferrer\" data-za-detail-view-id=\"1043\"><span class=\"invisible\">https:\/\/<\/span><span class=\"visible\">lore.kernel.org\/linux-b<\/span><span class=\"invisible\">lock\/20190116175003.17880-1-axboe@kernel.dk\/<\/span><\/a><\/u><\/li>\n<li><u><a class=\" external\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/lwn.net\/ml\/linux-fsdevel\/20190109160036.GK6310%40bombadil.infradead.org\/\" target=\"_blank\" rel=\"nofollow noopener noreferrer\" data-za-detail-view-id=\"1043\"><span class=\"invisible\">https:\/\/<\/span><span class=\"visible\">lwn.net\/ml\/linux-fsdeve<\/span><span class=\"invisible\">l\/20190109160036.GK6310@bombadil.infradead.org\/<\/span><\/a><\/u><\/li>\n<li><u><a class=\" external\" href=\"https:\/\/link.zhihu.com\/?target=http%3A\/\/git.kernel.dk\/cgit\/fio\/plain\/t\/io_uring.c\" target=\"_blank\" rel=\"nofollow noopener noreferrer\" data-za-detail-view-id=\"1043\"><span class=\"invisible\">http:\/\/<\/span><span class=\"visible\">git.kernel.dk\/cgit\/fio\/<\/span><span class=\"invisible\">plain\/t\/io_uring.c<\/span><\/a><\/u><\/li>\n<li><u><a class=\" external\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/lore.kernel.org\/linux-block\/20190211190049.7888-14-axboe%40kernel.dk\/\" target=\"_blank\" rel=\"nofollow noopener noreferrer\" data-za-detail-view-id=\"1043\"><span class=\"invisible\">https:\/\/<\/span><span class=\"visible\">lore.kernel.org\/linux-b<\/span><span class=\"invisible\">lock\/20190211190049.7888-14-axboe@kernel.dk\/<\/span><\/a><\/u><\/li>\n<li><u><a class=\" external\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/lwn.net\/Articles\/743714\/\" target=\"_blank\" rel=\"nofollow noopener noreferrer\" data-za-detail-view-id=\"1043\"><span class=\"invisible\">https:\/\/<\/span><span class=\"visible\">lwn.net\/Articles\/743714<\/span><span class=\"invisible\">\/<\/span><\/a><\/u><\/li>\n<li><a class=\" wrap external\" href=\"https:\/\/link.zhihu.com\/?target=http%3A\/\/git.kernel.dk\/cgit\/liburing\/tree\/man\/io_uring_setup.2\" target=\"_blank\" rel=\"nofollow noopener noreferrer\" data-za-detail-view-id=\"1043\">io_uring_setup.2\\man &#8211; liburing &#8211; io_uring library<\/a><\/li>\n<\/ol>\n<p><b>\u4f5c\u8005\u4ecb\u7ecd<\/b><\/p>\n<div class=\"Popover\">\n<div id=\"Popover11-toggle\" aria-haspopup=\"true\" aria-expanded=\"false\" aria-owns=\"Popover11-content\"><a class=\"UserLink-link\" href=\"https:\/\/www.zhihu.com\/people\/868bcd4de1b33b62cb96914a0d958479\" target=\"_blank\" rel=\"noopener\" data-za-detail-view-element_name=\"User\">@panic<\/a><\/div>\n<\/div>\n<p>\uff0cSmartX \u5b58\u50a8\u7814\u53d1\u5de5\u7a0b\u5e08\u3002<u><a class=\" wrap external\" href=\"https:\/\/link.zhihu.com\/?target=http%3A\/\/www.smartx.com\/\" target=\"_blank\" rel=\"nofollow noopener noreferrer\" data-za-detail-view-id=\"1043\">SmartX<\/a><\/u>\u00a0\u62e5\u6709\u56fd\u5185\u6700\u9876\u5c16\u7684\u5206\u5e03\u5f0f\u5b58\u50a8\u548c\u8d85\u878d\u5408\u67b6\u6784\u7814\u53d1\u56e2\u961f\uff0c\u662f\u56fd\u5185\u8d85\u878d\u5408\u9886\u57df\u7684\u6280\u672f\u9886\u5bfc\u8005\u3002<\/p>\n<p>&nbsp;<\/p>\n<\/div>\n<\/div>\n<div class=\"ContentItem-time\">\u53d1\u5e03\u4e8e 2019-04-17<\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u539f\u6587\u5730\u5740 503 \u4eba\u8d5e\u540c\u4e86\u8be5\u6587\u7ae0 \u5728\u6b63\u6587<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"colormag_page_container_layout":"default_layout","colormag_page_sidebar_layout":"default_layout","footnotes":""},"categories":[],"tags":[118],"class_list":["post-5483","post","type-post","status-publish","format-standard","hentry","tag-02-"],"_links":{"self":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/5483","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/comments?post=5483"}],"version-history":[{"count":0,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/5483\/revisions"}],"wp:attachment":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/media?parent=5483"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/categories?post=5483"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/tags?post=5483"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}