ceph/src/os/bluestore/NVMEDevice.h

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5   *
   6  * Copyright (C) 2015 XSky <haomai@xsky.com>
   7  *
   8  * Author: Haomai Wang <haomaiwang@gmail.com>
   9  *
  10  * This is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License version 2.1, as published by the Free Software
  13  * Foundation.  See file COPYING.
  14  *
  15  */
  16
  17 #ifndef CEPH_OS_BLUESTORE_NVMEDEVICE
  18 #define CEPH_OS_BLUESTORE_NVMEDEVICE
  19
  20 #include <queue>
  21 #include <map>
  22 #include <limits>
  23
  24 // since _Static_assert introduced in c11
  25 #define _Static_assert static_assert
  26
  27
  28 #include "include/interval_set.h"
  29 #include "common/ceph_time.h"
  30 #include "common/Mutex.h"
  31 #include "common/Cond.h"
  32 #include "BlockDevice.h"
  33
  34 enum class IOCommand {
  35   READ_COMMAND,
  36   WRITE_COMMAND,
  37   FLUSH_COMMAND
  38 };
  39
  40 class Task;
  41 class PerfCounters;
  42 class SharedDriverData;
  43
  44 class NVMEDevice : public BlockDevice {
  45   /**
  46    * points to pinned, physically contiguous memory region;
  47    * contains 4KB IDENTIFY structure for controller which is
  48    *  target for CONTROLLER IDENTIFY command during initialization
  49    */
  50   SharedDriverData *driver;
  51   string name;
  52
  53   uint64_t size;
  54   uint64_t block_size;
  55
  56   bool aio_stop;
  57
  58   struct BufferedExtents {
  59     struct Extent {
  60       uint64_t x_len;
  61       uint64_t x_off;
  62       const char *data;
  63       uint64_t data_len;
  64     };
  65     using Offset = uint64_t;
  66     map<Offset, Extent> buffered_extents;
  67     uint64_t left_edge = std::numeric_limits<uint64_t>::max();
  68     uint64_t right_edge = 0;
  69
  70     void verify() {
  71       interval_set<uint64_t> m;
  72       for (auto && it : buffered_extents) {
  73         assert(!m.intersects(it.first, it.second.x_len));
  74         m.insert(it.first, it.second.x_len);
  75       }
  76     }
  77
  78     void insert(uint64_t off, uint64_t len, const char *data) {
  79       auto it = buffered_extents.lower_bound(off);
  80       if (it != buffered_extents.begin()) {
  81         --it;
  82         if (it->first + it->second.x_len <= off)
  83           ++it;
  84       }
  85       uint64_t end = off + len;
  86       if (off < left_edge)
  87         left_edge = off;
  88       if (end > right_edge)
  89         right_edge = end;
  90       while (it != buffered_extents.end()) {
  91         if (it->first >= end)
  92           break;
  93         uint64_t extent_it_end = it->first + it->second.x_len;
  94         assert(extent_it_end >= off);
  95         if (it->first <= off) {
  96           if (extent_it_end > end) {
  97             //         <-     data    ->
  98             // <-            it           ->
  99             it->second.x_len -= (extent_it_end - off);
 100             buffered_extents[end] = Extent{
 101                 extent_it_end - end, it->second.x_off + it->second.x_len + len, it->second.data, it->second.data_len};
 102           } else {
 103             //         <-     data    ->
 104             // <-     it    ->
 105             assert(extent_it_end <= end);
 106             it->second.x_len -= (extent_it_end - off);
 107           }
 108           ++it;
 109         } else {
 110           assert(it->first > off);
 111           if (extent_it_end > end) {
 112             //  <-     data    ->
 113             //      <-           it          ->
 114             uint64_t overlap = end - it->first;
 115             buffered_extents[end] = Extent{
 116                 it->second.x_len - overlap, it->second.x_off + overlap, it->second.data, it->second.data_len};
 117           } else {
 118             //  <-     data    ->
 119             //      <- it ->
 120           }
 121           buffered_extents.erase(it++);
 122         }
 123       }
 124       buffered_extents[off] = Extent{
 125           len, 0, data, len};
 126
 127       if (0)
 128         verify();
 129     }
 130
 131     void memcpy_check(char *dst, uint64_t dst_raw_len, uint64_t dst_off,
 132                       map<Offset, Extent>::iterator &it, uint64_t src_off, uint64_t copylen) {
 133       if (0) {
 134         assert(dst_off + copylen <= dst_raw_len);
 135         assert(it->second.x_off + src_off + copylen <= it->second.data_len);
 136       }
 137       memcpy(dst + dst_off, it->second.data + it->second.x_off + src_off, copylen);
 138     }
 139
 140     uint64_t read_overlap(uint64_t off, uint64_t len, char *buf) {
 141       uint64_t end = off + len;
 142       if (end <= left_edge || off >= right_edge)
 143         return 0;
 144
 145       uint64_t copied = 0;
 146       auto it = buffered_extents.lower_bound(off);
 147       if (it != buffered_extents.begin()) {
 148         --it;
 149         if (it->first + it->second.x_len <= off)
 150           ++it;
 151       }
 152       uint64_t copy_len;
 153       while (it != buffered_extents.end()) {
 154         if (it->first >= end)
 155           break;
 156         uint64_t extent_it_end = it->first + it->second.x_len;
 157         assert(extent_it_end >= off);
 158         if (it->first >= off) {
 159           if (extent_it_end > end) {
 160             //  <-     data    ->
 161             //      <-           it          ->
 162             copy_len = len - (it->first - off);
 163             memcpy_check(buf, len, it->first - off, it, 0, copy_len);
 164           } else {
 165             //  <-     data    ->
 166             //      <- it ->
 167             copy_len = it->second.x_len;
 168             memcpy_check(buf, len, it->first - off, it, 0, copy_len);
 169           }
 170         } else {
 171           if (extent_it_end > end) {
 172             //         <-     data    ->
 173             // <-           it          ->
 174             copy_len = len;
 175             memcpy_check(buf, len, 0, it, off - it->first, copy_len);
 176           } else {
 177             //         <-     data    ->
 178             // <-     it    ->
 179             assert(extent_it_end <= end);
 180             copy_len = it->first + it->second.x_len - off;
 181             memcpy_check(buf, len, 0, it, off - it->first, copy_len);
 182           }
 183         }
 184         copied += copy_len;
 185         ++it;
 186       }
 187       return copied;
 188     }
 189
 190     void clear() {
 191       buffered_extents.clear();
 192       left_edge = std::numeric_limits<uint64_t>::max();
 193       right_edge = 0;
 194     }
 195   };
 196   Mutex buffer_lock;
 197   BufferedExtents buffered_extents;
 198   Task *buffered_task_head = nullptr;
 199
 200   static void init();
 201  public:
 202   SharedDriverData *get_driver() { return driver; }
 203
 204  public:
 205   aio_callback_t aio_callback;
 206   void *aio_callback_priv;
 207
 208   NVMEDevice(CephContext* cct, aio_callback_t cb, void *cbpriv);
 209
 210   bool supported_bdev_label() override { return false; }
 211
 212   void aio_submit(IOContext *ioc) override;
 213
 214   uint64_t get_size() const override {
 215     return size;
 216   }
 217   uint64_t get_block_size() const override {
 218     return block_size;
 219   }
 220
 221   int read(uint64_t off, uint64_t len, bufferlist *pbl,
 222            IOContext *ioc,
 223            bool buffered) override;
 224   int aio_read(
 225     uint64_t off,
 226     uint64_t len,
 227     bufferlist *pbl,
 228     IOContext *ioc) override;
 229   int aio_write(uint64_t off, bufferlist& bl,
 230                 IOContext *ioc,
 231                 bool buffered) override;
 232   int write(uint64_t off, bufferlist& bl, bool buffered) override;
 233   int flush() override;
 234   int read_random(uint64_t off, uint64_t len, char *buf, bool buffered) override;
 235
 236   // for managing buffered readers/writers
 237   int invalidate_cache(uint64_t off, uint64_t len) override;
 238   int open(const string& path) override;
 239   void close() override;
 240   int collect_metadata(string prefix, map<string,string> *pm) const override;
 241 };
 242
 243 #endif