os/bluestore: fix deep-scrub operation againest disk silent errors

Say a object who has data caches, but in a while later, caches' underlying
physical device has silent disk erros accidentally, then caches and physical
data are not same. In such case, deep-scrub operation still tries to read
caches firstly and won't do crc checksum, then deep-scrub won't find such
data corruptions timely.

Here introduce a new flag 'CEPH_OSD_OP_FLAG_BYPASS_CLEAN_CACHE' which tells
deep-scrub to bypass object caches. Note that we only bypass cache who is in
STATE_CLEAN state. For STATE_WRITING caches, currently they are not written
to physical device, so deep-scrub operation can not read physical device and
can read these dirty caches safely. Once they are in STATE_CLEAN state(or not
added to bluestore cache), next round deep-scurb can check them correctly.

As to above discussions, I refactor BlueStore::BufferSpace::read sightly,
adding a new 'flags' argument, whose value will be 0 or:
     enum {
       BYPASS_CLEAN_CACHE = 0x1,     // bypass clean cache
     };

flags 0: normal read, do not bypass clean or dirty cache
flags BYPASS_CLEAN_CACHE: bypass clean cache, currently only for deep-scrube
                        operation

Test:
   I deliberately corrupt a object with cache, with this patch, deep-scrub
   can find data error very timely.

Signed-off-by: Xiaoguang Wang <xiaoguang.wang@easystack.cn>
(cherry picked from commit a7f1af2)

Conflicts:
	src/include/rados.h
	src/os/bluestore/BlueStore.cc: trivial resolution

src/include/rados.h

@@ -463,6 +463,7 @@ enum {
 	CEPH_OSD_OP_FLAG_FADVISE_WILLNEED   = 0x10,/* data will be accessed in the near future */
 	CEPH_OSD_OP_FLAG_FADVISE_DONTNEED   = 0x20,/* data will not be accessed in the near future */
 	CEPH_OSD_OP_FLAG_FADVISE_NOCACHE   = 0x40, /* data will be accessed only once by this client */
+	CEPH_OSD_OP_FLAG_BYPASS_CLEAN_CACHE = 0x100, /* bypass ObjectStore cache, mainly for deep-scrub */
 };
 
 #define EOLDSNAPC    85  /* ORDERSNAP flag set; writer has old snapc*/

src/os/bluestore/BlueStore.cc

@@ -1302,7 +1302,8 @@ void BlueStore::BufferSpace::read(
   uint32_t offset,
   uint32_t length,
   BlueStore::ready_regions_t& res,
-  interval_set<uint32_t>& res_intervals)
+  interval_set<uint32_t>& res_intervals,
+  int flags)
 {
   res.clear();
   res_intervals.clear();
@@ -1316,7 +1317,13 @@ void BlueStore::BufferSpace::read(
          ++i) {
       Buffer *b = i->second.get();
       assert(b->end() > offset);
-      if (b->is_writing() || b->is_clean()) {
+
+      bool val = false;
+      if (flags & BYPASS_CLEAN_CACHE)
+        val = b->is_writing();
+      else
+        val = b->is_writing() || b->is_clean();
+      if (val) {
         if (b->offset < offset) {
 	  uint32_t skip = offset - b->offset;
 	  uint32_t l = MIN(length, b->length - skip);
@@ -6693,6 +6700,7 @@ int BlueStore::_do_read(
 {
   FUNCTRACE();
   int r = 0;
+  int read_cache_policy = 0; // do not bypass clean or dirty cache
 
   dout(20) << __func__ << " 0x" << std::hex << offset << "~" << length
            << " size 0x" << o->onode.size << " (" << std::dec
@@ -6727,6 +6735,13 @@ int BlueStore::_do_read(
 
   ready_regions_t ready_regions;
 
+  // for deep-scrub, we only read dirty cache and bypass clean cache in
+  // order to read underlying block device in case there are silent disk errors.
+  if (op_flags & CEPH_OSD_OP_FLAG_BYPASS_CLEAN_CACHE) {
+    dout(20) << __func__ << " will bypass cache and do direct read" << dendl;
+    read_cache_policy = BufferSpace::BYPASS_CLEAN_CACHE;
+  }
+
   // build blob-wise list to of stuff read (that isn't cached)
   blobs2read_t blobs2read;
   unsigned left = length;
@@ -6752,7 +6767,8 @@ int BlueStore::_do_read(
     ready_regions_t cache_res;
     interval_set<uint32_t> cache_interval;
     bptr->shared_blob->bc.read(
-      bptr->shared_blob->get_cache(), b_off, b_len, cache_res, cache_interval);
+      bptr->shared_blob->get_cache(), b_off, b_len, cache_res, cache_interval,
+      read_cache_policy);
     dout(20) << __func__ << "  blob " << *bptr << std::hex
 	     << " need 0x" << b_off << "~" << b_len
 	     << " cache has 0x" << cache_interval

src/os/bluestore/BlueStore.h

@@ -240,6 +240,10 @@ class BlueStore : public ObjectStore,
 
   /// map logical extent range (object) onto buffers
   struct BufferSpace {
+    enum {
+      BYPASS_CLEAN_CACHE = 0x1,  // bypass clean cache
+    };
+
     typedef boost::intrusive::list<
       Buffer,
       boost::intrusive::member_hook<
@@ -339,7 +343,8 @@ class BlueStore : public ObjectStore,
 
     void read(Cache* cache, uint32_t offset, uint32_t length,
 	      BlueStore::ready_regions_t& res,
-	      interval_set<uint32_t>& res_intervals);
+	      interval_set<uint32_t>& res_intervals,
+	      int flags = 0);
 
     void truncate(Cache* cache, uint32_t offset) {
       discard(cache, offset, (uint32_t)-1 - offset);

src/osd/ReplicatedBackend.cc

@@ -712,7 +712,8 @@ int ReplicatedBackend::be_deep_scrub(
   dout(10) << __func__ << " " << poid << " pos " << pos << dendl;
   int r;
   uint32_t fadvise_flags = CEPH_OSD_OP_FLAG_FADVISE_SEQUENTIAL |
-                           CEPH_OSD_OP_FLAG_FADVISE_DONTNEED;
+                           CEPH_OSD_OP_FLAG_FADVISE_DONTNEED |
+                           CEPH_OSD_OP_FLAG_BYPASS_CLEAN_CACHE;
 
   utime_t sleeptime;
   sleeptime.set_from_double(cct->_conf->osd_debug_deep_scrub_sleep);

src/osd/osd_types.cc

@@ -100,6 +100,9 @@ const char * ceph_osd_op_flag_name(unsigned flag)
     case CEPH_OSD_OP_FLAG_FADVISE_NOCACHE:
       name = "fadvise_nocache";
       break;
+    case CEPH_OSD_OP_FLAG_BYPASS_CLEAN_CACHE:
+      name = "bypass_clean_cache";
+      break;
     default:
       name = "???";
   };