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>

1 files changed, 19 insertions, 3 deletions

diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index 90f2a0e8690..25d09c8648f 100644
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -1306,7 +1306,8 @@ 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();
@@ -1320,7 +1321,13 @@ void BlueStore::BufferSpace::read(
          ++i) {
       Buffer *b = i->second.get();
       ceph_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);
@@ -7408,6 +7415,7 @@ int BlueStore::_do_read(
 {
   FUNCTRACE(cct);
   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
@@ -7442,6 +7450,13 @@ 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;
@@ -7467,7 +7482,8 @@ 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