Merge pull request #5685 from ceph/wip-12133

cephfs-data-scan: scrub tag filtering (#12133 and #12145)

src/cls/cephfs/cls_cephfs.cc

@@ -123,6 +123,72 @@ static int accumulate_inode_metadata(cls_method_context_t hctx,
   return 0;
 }
 
+// I want to select objects that have a name ending 00000000
+// and an xattr (scrub_tag) not equal to a specific value.
+// This is so special case that we can't really pretend it's
+// generic, so just fess up and call this the cephfs filter.
+class PGLSCephFSFilter : public PGLSFilter {
+protected:
+  std::string scrub_tag;
+public:
+  int init(bufferlist::iterator& params) {
+    try {
+      InodeTagFilterArgs args;
+      args.decode(params);
+      scrub_tag = args.scrub_tag;
+    } catch (buffer::error &e) {
+      return -EINVAL;
+    }
+
+    if (scrub_tag.empty()) {
+      xattr = "";
+    } else {
+      xattr = "_scrub_tag";
+    }
+
+    return 0;
+  }
+
+  virtual ~PGLSCephFSFilter() {}
+  virtual bool reject_empty_xattr() { return false; }
+  virtual bool filter(const hobject_t &obj, bufferlist& xattr_data,
+                      bufferlist& outdata);
+};
+
+bool PGLSCephFSFilter::filter(const hobject_t &obj,
+                             bufferlist& xattr_data, bufferlist& outdata)
+{
+  const std::string need_ending = ".00000000";
+  const std::string &obj_name = obj.oid.name;
+
+  if (obj_name.length() < need_ending.length()) {
+    return false;
+  }
+
+  const bool match = obj_name.compare (obj_name.length() - need_ending.length(), need_ending.length(), need_ending) == 0;
+  if (!match) {
+    return false;
+  }
+
+  if (!scrub_tag.empty() && xattr_data.length() > 0) {
+    std::string tag_ondisk;
+    bufferlist::iterator q = xattr_data.begin();
+    try {
+      ::decode(tag_ondisk, q);
+      if (tag_ondisk == scrub_tag)
+	return false;
+    } catch (const buffer::error &err) {
+    }
+  }
+
+  return true;
+}
+
+PGLSFilter *inode_tag_filter()
+{
+  return new PGLSCephFSFilter();
+}
+
 /**
  * initialize class
  *
@@ -139,5 +205,8 @@ void __cls_init()
   cls_register_cxx_method(h_class, "accumulate_inode_metadata",
 			  CLS_METHOD_WR | CLS_METHOD_RD,
 			  accumulate_inode_metadata, &h_accumulate_inode_metadata);
+
+  // A PGLS filter
+  cls_register_cxx_filter(h_class, "inode_tag", inode_tag_filter);
 }
 

src/cls/cephfs/cls_cephfs.h

@@ -108,6 +108,26 @@ public:
   }
 };
 
+class InodeTagFilterArgs
+{
+  public:
+    std::string scrub_tag;
+
+  void encode(bufferlist &bl) const
+  {
+    ENCODE_START(1, 1, bl);
+    ::encode(scrub_tag, bl);
+    ENCODE_FINISH(bl);
+  }
+
+  void decode(bufferlist::iterator &bl)
+  {
+    DECODE_START(1, bl);
+    ::decode(scrub_tag, bl);
+    DECODE_FINISH(bl);
+  }
+};
+
 class AccumulateResult
 {
 public:

src/cls/cephfs/cls_cephfs_client.cc

@@ -144,3 +144,18 @@ int ClsCephFSClient::fetch_inode_accumulate_result(
   return 0;
 }
 
+void ClsCephFSClient::build_tag_filter(
+          const std::string &scrub_tag,
+          bufferlist *out_bl)
+{
+  assert(out_bl != NULL);
+
+  // Leading part of bl is un-versioned string naming the filter
+  ::encode(std::string("cephfs.inode_tag"), *out_bl);
+
+  // Filter-specific part of the bl: in our case this is a versioned structure
+  InodeTagFilterArgs args;
+  args.scrub_tag = scrub_tag;
+  args.encode(*out_bl);
+}
+

src/cls/cephfs/cls_cephfs_client.h

@@ -22,5 +22,9 @@ class ClsCephFSClient
       inode_backtrace_t *backtrace,
       ceph_file_layout *layout,
       AccumulateResult *result);
+
+  static void build_tag_filter(
+      const std::string &scrub_tag,
+      bufferlist *out_bl);
 };
 

src/tools/cephfs/DataScan.cc

@@ -52,7 +52,7 @@ bool DataScan::parse_kwarg(
   }
 
   const std::string arg(*i);
-  const std::string val(*(++i));
+  const std::string val(*(i + 1));
 
   if (arg == std::string("--output-dir")) {
     if (driver != NULL) {
@@ -60,6 +60,7 @@ bool DataScan::parse_kwarg(
       *r = -EINVAL;
       return false;
     }
+    dout(4) << "Using local file output to '" << val << "'" << dendl;
     driver = new LocalFileDriver(val, data_io);
     return true;
   } else if (arg == std::string("-n")) {
@@ -80,6 +81,10 @@ bool DataScan::parse_kwarg(
       return false;
     }
     return true;
+  } else if (arg == std::string("--filter-tag")) {
+    filter_tag = val;
+    dout(10) << "Applying tag filter: '" << filter_tag << "'" << dendl;
+    return true;
   } else {
     return false;
   }
@@ -156,6 +161,7 @@ int DataScan::main(const std::vector<const char*> &args)
     driver = new MetadataDriver();
     driver->set_force_corrupt(force_corrupt);
     driver->set_force_init(force_init);
+    dout(4) << "Using metadata pool output" << dendl;
   }
 
   dout(4) << "connecting to RADOS..." << dendl;
@@ -447,7 +453,28 @@ int DataScan::scan_inodes()
   float progress = 0.0;
   librados::NObjectIterator i = data_io.nobjects_begin(n, m);
 #else
-  librados::NObjectIterator i = data_io.nobjects_begin();
+  librados::NObjectIterator i;
+  bool legacy_filtering = false;
+
+  bufferlist filter_bl;
+  ClsCephFSClient::build_tag_filter(filter_tag, &filter_bl);
+
+  // try/catch to deal with older OSDs that don't support
+  // the cephfs pgls filtering mode
+  try {
+    i = data_io.nobjects_begin(filter_bl);
+    dout(4) << "OSDs accepted cephfs object filtering" << dendl;
+  } catch (const std::runtime_error &e) {
+    // A little unfriendly, librados raises std::runtime_error
+    // on pretty much any unhandled I/O return value, such as
+    // the OSD saying -EINVAL because of our use of a filter
+    // mode that it doesn't know about.
+    std::cerr << "OSDs do not support cephfs object filtering: using "
+                 "(slower) fallback mode" << std::endl;
+    legacy_filtering = true;
+    i = data_io.nobjects_begin();
+  }
+
 #endif
   librados::NObjectIterator i_end = data_io.nobjects_end();
 
@@ -484,10 +511,38 @@ int DataScan::scan_inodes()
       continue;
     }
 
-    // We are only interested in 0th objects during this phase: we touched
-    // the other objects during scan_extents
-    if (obj_name_offset != 0) {
-      continue;
+    if (legacy_filtering) {
+      dout(20) << "Applying filter to " << oid << dendl;
+
+      // We are only interested in 0th objects during this phase: we touched
+      // the other objects during scan_extents
+      if (obj_name_offset != 0) {
+        dout(20) << "Non-zeroth object" << dendl;
+        continue;
+      }
+
+      bufferlist scrub_tag_bl;
+      int r = data_io.getxattr(oid, "scrub_tag", scrub_tag_bl);
+      if (r >= 0) {
+        std::string read_tag;
+        bufferlist::iterator q = scrub_tag_bl.begin();
+	try {
+	  ::decode(read_tag, q);
+	  if (read_tag == filter_tag) {
+	    dout(20) << "skipping " << oid << " because it has the filter_tag"
+		     << dendl;
+	    continue;
+	  }
+	} catch (const buffer::error &err) {
+	}
+	dout(20) << "read non-matching tag '" << read_tag << "'" << dendl;
+      } else {
+        dout(20) << "no tag read (" << r << ")" << dendl;
+      }
+
+    } else {
+      assert(obj_name_offset == 0);
+      dout(20) << "OSD matched oid " << oid << dendl;
     }
 
     AccumulateResult accum_res;
@@ -496,7 +551,11 @@ int DataScan::scan_inodes()
     int r = ClsCephFSClient::fetch_inode_accumulate_result(
         data_io, oid, &backtrace, &loaded_layout, &accum_res);
 
-    if (r < 0) {
+    if (r == -EINVAL) {
+      dout(4) << "Accumulated metadata missing from '"
+              << oid << ", did you run scan_extents?" << dendl;
+      continue;
+    } else  if (r < 0) {
       dout(4) << "Unexpected error loading accumulated metadata from '"
               << oid << "': " << cpp_strerror(r) << dendl;
       // FIXME: this creates situation where if a client has a corrupt

src/tools/cephfs/DataScan.h

@@ -237,6 +237,8 @@ class DataScan : public MDSUtility
     bool force_corrupt;
     // Overwrite root objects even if they exist
     bool force_init;
+    // Only scan inodes without this scrub tag
+    string filter_tag;
 
     /**
      * @param r set to error on valid key with invalid value