kraken: rgw: rgw_file: fix marker computation

src/common/cohort_lru.h

@@ -143,6 +143,12 @@ namespace cohort {
 	    continue;
 	  // XXXX if object at LRU has refcnt==1, take it
 	  Object* o = &(lane.q.back());
+#if 0 /* XXX save for refactor */
+	  std::cout << __func__
+		    << " " << o
+		    << " refcnt: " << o->lru_refcnt
+		    << std::endl;
+#endif
 	  if (can_reclaim(o)) {
 	    ++(o->lru_refcnt);
 	    o->lru_flags |= FLAG_EVICTING;
@@ -213,6 +219,18 @@ namespace cohort {
 	    delete o;
 	  }
 	  lane.lock.unlock();
+	} else if (unlikely(refcnt == SENTINEL_REFCNT)) {
+	  Lane& lane = lane_of(o);
+	  lane.lock.lock();
+	  refcnt = o->lru_refcnt.load();
+	  if (likely(refcnt == SENTINEL_REFCNT)) {
+	    /* move to LRU */
+	    Object::Queue::iterator it =
+	      Object::Queue::s_iterator_to(*o);
+	    lane.q.erase(it);
+	    lane.q.push_back(*o);
+	  }
+	  lane.lock.unlock();
 	}
       } /* unref */
 

src/common/config_opts.h

@@ -1465,6 +1465,9 @@ OPTION(rgw_nfs_lru_lanes, OPT_INT, 5)
 OPTION(rgw_nfs_lru_lane_hiwat, OPT_INT, 911)
 OPTION(rgw_nfs_fhcache_partitions, OPT_INT, 3)
 OPTION(rgw_nfs_fhcache_size, OPT_INT, 2017) /* 3*2017=6051 */
+OPTION(rgw_nfs_namespace_expire_secs, OPT_INT, 300) /* namespace invalidate
+						     * timer */
+OPTION(rgw_nfs_max_gc, OPT_INT, 300) /* max gc events per cycle */
 OPTION(rgw_nfs_write_completion_interval_s, OPT_INT, 10) /* stateless (V3)
 							  * commit
 							  * delay */

src/include/rados/rgw_file.h

@@ -27,7 +27,7 @@ extern "C" {
 
 #define LIBRGW_FILE_VER_MAJOR 1
 #define LIBRGW_FILE_VER_MINOR 1
-#define LIBRGW_FILE_VER_EXTRA 0
+#define LIBRGW_FILE_VER_EXTRA 3
 
 #define LIBRGW_FILE_VERSION(maj, min, extra) ((maj << 16) + (min << 8) + extra)
 #define LIBRGW_FILE_VERSION_CODE LIBRGW_FILE_VERSION(LIBRGW_FILE_VER_MAJOR, LIBRGW_FILE_VER_MINOR, LIBRGW_FILE_VER_EXTRA)
@@ -36,7 +36,8 @@ extern "C" {
  * object types
  */
 enum rgw_fh_type {
-  RGW_FS_TYPE_FILE = 0,
+  RGW_FS_TYPE_NIL = 0,
+  RGW_FS_TYPE_FILE,
   RGW_FS_TYPE_DIRECTORY,
 };
 
@@ -91,6 +92,12 @@ void rgwfile_version(int *major, int *minor, int *extra);
 */
 #define RGW_LOOKUP_FLAG_NONE    0x0000
 #define RGW_LOOKUP_FLAG_CREATE  0x0001
+#define RGW_LOOKUP_FLAG_RCB     0x0002 /* readdir callback hint */
+#define RGW_LOOKUP_FLAG_DIR     0x0004
+#define RGW_LOOKUP_FLAG_FILE    0x0008
+
+#define RGW_LOOKUP_TYPE_FLAGS \
+  (RGW_LOOKUP_FLAG_DIR|RGW_LOOKUP_FLAG_FILE)
 
 int rgw_lookup(struct rgw_fs *rgw_fs,
 	      struct rgw_file_handle *parent_fh, const char *path,
@@ -119,6 +126,16 @@ int rgw_mount(librgw_t rgw, const char *uid, const char *key,
 	      const char *secret, struct rgw_fs **rgw_fs,
 	      uint32_t flags);
 
+/*
+ register invalidate callbacks
+*/
+#define RGW_REG_INVALIDATE_FLAG_NONE    0x0000
+
+typedef void (*rgw_fh_callback_t)(void *handle, struct rgw_fh_hk fh_hk);
+
+int rgw_register_invalidate(struct rgw_fs *rgw_fs, rgw_fh_callback_t cb,
+			    void *arg, uint32_t flags);
+
 /*
  detach rgw namespace
 */
@@ -189,7 +206,8 @@ int rgw_unlink(struct rgw_fs *rgw_fs,
 /*
     read  directory content
 */
-typedef bool (*rgw_readdir_cb)(const char *name, void *arg, uint64_t offset);
+typedef bool (*rgw_readdir_cb)(const char *name, void *arg, uint64_t offset,
+			       uint32_t flags);
 
 #define RGW_READDIR_FLAG_NONE      0x0000
 #define RGW_READDIR_FLAG_DOTDOT    0x0001 /* send dot names */

src/rgw/librgw.cc

@@ -13,6 +13,7 @@
  */
 #include <sys/types.h>
 #include <string.h>
+#include <chrono>
 
 #include "include/types.h"
 #include "include/rados/librgw.h"
@@ -80,6 +81,8 @@ namespace rgw {
     m_tp.drain(&req_wq);
   }
 
+#define MIN_EXPIRE_S 120
+
   void RGWLibProcess::run()
   {
     /* write completion interval */
@@ -92,6 +95,14 @@ namespace rgw {
     /* gc loop */
     while (! shutdown) {
       lsubdout(cct, rgw, 5) << "RGWLibProcess GC" << dendl;
+
+      /* dirent invalidate timeout--basically, the upper-bound on
+       * inconsistency with the S3 namespace */
+      auto expire_s = cct->_conf->rgw_nfs_namespace_expire_secs;
+
+      /* delay between gc cycles */
+      auto delay_s = std::max(1, std::min(MIN_EXPIRE_S, expire_s/2));
+
       unique_lock uniq(mtx);
     restart:
       int cur_gen = gen;
@@ -106,7 +117,7 @@ namespace rgw {
 	  goto restart; /* invalidated */
       }
       uniq.unlock();
-      std::this_thread::sleep_for(std::chrono::seconds(120));
+      std::this_thread::sleep_for(std::chrono::seconds(delay_s));
     }
   }