From: Matt Benjamin <mbenjamin@redhat.com>
Date: Fri, 19 Jan 2018 18:05:27 +0000 (-0500)
Subject: rgw_file: alternate fix deadlock on lru eviction
X-Git-Tag: v10.2.11~60^2
X-Git-Url: http://git.apps.os.sepia.ceph.com/?a=commitdiff_plain;h=dc48e5bc0a2c7446503e471e448adbddcdc7e29e;p=ceph.git

rgw_file: alternate fix deadlock on lru eviction

This change is an alternate fix for two problems found and fixed
by Yao Zongyou <yaozongyou@vip.qq.com>.

The deadlock can be avoided just by not taking it in the recycle
case, which invariantly holds the lock.

The invalidation of the insert iterator by the recyle-path unlink
we'd like to handle as a condition in order to preserve the cached
insertion point optimization we get in the common case.  (The
original behavior was, indeed, incorrect.)

Based on feedback from Yao, removed the RGWFileHandle dtor version
of the unlink check, which I think happened twice.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>
(cherry picked from commit 3cf0880f86b8f7911139c4e3d672cf47420c8f49)
---

diff --git a/src/common/cohort_lru.h b/src/common/cohort_lru.h
index 0121eda9dbe22..5c8850d9d9b73 100644
--- a/src/common/cohort_lru.h
+++ b/src/common/cohort_lru.h
@@ -39,6 +39,7 @@ namespace cohort {
     /* public flag values */
     constexpr uint32_t FLAG_NONE = 0x0000;
     constexpr uint32_t FLAG_INITIAL = 0x0001;
+    constexpr uint32_t FLAG_RECYCLE = 0x0002;
 
     enum class Edge : std::uint8_t
     {
@@ -229,12 +230,14 @@ namespace cohort {
 	}
       } /* unref */
 
-      Object* insert(ObjectFactory* fac, Edge edge, uint32_t flags) {
+      Object* insert(ObjectFactory* fac, Edge edge, uint32_t& flags) {
 	/* use supplied functor to re-use an evicted object, or
 	 * allocate a new one of the descendant type */
 	Object* o = evict_block();
-	if (o)
+	if (o) {
 	  fac->recycle(o); /* recycle existing object */
+	  flags |= FLAG_RECYCLE;
+	}
 	else
 	  o = fac->alloc(); /* get a new one */
 
diff --git a/src/rgw/rgw_file.cc b/src/rgw/rgw_file.cc
index 246500da7147d..fcc540d7e3332 100644
--- a/src/rgw/rgw_file.cc
+++ b/src/rgw/rgw_file.cc
@@ -870,10 +870,6 @@ namespace rgw {
   }
 
   RGWFileHandle::~RGWFileHandle() {
-    /* in the non-delete case, handle may still be in handle table */
-    if (fh_hook.is_linked()) {
-      fs->fh_cache.remove(fh.fh_hk.object, this, FHCache::FLAG_LOCK);
-    }
     /* cond-unref parent */
     if (parent && (! parent->is_root())) {
       /* safe because if parent->unref causes its deletion,
@@ -911,9 +907,11 @@ namespace rgw {
     lsubdout(fs->get_context(), rgw, 17)
       << __func__ << " " << *this
       << dendl;
-    /* remove if still in fh_cache */
+    /* in the non-delete case, handle may still be in handle table */
     if (fh_hook.is_linked()) {
-      fs->fh_cache.remove(fh.fh_hk.object, this, FHCache::FLAG_LOCK);
+      /* in this case, we are being called from a context which holds
+       * the partition lock */
+      fs->fh_cache.remove(fh.fh_hk.object, this, FHCache::FLAG_NONE);
     }
     return true;
   } /* RGWFileHandle::reclaim */
diff --git a/src/rgw/rgw_file.h b/src/rgw/rgw_file.h
index 0cfa5b331cb4d..5863e0eed6c5e 100644
--- a/src/rgw/rgw_file.h
+++ b/src/rgw/rgw_file.h
@@ -1004,16 +1004,25 @@ namespace rgw {
 	/* make or re-use handle */
 	RGWFileHandle::Factory prototype(this, get_inst(), parent, fhk,
 					 obj_name, CREATE_FLAGS(flags));
+	uint32_t iflags{cohort::lru::FLAG_INITIAL};
 	fh = static_cast<RGWFileHandle*>(
 	  fh_lru.insert(&prototype,
 			cohort::lru::Edge::MRU,
-			cohort::lru::FLAG_INITIAL));
+			iflags));
 	if (fh) {
 	  /* lock fh (LATCHED) */
 	  if (flags & RGWFileHandle::FLAG_LOCK)
 	    fh->mtx.lock();
-	  /* inserts, releasing latch */
-	  fh_cache.insert_latched(fh, lat, RGWFileHandle::FHCache::FLAG_UNLOCK);
+	  if (likely(! (iflags & cohort::lru::FLAG_RECYCLE))) {
+	    /* inserts at cached insert iterator, releasing latch */
+	    fh_cache.insert_latched(
+	      fh, lat, RGWFileHandle::FHCache::FLAG_UNLOCK);
+	  } else {
+	    /* recycle step invalidates Latch */
+	    fh_cache.insert(
+	      fhk.fh_hk.object, fh, RGWFileHandle::FHCache::FLAG_NONE);
+	    lat.lock->unlock(); /* !LATCHED */
+	  }
 	  get<1>(fhr) |= RGWFileHandle::FLAG_CREATE;
 	  /* ref parent (non-initial ref cannot fail on valid object) */
 	  if (! parent->is_root()) {