diff --git a/gn/tests.gni b/gn/tests.gni
index d888c6a..2cf9d33 100644
--- a/gn/tests.gni
+++ b/gn/tests.gni
@@ -158,6 +158,7 @@
   "$_tests/PremulAlphaRoundTripTest.cpp",
   "$_tests/PrimitiveProcessorTest.cpp",
   "$_tests/ProxyConversionTest.cpp",
+  "$_tests/ProxyRefTest.cpp",
   "$_tests/ProxyTest.cpp",
   "$_tests/QuickRejectTest.cpp",
   "$_tests/RandomTest.cpp",
diff --git a/include/gpu/GrGpuResource.h b/include/gpu/GrGpuResource.h
index 4597464..29d33df 100644
--- a/include/gpu/GrGpuResource.h
+++ b/include/gpu/GrGpuResource.h
@@ -93,6 +93,8 @@
     bool internalHasRef() const { return SkToBool(fRefCnt); }
 
 private:
+    friend class GrIORefProxy; // needs to forward on wrapped IO calls
+
     void addPendingRead() const {
         this->validate();
         ++fPendingReads;
diff --git a/include/private/GrSurfaceProxy.h b/include/private/GrSurfaceProxy.h
index dd5ece1..1ff6df8 100644
--- a/include/private/GrSurfaceProxy.h
+++ b/include/private/GrSurfaceProxy.h
@@ -50,13 +50,29 @@
 
     void validate() const {
 #ifdef SK_DEBUG    
-        SkASSERT(fRefCnt >= 0);
+        SkASSERT(fRefCnt >= 1);
+        SkASSERT(fPendingReads >= 0);
+        SkASSERT(fPendingWrites >= 0);
+        SkASSERT(fRefCnt + fPendingReads + fPendingWrites >= 1);
+
+        if (fTarget) {
+            SkASSERT(!fPendingReads && !fPendingWrites);
+            // The backing GrSurface can have more refs than the proxy if the proxy
+            // started off wrapping an external resource (that came in with refs).
+            // The GrSurface should never have fewer refs than the proxy however.
+            SkASSERT(fTarget->fRefCnt >= fRefCnt);
+        }
 #endif
     }
 
+    int32_t getProxyRefCnt_TestOnly() const;
+    int32_t getBackingRefCnt_TestOnly() const;
+    int32_t getPendingReadCnt_TestOnly() const;
+    int32_t getPendingWriteCnt_TestOnly() const;
+
 protected:
-    GrIORefProxy() : fRefCnt(1), fTarget(nullptr) {}
-    GrIORefProxy(sk_sp<GrSurface> surface) : fRefCnt(1) {
+    GrIORefProxy() : fTarget(nullptr), fRefCnt(1), fPendingReads(0), fPendingWrites(0) {}
+    GrIORefProxy(sk_sp<GrSurface> surface) : fRefCnt(1), fPendingReads(0), fPendingWrites(0) {
         // Since we're manually forwarding on refs & unrefs we don't want sk_sp doing
         // anything extra.
         fTarget = surface.release();
@@ -66,13 +82,76 @@
         // have forwarded on the unref call that got use here.
     }
 
-    // TODO: add the IO ref counts. Although if we can delay shader creation to flush time
-    // we may not even need to do that.
-    mutable int32_t fRefCnt;
+    // This GrIORefProxy was deferred before but has just been instantiated. To
+    // make all the reffing & unreffing work out we now need to transfer any deferred
+    // refs & unrefs to the new GrSurface
+    void transferRefs() {
+        SkASSERT(fTarget);
+
+        fTarget->fRefCnt += (fRefCnt-1); // don't xfer the proxy's creation ref
+        fTarget->fPendingReads += fPendingReads;
+        fTarget->fPendingWrites += fPendingWrites;
+
+        fPendingReads = 0;
+        fPendingWrites = 0;
+    }
 
     // For deferred proxies this will be null. For wrapped proxies it will point to the
     // wrapped resource.
     GrSurface* fTarget;
+
+private:
+    // This class is used to manage conversion of refs to pending reads/writes.
+    friend class GrGpuResourceRef;
+    template <typename, GrIOType> friend class GrPendingIOResource;
+
+    void addPendingRead() const {
+        this->validate();
+
+        if (fTarget) {
+            fTarget->addPendingRead();
+            return;
+        }
+
+        ++fPendingReads;
+    }
+
+    void completedRead() const {
+        this->validate();
+
+        if (fTarget) {
+            fTarget->completedRead();
+            return;
+        }
+    
+        SkFAIL("How was the read completed if the Proxy hasn't been instantiated?");
+    }
+
+    void addPendingWrite() const {
+        this->validate();
+
+        if (fTarget) {
+            fTarget->addPendingWrite();
+            return;
+        }
+
+        ++fPendingWrites;
+    }
+
+    void completedWrite() const {
+        this->validate();
+
+        if (fTarget) {
+            fTarget->completedWrite();
+            return;
+        }
+    
+        SkFAIL("How was the write completed if the Proxy hasn't been instantiated?");
+    }
+
+    mutable int32_t fRefCnt;
+    mutable int32_t fPendingReads;
+    mutable int32_t fPendingWrites;
 };
 
 class GrSurfaceProxy : public GrIORefProxy {
diff --git a/src/gpu/GrSurfaceProxy.cpp b/src/gpu/GrSurfaceProxy.cpp
index 831cd32..9de41ef 100644
--- a/src/gpu/GrSurfaceProxy.cpp
+++ b/src/gpu/GrSurfaceProxy.cpp
@@ -43,6 +43,8 @@
         return nullptr;
     }
 
+    this->INHERITED::transferRefs();
+
 #ifdef SK_DEBUG
     if (kInvalidGpuMemorySize != this->getRawGpuMemorySize_debugOnly()) {
         SkASSERT(fTarget->gpuMemorySize() <= this->getRawGpuMemorySize_debugOnly());    
diff --git a/tests/ProxyRefTest.cpp b/tests/ProxyRefTest.cpp
new file mode 100644
index 0000000..4f6a3e3
--- /dev/null
+++ b/tests/ProxyRefTest.cpp
@@ -0,0 +1,205 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+// This is a GPU-backend specific test.
+
+#include "Test.h"
+
+#if SK_SUPPORT_GPU
+#include "GrSurfaceProxy.h"
+#include "GrTextureProxy.h"
+#include "GrRenderTargetPriv.h"
+#include "GrRenderTargetProxy.h"
+
+static const int kWidthHeight = 128;
+
+int32_t GrIORefProxy::getProxyRefCnt_TestOnly() const {
+    return fRefCnt;
+}
+
+int32_t GrIORefProxy::getBackingRefCnt_TestOnly() const {
+    if (fTarget) {
+        return fTarget->fRefCnt;
+    }
+
+    return fRefCnt;
+}
+
+int32_t GrIORefProxy::getPendingReadCnt_TestOnly() const {
+    if (fTarget) {
+        SkASSERT(!fPendingReads);
+        return fTarget->fPendingReads;
+    }
+
+    return fPendingReads;
+}
+
+int32_t GrIORefProxy::getPendingWriteCnt_TestOnly() const {
+    if (fTarget) {
+        SkASSERT(!fPendingWrites);
+        return fTarget->fPendingWrites;
+    }
+
+    return fPendingWrites;
+}
+
+static void check_refs(skiatest::Reporter* reporter,
+                       GrSurfaceProxy* proxy,
+                       int32_t expectedProxyRefs,
+                       int32_t expectedBackingRefs,
+                       int32_t expectedNumReads,
+                       int32_t expectedNumWrites) {
+    REPORTER_ASSERT(reporter, proxy->getProxyRefCnt_TestOnly() == expectedProxyRefs);
+    REPORTER_ASSERT(reporter, proxy->getBackingRefCnt_TestOnly() == expectedBackingRefs);
+    REPORTER_ASSERT(reporter, proxy->getPendingReadCnt_TestOnly() == expectedNumReads);
+    REPORTER_ASSERT(reporter, proxy->getPendingWriteCnt_TestOnly() == expectedNumWrites);
+
+    SkASSERT(proxy->getProxyRefCnt_TestOnly() == expectedProxyRefs);
+    SkASSERT(proxy->getBackingRefCnt_TestOnly() == expectedBackingRefs);
+    SkASSERT(proxy->getPendingReadCnt_TestOnly() == expectedNumReads);
+    SkASSERT(proxy->getPendingWriteCnt_TestOnly() == expectedNumWrites);
+}
+
+static sk_sp<GrSurfaceProxy> make_deferred(const GrCaps& caps, GrTextureProvider* provider) {
+    GrSurfaceDesc desc;
+    desc.fFlags = kRenderTarget_GrSurfaceFlag;
+    desc.fWidth = kWidthHeight;
+    desc.fHeight = kWidthHeight;
+    desc.fConfig = kRGBA_8888_GrPixelConfig;
+
+    return GrSurfaceProxy::MakeDeferred(caps, desc, SkBackingFit::kApprox, SkBudgeted::kYes);
+}
+
+static sk_sp<GrSurfaceProxy> make_wrapped(const GrCaps& caps, GrTextureProvider* provider) {
+    GrSurfaceDesc desc;
+    desc.fFlags = kRenderTarget_GrSurfaceFlag;
+    desc.fWidth = kWidthHeight;
+    desc.fHeight = kWidthHeight;
+    desc.fConfig = kRGBA_8888_GrPixelConfig;
+
+    sk_sp<GrTexture> tex(provider->createTexture(desc, SkBudgeted::kNo));
+
+    // Flush the IOWrite from the initial discard or it will confuse the later ref count checks
+    tex->flushWrites();
+
+    return GrSurfaceProxy::MakeWrapped(std::move(tex));
+}
+
+DEF_GPUTEST_FOR_RENDERING_CONTEXTS(ProxyRefTest, reporter, ctxInfo) {
+    GrTextureProvider* provider = ctxInfo.grContext()->textureProvider();
+    const GrCaps& caps = *ctxInfo.grContext()->caps();
+
+    for (auto make : { make_deferred, make_wrapped }) {
+        // A single write
+        {
+            sk_sp<GrSurfaceProxy> sProxy((*make)(caps, provider));
+
+            GrPendingIOResource<GrSurfaceProxy, kWrite_GrIOType> fWrite(sProxy.get());
+
+            check_refs(reporter, sProxy.get(), 1, 1, 0, 1);
+
+            // In the deferred case, the discard batch created on instantiation adds an
+            // extra ref and write
+            bool proxyGetsDiscardRef = !sProxy->isWrapped_ForTesting() &&
+                                       caps.discardRenderTargetSupport();
+	    int expectedWrites = proxyGetsDiscardRef ? 2 : 1;
+
+            sProxy->instantiate(provider);
+
+            // In the deferred case, this checks that the refs transfered to the GrSurface
+            check_refs(reporter, sProxy.get(), 1, 1, 0, expectedWrites);
+        }
+
+        // A single read
+        {
+            sk_sp<GrSurfaceProxy> sProxy((*make)(caps, provider));
+
+            GrPendingIOResource<GrSurfaceProxy, kRead_GrIOType> fRead(sProxy.get());
+
+            check_refs(reporter, sProxy.get(), 1, 1, 1, 0);
+
+            // In the deferred case, the discard batch created on instantiation adds an
+            // extra ref and write
+            bool proxyGetsDiscardRef = !sProxy->isWrapped_ForTesting() &&
+                                       caps.discardRenderTargetSupport();
+            int expectedWrites = proxyGetsDiscardRef ? 1 : 0;
+
+            sProxy->instantiate(provider);
+
+            // In the deferred case, this checks that the refs transfered to the GrSurface
+            check_refs(reporter, sProxy.get(), 1, 1, 1, expectedWrites);
+        }
+
+        // A single read/write pair
+        {
+            sk_sp<GrSurfaceProxy> sProxy((*make)(caps, provider));
+
+            GrPendingIOResource<GrSurfaceProxy, kRW_GrIOType> fRW(sProxy.get());
+
+            check_refs(reporter, sProxy.get(), 1, 1, 1, 1);
+
+            // In the deferred case, the discard batch created on instantiation adds an
+            // extra ref and write
+            bool proxyGetsDiscardRef = !sProxy->isWrapped_ForTesting() &&
+                                       caps.discardRenderTargetSupport();
+            int expectedWrites = proxyGetsDiscardRef ? 2 : 1;
+
+            sProxy->instantiate(provider);
+
+            // In the deferred case, this checks that the refs transfered to the GrSurface
+            check_refs(reporter, sProxy.get(), 1, 1, 1, expectedWrites);
+        }
+
+        // Multiple normal refs
+        {
+            sk_sp<GrSurfaceProxy> sProxy((*make)(caps, provider));
+            sProxy->ref();
+            sProxy->ref();
+
+            check_refs(reporter, sProxy.get(), 3, 3, 0, 0);
+
+            bool proxyGetsDiscardRef = !sProxy->isWrapped_ForTesting() &&
+                                       caps.discardRenderTargetSupport();
+            int expectedWrites = proxyGetsDiscardRef ? 1 : 0;
+
+            sProxy->instantiate(provider);
+
+            // In the deferred case, this checks that the refs transfered to the GrSurface
+            check_refs(reporter, sProxy.get(), 3, 3, 0, expectedWrites);
+
+            sProxy->unref();
+            sProxy->unref();
+        }
+
+        // Continue using (reffing) proxy after instantiation
+        {
+            sk_sp<GrSurfaceProxy> sProxy((*make)(caps, provider));
+            sProxy->ref();
+
+            GrPendingIOResource<GrSurfaceProxy, kWrite_GrIOType> fWrite(sProxy.get());
+
+            check_refs(reporter, sProxy.get(), 2, 2, 0, 1);
+
+            bool proxyGetsDiscardRef = !sProxy->isWrapped_ForTesting() &&
+                                       caps.discardRenderTargetSupport();
+            int expectedWrites = proxyGetsDiscardRef ? 2 : 1;
+
+            sProxy->instantiate(provider);
+
+            // In the deferred case, this checks that the refs transfered to the GrSurface
+            check_refs(reporter, sProxy.get(), 2, 2, 0, expectedWrites);
+
+            sProxy->unref();
+            check_refs(reporter, sProxy.get(), 1, 1, 0, expectedWrites);
+
+            GrPendingIOResource<GrSurfaceProxy, kRead_GrIOType> fRead(sProxy.get());
+            check_refs(reporter, sProxy.get(), 1, 1, 1, expectedWrites);
+        }
+    }
+}
+
+#endif