Merge pull request #18438 from vespa-engine/balder/implement_malloc_usable_size

Implement _malloc_usable_size

14 files changed, 87 insertions, 93 deletions

diff --git a/vespamalloc/src/tests/test1/new_test.cpp b/vespamalloc/src/tests/test1/new_test.cpp
index a8392363f6f..9c291b3842e 100644
--- a/vespamalloc/src/tests/test1/new_test.cpp
+++ b/vespamalloc/src/tests/test1/new_test.cpp
@@ -1,6 +1,8 @@
 // Copyright 2017 Yahoo Holdings. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
 #include <vespa/vespalib/testkit/testapp.h>
 #include <vespa/log/log.h>
+#include <malloc.h>
+#include <dlfcn.h>
 
 LOG_SETUP("new_test");
 
@@ -105,5 +107,35 @@ TEST("verify new with alignment = 64 with single element") {
     LOG(info, "&s=%p", s.get());
 }
 
+void verify_vespamalloc_usable_size() {
+    struct AllocInfo { size_t requested; size_t usable;};
+    AllocInfo allocInfo[] = {{0x7, 0x20}, {0x27, 0x40}, {0x47, 0x80}, {0x87, 0x100}, {0x107, 0x200}, {0x207, 0x400},
+                            {0x407, 0x800}, {0x807, 0x1000}, {0x1007, 0x2000}, {0x2007, 0x4000}, {0x4007, 0x8000},
+                            {0x8007, 0x10000}, {0x10007, 0x20000}, {0x20007, 0x40000}, {0x40007, 0x80000}, {0x80007, 0x100000},
+                            {0x100007, 0x200000}, {0x200007, 0x400000}, {0x400007, 0x600000}};
+    for (const AllocInfo & info : allocInfo) {
+        std::unique_ptr<char[]> buf = std::make_unique<char[]>(info.requested);
+        size_t usable_size = malloc_usable_size(buf.get());
+        EXPECT_EQUAL(info.usable, usable_size);
+    }
+}
+
+TEST("verify malloc_usable_size is sane") {
+    constexpr size_t SZ = 33;
+    std::unique_ptr<char[]> buf = std::make_unique<char[]>(SZ);
+    size_t usable_size = malloc_usable_size(buf.get());
+    if (dlsym(RTLD_NEXT, "is_vespamallocd") != nullptr) {
+        // Debug variants will never have more memory available as there is pre/postamble for error detection.
+        EXPECT_EQUAL(SZ, usable_size);
+    } else if (dlsym(RTLD_NEXT, "is_vespamalloc") != nullptr) {
+        // Normal production vespamalloc will round up
+        EXPECT_EQUAL(64u, usable_size);
+        verify_vespamalloc_usable_size();
+    } else {
+        // Non vespamalloc implementations we can not say anything about
+        EXPECT_GREATER_EQUAL(usable_size, SZ);
+    }
+}
+
 
 TEST_MAIN() { TEST_RUN_ALL(); }
diff --git a/vespamalloc/src/vespamalloc/malloc/globalpool.h b/vespamalloc/src/vespamalloc/malloc/globalpool.h
index bc13231bb85..872fd9b936e 100644
--- a/vespamalloc/src/vespamalloc/malloc/globalpool.h
+++ b/vespamalloc/src/vespamalloc/malloc/globalpool.h
@@ -28,6 +28,7 @@ public:
     void enableThreadSupport() __attribute__((noinline));
 
     static void setParams(size_t alwaysReuseLimit, size_t threadCacheLimit);
+    static size_t computeExactSize(size_t sz);
 
     void info(FILE * os, size_t level=0) __attribute__((noinline));
 private:
diff --git a/vespamalloc/src/vespamalloc/malloc/globalpool.hpp b/vespamalloc/src/vespamalloc/malloc/globalpool.hpp
index a3930c7cca1..69f2a2cf820 100644
--- a/vespamalloc/src/vespamalloc/malloc/globalpool.hpp
+++ b/vespamalloc/src/vespamalloc/malloc/globalpool.hpp
@@ -122,11 +122,17 @@ AllocPoolT<MemBlockPtrT>::exchangeAlloc(SizeClassT sc, typename AllocPoolT<MemBl
 }
 
 template <typename MemBlockPtrT>
+size_t
+AllocPoolT<MemBlockPtrT>::computeExactSize(size_t sz) {
+    return (((sz + (_alwaysReuseLimit - 1)) / _alwaysReuseLimit) * _alwaysReuseLimit);
+}
+
+template <typename MemBlockPtrT>
 typename AllocPoolT<MemBlockPtrT>::ChunkSList *
 AllocPoolT<MemBlockPtrT>::exactAlloc(size_t exactSize, SizeClassT sc,
                                      typename AllocPoolT<MemBlockPtrT>::ChunkSList * csl)
 {
-    size_t adjustedSize((( exactSize + (_alwaysReuseLimit - 1))/_alwaysReuseLimit)*_alwaysReuseLimit);
+    size_t adjustedSize = computeExactSize(exactSize);
     void *exactBlock = _dataSegment.getBlock(adjustedSize, sc);
     MemBlockPtrT mem(exactBlock, MemBlockPtrT::unAdjustSize(adjustedSize));
     csl->add(mem);
diff --git a/vespamalloc/src/vespamalloc/malloc/malloc.cpp b/vespamalloc/src/vespamalloc/malloc/malloc.cpp
index 4a810c1ac5f..8efff973fa1 100644
--- a/vespamalloc/src/vespamalloc/malloc/malloc.cpp
+++ b/vespamalloc/src/vespamalloc/malloc/malloc.cpp
@@ -31,4 +31,11 @@ dumpInfo(size_t level)
 
 }
 
+extern "C" {
+
+int is_vespamalloc() __attribute__((visibility ("default")));
+int is_vespamalloc() { return 1; }
+
+}
+
 #include <vespamalloc/malloc/overload.h>
diff --git a/vespamalloc/src/vespamalloc/malloc/malloc.h b/vespamalloc/src/vespamalloc/malloc/malloc.h
index 49a7b765cc8..51efe28eafc 100644
--- a/vespamalloc/src/vespamalloc/malloc/malloc.h
+++ b/vespamalloc/src/vespamalloc/malloc/malloc.h
@@ -39,6 +39,9 @@ public:
     }
     size_t getMinSizeForAlignment(size_t align, size_t sz) const { return MemBlockPtrT::getMinSizeForAlignment(align, sz); }
     size_t sizeClass(const void *ptr) const { return _segment.sizeClass(ptr); }
+    size_t usable_size(void *ptr) const {
+        return MemBlockPtrT::usable_size(ptr, _segment.getMaxSize(ptr));
+    }
 
     void *calloc(size_t nelm, size_t esz) {
         void * ptr = malloc(nelm * esz);
diff --git a/vespamalloc/src/vespamalloc/malloc/mallocd.cpp b/vespamalloc/src/vespamalloc/malloc/mallocd.cpp
index 47c12b4f186..038715e4231 100644
--- a/vespamalloc/src/vespamalloc/malloc/mallocd.cpp
+++ b/vespamalloc/src/vespamalloc/malloc/mallocd.cpp
@@ -29,4 +29,11 @@ template void MemBlockBoundsCheckBaseT<20, 0>::dumpInfo(size_t);
 
 }
 
+extern "C" {
+
+int is_vespamallocd() __attribute__((visibility ("default")));
+int is_vespamallocd() { return 1; }
+
+}
+
 #include <vespamalloc/malloc/overload.h>
diff --git a/vespamalloc/src/vespamalloc/malloc/mallocdst16.cpp b/vespamalloc/src/vespamalloc/malloc/mallocdst16.cpp
index 0946f51d75b..67d3869ed34 100644
--- a/vespamalloc/src/vespamalloc/malloc/mallocdst16.cpp
+++ b/vespamalloc/src/vespamalloc/malloc/mallocdst16.cpp
@@ -29,4 +29,10 @@ template void MemBlockBoundsCheckBaseT<20, 16>::dumpInfo(size_t);
 
 }
 
+extern "C" {
+
+int is_vespamallocdst16() __attribute__((visibility ("default")));
+int is_vespamallocdst16() { return 1; }
+
+}
 #include <vespamalloc/malloc/overload.h>
diff --git a/vespamalloc/src/vespamalloc/malloc/mallocdst16_nl.cpp b/vespamalloc/src/vespamalloc/malloc/mallocdst16_nl.cpp
index df009f3737b..e1bd28cac48 100644
--- a/vespamalloc/src/vespamalloc/malloc/mallocdst16_nl.cpp
+++ b/vespamalloc/src/vespamalloc/malloc/mallocdst16_nl.cpp
@@ -15,4 +15,11 @@ template void MemBlockBoundsCheckBaseT<20, 16>::dumpInfo(size_t);
 
 }
 
+extern "C" {
+
+int is_vespamallocdst16_nl() __attribute__((visibility ("default")));
+int is_vespamallocdst16_nl() { return 1; }
+
+}
+
 #include <vespamalloc/malloc/overload.h>
diff --git a/vespamalloc/src/vespamalloc/malloc/memblock.h b/vespamalloc/src/vespamalloc/malloc/memblock.h
index 3f55e33b23e..8ad56921b37 100644
--- a/vespamalloc/src/vespamalloc/malloc/memblock.h
+++ b/vespamalloc/src/vespamalloc/malloc/memblock.h
@@ -34,6 +34,9 @@ public:
     void setThreadId(uint32_t ) { }
     void free()               { }
     size_t size()       const { return 0; }
+    static size_t usable_size(void *, size_t classSize) {
+        return classSize;
+    }
     bool allocated()    const { return false; }
     uint32_t threadId()      const { return 0; }
     void info(FILE *, unsigned level=0) const  { (void) level; }
diff --git a/vespamalloc/src/vespamalloc/malloc/memblockboundscheck.h b/vespamalloc/src/vespamalloc/malloc/memblockboundscheck.h
index 90445666063..7ffac7c2108 100644
--- a/vespamalloc/src/vespamalloc/malloc/memblockboundscheck.h
+++ b/vespamalloc/src/vespamalloc/malloc/memblockboundscheck.h
@@ -103,6 +103,9 @@ public:
             && ((p[3] == ALLOC_MAGIC) || (p[3] == FREE_MAGIC))
             && *(reinterpret_cast<const unsigned *> ((const char*)_ptr + size() + alignment() + StackTraceLen*sizeof(void *))) == TAIL_MAGIC;
     }
+    static size_t usable_size(void *ptr, size_t ) {
+        return MemBlockBoundsCheckBaseT(ptr).size();
+    }
     bool validAlloc1() const {
         unsigned *p((unsigned*)_ptr);
         return validCommon() && (p[3] == ALLOC_MAGIC);
diff --git a/vespamalloc/src/vespamalloc/malloc/overload.h b/vespamalloc/src/vespamalloc/malloc/overload.h
index ca3a76b705b..1ec37c37aa1 100644
--- a/vespamalloc/src/vespamalloc/malloc/overload.h
+++ b/vespamalloc/src/vespamalloc/malloc/overload.h
@@ -178,6 +178,11 @@ void free(void * ptr) {
     if (ptr) { vespamalloc::_GmemP->free(ptr); }
 }
 
+size_t malloc_usable_size(void *) __THROW __attribute__((visibility ("default")));
+size_t malloc_usable_size (void * ptr) __THROW  {
+    return (ptr) ? vespamalloc::_GmemP->usable_size(ptr) : 0;
+}
+
 #define ALIAS(x) __attribute__ ((weak, alias (x), visibility ("default")))
 #ifdef __clang__
 void* __libc_malloc(size_t sz)                       __THROW __attribute__((malloc, alloc_size(1))) ALIAS("malloc");
@@ -198,93 +203,9 @@ void  __libc_free(void* ptr)                         __THROW __attribute__((leaf
 void  __libc_cfree(void* ptr)                        __THROW __attribute__((leaf)) ALIAS("cfree");
 #endif
 struct mallinfo __libc_mallinfo()                    __THROW  ALIAS("mallinfo");
+size_t  __libc_malloc_usable_size(void *ptr)         __THROW  ALIAS("malloc_usable_size");
 void* __libc_memalign(size_t align, size_t s)        __THROW __attribute__((leaf, malloc, alloc_size(2))) ALIAS("memalign");
 int   __posix_memalign(void** r, size_t a, size_t s) __THROW __nonnull((1)) ALIAS("posix_memalign");
 #undef ALIAS
 
-#if 0
-#include <dlfcn.h>
-
-typedef void * (*dlopen_function) (const char *filename, int flag);
-
-extern "C" VESPA_DLL_EXPORT void * local_dlopen(const char *filename, int flag) __asm__("dlopen");
-
-VESPA_DLL_EXPORT void * local_dlopen(const char *filename, int flag)
-{
-    // A pointer to the library version of dlopen.
-    static dlopen_function real_dlopen = nullptr;
-
-    const char * dlopenName = "dlopen";
-
-    if (real_dlopen == nullptr) {
-        real_dlopen = (dlopen_function) dlsym (RTLD_NEXT, dlopenName);
-        if (real_dlopen == nullptr) {
-            fprintf (stderr, "Could not find the dlopen function!\n");
-            abort();
-        }
-    }
-    //flag = (flag & ~RTLD_DEEPBIND & ~RTLD_NOW) | RTLD_LAZY;
-    //fprintf(stderr, "modified dlopen('%s', %0x)\n", filename, flag);
-    void * handle = real_dlopen(filename, flag);
-    fprintf(stderr, "dlopen('%s', %0x) = %p\n", filename, flag, handle);
-    return handle;
-}
-
-typedef int (*dlclose_function) (void * handle);
-extern "C" VESPA_DLL_EXPORT int local_dlclose(void * handle) __asm__("dlclose");
-VESPA_DLL_EXPORT int local_dlclose(void * handle)
-{
-    // A pointer to the library version of dlclose.
-    static dlclose_function real_dlclose = nullptr;
-
-    const char * dlcloseName = "dlclose";
-
-    if (real_dlclose == nullptr) {
-        real_dlclose = (dlclose_function) dlsym (RTLD_NEXT, dlcloseName);
-        if (real_dlclose == nullptr) {
-            fprintf (stderr, "Could not find the dlclose function!\n");
-            abort();
-        }
-    }
-    int retval = real_dlclose(handle);
-    fprintf(stderr, "dlclose(%p) = %d\n", handle, retval);
-    return retval;
-}
-
-typedef void * (*dlsym_function) (void * handle, const char * symbol);
-extern "C" VESPA_DLL_EXPORT void * local_dlsym(void * handle, const char * symbol) __asm__("dlsym");
-VESPA_DLL_EXPORT void * local_dlsym(void * handle, const char * symbol)
-{
-    // A pointer to the library version of dlsym.
-    static dlsym_function real_dlsym = nullptr;
-
-    const char * dlsymName = "dlsym";
-
-    if (real_dlsym == nullptr) {
-        real_dlsym = (dlsym_function) dlvsym (RTLD_NEXT, dlsymName, "GLIBC_2.2.5");
-        if (real_dlsym == nullptr) {
-            fprintf (stderr, "Could not find the dlsym function!\n");
-            abort();
-        }
-    }
-    if (handle == RTLD_NEXT) {
-        fprintf(stderr, "dlsym(RTLD_NEXT, %s)\n", symbol);
-    } else if (handle == RTLD_DEFAULT) {
-        fprintf(stderr, "dlsym(RTLD_DEFAULT, %s)\n", symbol);
-    } else {
-        fprintf(stderr, "dlsym(%p, %s)\n", handle, symbol);
-    }
-    void * retval = real_dlsym(handle, symbol);
-    if (handle == RTLD_NEXT) {
-        fprintf(stderr, "dlsym(RTLD_NEXT, %s) = %p\n", symbol, retval);
-    } else if (handle == RTLD_DEFAULT) {
-        fprintf(stderr, "dlsym(RTLD_DEFAULT, %s) = %p\n", symbol, retval);
-    } else {
-        fprintf(stderr, "dlsym(%p, %s) = %p\n", handle, symbol, retval);
-    }
-    return retval;
-}
-
-#endif
-
 }
diff --git a/vespamalloc/src/vespamalloc/malloc/threadlist.h b/vespamalloc/src/vespamalloc/malloc/threadlist.h
index 95ff072d035..9edf086ca96 100644
--- a/vespamalloc/src/vespamalloc/malloc/threadlist.h
+++ b/vespamalloc/src/vespamalloc/malloc/threadlist.h
@@ -15,8 +15,8 @@ template <typename MemBlockPtrT, typename ThreadStatT>
 class ThreadListT
 {
 public:
-    typedef ThreadPoolT<MemBlockPtrT, ThreadStatT > ThreadPool;
-    typedef AllocPoolT<MemBlockPtrT> AllocPool;
+    using ThreadPool = ThreadPoolT<MemBlockPtrT, ThreadStatT >;
+    using AllocPool = AllocPoolT<MemBlockPtrT>;
     ThreadListT(AllocPool & pool);
     ~ThreadListT();
     void setParams(size_t alwayReuseLimit, size_t threadCacheLimit) {
diff --git a/vespamalloc/src/vespamalloc/malloc/threadpool.h b/vespamalloc/src/vespamalloc/malloc/threadpool.h
index 1846bbdab4d..d4a754532ff 100644
--- a/vespamalloc/src/vespamalloc/malloc/threadpool.h
+++ b/vespamalloc/src/vespamalloc/malloc/threadpool.h
@@ -62,7 +62,7 @@ private:
         ChunkSList *_freeTo;
     };
     void mallocHelper(size_t exactSize, SizeClassT sc, AllocFree & af, MemBlockPtrT & mem) __attribute__ ((noinline));
-    bool alwaysReuse(SizeClassT sc) { return sc > _alwaysReuseSCLimit; }
+    static bool alwaysReuse(SizeClassT sc) { return sc > _alwaysReuseSCLimit; }
 
     AllocPool   * _allocPool;
     AllocFree     _memList[NUM_SIZE_CLASSES];
diff --git a/vespamalloc/src/vespamalloc/malloc/threadpool.hpp b/vespamalloc/src/vespamalloc/malloc/threadpool.hpp
index 4b5277a6396..b28aa144c34 100644
--- a/vespamalloc/src/vespamalloc/malloc/threadpool.hpp
+++ b/vespamalloc/src/vespamalloc/malloc/threadpool.hpp
@@ -62,7 +62,7 @@ mallocHelper(size_t exactSize,
         af._allocFrom->sub(mem);
         PARANOID_CHECK2( if (!mem.ptr()) { *(int *)0 = 0; } );
     } else {
-        if ( ! this->alwaysReuse(sc) ) {
+        if ( ! alwaysReuse(sc) ) {
             af._allocFrom = _allocPool->exchangeAlloc(sc, af._allocFrom);
             _stat[sc].incExchangeAlloc();
             if (af._allocFrom) {
@@ -93,9 +93,7 @@ ThreadPoolT<MemBlockPtrT, ThreadStatT>::ThreadPoolT() :
 }
 
 template <typename MemBlockPtrT, typename ThreadStatT >
-ThreadPoolT<MemBlockPtrT, ThreadStatT>::~ThreadPoolT()
-{
-}
+ThreadPoolT<MemBlockPtrT, ThreadStatT>::~ThreadPoolT() = default;
 
 template <typename MemBlockPtrT, typename ThreadStatT >
 void ThreadPoolT<MemBlockPtrT, ThreadStatT>::malloc(size_t sz, MemBlockPtrT & mem)