MDEV-10814: Innodb large allocations - madvise - Don't dump

Note: Linux only

Core dumps of large buffer pool pages take time and space
and pose potential data expose in scenarios where data-at-rest
encryption is deployed.

Here we use madvise(MADV_DONT_DUMP) on large memory allocations
used by the innodb buffer pool, log_sys and recv_sys. The effect
of this system call is that these memory areas will not appear in
a core dump. Data from these buffers is rarely useful in fault
diagnosis.

log_sys and recv_sys structures now use large memory allocations
for their large buffer.

Debug builds don't include the madvise syscall and as such will
include full core dumps.

A function, buf_madvise_do_dump, is added but never called. It
is there to be called from a debugger to re-enable the core
dumping of all of these pages if for some reason the entire
contents of these buffers are needed.

Idea thanks to Hartmut Holzgraefe

storage/innobase/buf/buf0buf.cc

@@ -1173,6 +1173,57 @@ buf_page_is_corrupted(
 }
 
 #ifndef UNIV_INNOCHECKSUM
+
+#if defined(DBUG_OFF) && defined(HAVE_MADVISE) &&  defined(MADV_DODUMP)
+/** Enable buffers to be dumped to core files
+
+A convience function, not called anyhwere directly however
+it is left available for gdb or any debugger to call
+in the event that you want all of the memory to be dumped
+to a core file.
+
+Returns number of errors found in madvise calls. */
+int
+buf_madvise_do_dump()
+{
+	int ret= 0;
+	buf_pool_t*	buf_pool;
+	ulint		n;
+	buf_chunk_t*	chunk;
+
+	/* mirrors allocation in log_sys_init() */
+	if (log_sys->buf)
+	{
+		ret+= madvise(log_sys->first_in_use ? log_sys->buf
+						    : log_sys->buf - log_sys->buf_size,
+			      log_sys->buf_size,
+			      MADV_DODUMP);
+	}
+	/* mirrors recv_sys_init() */
+	if (recv_sys->buf)
+	{
+		ret+= madvise(recv_sys->buf, recv_sys->len, MADV_DODUMP);
+	}
+
+	buf_pool_mutex_enter_all();
+
+	for (int i= 0; i < srv_buf_pool_instances; i++)
+	{
+		buf_pool = buf_pool_from_array(i);
+		chunk = buf_pool->chunks;
+
+		for (int n = buf_pool->n_chunks; n--; chunk++)
+		{
+			ret+= madvise(chunk->mem, chunk->mem_size(), MADV_DODUMP);
+		}
+	}
+
+	buf_pool_mutex_exit_all();
+
+	return ret;
+}
+#endif
+
 /** Dump a page to stderr.
 @param[in]	read_buf	database page
 @param[in]	page_size	page size */
@@ -1502,7 +1553,7 @@ buf_chunk_init(
 	DBUG_EXECUTE_IF("ib_buf_chunk_init_fails", return(NULL););
 
 	chunk->mem = buf_pool->allocator.allocate_large(mem_size,
-							&chunk->mem_pfx);
+							&chunk->mem_pfx, true);
 
 	if (UNIV_UNLIKELY(chunk->mem == NULL)) {
 
@@ -1796,7 +1847,8 @@ buf_pool_init_instance(
 					}
 
 					buf_pool->allocator.deallocate_large(
-						chunk->mem, &chunk->mem_pfx);
+						chunk->mem, &chunk->mem_pfx, chunk->mem_size(),
+						true);
 				}
 				ut_free(buf_pool->chunks);
 				buf_pool_mutex_exit(buf_pool);
@@ -1943,7 +1995,7 @@ buf_pool_free_instance(
 		}
 
 		buf_pool->allocator.deallocate_large(
-			chunk->mem, &chunk->mem_pfx);
+			chunk->mem, &chunk->mem_pfx, true);
 	}
 
 	for (ulint i = BUF_FLUSH_LRU; i < BUF_FLUSH_N_TYPES; ++i) {
@@ -2819,7 +2871,7 @@ buf_pool_resize()
 				}
 
 				buf_pool->allocator.deallocate_large(
-					chunk->mem, &chunk->mem_pfx);
+					chunk->mem, &chunk->mem_pfx, true);
 
 				sum_freed += chunk->size;
 

storage/innobase/include/log0log.h

@@ -614,15 +614,15 @@ struct log_t{
 					mtr_commit and still ensure that
 					insertions in the flush_list happen
 					in the LSN order. */
-	byte*		buf_ptr;	/*!< unaligned log buffer, which should
-					be of double of buf_size */
-	byte*		buf;		/*!< log buffer currently in use;
-					this could point to either the first
-					half of the aligned(buf_ptr) or the
+	byte*		buf;		/*!< Memory of double the buf_size is
+					allocated here. This pointer will change
+					however to either the first half or the
 					second half in turns, so that log
 					write/flush to disk don't block
 					concurrent mtrs which will write
-					log to this buffer */
+					log to this buffer. Care to switch back
+					to the first half before freeing/resizing
+					must be undertaken. */
 	bool		first_in_use;	/*!< true if buf points to the first
 					half of the aligned(buf_ptr), false
 					if the second half */

storage/innobase/include/log0recv.h

@@ -217,6 +217,7 @@ struct recv_sys_t{
 				/*!< this is TRUE when a log rec application
 				batch is running */
 	byte*		buf;	/*!< buffer for parsing log records */
+	size_t		buf_size;	/*!< size of buf */
 	ulint		len;	/*!< amount of data in buf */
 	lsn_t		parse_start_lsn;
 				/*!< this is the lsn from which we were able to

storage/innobase/include/ut0new.h

@@ -129,6 +129,10 @@ with (), thus:
 #include <string.h> /* strlen(), strrchr(), strncmp() */
 
 #include "my_global.h" /* needed for headers from mysql/psi/ */
+#if !defined(DBUG_OFF) && defined(HAVE_MADVISE)
+#include <sys/mman.h>
+#endif
+
 /* JAN: TODO: missing 5.7 header */
 #ifdef HAVE_MYSQL_MEMORY_H
 #include "mysql/psi/mysql_memory.h" /* PSI_MEMORY_CALL() */
@@ -234,6 +238,45 @@ struct ut_new_pfx_t {
 #endif
 };
 
+static void ut_allocate_trace_dontdump(void *		ptr,
+			size_t	bytes,
+			bool		dontdump,
+			ut_new_pfx_t*	pfx,
+			const char*	file)
+{
+	ut_a(ptr != NULL);
+
+#if defined(DBUG_OFF) && defined(HAVE_MADVISE) && defined(MADV_DONTDUMP)
+	if (dontdump && madvise(ptr, bytes, MADV_DONTDUMP)) {
+		ib::warn() << "Failed to set memory to DONTDUMP: "
+			   << strerror(errno)
+			   << " ptr " << ptr
+			   << " size " << bytes;
+	}
+#endif
+	if (pfx != NULL) {
+#ifdef UNIV_PFS_MEMORY
+		allocate_trace(bytes, file, pfx);
+#endif /* UNIV_PFS_MEMORY */
+		pfx->m_size = bytes;
+	}
+}
+
+static void ut_dodump(void* ptr, size_t m_size)
+{
+	if (ptr == NULL) {
+		return;
+	}
+#if defined(DBUG_OFF) && defined(HAVE_MADVISE) && defined(MADV_DODUMP)
+	if (madvise(ptr, m_size, MADV_DODUMP)) {
+		ib::warn() << "Failed to set memory to DODUMP: "
+			   << strerror(errno)
+			   << " ptr " << ptr
+			   << " size " << m_size;
+	}
+#endif
+}
+
 /** Allocator class for allocating memory from inside std::* containers.
 @tparam	T		type of allocated object
 @tparam oom_fatal	whether to commit suicide when running out of memory */
@@ -294,6 +337,7 @@ class ut_allocator {
 	@param[in]	file		file name of the caller
 	@param[in]	set_to_zero	if true, then the returned memory is
 	initialized with 0x0 bytes.
+	@param[in]	throw_on_error	if true, raize exception if too big
 	@return pointer to the allocated memory */
 	pointer
 	allocate(
@@ -566,6 +610,8 @@ class ut_allocator {
 	/** Allocate a large chunk of memory that can hold 'n_elements'
 	objects of type 'T' and trace the allocation.
 	@param[in]	n_elements	number of elements
+	@param[in]	dontdump	if true, advise the OS is not to core
+	dump this memory.
 	@param[out]	pfx		storage for the description of the
 	allocated memory. The caller must provide space for this one and keep
 	it until the memory is no longer needed and then pass it to
@@ -574,7 +620,8 @@ class ut_allocator {
 	pointer
 	allocate_large(
 		size_type	n_elements,
-		ut_new_pfx_t*	pfx)
+		ut_new_pfx_t*	pfx,
+		bool		dontdump = false)
 	{
 		if (n_elements == 0 || n_elements > max_size()) {
 			return(NULL);
@@ -585,13 +632,11 @@ class ut_allocator {
 		pointer	ptr = reinterpret_cast<pointer>(
 			os_mem_alloc_large(&n_bytes));
 
-#ifdef UNIV_PFS_MEMORY
-		if (ptr != NULL) {
-			allocate_trace(n_bytes, NULL, pfx);
+		if (ptr == NULL) {
+			return NULL;
 		}
-#else
-		pfx->m_size = n_bytes;
-#endif /* UNIV_PFS_MEMORY */
+
+		ut_allocate_trace_dontdump(ptr, n_bytes, dontdump, pfx, NULL);
 
 		return(ptr);
 	}
@@ -600,17 +645,26 @@ class ut_allocator {
 	deallocation.
 	@param[in,out]	ptr	pointer to memory to free
 	@param[in]	pfx	descriptor of the memory, as returned by
-	allocate_large(). */
+	allocate_large().
+	@param[in]      dodump  if true, advise the OS to include this
+	memory again if a core dump occurs. */
 	void
 	deallocate_large(
 		pointer			ptr,
-		const ut_new_pfx_t*	pfx)
+		const ut_new_pfx_t*	pfx,
+		size_t			size,
+		bool			dodump = false)
 	{
+		if (dodump) {
+			ut_dodump(ptr, size);
+		}
 #ifdef UNIV_PFS_MEMORY
-		deallocate_trace(pfx);
+		if (pfx) {
+			deallocate_trace(pfx);
+		}
 #endif /* UNIV_PFS_MEMORY */
 
-		os_mem_free_large(ptr, pfx->m_size);
+		os_mem_free_large(ptr, size);
 	}
 
 #ifdef UNIV_PFS_MEMORY
@@ -842,6 +896,10 @@ ut_delete_array(
 	ut_allocator<byte>(key).allocate( \
 		n_bytes, NULL, __FILE__, false, false))
 
+#define ut_malloc_dontdump(n_bytes) static_cast<void*>( \
+	ut_allocator<byte>(PSI_NOT_INSTRUMENTED).allocate_large( \
+		n_bytes, true))
+
 #define ut_zalloc(n_bytes, key)		static_cast<void*>( \
 	ut_allocator<byte>(key).allocate( \
 		n_bytes, NULL, __FILE__, true, false))
@@ -865,6 +923,10 @@ ut_delete_array(
 #define ut_free(ptr)	ut_allocator<byte>(PSI_NOT_INSTRUMENTED).deallocate( \
 	reinterpret_cast<byte*>(ptr))
 
+#define ut_free_dodump(ptr, size) static_cast<void*>( \
+	ut_allocator<byte>(PSI_NOT_INSTRUMENTED).deallocate_large( \
+		ptr, NULL, size, true))
+
 #else /* UNIV_PFS_MEMORY */
 
 /* Fallbacks when memory tracing is disabled at compile time. */
@@ -887,6 +949,14 @@ ut_delete_array(
 
 #define ut_malloc_nokey(n_bytes)	::malloc(n_bytes)
 
+static inline void *ut_malloc_dontdump(size_t n_bytes)
+{
+	void *ptr = os_mem_alloc_large(&n_bytes);
+
+	ut_allocate_trace_dontdump(ptr, n_bytes, true, NULL, NULL);
+	return ptr;
+}
+
 #define ut_zalloc_nokey(n_bytes)	::calloc(1, n_bytes)
 
 #define ut_zalloc_nokey_nofatal(n_bytes)	::calloc(1, n_bytes)
@@ -895,6 +965,12 @@ ut_delete_array(
 
 #define ut_free(ptr)			::free(ptr)
 
+static inline void ut_free_dodump(void *ptr, size_t size)
+{
+	ut_dodump(ptr, size);
+	os_mem_free_large(ptr, size);
+}
+
 #endif /* UNIV_PFS_MEMORY */
 
 #endif /* ut0new_h */

storage/innobase/log/log0log.cc

@@ -226,16 +226,18 @@ log_buffer_extend(
 	log_sys->buf_free -= move_start;
 	log_sys->buf_next_to_write -= move_start;
 
+	/* free previous after getting the right address */
+	if (!log_sys->first_in_use) {
+		log_sys->buf -= log_sys->buf_size;
+	}
+	ut_free_dodump(log_sys->buf, log_sys->buf_size * 2);
+
 	/* reallocate log buffer */
 	srv_log_buffer_size = len / UNIV_PAGE_SIZE + 1;
-	ut_free(log_sys->buf_ptr);
-
 	log_sys->buf_size = LOG_BUFFER_SIZE;
 
-	log_sys->buf_ptr = static_cast<byte*>(
-		ut_zalloc_nokey(log_sys->buf_size * 2 + OS_FILE_LOG_BLOCK_SIZE));
 	log_sys->buf = static_cast<byte*>(
-		ut_align(log_sys->buf_ptr, OS_FILE_LOG_BLOCK_SIZE));
+		ut_malloc_dontdump(log_sys->buf_size * 2));
 
 	log_sys->first_in_use = true;
 
@@ -723,10 +725,8 @@ log_sys_init()
 
 	log_sys->buf_size = LOG_BUFFER_SIZE;
 
-	log_sys->buf_ptr = static_cast<byte*>(
-		ut_zalloc_nokey(log_sys->buf_size * 2 + OS_FILE_LOG_BLOCK_SIZE));
 	log_sys->buf = static_cast<byte*>(
-		ut_align(log_sys->buf_ptr, OS_FILE_LOG_BLOCK_SIZE));
+		ut_malloc_dontdump(log_sys->buf_size * 2));
 
 	log_sys->first_in_use = true;
 
@@ -1085,12 +1085,12 @@ log_buffer_switch()
 						 OS_FILE_LOG_BLOCK_SIZE);
 
 	if (log_sys->first_in_use) {
-		ut_ad(log_sys->buf == ut_align(log_sys->buf_ptr,
+		ut_ad(log_sys->buf == ut_align(log_sys->buf,
 					       OS_FILE_LOG_BLOCK_SIZE));
 		log_sys->buf += log_sys->buf_size;
 	} else {
 		log_sys->buf -= log_sys->buf_size;
-		ut_ad(log_sys->buf == ut_align(log_sys->buf_ptr,
+		ut_ad(log_sys->buf == ut_align(log_sys->buf,
 					       OS_FILE_LOG_BLOCK_SIZE));
 	}
 
@@ -2254,8 +2254,10 @@ log_shutdown()
 {
 	log_group_close_all();
 
-	ut_free(log_sys->buf_ptr);
-	log_sys->buf_ptr = NULL;
+	if (!log_sys->first_in_use) {
+		log_sys->buf -= log_sys->buf_size;
+	}
+	ut_free_dodump(log_sys->buf, log_sys->buf_size * 2);
 	log_sys->buf = NULL;
 	ut_free(log_sys->checkpoint_buf_ptr);
 	log_sys->checkpoint_buf_ptr = NULL;

storage/innobase/log/log0recv.cc

@@ -438,7 +438,9 @@ recv_sys_close()
 			os_event_destroy(recv_sys->flush_end);
 		}
 
-		ut_free(recv_sys->buf);
+		if (recv_sys->buf != NULL) {
+			ut_free_dodump(recv_sys->buf, recv_sys->buf_size);
+		}
 
 		ut_ad(!recv_writer_thread_active);
 		mutex_free(&recv_sys->writer_mutex);
@@ -553,7 +555,8 @@ recv_sys_init()
 	}
 
 	recv_sys->buf = static_cast<byte*>(
-		ut_malloc_nokey(RECV_PARSING_BUF_SIZE));
+		ut_malloc_dontdump(RECV_PARSING_BUF_SIZE));
+	recv_sys->buf_size = RECV_PARSING_BUF_SIZE;
 
 	recv_sys->addr_hash = hash_create(size / 512);
 	recv_sys->progress_time = ut_time();
@@ -588,8 +591,9 @@ recv_sys_debug_free(void)
 
 	hash_table_free(recv_sys->addr_hash);
 	mem_heap_free(recv_sys->heap);
-	ut_free(recv_sys->buf);
+	ut_free_dodump(recv_sys->buf, recv_sys->buf_size);
 
+	recv_sys->buf_size = 0;
 	recv_sys->buf = NULL;
 	recv_sys->heap = NULL;
 	recv_sys->addr_hash = NULL;

storage/innobase/row/row0log.cc

@@ -71,6 +71,7 @@ enum row_op {
 /** Log block for modifications during online ALTER TABLE */
 struct row_log_buf_t {
 	byte*		block;	/*!< file block buffer */
+	size_t		size; /*!< length of block in bytes */
 	ut_new_pfx_t	block_pfx; /*!< opaque descriptor of "block". Set
 				by ut_allocator::allocate_large() and fed to
 				ut_allocator::deallocate_large(). */
@@ -265,6 +266,7 @@ row_log_block_allocate(
 		if (log_buf.block == NULL) {
 			DBUG_RETURN(false);
 		}
+		log_buf.size = srv_sort_buf_size;
 	}
 	DBUG_RETURN(true);
 }
@@ -279,7 +281,7 @@ row_log_block_free(
 	DBUG_ENTER("row_log_block_free");
 	if (log_buf.block != NULL) {
 		ut_allocator<byte>(mem_key_row_log_buf).deallocate_large(
-			log_buf.block, &log_buf.block_pfx);
+			log_buf.block, &log_buf.block_pfx, log_buf.size);
 		log_buf.block = NULL;
 	}
 	DBUG_VOID_RETURN;

storage/innobase/row/row0merge.cc

@@ -4619,6 +4619,7 @@ row_merge_build_indexes(
 	merge_file_t*		merge_files;
 	row_merge_block_t*	block;
 	ut_new_pfx_t		block_pfx;
+	size_t			block_size;
 	ut_new_pfx_t		crypt_pfx;
 	row_merge_block_t*	crypt_block = NULL;
 	ulint			i;
@@ -4654,7 +4655,8 @@ row_merge_build_indexes(
 
 	/* This will allocate "3 * srv_sort_buf_size" elements of type
 	row_merge_block_t. The latter is defined as byte. */
-	block = alloc.allocate_large(3 * srv_sort_buf_size, &block_pfx);
+	block_size = 3 * srv_sort_buf_size;
+	block = alloc.allocate_large(block_size, &block_pfx);
 
 	if (block == NULL) {
 		DBUG_RETURN(DB_OUT_OF_MEMORY);
@@ -4664,7 +4666,7 @@ row_merge_build_indexes(
 
 	if (log_tmp_is_encrypted()) {
 		crypt_block = static_cast<row_merge_block_t*>(
-			alloc.allocate_large(3 * srv_sort_buf_size,
+			alloc.allocate_large(block_size,
 					     &crypt_pfx));
 
 		if (crypt_block == NULL) {
@@ -5035,10 +5037,10 @@ row_merge_build_indexes(
 
 	ut_free(merge_files);
 
-	alloc.deallocate_large(block, &block_pfx);
+	alloc.deallocate_large(block, &block_pfx, block_size);
 
 	if (crypt_block) {
-		alloc.deallocate_large(crypt_block, &crypt_pfx);
+		alloc.deallocate_large(crypt_block, &crypt_pfx, block_size);
 	}
 
 	DICT_TF2_FLAG_UNSET(new_table, DICT_TF2_FTS_ADD_DOC_ID);