Suspended refs inode

CMakeLists.txt

@@ -35,7 +35,7 @@ project(pmemfile C CXX)
 set(CMAKE_DISABLE_IN_SOURCE_BUILD ON)
 
 set(VERSION_MAJOR 0)
-set(VERSION_MINOR 4)
+set(VERSION_MINOR 5)
 set(VERSION_PATCH 0)
 set(VERSION ${VERSION_MAJOR}.${VERSION_MINOR}.${VERSION_PATCH})
 

include/libpmemfile-posix.h

@@ -406,8 +406,10 @@ PMEMfile *pmemfile_open_parent(PMEMfilepool *pfp, PMEMfile *at,
 
 const char *pmemfile_errormsg(void);
 
-int pmemfile_pool_resume(PMEMfilepool *pfp, const char *pathname);
-int pmemfile_pool_suspend(PMEMfilepool *pfp);
+int pmemfile_pool_resume(PMEMfilepool *pfp, const char *pool_path,
+		unsigned at_root, const char * const *paths, int flags);
+int pmemfile_pool_suspend(PMEMfilepool *pfp, unsigned at_root,
+		const char *const * paths, int flags);
 
 #include "libpmemfile-posix-stubs.h"
 

src/libpmemfile-posix/data.c

@@ -979,3 +979,66 @@ vinode_remove_interval(PMEMfilepool *pfp, struct pmemfile_vinode *vinode,
 
 	return deallocated_space;
 }
+
+/*
+ * vinode_read -- reads file
+ */
+size_t
+vinode_read(PMEMfilepool *pfp, struct pmemfile_vinode *vinode, size_t offset,
+		struct pmemfile_block_desc **last_block, char *buf,
+		size_t count)
+{
+	uint64_t size = inode_get_size(vinode->inode);
+
+	/*
+	 * Start reading at offset, stop reading
+	 * when end of file is reached, or count bytes were read.
+	 * The following two branches compute how many bytes are
+	 * going to be read.
+	 */
+	if (offset >= size)
+		return 0; /* EOF already */
+
+	if (size - offset < count)
+		count = size - offset;
+
+	struct pmemfile_block_desc *block =
+		find_closest_block_with_hint(vinode, offset, *last_block);
+
+	block = iterate_on_file_range(pfp, vinode, block, offset,
+			count, buf, read_from_blocks);
+
+	if (block)
+		*last_block = block;
+
+	return count;
+}
+
+/*
+ * vinode_write -- writes to file
+ */
+void
+vinode_write(PMEMfilepool *pfp, struct pmemfile_vinode *vinode, size_t offset,
+		struct pmemfile_block_desc **last_block,
+		const char *buf, size_t count)
+{
+	ASSERT(count > 0);
+
+	/*
+	 * Two steps:
+	 * - Zero Fill some new blocks, in case the file is extended by
+	 *   writing to the file after seeking past file size ( optionally )
+	 * - Copy the data from the users buffer
+	 */
+
+	/* All blocks needed for writing are properly allocated at this point */
+
+	struct pmemfile_block_desc *block =
+		find_closest_block_with_hint(vinode, offset, *last_block);
+
+	block = iterate_on_file_range(pfp, vinode, block, offset,
+			count, (char *)buf, write_to_blocks);
+
+	if (block)
+		*last_block = block;
+}

src/libpmemfile-posix/data.h

@@ -62,4 +62,12 @@ struct pmemfile_block_desc *iterate_on_file_range(PMEMfilepool *pfp,
 		struct pmemfile_block_desc *starting_block, uint64_t offset,
 		uint64_t len, char *buf, enum cpy_direction dir);
 
+void vinode_write(PMEMfilepool *pfp, struct pmemfile_vinode *vinode,
+		size_t offset, struct pmemfile_block_desc **last_block,
+		const char *buf, size_t count);
+
+size_t vinode_read(PMEMfilepool *pfp, struct pmemfile_vinode *vinode,
+		size_t offset, struct pmemfile_block_desc **last_block,
+		char *buf, size_t count);
+
 #endif

src/libpmemfile-posix/inode.c

@@ -728,10 +728,6 @@ vinode_suspend(PMEMfilepool *pfp, struct pmemfile_vinode *vinode)
 		*tm = vinode->atime;
 	}
 
-	_inode_array_add(pfp, pfp->super->suspended_inodes, vinode->tinode,
-			&vinode->suspended.arr, &vinode->suspended.idx,
-			INODE_ARRAY_NOLOCK);
-
 	if (vinode->blocks) {
 		offset_map_delete(vinode->blocks);
 		vinode->blocks = NULL;
@@ -749,49 +745,17 @@ add_off(void *ptr, uintptr_t off)
 	return (void *)((uintptr_t)ptr + off);
 }
 
-/*
- * inode_resume -- restores persistent part of inode after suspend
- */
-void
-inode_resume(PMEMfilepool *pfp, struct pmemfile_vinode *vinode,
-		PMEMobjpool *old_pop)
-{
-	struct inode_suspend_info suspended = vinode->suspended;
-	struct pmemfile_inode *inode = vinode->inode;
-
-	ASSERT(vinode->suspended.arr != NULL);
-
-	if (pfp->pop != old_pop) {
-		uintptr_t diff = (uintptr_t)pfp->pop - (uintptr_t)old_pop;
-
-		suspended.arr = add_off(suspended.arr, diff);
-		inode = add_off(inode, diff);
-	}
-
-	ASSERT(inode->suspended_references > 0);
-
-	TX_ADD_DIRECT(&inode->suspended_references);
-	inode->suspended_references--;
-
-	_inode_array_unregister(pfp, suspended.arr, suspended.idx,
-			INODE_ARRAY_NOLOCK);
-}
-
 /*
  * vinode_resume -- restores runtime part of inode after suspend
  */
 void
 vinode_resume(PMEMfilepool *pfp, struct pmemfile_vinode *vinode,
 		PMEMobjpool *old_pop)
 {
-	vinode->suspended.arr = NULL;
-	vinode->suspended.idx = 0;
-
 	if (pfp->pop != old_pop) {
 		uintptr_t diff = (uintptr_t)pfp->pop - (uintptr_t)old_pop;
 
 		vinode->inode = add_off(vinode->inode, diff);
-
 		if (vinode->orphaned.arr)
 			vinode->orphaned.arr =
 					add_off(vinode->orphaned.arr, diff);

src/libpmemfile-posix/inode.h

@@ -354,6 +354,11 @@ static inline bool vinode_is_longsymlink(struct pmemfile_vinode *vinode)
 
 const char *get_symlink(PMEMfilepool *pfp, struct pmemfile_vinode *vinode);
 
+static inline bool inode_has_suspended_refs(const struct pmemfile_inode *inode)
+{
+	return (inode_get_flags(inode) & PMEMFILE_I_SUSPENDED_REF) != 0;
+}
+
 struct pmemfile_cred;
 TOID(struct pmemfile_inode) inode_alloc(PMEMfilepool *pfp,
 		struct pmemfile_cred *cred, uint64_t flags);
@@ -404,9 +409,10 @@ blockp_as_oid(struct pmemfile_block_desc *block)
 int vinode_rdlock_with_block_tree(PMEMfilepool *, struct pmemfile_vinode *);
 
 void vinode_suspend(PMEMfilepool *pfp, struct pmemfile_vinode *vinode);
-void inode_resume(PMEMfilepool *pfp, struct pmemfile_vinode *vinode,
-		PMEMobjpool *old_pop);
 void vinode_resume(PMEMfilepool *pfp, struct pmemfile_vinode *vinode,
 		PMEMobjpool *old_pop);
 
+/* 2 for "0x" 16 for pool_uuid 1 for ":" 2 for "0x" 16 for offset 1 for "\n" */
+#define SUSPENDED_INODE_LINE_LENGTH (2 + 16 + 1 + 2 + 16 + 1)
+
 #endif

src/libpmemfile-posix/layout.h

@@ -237,6 +237,15 @@ struct pmemfile_inode {
  */
 COMPILE_ERROR_ON(sizeof(union pmemfile_inode_slots) != 8);
 
+/*
+ * Most constants used with the flags field of pmemfile_inode are defined in the
+ * public header.
+ *
+ * Use the most significant 16 bits for flags only used internally. Hopefully
+ * this is not going to conflict with any flags in a Kernel API in the future.
+ */
+#define PMEMFILE_I_SUSPENDED_REF (UINT64_C(1) << 48)
+
 COMPILE_ERROR_ON(sizeof(struct pmemfile_inode) != PMEMFILE_INODE_SIZE);
 
 #define PMEMFILE_INODE_ARRAY_VERSION(a) ((uint32_t)0x00414E49 | \
@@ -294,9 +303,6 @@ struct pmemfile_super {
 	/* list of arrays of inodes that were deleted, but are still opened */
 	TOID(struct pmemfile_inode_array) orphaned_inodes;
 
-	/* list of arrays of inodes that are suspended */
-	TOID(struct pmemfile_inode_array) suspended_inodes;
-
 	/*
 	 * The array of root directories. Each one of them is a root of a
 	 * separate directory tree. The path "/" resolves to root #0, all other
@@ -308,7 +314,6 @@ struct pmemfile_super {
 	char padding[PMEMFILE_SUPER_SIZE
 			- 8  /* version */
 			- 16 * (PMEMFILE_ROOT_COUNT) /* toid */
-			- 16 /* toid */
 			- 16 /* toid */];
 };