XArray updates for 5.7-rc1

- Fix two bugs which affected multi-index entries larger than 2^26 indices
  - Fix some documentation
  - Remove unused IDA macros
  - Add a small optimisation for tiny configurations
  - Fix a bug which could cause an RCU walker to terminate a marked walk early
 -----BEGIN PGP SIGNATURE-----
 
 iQEzBAABCgAdFiEEejHryeLBw/spnjHrDpNsjXcpgj4FAl6DT1YACgkQDpNsjXcp
 gj4GqAf/fVxinQwhlxF2UGVho67JgXpQklHtJi7uLUbdxEqueU1miz9W5rLfRUs5
 JP8+7ahwkV4lfphzf8FhfGkQmc3aWFsPLjpicnH1y0Mei0C24mYZgAO0ZA+S1o+5
 HZJ82jb8QtAHrhlnHZcwC20Q0bq5Q/HWVMfdv9YmZlBV0U4E/lUaZtpSc1T0Hcjt
 7faT8QWeudEwraklp0lRq2AFVJainbH4Xcyr7EHMcZySRfk7p6bgLlQ6QbHbjcsh
 IdtbbR/xOkC9eLX5WtDSzVg1q5AS3l0Fegu14A+/OuyXnsJYe1vF43jw9xId/V2u
 Wn+GRbeBxV3flQWK+8PoVvlQCsY4fw==
 =O2iY
 -----END PGP SIGNATURE-----

Merge tag 'xarray-5.7' of git://git.infradead.org/users/willy/linux-dax

Pull XArray updates from Matthew Wilcox:

 - Fix two bugs which affected multi-index entries larger than 2^26
   indices

 - Fix some documentation

 - Remove unused IDA macros

 - Add a small optimisation for tiny configurations

 - Fix a bug which could cause an RCU walker to terminate a marked walk
   early

* tag 'xarray-5.7' of git://git.infradead.org/users/willy/linux-dax:
  xarray: Fix early termination of xas_for_each_marked
  radix tree test suite: Support kmem_cache alignment
  XArray: Optimise xas_sibling() if !CONFIG_XARRAY_MULTI
  ida: remove abandoned macros
  XArray: Fix incorrect comment in header file
  XArray: Fix xas_pause for large multi-index entries
  XArray: Fix xa_find_next for large multi-index entries

include/linux/xarray.h

@@ -32,8 +32,8 @@
  * The following internal entries have a special meaning:
  *
  * 0-62: Sibling entries
- * 256: Zero entry
- * 257: Retry entry
+ * 256: Retry entry
+ * 257: Zero entry
  *
  * Errors are also represented as internal entries, but use the negative
  * space (-4094 to -2).  They're never stored in the slots array; only
@@ -1648,6 +1648,7 @@ static inline void *xas_next_marked(struct xa_state *xas, unsigned long max,
 								xa_mark_t mark)
 {
 	struct xa_node *node = xas->xa_node;
+	void *entry;
 	unsigned int offset;
 
 	if (unlikely(xas_not_node(node) || node->shift))
@@ -1659,7 +1660,10 @@ static inline void *xas_next_marked(struct xa_state *xas, unsigned long max,
 		return NULL;
 	if (offset == XA_CHUNK_SIZE)
 		return xas_find_marked(xas, max, mark);
-	return xa_entry(xas->xa, node, offset);
+	entry = xa_entry(xas->xa, node, offset);
+	if (!entry)
+		return xas_find_marked(xas, max, mark);
+	return entry;
 }
 
 /*

lib/radix-tree.c

@@ -55,14 +55,6 @@ struct kmem_cache *radix_tree_node_cachep;
 						RADIX_TREE_MAP_SHIFT))
 #define IDR_PRELOAD_SIZE	(IDR_MAX_PATH * 2 - 1)
 
-/*
- * The IDA is even shorter since it uses a bitmap at the last level.
- */
-#define IDA_INDEX_BITS		(8 * sizeof(int) - 1 - ilog2(IDA_BITMAP_BITS))
-#define IDA_MAX_PATH		(DIV_ROUND_UP(IDA_INDEX_BITS, \
-						RADIX_TREE_MAP_SHIFT))
-#define IDA_PRELOAD_SIZE	(IDA_MAX_PATH * 2 - 1)
-
 /*
  * Per-cpu pool of preloaded nodes
  */

lib/test_xarray.c

@@ -12,6 +12,9 @@
 static unsigned int tests_run;
 static unsigned int tests_passed;
 
+static const unsigned int order_limit =
+		IS_ENABLED(CONFIG_XARRAY_MULTI) ? BITS_PER_LONG : 1;
+
 #ifndef XA_DEBUG
 # ifdef __KERNEL__
 void xa_dump(const struct xarray *xa) { }
@@ -959,6 +962,20 @@ static noinline void check_multi_find_2(struct xarray *xa)
 	}
 }
 
+static noinline void check_multi_find_3(struct xarray *xa)
+{
+	unsigned int order;
+
+	for (order = 5; order < order_limit; order++) {
+		unsigned long index = 1UL << (order - 5);
+
+		XA_BUG_ON(xa, !xa_empty(xa));
+		xa_store_order(xa, 0, order - 4, xa_mk_index(0), GFP_KERNEL);
+		XA_BUG_ON(xa, xa_find_after(xa, &index, ULONG_MAX, XA_PRESENT));
+		xa_erase_index(xa, 0);
+	}
+}
+
 static noinline void check_find_1(struct xarray *xa)
 {
 	unsigned long i, j, k;
@@ -1081,6 +1098,7 @@ static noinline void check_find(struct xarray *xa)
 	for (i = 2; i < 10; i++)
 		check_multi_find_1(xa, i);
 	check_multi_find_2(xa);
+	check_multi_find_3(xa);
 }
 
 /* See find_swap_entry() in mm/shmem.c */
@@ -1138,6 +1156,42 @@ static noinline void check_find_entry(struct xarray *xa)
 	XA_BUG_ON(xa, !xa_empty(xa));
 }
 
+static noinline void check_pause(struct xarray *xa)
+{
+	XA_STATE(xas, xa, 0);
+	void *entry;
+	unsigned int order;
+	unsigned long index = 1;
+	unsigned int count = 0;
+
+	for (order = 0; order < order_limit; order++) {
+		XA_BUG_ON(xa, xa_store_order(xa, index, order,
+					xa_mk_index(index), GFP_KERNEL));
+		index += 1UL << order;
+	}
+
+	rcu_read_lock();
+	xas_for_each(&xas, entry, ULONG_MAX) {
+		XA_BUG_ON(xa, entry != xa_mk_index(1UL << count));
+		count++;
+	}
+	rcu_read_unlock();
+	XA_BUG_ON(xa, count != order_limit);
+
+	count = 0;
+	xas_set(&xas, 0);
+	rcu_read_lock();
+	xas_for_each(&xas, entry, ULONG_MAX) {
+		XA_BUG_ON(xa, entry != xa_mk_index(1UL << count));
+		count++;
+		xas_pause(&xas);
+	}
+	rcu_read_unlock();
+	XA_BUG_ON(xa, count != order_limit);
+
+	xa_destroy(xa);
+}
+
 static noinline void check_move_tiny(struct xarray *xa)
 {
 	XA_STATE(xas, xa, 0);
@@ -1646,6 +1700,7 @@ static int xarray_checks(void)
 	check_xa_alloc();
 	check_find(&array);
 	check_find_entry(&array);
+	check_pause(&array);
 	check_account(&array);
 	check_destroy(&array);
 	check_move(&array);

lib/xarray.c

@@ -970,7 +970,7 @@ void xas_pause(struct xa_state *xas)
 
 	xas->xa_node = XAS_RESTART;
 	if (node) {
-		unsigned int offset = xas->xa_offset;
+		unsigned long offset = xas->xa_offset;
 		while (++offset < XA_CHUNK_SIZE) {
 			if (!xa_is_sibling(xa_entry(xas->xa, node, offset)))
 				break;
@@ -1208,6 +1208,8 @@ void *xas_find_marked(struct xa_state *xas, unsigned long max, xa_mark_t mark)
 		}
 
 		entry = xa_entry(xas->xa, xas->xa_node, xas->xa_offset);
+		if (!entry && !(xa_track_free(xas->xa) && mark == XA_FREE_MARK))
+			continue;
 		if (!xa_is_node(entry))
 			return entry;
 		xas->xa_node = xa_to_node(entry);
@@ -1836,10 +1838,11 @@ static bool xas_sibling(struct xa_state *xas)
 	struct xa_node *node = xas->xa_node;
 	unsigned long mask;
 
-	if (!node)
+	if (!IS_ENABLED(CONFIG_XARRAY_MULTI) || !node)
 		return false;
 	mask = (XA_CHUNK_SIZE << node->shift) - 1;
-	return (xas->xa_index & mask) > (xas->xa_offset << node->shift);
+	return (xas->xa_index & mask) >
+		((unsigned long)xas->xa_offset << node->shift);
 }
 
 /**

tools/testing/radix-tree/Makefile

@@ -7,8 +7,8 @@ LDLIBS+= -lpthread -lurcu
 TARGETS = main idr-test multiorder xarray
 CORE_OFILES := xarray.o radix-tree.o idr.o linux.o test.o find_bit.o bitmap.o
 OFILES = main.o $(CORE_OFILES) regression1.o regression2.o regression3.o \
-	 regression4.o \
-	 tag_check.o multiorder.o idr-test.o iteration_check.o benchmark.o
+	 regression4.o tag_check.o multiorder.o idr-test.o iteration_check.o \
+	 iteration_check_2.o benchmark.o
 
 ifndef SHIFT
 	SHIFT=3

tools/testing/radix-tree/iteration_check_2.c

@@ -0,0 +1,87 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * iteration_check_2.c: Check that deleting a tagged entry doesn't cause
+ * an RCU walker to finish early.
+ * Copyright (c) 2020 Oracle
+ * Author: Matthew Wilcox <willy@infradead.org>
+ */
+#include <pthread.h>
+#include "test.h"
+
+static volatile bool test_complete;
+
+static void *iterator(void *arg)
+{
+	XA_STATE(xas, arg, 0);
+	void *entry;
+
+	rcu_register_thread();
+
+	while (!test_complete) {
+		xas_set(&xas, 0);
+		rcu_read_lock();
+		xas_for_each_marked(&xas, entry, ULONG_MAX, XA_MARK_0)
+			;
+		rcu_read_unlock();
+		assert(xas.xa_index >= 100);
+	}
+
+	rcu_unregister_thread();
+	return NULL;
+}
+
+static void *throbber(void *arg)
+{
+	struct xarray *xa = arg;
+
+	rcu_register_thread();
+
+	while (!test_complete) {
+		int i;
+
+		for (i = 0; i < 100; i++) {
+			xa_store(xa, i, xa_mk_value(i), GFP_KERNEL);
+			xa_set_mark(xa, i, XA_MARK_0);
+		}
+		for (i = 0; i < 100; i++)
+			xa_erase(xa, i);
+	}
+
+	rcu_unregister_thread();
+	return NULL;
+}
+
+void iteration_test2(unsigned test_duration)
+{
+	pthread_t threads[2];
+	DEFINE_XARRAY(array);
+	int i;
+
+	printv(1, "Running iteration test 2 for %d seconds\n", test_duration);
+
+	test_complete = false;
+
+	xa_store(&array, 100, xa_mk_value(100), GFP_KERNEL);
+	xa_set_mark(&array, 100, XA_MARK_0);
+
+	if (pthread_create(&threads[0], NULL, iterator, &array)) {
+		perror("create iterator thread");
+		exit(1);
+	}
+	if (pthread_create(&threads[1], NULL, throbber, &array)) {
+		perror("create throbber thread");
+		exit(1);
+	}
+
+	sleep(test_duration);
+	test_complete = true;
+
+	for (i = 0; i < 2; i++) {
+		if (pthread_join(threads[i], NULL)) {
+			perror("pthread_join");
+			exit(1);
+		}
+	}
+
+	xa_destroy(&array);
+}

tools/testing/radix-tree/linux.c

@@ -19,37 +19,44 @@ int test_verbose;
 
 struct kmem_cache {
 	pthread_mutex_t lock;
-	int size;
+	unsigned int size;
+	unsigned int align;
 	int nr_objs;
 	void *objs;
 	void (*ctor)(void *);
 };
 
-void *kmem_cache_alloc(struct kmem_cache *cachep, int flags)
+void *kmem_cache_alloc(struct kmem_cache *cachep, int gfp)
 {
-	struct radix_tree_node *node;
+	void *p;
 
-	if (!(flags & __GFP_DIRECT_RECLAIM))
+	if (!(gfp & __GFP_DIRECT_RECLAIM))
 		return NULL;
 
 	pthread_mutex_lock(&cachep->lock);
 	if (cachep->nr_objs) {
+		struct radix_tree_node *node = cachep->objs;
 		cachep->nr_objs--;
-		node = cachep->objs;
 		cachep->objs = node->parent;
 		pthread_mutex_unlock(&cachep->lock);
 		node->parent = NULL;
+		p = node;
 	} else {
 		pthread_mutex_unlock(&cachep->lock);
-		node = malloc(cachep->size);
+		if (cachep->align)
+			posix_memalign(&p, cachep->align, cachep->size);
+		else
+			p = malloc(cachep->size);
 		if (cachep->ctor)
-			cachep->ctor(node);
+			cachep->ctor(p);
+		else if (gfp & __GFP_ZERO)
+			memset(p, 0, cachep->size);
 	}
 
 	uatomic_inc(&nr_allocated);
 	if (kmalloc_verbose)
-		printf("Allocating %p from slab\n", node);
-	return node;
+		printf("Allocating %p from slab\n", p);
+	return p;
 }
 
 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
@@ -59,7 +66,7 @@ void kmem_cache_free(struct kmem_cache *cachep, void *objp)
 	if (kmalloc_verbose)
 		printf("Freeing %p to slab\n", objp);
 	pthread_mutex_lock(&cachep->lock);
-	if (cachep->nr_objs > 10) {
+	if (cachep->nr_objs > 10 || cachep->align) {
 		memset(objp, POISON_FREE, cachep->size);
 		free(objp);
 	} else {
@@ -98,13 +105,14 @@ void kfree(void *p)
 }
 
 struct kmem_cache *
-kmem_cache_create(const char *name, size_t size, size_t offset,
-	unsigned long flags, void (*ctor)(void *))
+kmem_cache_create(const char *name, unsigned int size, unsigned int align,
+		unsigned int flags, void (*ctor)(void *))
 {
 	struct kmem_cache *ret = malloc(sizeof(*ret));
 
 	pthread_mutex_init(&ret->lock, NULL);
 	ret->size = size;
+	ret->align = align;
 	ret->nr_objs = 0;
 	ret->objs = NULL;
 	ret->ctor = ctor;

tools/testing/radix-tree/linux/slab.h

@@ -20,8 +20,8 @@ static inline void *kzalloc(size_t size, gfp_t gfp)
 void *kmem_cache_alloc(struct kmem_cache *cachep, int flags);
 void kmem_cache_free(struct kmem_cache *cachep, void *objp);
 
-struct kmem_cache *
-kmem_cache_create(const char *name, size_t size, size_t offset,
-	unsigned long flags, void (*ctor)(void *));
+struct kmem_cache *kmem_cache_create(const char *name, unsigned int size,
+			unsigned int align, unsigned int flags,
+			void (*ctor)(void *));
 
 #endif		/* SLAB_H */

tools/testing/radix-tree/main.c

@@ -311,6 +311,7 @@ int main(int argc, char **argv)
 	regression4_test();
 	iteration_test(0, 10 + 90 * long_run);
 	iteration_test(7, 10 + 90 * long_run);
+	iteration_test2(10 + 90 * long_run);
 	single_thread_tests(long_run);
 
 	/* Free any remaining preallocated nodes */

tools/testing/radix-tree/test.h

@@ -34,6 +34,7 @@ void xarray_tests(void);
 void tag_check(void);
 void multiorder_checks(void);
 void iteration_test(unsigned order, unsigned duration);
+void iteration_test2(unsigned duration);
 void benchmark(void);
 void idr_checks(void);
 void ida_tests(void);