generic: test DAX DMA vs truncate/hole-punch
authorRoss Zwisler <ross.zwisler@linux.intel.com>
Fri, 27 Jul 2018 21:12:52 +0000 (15:12 -0600)
committerEryu Guan <guaneryu@gmail.com>
Sun, 29 Jul 2018 05:43:57 +0000 (13:43 +0800)
This adds a regression test for the following series:

[PATCH v4 0/2] ext4: fix DAX dma vs truncate/hole-punch
https://lists.01.org/pipermail/linux-nvdimm/2018-July/016842.html

which adds synchronization between DAX DMA in ext4 and truncate/hole-punch.
The intention of the test is to test those specific changes, but it runs
fine both with XFS and without DAX so I've put it in the generic tests
instead of ext4 and not restricted it to only DAX configurations.

When run with v4.18-rc6 + DAX + ext4, this test will hit the following
WARN_ON_ONCE() in dax_disassociate_entry():

WARN_ON_ONCE(trunc && page_ref_count(page) > 1);

If you change this to a WARN_ON() instead, you can see that each of the
four paths being exercised in this test hits that condition many times in
the one second that the subtest is being run.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Eryu Guan <guaneryu@gmail.com>
Signed-off-by: Eryu Guan <guaneryu@gmail.com>
.gitignore
src/Makefile
src/t_mmap_collision.c [new file with mode: 0644]
tests/generic/503 [new file with mode: 0755]
tests/generic/503.out [new file with mode: 0644]
tests/generic/group

index efc73a7c5c83fc539d0805137830b4eeac47816a..ea1aac8afa09968cad79ba15e0369d62be7b8cda 100644 (file)
 /src/t_holes
 /src/t_immutable
 /src/t_locks_execve
 /src/t_holes
 /src/t_immutable
 /src/t_locks_execve
+/src/t_mmap_collision
 /src/t_mmap_cow_race
 /src/t_mmap_dio
 /src/t_mmap_fallocate
 /src/t_mmap_cow_race
 /src/t_mmap_dio
 /src/t_mmap_fallocate
index 9e971bccae5f3a6d2406cc44e8faea567caa6563..418265857be5c5994b6fc671f930939d84d2c086 100644 (file)
@@ -16,7 +16,7 @@ TARGETS = dirstress fill fill2 getpagesize holes lstat64 \
        holetest t_truncate_self t_mmap_dio af_unix t_mmap_stale_pmd \
        t_mmap_cow_race t_mmap_fallocate fsync-err t_mmap_write_ro \
        t_ext4_dax_journal_corruption t_ext4_dax_inline_corruption \
        holetest t_truncate_self t_mmap_dio af_unix t_mmap_stale_pmd \
        t_mmap_cow_race t_mmap_fallocate fsync-err t_mmap_write_ro \
        t_ext4_dax_journal_corruption t_ext4_dax_inline_corruption \
-       t_ofd_locks t_locks_execve
+       t_ofd_locks t_locks_execve t_mmap_collision
 
 LINUX_TARGETS = xfsctl bstat t_mtab getdevicesize preallo_rw_pattern_reader \
        preallo_rw_pattern_writer ftrunc trunc fs_perms testx looptest \
 
 LINUX_TARGETS = xfsctl bstat t_mtab getdevicesize preallo_rw_pattern_reader \
        preallo_rw_pattern_writer ftrunc trunc fs_perms testx looptest \
diff --git a/src/t_mmap_collision.c b/src/t_mmap_collision.c
new file mode 100644 (file)
index 0000000..d547bc0
--- /dev/null
@@ -0,0 +1,235 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (c) 2018 Intel Corporation.
+ *
+ * As of kernel version 4.18-rc6 Linux has an issue with ext4+DAX where DMA
+ * and direct I/O operations aren't synchronized with respect to operations
+ * which can change the block mappings of an inode.  This means that we can
+ * schedule an I/O for an inode and have the block mapping for that inode
+ * change before the I/O is actually complete.  So, blocks which were once
+ * allocated to a given inode and then freed could still have I/O operations
+ * happening to them.  If these blocks have also been reallocated to a
+ * different inode, this interaction can lead to data corruption.
+ *
+ * This test exercises four of the paths in ext4 which hit this issue.
+ */
+#include <errno.h>
+#include <fcntl.h>
+#include <pthread.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#define PAGE(a) ((a)*0x1000)
+#define FILE_SIZE PAGE(4)
+
+void *dax_data;
+int nodax_fd;
+int dax_fd;
+bool done;
+
+#define err_exit(op)                                                          \
+{                                                                             \
+       fprintf(stderr, "%s %s: %s\n", __func__, op, strerror(errno));        \
+       exit(1);                                                              \
+}
+
+#if defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
+void punch_hole_fn(void *ptr)
+{
+       ssize_t read;
+       int rc;
+
+       while (!done) {
+               read = 0;
+
+               do {
+                       rc = pread(nodax_fd, dax_data + read, FILE_SIZE - read,
+                                       read);
+                       if (rc > 0)
+                               read += rc;
+               } while (rc > 0);
+
+               if (read != FILE_SIZE || rc != 0)
+                       err_exit("pread");
+
+               rc = fallocate(dax_fd,
+                               FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
+                               0, FILE_SIZE);
+               if (rc < 0)
+                       err_exit("fallocate");
+
+               usleep(rand() % 1000);
+       }
+}
+#else
+void punch_hole_fn(void *ptr) { }
+#endif
+
+#if defined(FALLOC_FL_ZERO_RANGE) && defined(FALLOC_FL_KEEP_SIZE)
+void zero_range_fn(void *ptr)
+{
+       ssize_t read;
+       int rc;
+
+       while (!done) {
+               read = 0;
+
+               do {
+                       rc = pread(nodax_fd, dax_data + read, FILE_SIZE - read,
+                                       read);
+                       if (rc > 0)
+                               read += rc;
+               } while (rc > 0);
+
+               if (read != FILE_SIZE || rc != 0)
+                       err_exit("pread");
+
+               rc = fallocate(dax_fd,
+                               FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE,
+                               0, FILE_SIZE);
+               if (rc < 0)
+                       err_exit("fallocate");
+
+               usleep(rand() % 1000);
+       }
+}
+#else
+void zero_range_fn(void *ptr) { }
+#endif
+
+void truncate_down_fn(void *ptr)
+{
+       ssize_t read;
+       int rc;
+
+       while (!done) {
+               read = 0;
+
+               if (ftruncate(dax_fd, 0) < 0)
+                       err_exit("ftruncate");
+               if (fallocate(dax_fd, 0, 0, FILE_SIZE) < 0)
+                       err_exit("fallocate");
+
+               do {
+                       rc = pread(nodax_fd, dax_data + read, FILE_SIZE - read,
+                                       read);
+                       if (rc > 0)
+                               read += rc;
+               } while (rc > 0);
+
+               /*
+                * For this test we ignore errors from pread().  These errors
+                * can happen if we try and read while the other thread has
+                * made the file size 0.
+                */
+
+               usleep(rand() % 1000);
+       }
+}
+
+#ifdef FALLOC_FL_COLLAPSE_RANGE
+void collapse_range_fn(void *ptr)
+{
+       ssize_t read;
+       int rc;
+
+       while (!done) {
+               read = 0;
+
+               if (fallocate(dax_fd, 0, 0, FILE_SIZE) < 0)
+                       err_exit("fallocate 1");
+               if (fallocate(dax_fd, FALLOC_FL_COLLAPSE_RANGE, 0, PAGE(1)) < 0)
+                       err_exit("fallocate 2");
+               if (fallocate(dax_fd, 0, 0, FILE_SIZE) < 0)
+                       err_exit("fallocate 3");
+
+               do {
+                       rc = pread(nodax_fd, dax_data + read, FILE_SIZE - read,
+                                       read);
+                       if (rc > 0)
+                               read += rc;
+               } while (rc > 0);
+
+               /* For this test we ignore errors from pread. */
+
+               usleep(rand() % 1000);
+       }
+}
+#else
+void collapse_range_fn(void *ptr) { }
+#endif
+
+void run_test(void (*test_fn)(void *))
+{
+       const int NUM_THREADS = 2;
+       pthread_t worker_thread[NUM_THREADS];
+       int i;
+
+       done = 0;
+       for (i = 0; i < NUM_THREADS; i++)
+               pthread_create(&worker_thread[i], NULL, (void*)test_fn, NULL);
+
+       sleep(1);
+       done = 1;
+
+       for (i = 0; i < NUM_THREADS; i++)
+               pthread_join(worker_thread[i], NULL);
+}
+
+int main(int argc, char *argv[])
+{
+       int err;
+
+       if (argc != 3) {
+               printf("Usage: %s <dax file> <non-dax file>\n",
+                               basename(argv[0]));
+               exit(0);
+       }
+
+       dax_fd = open(argv[1], O_RDWR|O_CREAT, S_IRUSR|S_IWUSR);
+       if (dax_fd < 0)
+               err_exit("dax_fd open");
+
+       nodax_fd = open(argv[2], O_RDWR|O_CREAT|O_DIRECT, S_IRUSR|S_IWUSR);
+       if (nodax_fd < 0)
+               err_exit("nodax_fd open");
+
+       if (ftruncate(dax_fd, 0) < 0)
+               err_exit("dax_fd ftruncate");
+       if (fallocate(dax_fd, 0, 0, FILE_SIZE) < 0)
+               err_exit("dax_fd fallocate");
+
+       if (ftruncate(nodax_fd, 0) < 0)
+               err_exit("nodax_fd ftruncate");
+       if (fallocate(nodax_fd, 0, 0, FILE_SIZE) < 0)
+               err_exit("nodax_fd fallocate");
+
+       dax_data = mmap(NULL, FILE_SIZE, PROT_READ|PROT_WRITE, MAP_SHARED,
+                       dax_fd, 0);
+       if (dax_data == MAP_FAILED)
+               err_exit("mmap");
+
+       run_test(&punch_hole_fn);
+       run_test(&zero_range_fn);
+       run_test(&truncate_down_fn);
+       run_test(&collapse_range_fn);
+
+       if (munmap(dax_data, FILE_SIZE) != 0)
+               err_exit("munmap");
+
+       err = close(dax_fd);
+       if (err < 0)
+               err_exit("dax_fd close");
+
+       err = close(nodax_fd);
+       if (err < 0)
+               err_exit("nodax_fd close");
+
+       return 0;
+}
diff --git a/tests/generic/503 b/tests/generic/503
new file mode 100755 (executable)
index 0000000..31d7218
--- /dev/null
@@ -0,0 +1,64 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2018 Intel Corporation.  All Rights Reserved.
+#
+# FS QA Test No. 503
+#
+# This is a regression test for kernel patch:
+#   ext4: handle layout changes to pinned DAX mapping
+#
+# This test exercises each of the DAX paths in ext4 which remove blocks from
+# an inode's block map.  This includes things like hole punch, truncate down,
+# etc.  This test was written to regression test errors seen with an ext4 +
+# DAX setup, but the test runs fine with or without DAX and with XFS so we
+# don't require the DAX mount option or a specific filesystem for the test.
+
+seq=`basename $0`
+seqres=$RESULT_DIR/$seq
+echo "QA output created by $seq"
+
+here=`pwd`
+tmp=/tmp/$$
+status=1       # failure is the default!
+trap "_cleanup; exit \$status" 0 1 2 3 15
+
+_cleanup()
+{
+       cd /
+       rm -f $tmp.*
+}
+
+# get standard environment, filters and checks
+. ./common/rc
+. ./common/filter
+
+# remove previous $seqres.full before test
+rm -f $seqres.full
+
+# Modify as appropriate.
+_supported_fs generic
+_supported_os Linux
+_require_test
+_require_scratch
+_require_test_program "t_mmap_collision"
+_require_xfs_io_command "falloc"
+_require_xfs_io_command "fpunch"
+_require_xfs_io_command "fcollapse"
+_require_xfs_io_command "fzero"
+
+_scratch_mkfs >> $seqres.full 2>&1
+# To get the failure we turn off DAX on our SCRATCH_MNT so we can get O_DIRECT
+# behavior.  We will continue to use unmodified mount options for the test
+# TEST_DIR.  The failures fixed by the above mentioned kernel patch trigger
+# when those mount options include "-o dax", but the test runs fine without
+# that option so we don't require it.
+export MOUNT_OPTIONS=""
+_scratch_mount >> $seqres.full 2>&1
+
+# real QA test starts here
+$here/src/t_mmap_collision $TEST_DIR/testfile $SCRATCH_MNT/testfile
+
+# success, all done
+echo "Silence is golden"
+status=0
+exit
diff --git a/tests/generic/503.out b/tests/generic/503.out
new file mode 100644 (file)
index 0000000..6c1400a
--- /dev/null
@@ -0,0 +1,2 @@
+QA output created by 503
+Silence is golden
index d0b7dcf642aabef2d6efab7a19f05b76975e283b..79a5705a87bbf5acd851c3bf73dfc7df69651249 100644 (file)
 500 auto thin trim
 501 auto quick clone log
 502 auto quick log
 500 auto thin trim
 501 auto quick clone log
 502 auto quick log
+503 auto quick dax punch collapse zero