From: Ross Zwisler Subject: [fstests PATCH v3] generic/999: test DAX DMA vs truncate/hole-punch Date: Fri, 27 Jul 2018 15:12:52 -0600 Message-ID: <20180727211252.14895-1-zwisler@kernel.org> References: <20180714125307.GF2830@desktop> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Eryu Guan , fstests-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, Dave Chinner , Jan Kara , Dan Williams , Christoph Hellwig , linux-nvdimm-hn68Rpc1hR1g9hUCZPvPmw@public.gmane.org, linux-ext4-u79uwXL29TY76Z2rM5mHXA@public.gmane.org Return-path: in-reply-to: <20180714125307.GF2830@desktop>> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: linux-nvdimm-bounces-hn68Rpc1hR1g9hUCZPvPmw@public.gmane.org Sender: "Linux-nvdimm" List-Id: linux-ext4.vger.kernel.org From: Ross Zwisler This adds a regression test for the following series: 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 --- Changes since v2: - Added detailed description to tests/generic/999 explaining the purpose of the test (Eryu). - Added _require_xfs_io_command for falloc, fpunch, fcollapse and fzero to account for filesystems that don't support these commands (Eryu). - Incorporated other feedback from Eryu. Thanks, Eryu, for the review. --- .gitignore | 1 + src/Makefile | 2 +- src/t_mmap_collision.c | 235 +++++++++++++++++++++++++++++++++++++++++++++++++ tests/generic/999 | 64 ++++++++++++++ tests/generic/999.out | 2 + tests/generic/group | 1 + 6 files changed, 304 insertions(+), 1 deletion(-) create mode 100644 src/t_mmap_collision.c create mode 100755 tests/generic/999 create mode 100644 tests/generic/999.out diff --git a/.gitignore b/.gitignore index efc73a7c..ea1aac8a 100644 --- a/.gitignore +++ b/.gitignore @@ -125,6 +125,7 @@ /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 diff --git a/src/Makefile b/src/Makefile index 9e971bcc..41826585 100644 --- a/src/Makefile +++ b/src/Makefile @@ -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 \ - 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 \ diff --git a/src/t_mmap_collision.c b/src/t_mmap_collision.c new file mode 100644 index 00000000..d547bc05 --- /dev/null +++ b/src/t_mmap_collision.c @@ -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 +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#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 \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/999 b/tests/generic/999 new file mode 100755 index 00000000..0681b075 --- /dev/null +++ b/tests/generic/999 @@ -0,0 +1,64 @@ +#! /bin/bash +# SPDX-License-Identifier: GPL-2.0 +# Copyright (c) 2018 Intel Corporation. All Rights Reserved. +# +# FS QA Test generic/999 +# +# 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/999.out b/tests/generic/999.out new file mode 100644 index 00000000..3b276ca8 --- /dev/null +++ b/tests/generic/999.out @@ -0,0 +1,2 @@ +QA output created by 999 +Silence is golden diff --git a/tests/generic/group b/tests/generic/group index d0b7dcf6..92016cf4 100644 --- a/tests/generic/group +++ b/tests/generic/group @@ -505,3 +505,4 @@ 500 auto thin trim 501 auto quick clone log 502 auto quick log +999 auto quick dax punch collapse zero -- 2.14.4