From 90bd070aae6c4fb5d302f9c4b9c88be60c8197ec Mon Sep 17 00:00:00 2001 From: Wengang Wang Date: Fri, 9 Apr 2021 13:27:29 -0700 Subject: ocfs2: fix deadlock between setattr and dio_end_io_write The following deadlock is detected: truncate -> setattr path is waiting for pending direct IO to be done (inode->i_dio_count become zero) with inode->i_rwsem held (down_write). PID: 14827 TASK: ffff881686a9af80 CPU: 20 COMMAND: "ora_p005_hrltd9" #0 __schedule at ffffffff818667cc #1 schedule at ffffffff81866de6 #2 inode_dio_wait at ffffffff812a2d04 #3 ocfs2_setattr at ffffffffc05f322e [ocfs2] #4 notify_change at ffffffff812a5a09 #5 do_truncate at ffffffff812808f5 #6 do_sys_ftruncate.constprop.18 at ffffffff81280cf2 #7 sys_ftruncate at ffffffff81280d8e #8 do_syscall_64 at ffffffff81003949 #9 entry_SYSCALL_64_after_hwframe at ffffffff81a001ad dio completion path is going to complete one direct IO (decrement inode->i_dio_count), but before that it hung at locking inode->i_rwsem: #0 __schedule+700 at ffffffff818667cc #1 schedule+54 at ffffffff81866de6 #2 rwsem_down_write_failed+536 at ffffffff8186aa28 #3 call_rwsem_down_write_failed+23 at ffffffff8185a1b7 #4 down_write+45 at ffffffff81869c9d #5 ocfs2_dio_end_io_write+180 at ffffffffc05d5444 [ocfs2] #6 ocfs2_dio_end_io+85 at ffffffffc05d5a85 [ocfs2] #7 dio_complete+140 at ffffffff812c873c #8 dio_aio_complete_work+25 at ffffffff812c89f9 #9 process_one_work+361 at ffffffff810b1889 #10 worker_thread+77 at ffffffff810b233d #11 kthread+261 at ffffffff810b7fd5 #12 ret_from_fork+62 at ffffffff81a0035e Thus above forms ABBA deadlock. The same deadlock was mentioned in upstream commit 28f5a8a7c033 ("ocfs2: should wait dio before inode lock in ocfs2_setattr()"). It seems that that commit only removed the cluster lock (the victim of above dead lock) from the ABBA deadlock party. End-user visible effects: Process hang in truncate -> ocfs2_setattr path and other processes hang at ocfs2_dio_end_io_write path. This is to fix the deadlock itself. It removes inode_lock() call from dio completion path to remove the deadlock and add ip_alloc_sem lock in setattr path to synchronize the inode modifications. [wen.gang.wang@oracle.com: remove the "had_alloc_lock" as suggested] Link: https://lkml.kernel.org/r/20210402171344.1605-1-wen.gang.wang@oracle.com Link: https://lkml.kernel.org/r/20210331203654.3911-1-wen.gang.wang@oracle.com Signed-off-by: Wengang Wang Reviewed-by: Joseph Qi Cc: Mark Fasheh Cc: Joel Becker Cc: Junxiao Bi Cc: Changwei Ge Cc: Gang He Cc: Jun Piao Cc: Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- fs/ocfs2/aops.c | 11 +---------- fs/ocfs2/file.c | 8 ++++++-- 2 files changed, 7 insertions(+), 12 deletions(-) (limited to 'fs') diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c index 3bfb4147895a..ad20403b383f 100644 --- a/fs/ocfs2/aops.c +++ b/fs/ocfs2/aops.c @@ -2295,7 +2295,7 @@ static int ocfs2_dio_end_io_write(struct inode *inode, struct ocfs2_alloc_context *meta_ac = NULL; handle_t *handle = NULL; loff_t end = offset + bytes; - int ret = 0, credits = 0, locked = 0; + int ret = 0, credits = 0; ocfs2_init_dealloc_ctxt(&dealloc); @@ -2306,13 +2306,6 @@ static int ocfs2_dio_end_io_write(struct inode *inode, !dwc->dw_orphaned) goto out; - /* ocfs2_file_write_iter will get i_mutex, so we need not lock if we - * are in that context. */ - if (dwc->dw_writer_pid != task_pid_nr(current)) { - inode_lock(inode); - locked = 1; - } - ret = ocfs2_inode_lock(inode, &di_bh, 1); if (ret < 0) { mlog_errno(ret); @@ -2393,8 +2386,6 @@ out: if (meta_ac) ocfs2_free_alloc_context(meta_ac); ocfs2_run_deallocs(osb, &dealloc); - if (locked) - inode_unlock(inode); ocfs2_dio_free_write_ctx(inode, dwc); return ret; diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c index 6611c64ca0be..5edc1d0cf115 100644 --- a/fs/ocfs2/file.c +++ b/fs/ocfs2/file.c @@ -1245,22 +1245,24 @@ int ocfs2_setattr(struct user_namespace *mnt_userns, struct dentry *dentry, goto bail_unlock; } } + down_write(&OCFS2_I(inode)->ip_alloc_sem); handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS + 2 * ocfs2_quota_trans_credits(sb)); if (IS_ERR(handle)) { status = PTR_ERR(handle); mlog_errno(status); - goto bail_unlock; + goto bail_unlock_alloc; } status = __dquot_transfer(inode, transfer_to); if (status < 0) goto bail_commit; } else { + down_write(&OCFS2_I(inode)->ip_alloc_sem); handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); if (IS_ERR(handle)) { status = PTR_ERR(handle); mlog_errno(status); - goto bail_unlock; + goto bail_unlock_alloc; } } @@ -1273,6 +1275,8 @@ int ocfs2_setattr(struct user_namespace *mnt_userns, struct dentry *dentry, bail_commit: ocfs2_commit_trans(osb, handle); +bail_unlock_alloc: + up_write(&OCFS2_I(inode)->ip_alloc_sem); bail_unlock: if (status && inode_locked) { ocfs2_inode_unlock_tracker(inode, 1, &oh, had_lock); -- cgit v1.2.3 From df41872b68601059dd4a84858952dcae58acd331 Mon Sep 17 00:00:00 2001 From: Jack Qiu Date: Fri, 9 Apr 2021 13:27:35 -0700 Subject: fs: direct-io: fix missing sdio->boundary I encountered a hung task issue, but not a performance one. I run DIO on a device (need lba continuous, for example open channel ssd), maybe hungtask in below case: DIO: Checkpoint: get addr A(at boundary), merge into BIO, no submit because boundary missing flush dirty data(get addr A+1), wait IO(A+1) writeback timeout, because DIO(A) didn't submit get addr A+2 fail, because checkpoint is doing dio_send_cur_page() may clear sdio->boundary, so prevent it from missing a boundary. Link: https://lkml.kernel.org/r/20210322042253.38312-1-jack.qiu@huawei.com Fixes: b1058b981272 ("direct-io: submit bio after boundary buffer is added to it") Signed-off-by: Jack Qiu Reviewed-by: Jan Kara Cc: Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- fs/direct-io.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) (limited to 'fs') diff --git a/fs/direct-io.c b/fs/direct-io.c index b61491bf3166..b2e86e739d7a 100644 --- a/fs/direct-io.c +++ b/fs/direct-io.c @@ -812,6 +812,7 @@ submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, struct buffer_head *map_bh) { int ret = 0; + int boundary = sdio->boundary; /* dio_send_cur_page may clear it */ if (dio->op == REQ_OP_WRITE) { /* @@ -850,10 +851,10 @@ submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page, sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits; out: /* - * If sdio->boundary then we want to schedule the IO now to + * If boundary then we want to schedule the IO now to * avoid metadata seeks. */ - if (sdio->boundary) { + if (boundary) { ret = dio_send_cur_page(dio, sdio, map_bh); if (sdio->bio) dio_bio_submit(dio, sdio); -- cgit v1.2.3 From 53b74fa990bf76f290aa5930abfcf37424a1a865 Mon Sep 17 00:00:00 2001 From: Naohiro Aota Date: Thu, 8 Apr 2021 17:25:28 +0900 Subject: btrfs: zoned: move superblock logging zone location Moves the location of the superblock logging zones. The new locations of the logging zones are now determined based on fixed block addresses instead of on fixed zone numbers. The old placement method based on fixed zone numbers causes problems when one needs to inspect a file system image without access to the drive zone information. In such case, the super block locations cannot be reliably determined as the zone size is unknown. By locating the superblock logging zones using fixed addresses, we can scan a dumped file system image without the zone information since a super block copy will always be present at or after the fixed known locations. Introduce the following three pairs of zones containing fixed offset locations, regardless of the device zone size. - primary superblock: offset 0B (and the following zone) - first copy: offset 512G (and the following zone) - Second copy: offset 4T (4096G, and the following zone) If a logging zone is outside of the disk capacity, we do not record the superblock copy. The first copy position is much larger than for a non-zoned filesystem, which is at 64M. This is to avoid overlapping with the log zones for the primary superblock. This higher location is arbitrary but allows supporting devices with very large zone sizes, plus some space around in between. Such large zone size is unrealistic and very unlikely to ever be seen in real devices. Currently, SMR disks have a zone size of 256MB, and we are expecting ZNS drives to be in the 1-4GB range, so this limit gives us room to breathe. For now, we only allow zone sizes up to 8GB. The maximum zone size that would still fit in the space is 256G. The fixed location addresses are somewhat arbitrary, with the intent of maintaining superblock reliability for smaller and larger devices, with the preference for the latter. For this reason, there are two superblocks under the first 1T. This should cover use cases for physical devices and for emulated/device-mapper devices. The superblock logging zones are reserved for superblock logging and never used for data or metadata blocks. Note that we only reserve the two zones per primary/copy actually used for superblock logging. We do not reserve the ranges of zones possibly containing superblocks with the largest supported zone size (0-16GB, 512G-528GB, 4096G-4112G). The zones containing the fixed location offsets used to store superblocks on a non-zoned volume are also reserved to avoid confusion. Signed-off-by: Naohiro Aota Signed-off-by: David Sterba --- fs/btrfs/zoned.c | 53 ++++++++++++++++++++++++++++++++++++++++++----------- 1 file changed, 42 insertions(+), 11 deletions(-) (limited to 'fs') diff --git a/fs/btrfs/zoned.c b/fs/btrfs/zoned.c index 43948bd40e02..ba7a303300a3 100644 --- a/fs/btrfs/zoned.c +++ b/fs/btrfs/zoned.c @@ -21,9 +21,30 @@ /* Pseudo write pointer value for conventional zone */ #define WP_CONVENTIONAL ((u64)-2) +/* + * Location of the first zone of superblock logging zone pairs. + * + * - primary superblock: 0B (zone 0) + * - first copy: 512G (zone starting at that offset) + * - second copy: 4T (zone starting at that offset) + */ +#define BTRFS_SB_LOG_PRIMARY_OFFSET (0ULL) +#define BTRFS_SB_LOG_FIRST_OFFSET (512ULL * SZ_1G) +#define BTRFS_SB_LOG_SECOND_OFFSET (4096ULL * SZ_1G) + +#define BTRFS_SB_LOG_FIRST_SHIFT const_ilog2(BTRFS_SB_LOG_FIRST_OFFSET) +#define BTRFS_SB_LOG_SECOND_SHIFT const_ilog2(BTRFS_SB_LOG_SECOND_OFFSET) + /* Number of superblock log zones */ #define BTRFS_NR_SB_LOG_ZONES 2 +/* + * Maximum supported zone size. Currently, SMR disks have a zone size of + * 256MiB, and we are expecting ZNS drives to be in the 1-4GiB range. We do not + * expect the zone size to become larger than 8GiB in the near future. + */ +#define BTRFS_MAX_ZONE_SIZE SZ_8G + static int copy_zone_info_cb(struct blk_zone *zone, unsigned int idx, void *data) { struct blk_zone *zones = data; @@ -111,23 +132,22 @@ static int sb_write_pointer(struct block_device *bdev, struct blk_zone *zones, } /* - * The following zones are reserved as the circular buffer on ZONED btrfs. - * - The primary superblock: zones 0 and 1 - * - The first copy: zones 16 and 17 - * - The second copy: zones 1024 or zone at 256GB which is minimum, and - * the following one + * Get the first zone number of the superblock mirror */ static inline u32 sb_zone_number(int shift, int mirror) { - ASSERT(mirror < BTRFS_SUPER_MIRROR_MAX); + u64 zone; + ASSERT(mirror < BTRFS_SUPER_MIRROR_MAX); switch (mirror) { - case 0: return 0; - case 1: return 16; - case 2: return min_t(u64, btrfs_sb_offset(mirror) >> shift, 1024); + case 0: zone = 0; break; + case 1: zone = 1ULL << (BTRFS_SB_LOG_FIRST_SHIFT - shift); break; + case 2: zone = 1ULL << (BTRFS_SB_LOG_SECOND_SHIFT - shift); break; } - return 0; + ASSERT(zone <= U32_MAX); + + return (u32)zone; } /* @@ -300,10 +320,21 @@ int btrfs_get_dev_zone_info(struct btrfs_device *device) zone_sectors = bdev_zone_sectors(bdev); } - nr_sectors = bdev_nr_sectors(bdev); /* Check if it's power of 2 (see is_power_of_2) */ ASSERT(zone_sectors != 0 && (zone_sectors & (zone_sectors - 1)) == 0); zone_info->zone_size = zone_sectors << SECTOR_SHIFT; + + /* We reject devices with a zone size larger than 8GB */ + if (zone_info->zone_size > BTRFS_MAX_ZONE_SIZE) { + btrfs_err_in_rcu(fs_info, + "zoned: %s: zone size %llu larger than supported maximum %llu", + rcu_str_deref(device->name), + zone_info->zone_size, BTRFS_MAX_ZONE_SIZE); + ret = -EINVAL; + goto out; + } + + nr_sectors = bdev_nr_sectors(bdev); zone_info->zone_size_shift = ilog2(zone_info->zone_size); zone_info->max_zone_append_size = (u64)queue_max_zone_append_sectors(queue) << SECTOR_SHIFT; -- cgit v1.2.3 From c7d95613c7d6e003969722a290397b8271bdad17 Mon Sep 17 00:00:00 2001 From: Pavel Begunkov Date: Tue, 13 Apr 2021 11:43:00 +0100 Subject: io_uring: fix early sqd_list removal sqpoll hangs [ 245.463317] INFO: task iou-sqp-1374:1377 blocked for more than 122 seconds. [ 245.463334] task:iou-sqp-1374 state:D flags:0x00004000 [ 245.463345] Call Trace: [ 245.463352] __schedule+0x36b/0x950 [ 245.463376] schedule+0x68/0xe0 [ 245.463385] __io_uring_cancel+0xfb/0x1a0 [ 245.463407] do_exit+0xc0/0xb40 [ 245.463423] io_sq_thread+0x49b/0x710 [ 245.463445] ret_from_fork+0x22/0x30 It happens when sqpoll forgot to run park_task_work and goes to exit, then exiting user may remove ctx from sqd_list, and so corresponding io_sq_thread() -> io_uring_cancel_sqpoll() won't be executed. Hopefully it just stucks in do_exit() in this case. Fixes: dbe1bdbb39db ("io_uring: handle signals for IO threads like a normal thread") Reported-by: Joakim Hassila Signed-off-by: Pavel Begunkov Signed-off-by: Jens Axboe --- fs/io_uring.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) (limited to 'fs') diff --git a/fs/io_uring.c b/fs/io_uring.c index bd14327c8e7e..dff34975d86b 100644 --- a/fs/io_uring.c +++ b/fs/io_uring.c @@ -6754,6 +6754,9 @@ static int io_sq_thread(void *data) current->flags |= PF_NO_SETAFFINITY; mutex_lock(&sqd->lock); + /* a user may had exited before the thread started */ + io_run_task_work_head(&sqd->park_task_work); + while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) { int ret; bool cap_entries, sqt_spin, needs_sched; @@ -6770,10 +6773,10 @@ static int io_sq_thread(void *data) } cond_resched(); mutex_lock(&sqd->lock); - if (did_sig) - break; io_run_task_work(); io_run_task_work_head(&sqd->park_task_work); + if (did_sig) + break; timeout = jiffies + sqd->sq_thread_idle; continue; } -- cgit v1.2.3