Skip to content

Commit ae2da51

Browse files
ioworker0akpm00
ioworker0
authored and
akpm00
committed
locking/rwsem: make owner helpers globally available
Patch series "extend hung task blocker tracking to rwsems". Inspired by mutex blocker tracking[1], and having already extended it to semaphores, let's now add support for reader-writer semaphores (rwsems). The approach is simple: when a task enters TASK_UNINTERRUPTIBLE while waiting for an rwsem, we just call hung_task_set_blocker(). The hung task detector can then query the rwsem's owner to identify the lock holder. Tracking works reliably for writers, as there can only be a single writer holding the lock, and its task struct is stored in the owner field. The main challenge lies with readers. The owner field points to only one of many concurrent readers, so we might lose track of the blocker if that specific reader unlocks, even while others remain. This is not a significant issue, however. In practice, long-lasting lock contention is almost always caused by a writer. Therefore, reliably tracking the writer is the primary goal of this patch series ;) With this change, the hung task detector can now show blocker task's info like below: [Fri Jun 27 15:21:34 2025] INFO: task cat:28631 blocked for more than 122 seconds. [Fri Jun 27 15:21:34 2025] Tainted: G S 6.16.0-rc3 #8 [Fri Jun 27 15:21:34 2025] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [Fri Jun 27 15:21:34 2025] task:cat state:D stack:0 pid:28631 tgid:28631 ppid:28501 task_flags:0x400000 flags:0x00004000 [Fri Jun 27 15:21:34 2025] Call Trace: [Fri Jun 27 15:21:34 2025] <TASK> [Fri Jun 27 15:21:34 2025] __schedule+0x7c7/0x1930 [Fri Jun 27 15:21:34 2025] ? __pfx___schedule+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? policy_nodemask+0x215/0x340 [Fri Jun 27 15:21:34 2025] ? _raw_spin_lock_irq+0x8a/0xe0 [Fri Jun 27 15:21:34 2025] ? __pfx__raw_spin_lock_irq+0x10/0x10 [Fri Jun 27 15:21:34 2025] schedule+0x6a/0x180 [Fri Jun 27 15:21:34 2025] schedule_preempt_disabled+0x15/0x30 [Fri Jun 27 15:21:34 2025] rwsem_down_read_slowpath+0x55e/0xe10 [Fri Jun 27 15:21:34 2025] ? __pfx_rwsem_down_read_slowpath+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? __pfx___might_resched+0x10/0x10 [Fri Jun 27 15:21:34 2025] down_read+0xc9/0x230 [Fri Jun 27 15:21:34 2025] ? __pfx_down_read+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? __debugfs_file_get+0x14d/0x700 [Fri Jun 27 15:21:34 2025] ? __pfx___debugfs_file_get+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? handle_pte_fault+0x52a/0x710 [Fri Jun 27 15:21:34 2025] ? selinux_file_permission+0x3a9/0x590 [Fri Jun 27 15:21:34 2025] read_dummy_rwsem_read+0x4a/0x90 [Fri Jun 27 15:21:34 2025] full_proxy_read+0xff/0x1c0 [Fri Jun 27 15:21:34 2025] ? rw_verify_area+0x6d/0x410 [Fri Jun 27 15:21:34 2025] vfs_read+0x177/0xa50 [Fri Jun 27 15:21:34 2025] ? __pfx_vfs_read+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? fdget_pos+0x1cf/0x4c0 [Fri Jun 27 15:21:34 2025] ksys_read+0xfc/0x1d0 [Fri Jun 27 15:21:34 2025] ? __pfx_ksys_read+0x10/0x10 [Fri Jun 27 15:21:34 2025] do_syscall_64+0x66/0x2d0 [Fri Jun 27 15:21:34 2025] entry_SYSCALL_64_after_hwframe+0x76/0x7e [Fri Jun 27 15:21:34 2025] RIP: 0033:0x7f3f8faefb40 [Fri Jun 27 15:21:34 2025] RSP: 002b:00007ffdeda5ab98 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 [Fri Jun 27 15:21:34 2025] RAX: ffffffffffffffda RBX: 0000000000010000 RCX: 00007f3f8faefb40 [Fri Jun 27 15:21:34 2025] RDX: 0000000000010000 RSI: 00000000010fa000 RDI: 0000000000000003 [Fri Jun 27 15:21:34 2025] RBP: 00000000010fa000 R08: 0000000000000000 R09: 0000000000010fff [Fri Jun 27 15:21:34 2025] R10: 00007ffdeda59fe0 R11: 0000000000000246 R12: 00000000010fa000 [Fri Jun 27 15:21:34 2025] R13: 0000000000000003 R14: 0000000000000000 R15: 0000000000000fff [Fri Jun 27 15:21:34 2025] </TASK> [Fri Jun 27 15:21:34 2025] INFO: task cat:28631 <reader> blocked on an rw-semaphore likely owned by task cat:28630 <writer> [Fri Jun 27 15:21:34 2025] task:cat state:S stack:0 pid:28630 tgid:28630 ppid:28501 task_flags:0x400000 flags:0x00004000 [Fri Jun 27 15:21:34 2025] Call Trace: [Fri Jun 27 15:21:34 2025] <TASK> [Fri Jun 27 15:21:34 2025] __schedule+0x7c7/0x1930 [Fri Jun 27 15:21:34 2025] ? __pfx___schedule+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? __mod_timer+0x304/0xa80 [Fri Jun 27 15:21:34 2025] schedule+0x6a/0x180 [Fri Jun 27 15:21:34 2025] schedule_timeout+0xfb/0x230 [Fri Jun 27 15:21:34 2025] ? __pfx_schedule_timeout+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? __pfx_process_timeout+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? down_write+0xc4/0x140 [Fri Jun 27 15:21:34 2025] msleep_interruptible+0xbe/0x150 [Fri Jun 27 15:21:34 2025] read_dummy_rwsem_write+0x54/0x90 [Fri Jun 27 15:21:34 2025] full_proxy_read+0xff/0x1c0 [Fri Jun 27 15:21:34 2025] ? rw_verify_area+0x6d/0x410 [Fri Jun 27 15:21:34 2025] vfs_read+0x177/0xa50 [Fri Jun 27 15:21:34 2025] ? __pfx_vfs_read+0x10/0x10 [Fri Jun 27 15:21:34 2025] ? fdget_pos+0x1cf/0x4c0 [Fri Jun 27 15:21:34 2025] ksys_read+0xfc/0x1d0 [Fri Jun 27 15:21:34 2025] ? __pfx_ksys_read+0x10/0x10 [Fri Jun 27 15:21:34 2025] do_syscall_64+0x66/0x2d0 [Fri Jun 27 15:21:34 2025] entry_SYSCALL_64_after_hwframe+0x76/0x7e [Fri Jun 27 15:21:34 2025] RIP: 0033:0x7f8f288efb40 [Fri Jun 27 15:21:34 2025] RSP: 002b:00007ffffb631038 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 [Fri Jun 27 15:21:34 2025] RAX: ffffffffffffffda RBX: 0000000000010000 RCX: 00007f8f288efb40 [Fri Jun 27 15:21:34 2025] RDX: 0000000000010000 RSI: 000000002a4b5000 RDI: 0000000000000003 [Fri Jun 27 15:21:34 2025] RBP: 000000002a4b5000 R08: 0000000000000000 R09: 0000000000010fff [Fri Jun 27 15:21:34 2025] R10: 00007ffffb630460 R11: 0000000000000246 R12: 000000002a4b5000 [Fri Jun 27 15:21:34 2025] R13: 0000000000000003 R14: 0000000000000000 R15: 0000000000000fff [Fri Jun 27 15:21:34 2025] </TASK> This patch (of 3): In preparation for extending blocker tracking to support rwsems, make the rwsem_owner() and is_rwsem_reader_owned() helpers globally available for determining if the blocker is a writer or one of the readers. Additionally, a stale owner pointer in a reader-owned rwsem can lead to false positives in blocker tracking when CONFIG_DETECT_HUNG_TASK_BLOCKER is enabled. To mitigate this, clear the owner field on the reader unlock path, similar to what CONFIG_DEBUG_RWSEMS does. A NULL owner is better than a stale one for diagnostics. Link: https://lkml.kernel.org/r/[email protected] Link: https://lkml.kernel.org/r/[email protected] Link: https://lore.kernel.org/all/174046694331.2194069.15472952050240807469.stgit@mhiramat.tok.corp.google.com/ [1] Signed-off-by: Lance Yang <[email protected]> Reviewed-by: Masami Hiramatsu (Google) <[email protected]> Cc: Anna Schumaker <[email protected]> Cc: Boqun Feng <[email protected]> Cc: Ingo Molnar <[email protected]> Cc: Joel Granados <[email protected]> Cc: John Stultz <[email protected]> Cc: Kent Overstreet <[email protected]> Cc: Mingzhe Yang <[email protected]> Cc: Peter Zijlstra <[email protected]> Cc: Sergey Senozhatsky <[email protected]> Cc: Steven Rostedt <[email protected]> Cc: Tomasz Figa <[email protected]> Cc: Waiman Long <[email protected]> Cc: Will Deacon <[email protected]> Cc: Yongliang Gao <[email protected]> Cc: Zi Li <[email protected]> Signed-off-by: Andrew Morton <[email protected]>
1 parent 249e7ce commit ae2da51

File tree

2 files changed

+19
-7
lines changed

2 files changed

+19
-7
lines changed

‎include/linux/rwsem.h

Lines changed: 12 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -132,6 +132,18 @@ static inline int rwsem_is_contended(struct rw_semaphore *sem)
132132
return !list_empty(&sem->wait_list);
133133
}
134134

135+
#if defined(CONFIG_DEBUG_RWSEMS) || defined(CONFIG_DETECT_HUNG_TASK_BLOCKER)
136+
/*
137+
* Return just the real task structure pointer of the owner
138+
*/
139+
extern struct task_struct *rwsem_owner(struct rw_semaphore *sem);
140+
141+
/*
142+
* Return true if the rwsem is owned by a reader.
143+
*/
144+
extern bool is_rwsem_reader_owned(struct rw_semaphore *sem);
145+
#endif
146+
135147
#else /* !CONFIG_PREEMPT_RT */
136148

137149
#include <linux/rwbase_rt.h>

‎kernel/locking/rwsem.c

Lines changed: 7 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -181,11 +181,11 @@ static inline void rwsem_set_reader_owned(struct rw_semaphore *sem)
181181
__rwsem_set_reader_owned(sem, current);
182182
}
183183

184-
#ifdef CONFIG_DEBUG_RWSEMS
184+
#if defined(CONFIG_DEBUG_RWSEMS) || defined(CONFIG_DETECT_HUNG_TASK_BLOCKER)
185185
/*
186186
* Return just the real task structure pointer of the owner
187187
*/
188-
static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
188+
struct task_struct *rwsem_owner(struct rw_semaphore *sem)
189189
{
190190
return (struct task_struct *)
191191
(atomic_long_read(&sem->owner) & ~RWSEM_OWNER_FLAGS_MASK);
@@ -194,7 +194,7 @@ static inline struct task_struct *rwsem_owner(struct rw_semaphore *sem)
194194
/*
195195
* Return true if the rwsem is owned by a reader.
196196
*/
197-
static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
197+
bool is_rwsem_reader_owned(struct rw_semaphore *sem)
198198
{
199199
/*
200200
* Check the count to see if it is write-locked.
@@ -207,10 +207,10 @@ static inline bool is_rwsem_reader_owned(struct rw_semaphore *sem)
207207
}
208208

209209
/*
210-
* With CONFIG_DEBUG_RWSEMS configured, it will make sure that if there
211-
* is a task pointer in owner of a reader-owned rwsem, it will be the
212-
* real owner or one of the real owners. The only exception is when the
213-
* unlock is done by up_read_non_owner().
210+
* With CONFIG_DEBUG_RWSEMS or CONFIG_DETECT_HUNG_TASK_BLOCKER configured,
211+
* it will make sure that the owner field of a reader-owned rwsem either
212+
* points to a real reader-owner(s) or gets cleared. The only exception is
213+
* when the unlock is done by up_read_non_owner().
214214
*/
215215
static inline void rwsem_clear_reader_owned(struct rw_semaphore *sem)
216216
{

0 commit comments

Comments
 (0)