Skip to contentSkip to conversation

printk: use printk_safe to handle printk() recursive calls

RFC PATCH Superseded Sergey Senozhatsky 6 participants 23 messages Oct 18, 2016 – Oct 20, 2016

Subsystems (4)

v3 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

v3 cover printk: use printk_safe to handle printk() recursive calls

Hello,

RFC

This patch set extends a lock-less NMI per-cpu buffers idea to
handle recursive printk() calls. The basic mechanism is pretty much the same -- at the beginning of a deadlock-prone section we switch to lock-less printk callback, and return back to a default printk implementation at the end; the messages are getting flushed to a logbuf buffer from a safer context.

Deadlock scenarios that printk_safe can handle:

a) printk recursion from logbuf_lock spin_lock section in printk() printk()

    raw_spin_lock(&logbuf_lock);
    WARN_ON(1);
    raw_spin_unlock(&logbuf_lock);

b) printk from sem->lock spin_lock section
printk()

    console_trylock()
      down_trylock()
        raw_spin_lock_irqsave(&sem->lock, flags);
        WARN_ON(1);
        raw_spin_unlock_irqrestore(&sem->lock, flags);

c) printk from logbuf_lock spin_lock section in console_unlock() printk()

    console_unlock()
      raw_spin_lock(&logbuf_lock);
      WARN_ON(1);
      raw_spin_unlock(&logbuf_lock);

d) printk from ->pi_lock from semaphore up
printk()

    console_unlock()
      up()
        try_to_wake_up()
          raw_spin_lock_irqsave(&p->pi_lock, flags);
          WARN_ON(1);
          raw_spin_unlock_irqrestore(&p->pi_lock, flags);

[[against next-20161018]]

v3: (review by Petr)
-- renamed to printk_safe
-- !NMI config build fix
-- report lost messages for both printk_sae and printk_nmi -- dropped recursion reporting patch
-- etc.

v2:
-- fix build error on !NMI configs, reported by Fengguang -- reworked the series based on Petr's feedback -- added a new patch to drop zap_locks()

Sergey Senozhatsky (6):
printk: use vprintk_func in vprintk()
printk: rename nmi.c and exported api
printk: introduce per-cpu safe_print seq buffer
printk: report lost messages in printk safe/nmi contexts
printk: use printk_safe buffers
printk: remove zap_locks() function

 arch/arm/kernel/smp.c                  |   4 +-
 include/linux/hardirq.h                |   4 +-
 include/linux/printk.h                 |  29 ++--
 init/Kconfig                           |  16 ++-
 init/main.c                            |   2 +-
 kernel/kexec_core.c                    |   2 +-
 kernel/panic.c                         |   4 +-
 kernel/printk/Makefile                 |   3 +-
 kernel/printk/internal.h               |  65 +++++----
 kernel/printk/printk.c                 | 116 ++++------------
 kernel/printk/{nmi.c => printk_safe.c} | 239 +++++++++++++++++++++++++--------
 lib/nmi_backtrace.c                    |   2 +-
 12 files changed, 277 insertions(+), 209 deletions(-)
 rename kernel/printk/{nmi.c => printk_safe.c} (54%)
Signature · 1 line
2.10.1.382.ga23ca1b
#1 lore raw

RFC PATCH v3 1/6 printk: use vprintk_func in vprintk()

vprintk(), just like printk(), better be using per-cpu printk_func instead of direct vprintk_emit() call. Just in case if vprintk() will ever be called from NMI, or from any other context that can deadlock in printk().

---
 kernel/printk/printk.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/printk/printk.c +1 -1
index d5e3973..be7f60b 100644
@@ -1917,7 +1917,7 @@ EXPORT_SYMBOL(vprintk_emit);
 
 asmlinkage int vprintk(const char *fmt, va_list args)
 {
-	return vprintk_emit(0, LOGLEVEL_DEFAULT, NULL, 0, fmt, args);
+	return vprintk_func(fmt, args);
 }
 EXPORT_SYMBOL(vprintk);
 
Signature · 1 line
2.10.1.382.ga23ca1b
reviewer Reviewed-by on v3 1/6
#2 lore raw

Re: [RFC][PATCHv3 1/6] printk: use vprintk_func in vprintk()

Quoted 6 lines · On Wed, 19 Oct 2016 00:40:40 +0900 Sergey Senozhatsky…

vprintk(), just like printk(), better be using per-cpu printk_func instead of direct vprintk_emit() call. Just in case if vprintk() will ever be called from NMI, or from any other context that can deadlock in printk().

-- Steve

Quoted 17 lines
---
 kernel/printk/printk.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/printk/printk.c +1 -1
index d5e3973..be7f60b 100644
@@ -1917,7 +1917,7 @@ EXPORT_SYMBOL(vprintk_emit);
 
 asmlinkage int vprintk(const char *fmt, va_list args)
 {
-	return vprintk_emit(0, LOGLEVEL_DEFAULT, NULL, 0, fmt, args);
+	return vprintk_func(fmt, args);
 }
 EXPORT_SYMBOL(vprintk);
 
#3 lore raw

Re: [RFC][PATCHv3 1/6] printk: use vprintk_func in vprintk()

Quoted 11 lines · Steven Rostedt
Quoted 6 lines · On Wed, 19 Oct 2016 00:40:40 +0900 Sergey Senozhatsky…

vprintk(), just like printk(), better be using per-cpu printk_func instead of direct vprintk_emit() call. Just in case if vprintk() will ever be called from NMI, or from any other context that can deadlock in printk().

Thanks.

	-ss
#4 lore raw

RFC PATCH v3 2/6 printk: rename nmi.c and exported api

In reply to Sergey Senozhatsky

A preparation patch for printk_safe work. No functional change.
- rename nmi.c to print_safe.c
- rename exported functions to have `safe' prefix.

---
 arch/arm/kernel/smp.c                  |  4 +-
 include/linux/hardirq.h                |  4 +-
 include/linux/printk.h                 | 20 +++++-----
 init/Kconfig                           | 16 ++++----
 init/main.c                            |  2 +-
 kernel/kexec_core.c                    |  2 +-
 kernel/panic.c                         |  4 +-
 kernel/printk/Makefile                 |  2 +-
 kernel/printk/{nmi.c => printk_safe.c} | 69 +++++++++++++++++-----------------
 lib/nmi_backtrace.c                    |  2 +-
 10 files changed, 64 insertions(+), 61 deletions(-)
 rename kernel/printk/{nmi.c => printk_safe.c} (77%)
arch/arm/kernel/smp.c +2 -2
index 7dd14e8..8b056e1 100644
@@ -644,11 +644,11 @@ void handle_IPI(int ipinr, struct pt_regs *regs)
 		break;
 
 	case IPI_CPU_BACKTRACE:
-		printk_nmi_enter();
+		printk_safe_nmi_enter();
 		irq_enter();
 		nmi_cpu_backtrace(regs);
 		irq_exit();
-		printk_nmi_exit();
+		printk_safe_nmi_exit();
 		break;
 
 	default:
include/linux/hardirq.h +2 -2
index c683996..14840aa 100644
@@ -61,7 +61,7 @@ extern void irq_exit(void);
 
 #define nmi_enter()						\
 	do {							\
-		printk_nmi_enter();				\
+		printk_safe_nmi_enter();			\
 		lockdep_off();					\
 		ftrace_nmi_enter();				\
 		BUG_ON(in_nmi());				\
@@ -78,7 +78,7 @@ extern void irq_exit(void);
 		preempt_count_sub(NMI_OFFSET + HARDIRQ_OFFSET);	\
 		ftrace_nmi_exit();				\
 		lockdep_on();					\
-		printk_nmi_exit();				\
+		printk_safe_nmi_exit();				\
 	} while (0)
 
 #endif /* LINUX_HARDIRQ_H */
include/linux/printk.h +10 -10
index eac1af8..c9a3080 100644
@@ -132,17 +132,17 @@ void early_printk(const char *s, ...) { }
 #endif
 
 #ifdef CONFIG_PRINTK_NMI
-extern void printk_nmi_init(void);
-extern void printk_nmi_enter(void);
-extern void printk_nmi_exit(void);
-extern void printk_nmi_flush(void);
-extern void printk_nmi_flush_on_panic(void);
+extern void printk_safe_init(void);
+extern void printk_safe_nmi_enter(void);
+extern void printk_safe_nmi_exit(void);
+extern void printk_safe_flush(void);
+extern void printk_safe_flush_on_panic(void);
 #else
-static inline void printk_nmi_init(void) { }
-static inline void printk_nmi_enter(void) { }
-static inline void printk_nmi_exit(void) { }
-static inline void printk_nmi_flush(void) { }
-static inline void printk_nmi_flush_on_panic(void) { }
+static inline void printk_safe_init(void) { }
+static inline void printk_safe_nmi_enter(void) { }
+static inline void printk_safe_nmi_exit(void) { }
+static inline void printk_safe_flush(void) { }
+static inline void printk_safe_flush_on_panic(void) { }
 #endif /* PRINTK_NMI */
 
 #ifdef CONFIG_PRINTK
init/Kconfig +9 -7
index 34407f1..77fe553 100644
@@ -875,17 +875,19 @@ config LOG_CPU_MAX_BUF_SHIFT
 		     13 =>   8 KB for each CPU
 		     12 =>   4 KB for each CPU
 
-config NMI_LOG_BUF_SHIFT
-	int "Temporary per-CPU NMI log buffer size (12 => 4KB, 13 => 8KB)"
+config PRINTK_SAFE_LOG_BUF_SHIFT
+	int "Temporary per-CPU printk log buffer size (12 => 4KB, 13 => 8KB)"
 	range 10 21
 	default 13
-	depends on PRINTK_NMI
+	depends on PRINTK
 	help
-	  Select the size of a per-CPU buffer where NMI messages are temporary
-	  stored. They are copied to the main log buffer in a safe context
-	  to avoid a deadlock. The value defines the size as a power of 2.
+	  Select the size of an alternate printk per-CPU buffer where messages
+	  printed from usafe contexts are temporary stored. One example would
+	  be NMI messages, another one - printk recursion. The messages are
+	  copied to the main log buffer in a safe context to avoid a deadlock.
+	  The value defines the size as a power of 2.
 
-	  NMI messages are rare and limited. The largest one is when
+	  Those messages are rare and limited. The largest one is when
 	  a backtrace is printed. It usually fits into 4KB. Select
 	  8KB if you want to be on the safe side.
 
init/main.c +1 -1
index 9af9274..17b0a49 100644
@@ -577,7 +577,7 @@ asmlinkage __visible void __init start_kernel(void)
 	timekeeping_init();
 	time_init();
 	sched_clock_postinit();
-	printk_nmi_init();
+	printk_safe_init();
 	perf_event_init();
 	profile_init();
 	call_function_init();
kernel/kexec_core.c +1 -1
index 5616755..297508a1 100644
@@ -914,7 +914,7 @@ void crash_kexec(struct pt_regs *regs)
 	old_cpu = atomic_cmpxchg(&panic_cpu, PANIC_CPU_INVALID, this_cpu);
 	if (old_cpu == PANIC_CPU_INVALID) {
 		/* This is the 1st CPU which comes here, so go ahead. */
-		printk_nmi_flush_on_panic();
+		printk_safe_flush_on_panic();
 		__crash_kexec(regs);
 
 		/*
kernel/panic.c +2 -2
index e6480e2..67c11fd 100644
@@ -188,7 +188,7 @@ void panic(const char *fmt, ...)
 	 * Bypass the panic_cpu check and call __crash_kexec directly.
 	 */
 	if (!_crash_kexec_post_notifiers) {
-		printk_nmi_flush_on_panic();
+		printk_safe_flush_on_panic();
 		__crash_kexec(NULL);
 
 		/*
@@ -213,7 +213,7 @@ void panic(const char *fmt, ...)
 	atomic_notifier_call_chain(&panic_notifier_list, 0, buf);
 
 	/* Call flush even twice. It tries harder with a single online CPU */
-	printk_nmi_flush_on_panic();
+	printk_safe_flush_on_panic();
 	kmsg_dump(KMSG_DUMP_PANIC);
 
 	/*
kernel/printk/Makefile +1 -1
index abb0042..6079281 100644
@@ -1,3 +1,3 @@
 obj-y	= printk.o
-obj-$(CONFIG_PRINTK_NMI)		+= nmi.o
+obj-$(CONFIG_PRINTK_NMI)		+= printk_safe.o
 obj-$(CONFIG_A11Y_BRAILLE_CONSOLE)	+= braille.o
kernel/printk/printk_safe.c renamed from kernel/printk/nmi.c +35 -34

191 lines not on this page · raw message

lib/nmi_backtrace.c +1 -1
index 7555475..5f7999e 100644
@@ -77,7 +77,7 @@ void nmi_trigger_cpumask_backtrace(const cpumask_t *mask,
 	 * Force flush any remote buffers that might be stuck in IRQ context
 	 * and therefore could not run their irq_work.
 	 */
-	printk_nmi_flush();
+	printk_safe_flush();
 
 	clear_bit_unlock(0, &backtrace_flag);
 	put_cpu();
Signature · 1 line
2.10.1.382.ga23ca1b
#7 lore raw

RFC PATCH v3 3/6 printk: introduce per-cpu safe_print seq buffer

In reply to Sergey Senozhatsky

This patch extends the idea of NMI per-cpu buffers to regions that may cause recursive printk() calls and possible deadlocks. Namely, printk() can't handle printk calls from schedule code or printk() calls from lock debugging code (spin_dump() for instance); because those may be called with sem->lock' already taken or any other critical' locks (p->pi_lock, etc.). An example of deadlock can be

vprintk_emit()
console_unlock()

   up()                        << raw_spin_lock_irqsave(&sem->lock, flags);
    wake_up_process()
     try_to_wake_up()
      ttwu_queue()
       ttwu_activate()
        activate_task()
         enqueue_task()
          enqueue_task_fair()
           cfs_rq_of()
            task_of()
             WARN_ON_ONCE(!entity_is_task(se))
              vprintk_emit()
               console_trylock()
                down_trylock()
                 raw_spin_lock_irqsave(&sem->lock, flags)
                 ^^^^ deadlock

and some other cases.

Just like in NMI implementation, the solution uses a per-cpu `printk_func' pointer to 'redirect' printk() calls to a 'safe' callback, that store messages in a per-cpu buffer and flushes them back to logbuf buffer later.

Usage example:

printk()
printk_safe_enter(flags)
//
// any printk() call from here will endup in vprintk_safe(), // that stores messages in a special per-CPU buffer. //
printk_safe_exit(flags)

The 'redirection' mechanism, though, has been reworked, as suggested by Petr Mladek. Instead of using a per-cpu @print_func callback we now keep a per-cpu printk-context variable and call either default or nmi vprintk function depending on its value. printk_nmi_entrer/exit and printk_safe_enter/exit, thus, just set/celar corresponding bits in printk-context functions.

The patch only adds printk_safe support, we don't use it yet.

---
 include/linux/printk.h      |  21 +++++--
 kernel/printk/Makefile      |   3 +-
 kernel/printk/internal.h    |  58 ++++++++++--------
 kernel/printk/printk.c      |   3 -
 kernel/printk/printk_safe.c | 140 ++++++++++++++++++++++++++++++++++----------
 5 files changed, 160 insertions(+), 65 deletions(-)
include/linux/printk.h +15 -6
index c9a3080..e8ede8f 100644
@@ -132,17 +132,11 @@ void early_printk(const char *s, ...) { }
 #endif
 
 #ifdef CONFIG_PRINTK_NMI
-extern void printk_safe_init(void);
 extern void printk_safe_nmi_enter(void);
 extern void printk_safe_nmi_exit(void);
-extern void printk_safe_flush(void);
-extern void printk_safe_flush_on_panic(void);
 #else
-static inline void printk_safe_init(void) { }
 static inline void printk_safe_nmi_enter(void) { }
 static inline void printk_safe_nmi_exit(void) { }
-static inline void printk_safe_flush(void) { }
-static inline void printk_safe_flush_on_panic(void) { }
 #endif /* PRINTK_NMI */
 
 #ifdef CONFIG_PRINTK
@@ -194,6 +188,9 @@ void __init setup_log_buf(int early);
 __printf(1, 2) void dump_stack_set_arch_desc(const char *fmt, ...);
 void dump_stack_print_info(const char *log_lvl);
 void show_regs_print_info(const char *log_lvl);
+extern void printk_safe_init(void);
+extern void printk_safe_flush(void);
+extern void printk_safe_flush_on_panic(void);
 #else
 static inline __printf(1, 0)
 int vprintk(const char *s, va_list args)
@@ -253,6 +250,18 @@ static inline void dump_stack_print_info(const char *log_lvl)
 static inline void show_regs_print_info(const char *log_lvl)
 {
 }
+
+static inline void printk_safe_init(void)
+{
+}
+
+static inline void printk_safe_flush(void)
+{
+}
+
+static inline void printk_safe_flush_on_panic(void)
+{
+}
 #endif
 
 extern asmlinkage void dump_stack(void) __cold;
kernel/printk/Makefile +1 -2
index 6079281..87ad14d 100644
@@ -1,3 +1,2 @@
-obj-y	= printk.o
-obj-$(CONFIG_PRINTK_NMI)		+= printk_safe.o
+obj-y	= printk_safe.o printk.o
 obj-$(CONFIG_A11Y_BRAILLE_CONSOLE)	+= braille.o
kernel/printk/internal.h +35 -23

75 lines not on this page · raw message

kernel/printk/printk.c +0 -3
index be7f60b..58cfc5c 100644
@@ -2024,9 +2024,6 @@ static size_t msg_print_text(const struct printk_log *msg, enum log_flags prev,
 static size_t cont_print_text(char *text, size_t size) { return 0; }
 static bool suppress_message_printing(int level) { return false; }
 
-/* Still needs to be defined for users */
-DEFINE_PER_CPU(printk_func_t, printk_func);
-
 #endif /* CONFIG_PRINTK */
 
 #ifdef CONFIG_EARLY_PRINTK
kernel/printk/printk_safe.c +109 -31

205 lines not on this page · raw message

Signature · 1 line
2.10.1.382.ga23ca1b
#8 lore raw

Re: [RFC][PATCHv3 3/6] printk: introduce per-cpu safe_print seq buffer

kernel/printk/printk_safe.c void printk_safe_flush_on_panic(void) · line 298

Quoted 4 lines · On Wed, 19 Oct 2016 00:40:42 +0900 Sergey Senozhatsky…
kernel/printk/printk_safe.c +4 -0
+
+/*
+ * Returns with local IRQs disabled.
+ * Can be preempted by NMI.

It does? How does it return with IRQs disabled?

kernel/printk/printk_safe.c void printk_safe_flush_on_panic(void) · line 308

Quoted 10 lines

+ */

kernel/printk/printk_safe.c +9 -0
+void __printk_safe_enter(void)
+{
+	this_cpu_inc(printk_safe_context);
+}
+
+/*
+ * Restores local IRQs state saved in printk_safe_enter().
+ * Can be preempted by NMI.
+ */

How does this return with irqs enabled?

-- Steve

Quoted 5 lines
kernel/printk/printk_safe.c +5 -0
+void __printk_safe_exit(void)
+{
+	this_cpu_dec(printk_safe_context);
+}
+
#9 lore raw

Re: [RFC][PATCHv3 3/6] printk: introduce per-cpu safe_print seq buffer

Quoted 7 lines · Steven Rostedt

kernel/printk/printk_safe.c void printk_safe_flush_on_panic(void) · line 298

Quoted 4 lines
kernel/printk/printk_safe.c +4 -0
+
+/*
+ * Returns with local IRQs disabled.
+ * Can be preempted by NMI.

It does? How does it return with IRQs disabled?

Quoted 12 lines

kernel/printk/printk_safe.c void printk_safe_flush_on_panic(void) · line 308

Quoted 10 lines

+ */

kernel/printk/printk_safe.c +9 -0
+void __printk_safe_enter(void)
+{
+	this_cpu_inc(printk_safe_context);
+}
+
+/*
+ * Restores local IRQs state saved in printk_safe_enter().
+ * Can be preempted by NMI.
+ */

How does this return with irqs enabled?

ah... leftovers. thanks, will correct.

	-ss
#10 lore raw

Re: [RFC][PATCHv3 3/6] printk: introduce per-cpu safe_print seq buffer

In reply to Sergey Senozhatsky

kernel/printk/Makefile line 1

Quoted 3 lines · Sergey Senozhatsky
kernel/printk/Makefile +1 -2
-obj-y	= printk.o
-obj-$(CONFIG_PRINTK_NMI)		+= printk_safe.o
+obj-y	= printk_safe.o printk.o

[..]

d'oh... this breaks !PRINTK configs. will fix in v4.

	-ss
#11 lore raw

RFC PATCH v3 4/6 printk: report lost messages in printk safe/nmi contexts

In reply to Sergey Senozhatsky

Account lost messages in pritk-safe and printk-safe-nmi contexts and report those numbers during printk_safe_flush().

---
 kernel/printk/internal.h    | 17 ---------------
 kernel/printk/printk.c      | 10 ---------
 kernel/printk/printk_safe.c | 50 +++++++++++++++++++++++++++++++++++++++++++--
 3 files changed, 48 insertions(+), 29 deletions(-)
kernel/printk/internal.h +0 -17
index 015f68f..375fe6d 100644
@@ -16,23 +16,6 @@
  */
 #include <linux/percpu.h>
 
-#ifdef CONFIG_PRINTK_NMI
-
-extern atomic_t nmi_message_lost;
-static inline int get_nmi_message_lost(void)
-{
-	return atomic_xchg(&nmi_message_lost, 0);
-}
-
-#else /* CONFIG_PRINTK_NMI */
-
-static inline int get_nmi_message_lost(void)
-{
-	return 0;
-}
-
-#endif /* CONFIG_PRINTK_NMI */
-
 #ifdef CONFIG_PRINTK
 
 #define PRINTK_SAFE_CONTEXT_MASK	0x7fffffff
kernel/printk/printk.c +0 -10
index 58cfc5c..a99414b 100644
@@ -1791,7 +1791,6 @@ asmlinkage int vprintk_emit(int facility, int level,
 	unsigned long flags;
 	int this_cpu;
 	int printed_len = 0;
-	int nmi_message_lost;
 	bool in_sched = false;
 	/* cpu currently holding logbuf_lock in this function */
 	static unsigned int logbuf_cpu = UINT_MAX;
@@ -1842,15 +1841,6 @@ asmlinkage int vprintk_emit(int facility, int level,
 					 strlen(recursion_msg));
 	}
 
-	nmi_message_lost = get_nmi_message_lost();
-	if (unlikely(nmi_message_lost)) {
-		text_len = scnprintf(textbuf, sizeof(textbuf),
-				     "BAD LUCK: lost %d message(s) from NMI context!",
-				     nmi_message_lost);
-		printed_len += log_store(0, 2, LOG_PREFIX|LOG_NEWLINE, 0,
-					 NULL, 0, textbuf, text_len);
-	}
-
 	/*
 	 * The printf needs to come first; we need the syslog
 	 * prefix which might be passed-in as a parameter.
kernel/printk/printk_safe.c +48 -2

78 lines not on this page · raw message

Signature · 1 line
2.10.1.382.ga23ca1b
#12 lore raw

RFC PATCH v3 5/6 printk: use printk_safe buffers

In reply to Sergey Senozhatsky

Use printk_safe per-CPU buffers in in printk recursion-prone blocks:
-- around logbuf_lock protected sections in vprintk_emit() and
console_unlock()
-- around down_trylock_console_sem() and up_console_sem()

Note that this solution addresses deadlocks caused by printk() recursive calls only.

Examples:

1) printk() from logbuf_lock spin_lock section

Assume the following code:
printk()

    raw_spin_lock(&logbuf_lock);
    WARN_ON(1);
    raw_spin_unlock(&logbuf_lock);

which now produces:

 ------------[ cut here ]------------
 WARNING: CPU: 0 PID: 366 at kernel/printk/printk.c:1811 vprintk_emit+0x1cd/0x438
 CPU: 0 PID: 366 Comm: bash
 Call Trace:
  [<ffffffff811e24d9>] dump_stack+0x68/0x92
  [<ffffffff810455cd>] __warn+0xc2/0xdd
  [<ffffffff81045658>] warn_slowpath_null+0x1d/0x1f
  [<ffffffff81091527>] vprintk_emit+0x1cd/0x438
  [<ffffffff810918fe>] vprintk_default+0x1d/0x1f
  [<ffffffff810fdf6a>] printk+0x48/0x50
  [..]
  [<ffffffff813caaaa>] entry_SYSCALL_64_fastpath+0x18/0xad
 ---[ end trace ]---

2) printk() from semaphore sem->lock spin_lock section

Assume the following code

  printk()
    console_trylock()
      down_trylock()
        raw_spin_lock_irqsave(&sem->lock, flags);
        WARN_ON(1);
        raw_spin_unlock_irqrestore(&sem->lock, flags);

which now produces:

 ------------[ cut here ]------------
 WARNING: CPU: 1 PID: 363 at kernel/locking/semaphore.c:141 down_trylock+0x3d/0x62
 CPU: 1 PID: 363 Comm: bash
 Call Trace:
  [<ffffffff811e24e0>] dump_stack+0x68/0x92
  [<ffffffff810455cd>] __warn+0xc2/0xdd
  [<ffffffff81045658>] warn_slowpath_null+0x1d/0x1f
  [<ffffffff810838df>] down_trylock+0x3d/0x62
  [<ffffffff8109177e>] ? vprintk_emit+0x3f9/0x414
  [<ffffffff810905cb>] console_trylock+0x31/0xeb
  [<ffffffff8109177e>] vprintk_emit+0x3f9/0x414
  [<ffffffff81091905>] vprintk_default+0x1d/0x1f
  [<ffffffff810fdf71>] printk+0x48/0x50
  [..]
  [<ffffffff813caaaa>] entry_SYSCALL_64_fastpath+0x18/0xad
 ---[ end trace ]---

3) printk() from console_unlock()

Assume the following code:

  printk()
    console_unlock()
      raw_spin_lock(&logbuf_lock);
      WARN_ON(1);
      raw_spin_unlock(&logbuf_lock);

which now produces:

 ------------[ cut here ]------------
 WARNING: CPU: 1 PID: 329 at kernel/printk/printk.c:2384 console_unlock+0x12d/0x559
 CPU: 1 PID: 329 Comm: bash
 Call Trace:
  [<ffffffff811f2380>] dump_stack+0x68/0x92
  [<ffffffff8103d649>] __warn+0xb8/0xd3
  [<ffffffff8103d6ca>] warn_slowpath_null+0x18/0x1a
  [<ffffffff8107f917>] console_unlock+0x12d/0x559
  [<ffffffff810772c4>] ? trace_hardirqs_on_caller+0x16d/0x189
  [<ffffffff810772ed>] ? trace_hardirqs_on+0xd/0xf
  [<ffffffff810800a6>] vprintk_emit+0x363/0x374
  [<ffffffff81080219>] vprintk_default+0x18/0x1a
  [<ffffffff810c7f77>] printk+0x43/0x4b
  [..]
  [<ffffffff814c2d6a>] entry_SYSCALL_64_fastpath+0x18/0xad
 ---[ end trace ]---

4) printk() from try_to_wake_up()

Assume the following code:

  printk()
    console_unlock()
      up()
        try_to_wake_up()
          raw_spin_lock_irqsave(&p->pi_lock, flags);
          WARN_ON(1);
          raw_spin_unlock_irqrestore(&p->pi_lock, flags);

which now produces:

 ------------[ cut here ]------------
 WARNING: CPU: 3 PID: 363 at kernel/sched/core.c:2028 try_to_wake_up+0x7f/0x4f7
 CPU: 3 PID: 363 Comm: bash
 Call Trace:
  [<ffffffff811e2557>] dump_stack+0x68/0x92
  [<ffffffff810455cd>] __warn+0xc2/0xdd
  [<ffffffff81045658>] warn_slowpath_null+0x1d/0x1f
  [<ffffffff8106c8d0>] try_to_wake_up+0x7f/0x4f7
  [<ffffffff8106cd5d>] wake_up_process+0x15/0x17
  [<ffffffff813c82c6>] __up.isra.0+0x56/0x63
  [<ffffffff810839a3>] up+0x32/0x42
  [<ffffffff8108f624>] __up_console_sem+0x37/0x55
  [<ffffffff810910ff>] console_unlock+0x21e/0x4c2
  [<ffffffff810917bf>] vprintk_emit+0x41c/0x462
  [<ffffffff81091971>] vprintk_default+0x1d/0x1f
  [<ffffffff810fdfdd>] printk+0x48/0x50
  [..]
  [<ffffffff813cab2a>] entry_SYSCALL_64_fastpath+0x18/0xad
 ---[ end trace ]---
---
 kernel/printk/printk.c | 42 +++++++++++++++++++++++++++---------------
 1 file changed, 27 insertions(+), 15 deletions(-)
kernel/printk/printk.c +27 -15
index a99414b..688c1a4 100644
@@ -213,17 +213,31 @@ static int nr_ext_console_drivers;
 
 static int __down_trylock_console_sem(unsigned long ip)
 {
-	if (down_trylock(&console_sem))
+	int lock_failed;
+	unsigned long flags;
+
+	printk_safe_enter(flags);
+	lock_failed = down_trylock(&console_sem);
+	printk_safe_exit(flags);
+
+	if (lock_failed)
 		return 1;
 	mutex_acquire(&console_lock_dep_map, 0, 1, ip);
 	return 0;
 }
 #define down_trylock_console_sem() __down_trylock_console_sem(_RET_IP_)
 
-#define up_console_sem() do { \
-	mutex_release(&console_lock_dep_map, 1, _RET_IP_);\
-	up(&console_sem);\
-} while (0)
+static void __up_console_sem(unsigned long ip)
+{
+	unsigned long flags;
+
+	mutex_release(&console_lock_dep_map, 1, ip);
+
+	printk_safe_enter(flags);
+	up(&console_sem);
+	printk_safe_exit(flags);
+}
+#define up_console_sem() __up_console_sem(_RET_IP_)
 
 /*
  * This is used for debugging the mess that is the VT code by
@@ -1803,7 +1817,7 @@ asmlinkage int vprintk_emit(int facility, int level,
 	boot_delay_msec(level);
 	printk_delay();
 
-	local_irq_save(flags);
+	printk_safe_enter(flags);
 	this_cpu = smp_processor_id();
 
 	/*
@@ -1819,13 +1833,12 @@ asmlinkage int vprintk_emit(int facility, int level,
 		 */
 		if (!oops_in_progress && !lockdep_recursing(current)) {
 			recursion_bug = true;
-			local_irq_restore(flags);
+			printk_safe_exit(flags);
 			return 0;
 		}
 		zap_locks();
 	}
 
-	lockdep_off();
 	/* This stops the holder of console_sem just where we want him */
 	raw_spin_lock(&logbuf_lock);
 	logbuf_cpu = this_cpu;
@@ -1885,12 +1898,10 @@ asmlinkage int vprintk_emit(int facility, int level,
 
 	logbuf_cpu = UINT_MAX;
 	raw_spin_unlock(&logbuf_lock);
-	lockdep_on();
-	local_irq_restore(flags);
+	printk_safe_exit(flags);
 
 	/* If called from the scheduler, we can not call up(). */
 	if (!in_sched) {
-		lockdep_off();
 		/*
 		 * Try to acquire and then immediately release the console
 		 * semaphore.  The release will print out buffers and wake up
@@ -1898,7 +1909,6 @@ asmlinkage int vprintk_emit(int facility, int level,
 		 */
 		if (console_trylock())
 			console_unlock();
-		lockdep_on();
 	}
 
 	return printed_len;
@@ -2377,7 +2387,8 @@ void console_unlock(void)
 		size_t len;
 		int level;
 
-		raw_spin_lock_irqsave(&logbuf_lock, flags);
+		printk_safe_enter(flags);
+		raw_spin_lock(&logbuf_lock);
 		if (seen_seq != log_next_seq) {
 			wake_klogd = true;
 			seen_seq = log_next_seq;
@@ -2438,7 +2449,7 @@ void console_unlock(void)
 		stop_critical_timings();	/* don't trace print latency */
 		call_console_drivers(level, ext_text, ext_len, text, len);
 		start_critical_timings();
-		local_irq_restore(flags);
+		printk_safe_exit(flags);
 
 		if (do_cond_resched)
 			cond_resched();
@@ -2461,7 +2472,8 @@ void console_unlock(void)
 	 */
 	raw_spin_lock(&logbuf_lock);
 	retry = console_seq != log_next_seq;
-	raw_spin_unlock_irqrestore(&logbuf_lock, flags);
+	raw_spin_unlock(&logbuf_lock);
+	printk_safe_exit(flags);
 
 	if (retry && console_trylock())
 		goto again;
Signature · 1 line
2.10.1.382.ga23ca1b
#13 lore raw

RFC PATCH v3 6/6 printk: remove zap_locks() function

In reply to Sergey Senozhatsky

We use printk-safe now which makes printk-recursion detection code in vprintk_emit() is unreachable. The tricky thing here is that, apart from detecting and reporting printk recursions, that code also used to zap_lockc() in case of panic. However, zap_locks() does not look to be needed anymore:

1) Since commit 08d78658f393 ("panic: release stale console lock to always get the logbuf printed out") panic flushing of logbuf' to console ignores the state of console_sem' by doing panic()

		console_trylock();
		console_unlock();

2) Since commit cf9b1106c81c ("printk/nmi: flush NMI messages on the system panic") panic attempts to zap the logbuf_lock' spin_lock to successfully flush nmi messages to logbuf'.

Basically, it seems that we either already do what zap_locks() used to do but in other places or we ignore the state of the lock. May be we still would want to do sema_init() in printk_safe_flush_on_panic(), just in case.

---
 kernel/printk/printk.c | 61 --------------------------------------------------
 1 file changed, 61 deletions(-)
kernel/printk/printk.c +0 -61
index 688c1a4..79df499 100644
@@ -1619,28 +1619,6 @@ static void call_console_drivers(int level,
 	}
 }
 
-/*
- * Zap console related locks when oopsing.
- * To leave time for slow consoles to print a full oops,
- * only zap at most once every 30 seconds.
- */
-static void zap_locks(void)
-{
-	static unsigned long oops_timestamp;
-
-	if (time_after_eq(jiffies, oops_timestamp) &&
-	    !time_after(jiffies, oops_timestamp + 30 * HZ))
-		return;
-
-	oops_timestamp = jiffies;
-
-	debug_locks_off();
-	/* If a crash is occurring, make sure we can't deadlock */
-	raw_spin_lock_init(&logbuf_lock);
-	/* And make sure that we print immediately */
-	sema_init(&console_sem, 1);
-}
-
 int printk_delay_msec __read_mostly;
 
 static inline void printk_delay(void)
@@ -1797,17 +1775,13 @@ asmlinkage int vprintk_emit(int facility, int level,
 			    const char *dict, size_t dictlen,
 			    const char *fmt, va_list args)
 {
-	static bool recursion_bug;
 	static char textbuf[LOG_LINE_MAX];
 	char *text = textbuf;
 	size_t text_len = 0;
 	enum log_flags lflags = 0;
 	unsigned long flags;
-	int this_cpu;
 	int printed_len = 0;
 	bool in_sched = false;
-	/* cpu currently holding logbuf_lock in this function */
-	static unsigned int logbuf_cpu = UINT_MAX;
 
 	if (level == LOGLEVEL_SCHED) {
 		level = LOGLEVEL_DEFAULT;
@@ -1818,42 +1792,8 @@ asmlinkage int vprintk_emit(int facility, int level,
 	printk_delay();
 
 	printk_safe_enter(flags);
-	this_cpu = smp_processor_id();
-
-	/*
-	 * Ouch, printk recursed into itself!
-	 */
-	if (unlikely(logbuf_cpu == this_cpu)) {
-		/*
-		 * If a crash is occurring during printk() on this CPU,
-		 * then try to get the crash message out but make sure
-		 * we can't deadlock. Otherwise just return to avoid the
-		 * recursion and return - but flag the recursion so that
-		 * it can be printed at the next appropriate moment:
-		 */
-		if (!oops_in_progress && !lockdep_recursing(current)) {
-			recursion_bug = true;
-			printk_safe_exit(flags);
-			return 0;
-		}
-		zap_locks();
-	}
-
 	/* This stops the holder of console_sem just where we want him */
 	raw_spin_lock(&logbuf_lock);
-	logbuf_cpu = this_cpu;
-
-	if (unlikely(recursion_bug)) {
-		static const char recursion_msg[] =
-			"BUG: recent printk recursion!";
-
-		recursion_bug = false;
-		/* emit KERN_CRIT message */
-		printed_len += log_store(0, 2, LOG_PREFIX|LOG_NEWLINE, 0,
-					 NULL, 0, recursion_msg,
-					 strlen(recursion_msg));
-	}
-
 	/*
 	 * The printf needs to come first; we need the syslog
 	 * prefix which might be passed-in as a parameter.
@@ -1896,7 +1836,6 @@ asmlinkage int vprintk_emit(int facility, int level,
 
 	printed_len += log_output(facility, level, lflags, dict, dictlen, text, text_len);
 
-	logbuf_cpu = UINT_MAX;
 	raw_spin_unlock(&logbuf_lock);
 	printk_safe_exit(flags);
 
Signature · 1 line
2.10.1.382.ga23ca1b
#15 lore raw
Quoted 7 lines · Joe Perches
Quoted 2 lines · Sergey Senozhatsky

This patch set extends a lock-less NMI per-cpu buffers idea to handle recursive printk() calls.

trivia:

recursive or reentrant?

a recursive one.

printk -> {foo} -> printk

reentrant printk() case was addressed by Petr Mladek's nmi patchset:

printk()

	foo
	-------> NMI
		printk()

	-ss
#16 lore raw
In reply to Sergey Senozhatsky
Quoted 10 lines · Sergey Senozhatsky

Hello,

RFC

This patch set extends a lock-less NMI per-cpu buffers idea to
handle recursive printk() calls. The basic mechanism is pretty much the same -- at the beginning of a deadlock-prone section we switch to lock-less printk callback, and return back to a default printk implementation at the end; the messages are getting flushed to a logbuf buffer from a safer context.

So I think you're not taking this far enough. You've also missed an entire class of deadlocks.

The first is that you still keep the logbuf. Having this global serialized thing is a source of fail. It would be much better to only keep per cpu stuff. _OR_ at the very least make the logbuf itself lockfree. So generate the printk entry local (so we know its size) then atomically reserve the logbuf entry and copy it over.

The entire class of deadlocks you've missed is that console->write() is a piece of crap too ;-) Even the bog standard 8250 serial console driver can do wakeups.

See for example:

https://marc.info/?l=linux-kernel&m=132446649823347&w=2

I've entirely given up on printk(), I'll post the 3 patches I carry to make things work for me.

#17 lore raw
Quoted 9 lines · Peter Zijlstra
Quoted 6 lines

This patch set extends a lock-less NMI per-cpu buffers idea to
handle recursive printk() calls. The basic mechanism is pretty much the same -- at the beginning of a deadlock-prone section we switch to lock-less printk callback, and return back to a default printk implementation at the end; the messages are getting flushed to a logbuf buffer from a safer context.

So I think you're not taking this far enough. You've also missed an entire class of deadlocks.

yes. this patch set addresses only printk() recursion. printk deadlocks in general are much more complicated and basically we don't even know how many locks we are about to take every time we call printk() (there may be serial console locks, etc we never know in advance).

Quoted 5 lines

The first is that you still keep the logbuf. Having this global serialized thing is a source of fail. It would be much better to only keep per cpu stuff. _OR_ at the very least make the logbuf itself lockfree. So generate the printk entry local (so we know its size) then atomically reserve the logbuf entry and copy it over.

logbuf spin_lock is not the only problematic lock. we _at least_ have console sem spin_lock and rq spin_lock:

	https://marc.info/?l=linux-kernel&m=145406977822244

	https://marc.info/?l=linux-kernel&m=146853438804534&w=2
Quoted 3 lines

The entire class of deadlocks you've missed is that console->write() is a piece of crap too ;-) Even the bog standard 8250 serial console driver can do wakeups.

yes. well, I didn't exactly "miss" it. I just don't have any working solution for those locks yet. so things like

SyS_ioctl
do_vfs_ioctl
tty_ioctl

    n_tty_ioctl
     tty_mode_ioctl
      set_termios
       tty_set_termios
        uart_set_termios
         uart_change_speed
          FOO_serial_set_termios
           spin_lock_irqsave(&port->lock)     // lock the output port
           ....
           !! WARN() or pr_err() or printk()
               vprintk_emit()
                /* console_trylock() */
                console_unlock()
                 call_console_drivers()
                  FOO_write()
                   spin_lock_irqsave(&port->lock)     // already locked

are out of this patch set's scope. another class of deadlocks is when we call printk under one of the locks that printk() can take later in order to print the message.

so in this particular patch set I make "printk() -> $FOO -> printk()" less damaging. which is just a first step.

we are also looking at things like might_printk()
https://marc.info/?l=linux-kernel&m=147607261116163

and even extending the lockdep
https://marc.info/?l=linux-kernel&m=147617830511381

not to miss out a DEFERRED_WARN patch set...
//hm, I can't find it online

Subject: [RFC 0/5] printk: Implement WARN_*DEFERRED()
Message-Id: <[email protected]>

Quoted 3 lines

See for example:

https://marc.info/?l=linux-kernel&m=132446649823347&w=2

interesting idea.

Quoted 2 lines

I've entirely given up on printk(), I'll post the 3 patches I carry to make things work for me.

yes, printk is surprisingly fragile and, some would _probably_ even say, broken.

a) we have numerous scenarios when it can deadlock
b) we have numerous scenarios when it can soft-lockup/hard-lockup the system
c) we don't have a bullet proof panic printk. zap_locks() resets logbuf and console_sem, but there may be serial console locks/etc.

	-ss
#18 lore raw
Quoted 5 lines · Sergey Senozhatsky

not to miss out a DEFERRED_WARN patch set...
//hm, I can't find it online

Subject: [RFC 0/5] printk: Implement WARN_*DEFERRED()
Message-Id: <[email protected]>

That never hit lkml, that msgid does not resolve on marc and I can only find a few replies in my local lkml archive.

Can't say I like the idea though. Its propagating the printk() failure up into more APIs instead of actually fixing it.

Esp. for WARN and the like you do not want DEFERRED, because DEFERRED means you'll likely never actually get to see the problem because we're dead by the time 'later' happens.

maintainer
#19 lore raw
In reply to Peter Zijlstra
Quoted 12 lines · Peter Zijlstra
Quoted 10 lines · Sergey Senozhatsky

Hello,

RFC

This patch set extends a lock-less NMI per-cpu buffers idea to
handle recursive printk() calls. The basic mechanism is pretty much the same -- at the beginning of a deadlock-prone section we switch to lock-less printk callback, and return back to a default printk implementation at the end; the messages are getting flushed to a logbuf buffer from a safer context.

Quoted 8 lines

So I think you're not taking this far enough. You've also missed an entire class of deadlocks.

The first is that you still keep the logbuf. Having this global serialized thing is a source of fail. It would be much better to only keep per cpu stuff. _OR_ at the very least make the logbuf itself lockfree. So generate the printk entry local (so we know its size) then atomically reserve the logbuf entry and copy it over.

This is close to what the lockless ring_buffer does. I mean kernel/trace/ring_buffer.c that is used by trace_printk. It has per-CPU buffers, does reservations, ...

Sadly the ring_buffer code is very tricky. Only a single reader is allowed at a time. And locating the information in the crash dump is a task for prisoners.

Note that it still does _not_ solve problems with the console output. I am not sure if going this way would be a real win.

Sigh, a genius idea would help.

Quoted 3 lines

The entire class of deadlocks you've missed is that console->write() is a piece of crap too ;-) Even the bog standard 8250 serial console driver can do wakeups.

I wonder if all the hard problems are actually related to the console handling.

There are problems with the single logbuffer but these should get eliminated by the NMI/safe temporary per-CPU buffers.

All might be easier if we always offload the console handling into a kthread or so and trigger it via the minimalist irq_work. It would kill huge bunch of possible deadlocks. It will even allow to get rid of printk_deferred() and the uncertainty where it is needed.

The penalty would be "slightly" delayed console output. But is it a real problem? It should not be a big deal when everything works. We could always try hard when panicking. And there always might be a fallback with that direct early_console().

Best Regards,
Petr

#20 lore raw
Quoted 18 lines · Petr Mladek
Quoted 3 lines · Peter Zijlstra

The entire class of deadlocks you've missed is that console->write() is a piece of crap too ;-) Even the bog standard 8250 serial console driver can do wakeups.

I wonder if all the hard problems are actually related to the console handling.

There are problems with the single logbuffer but these should get eliminated by the NMI/safe temporary per-CPU buffers.

All might be easier if we always offload the console handling into a kthread or so and trigger it via the minimalist irq_work. It would kill huge bunch of possible deadlocks. It will even allow to get rid of printk_deferred() and the uncertainty where it is needed.

Quoted 4 lines

The penalty would be "slightly" delayed console output. But is it a real problem? It should not be a big deal when everything works. We could always try hard when panicking. And there always might be a fallback with that direct early_console().

Right, so I've had to debug quite a few situations where that 'later' simply never happens...

I agree that such a scenario is basically all the regular console drivers are capable of though.

It might make sense to go there, but allow early_console to print on the go, keeping synchronous output available. We would still need the logbuf to become NMI safe wrt adding entries though.

It would however place hard requirements of early_console implementations; already violated by for instance the USB3-debug-port implementation I just looked at:

lkml.kernel.org/r/[email protected]

#21 lore raw
Quoted 2 lines · Peter Zijlstra
Quoted 1 line · Petr Mladek

On Tue 2016-10-18 19:07:54, Peter Zijlstra wrote:

[..]

Quoted 4 lines

It might make sense to go there, but allow early_console to print on the go, keeping synchronous output available. We would still need the logbuf to become NMI safe wrt adding entries though.

as a side (spam) note,

I sent a [TECH TOPIC] suggestion to ksummit-discuss a while ago
https://lists.linuxfoundation.org/pipermail/ksummit-discuss/2016-July/002740.html

and a later update
https://lists.linuxfoundation.org/pipermail/ksummit-discuss/2016-September/004006.html

which covers some of the printk problems. so I'm sure I can tell whether the suggestion has been approved or rejected, but hopefully we can find some time to sit down and talk about the issues.

	-ss
#22 lore raw
In reply to Sergey Senozhatsky

On (10/19/16 00:40), Sergey Senozhatsky wrote:
[..]

Quoted 23 lines

Deadlock scenarios that printk_safe can handle:

a) printk recursion from logbuf_lock spin_lock section in printk() printk()

    raw_spin_lock(&logbuf_lock);
    WARN_ON(1);
    raw_spin_unlock(&logbuf_lock);

b) printk from sem->lock spin_lock section
printk()

    console_trylock()
      down_trylock()
        raw_spin_lock_irqsave(&sem->lock, flags);
        WARN_ON(1);
        raw_spin_unlock_irqrestore(&sem->lock, flags);

c) printk from logbuf_lock spin_lock section in console_unlock() printk()

    console_unlock()
      raw_spin_lock(&logbuf_lock);
      WARN_ON(1);
      raw_spin_unlock(&logbuf_lock);
Quoted 8 lines

d) printk from ->pi_lock from semaphore up
printk()

    console_unlock()
      up()
        try_to_wake_up()
          raw_spin_lock_irqsave(&p->pi_lock, flags);
          WARN_ON(1);
          raw_spin_unlock_irqrestore(&p->pi_lock, flags);

and

e) anything from call_console_drivers() should be fine as well. console_unlock() big printing loop is covered by printk_safe, need to protect console_cont_flush()->call_console_drivers().

so printk "direct path" seems to be safe now:

printk -> .. -> console_unlock() -> call_console_drivers() -> serial_console_write()

anything that is "indirect" is not covered in this patch set.

	-ss