linux-brain/kernel/debug/debug_core.c

1156 lines
27 KiB
C
Raw Permalink Normal View History

/*
* Kernel Debug Core
*
* Maintainer: Jason Wessel <jason.wessel@windriver.com>
*
* Copyright (C) 2000-2001 VERITAS Software Corporation.
* Copyright (C) 2002-2004 Timesys Corporation
* Copyright (C) 2003-2004 Amit S. Kale <amitkale@linsyssoft.com>
* Copyright (C) 2004 Pavel Machek <pavel@ucw.cz>
* Copyright (C) 2004-2006 Tom Rini <trini@kernel.crashing.org>
* Copyright (C) 2004-2006 LinSysSoft Technologies Pvt. Ltd.
* Copyright (C) 2005-2009 Wind River Systems, Inc.
* Copyright (C) 2007 MontaVista Software, Inc.
* Copyright (C) 2008 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
*
* Contributors at various stages not listed above:
* Jason Wessel ( jason.wessel@windriver.com )
* George Anzinger <george@mvista.com>
* Anurekh Saxena (anurekh.saxena@timesys.com)
* Lake Stevens Instrument Division (Glenn Engel)
* Jim Kingdon, Cygnus Support.
*
* Original KGDB stub: David Grothe <dave@gcom.com>,
* Tigran Aivazian <tigran@sco.com>
*
* This file is licensed under the terms of the GNU General Public License
* version 2. This program is licensed "as is" without any warranty of any
* kind, whether express or implied.
*/
#define pr_fmt(fmt) "KGDB: " fmt
#include <linux/pid_namespace.h>
#include <linux/clocksource.h>
#include <linux/serial_core.h>
#include <linux/interrupt.h>
#include <linux/spinlock.h>
#include <linux/console.h>
#include <linux/threads.h>
#include <linux/uaccess.h>
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/ptrace.h>
#include <linux/string.h>
#include <linux/delay.h>
#include <linux/sched.h>
#include <linux/sysrq.h>
#include <linux/reboot.h>
#include <linux/init.h>
#include <linux/kgdb.h>
#include <linux/kdb.h>
#include <linux/nmi.h>
#include <linux/pid.h>
#include <linux/smp.h>
#include <linux/mm.h>
mm: per-thread vma caching This patch is a continuation of efforts trying to optimize find_vma(), avoiding potentially expensive rbtree walks to locate a vma upon faults. The original approach (https://lkml.org/lkml/2013/11/1/410), where the largest vma was also cached, ended up being too specific and random, thus further comparison with other approaches were needed. There are two things to consider when dealing with this, the cache hit rate and the latency of find_vma(). Improving the hit-rate does not necessarily translate in finding the vma any faster, as the overhead of any fancy caching schemes can be too high to consider. We currently cache the last used vma for the whole address space, which provides a nice optimization, reducing the total cycles in find_vma() by up to 250%, for workloads with good locality. On the other hand, this simple scheme is pretty much useless for workloads with poor locality. Analyzing ebizzy runs shows that, no matter how many threads are running, the mmap_cache hit rate is less than 2%, and in many situations below 1%. The proposed approach is to replace this scheme with a small per-thread cache, maximizing hit rates at a very low maintenance cost. Invalidations are performed by simply bumping up a 32-bit sequence number. The only expensive operation is in the rare case of a seq number overflow, where all caches that share the same address space are flushed. Upon a miss, the proposed replacement policy is based on the page number that contains the virtual address in question. Concretely, the following results are seen on an 80 core, 8 socket x86-64 box: 1) System bootup: Most programs are single threaded, so the per-thread scheme does improve ~50% hit rate by just adding a few more slots to the cache. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 50.61% | 19.90 | | patched | 73.45% | 13.58 | +----------------+----------+------------------+ 2) Kernel build: This one is already pretty good with the current approach as we're dealing with good locality. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 75.28% | 11.03 | | patched | 88.09% | 9.31 | +----------------+----------+------------------+ 3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 70.66% | 17.14 | | patched | 91.15% | 12.57 | +----------------+----------+------------------+ 4) Ebizzy: There's a fair amount of variation from run to run, but this approach always shows nearly perfect hit rates, while baseline is just about non-existent. The amounts of cycles can fluctuate between anywhere from ~60 to ~116 for the baseline scheme, but this approach reduces it considerably. For instance, with 80 threads: +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 1.06% | 91.54 | | patched | 99.97% | 14.18 | +----------------+----------+------------------+ [akpm@linux-foundation.org: fix nommu build, per Davidlohr] [akpm@linux-foundation.org: document vmacache_valid() logic] [akpm@linux-foundation.org: attempt to untangle header files] [akpm@linux-foundation.org: add vmacache_find() BUG_ON] [hughd@google.com: add vmacache_valid_mm() (from Oleg)] [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: adjust and enhance comments] Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Reviewed-by: Michel Lespinasse <walken@google.com> Cc: Oleg Nesterov <oleg@redhat.com> Tested-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 07:37:25 +09:00
#include <linux/vmacache.h>
#include <linux/rcupdate.h>
kgdb: Fix kgdb_roundup_cpus() for arches who used smp_call_function() When I had lockdep turned on and dropped into kgdb I got a nice splat on my system. Specifically it hit: DEBUG_LOCKS_WARN_ON(current->hardirq_context) Specifically it looked like this: sysrq: SysRq : DEBUG ------------[ cut here ]------------ DEBUG_LOCKS_WARN_ON(current->hardirq_context) WARNING: CPU: 0 PID: 0 at .../kernel/locking/lockdep.c:2875 lockdep_hardirqs_on+0xf0/0x160 CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.19.0 #27 pstate: 604003c9 (nZCv DAIF +PAN -UAO) pc : lockdep_hardirqs_on+0xf0/0x160 ... Call trace: lockdep_hardirqs_on+0xf0/0x160 trace_hardirqs_on+0x188/0x1ac kgdb_roundup_cpus+0x14/0x3c kgdb_cpu_enter+0x53c/0x5cc kgdb_handle_exception+0x180/0x1d4 kgdb_compiled_brk_fn+0x30/0x3c brk_handler+0x134/0x178 do_debug_exception+0xfc/0x178 el1_dbg+0x18/0x78 kgdb_breakpoint+0x34/0x58 sysrq_handle_dbg+0x54/0x5c __handle_sysrq+0x114/0x21c handle_sysrq+0x30/0x3c qcom_geni_serial_isr+0x2dc/0x30c ... ... irq event stamp: ...45 hardirqs last enabled at (...44): [...] __do_softirq+0xd8/0x4e4 hardirqs last disabled at (...45): [...] el1_irq+0x74/0x130 softirqs last enabled at (...42): [...] _local_bh_enable+0x2c/0x34 softirqs last disabled at (...43): [...] irq_exit+0xa8/0x100 ---[ end trace adf21f830c46e638 ]--- Looking closely at it, it seems like a really bad idea to be calling local_irq_enable() in kgdb_roundup_cpus(). If nothing else that seems like it could violate spinlock semantics and cause a deadlock. Instead, let's use a private csd alongside smp_call_function_single_async() to round up the other CPUs. Using smp_call_function_single_async() doesn't require interrupts to be enabled so we can remove the offending bit of code. In order to avoid duplicating this across all the architectures that use the default kgdb_roundup_cpus(), we'll add a "weak" implementation to debug_core.c. Looking at all the people who previously had copies of this code, there were a few variants. I've attempted to keep the variants working like they used to. Specifically: * For arch/arc we passed NULL to kgdb_nmicallback() instead of get_irq_regs(). * For arch/mips there was a bit of extra code around kgdb_nmicallback() NOTE: In this patch we will still get into trouble if we try to round up a CPU that failed to round up before. We'll try to round it up again and potentially hang when we try to grab the csd lock. That's not new behavior but we'll still try to do better in a future patch. Suggested-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Douglas Anderson <dianders@chromium.org> Cc: Vineet Gupta <vgupta@synopsys.com> Cc: Russell King <linux@armlinux.org.uk> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Richard Kuo <rkuo@codeaurora.org> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Paul Burton <paul.burton@mips.com> Cc: James Hogan <jhogan@kernel.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: Rich Felker <dalias@libc.org> Cc: "David S. Miller" <davem@davemloft.net> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Acked-by: Will Deacon <will.deacon@arm.com> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:26 +09:00
#include <linux/irq.h>
#include <asm/cacheflush.h>
#include <asm/byteorder.h>
#include <linux/atomic.h>
#include "debug_core.h"
static int kgdb_break_asap;
struct debuggerinfo_struct kgdb_info[NR_CPUS];
/**
* kgdb_connected - Is a host GDB connected to us?
*/
int kgdb_connected;
EXPORT_SYMBOL_GPL(kgdb_connected);
/* All the KGDB handlers are installed */
int kgdb_io_module_registered;
/* Guard for recursive entry */
static int exception_level;
struct kgdb_io *dbg_io_ops;
static DEFINE_SPINLOCK(kgdb_registration_lock);
/* Action for the reboot notifiter, a global allow kdb to change it */
static int kgdbreboot;
/* kgdb console driver is loaded */
static int kgdb_con_registered;
/* determine if kgdb console output should be used */
static int kgdb_use_con;
/* Flag for alternate operations for early debugging */
bool dbg_is_early = true;
/* Next cpu to become the master debug core */
int dbg_switch_cpu;
/* Use kdb or gdbserver mode */
int dbg_kdb_mode = 1;
module_param(kgdb_use_con, int, 0644);
module_param(kgdbreboot, int, 0644);
/*
* Holds information about breakpoints in a kernel. These breakpoints are
* added and removed by gdb.
*/
static struct kgdb_bkpt kgdb_break[KGDB_MAX_BREAKPOINTS] = {
[0 ... KGDB_MAX_BREAKPOINTS-1] = { .state = BP_UNDEFINED }
};
/*
* The CPU# of the active CPU, or -1 if none:
*/
atomic_t kgdb_active = ATOMIC_INIT(-1);
EXPORT_SYMBOL_GPL(kgdb_active);
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
static DEFINE_RAW_SPINLOCK(dbg_master_lock);
static DEFINE_RAW_SPINLOCK(dbg_slave_lock);
/*
* We use NR_CPUs not PERCPU, in case kgdb is used to debug early
* bootup code (which might not have percpu set up yet):
*/
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
static atomic_t masters_in_kgdb;
static atomic_t slaves_in_kgdb;
static atomic_t kgdb_break_tasklet_var;
atomic_t kgdb_setting_breakpoint;
struct task_struct *kgdb_usethread;
struct task_struct *kgdb_contthread;
int kgdb_single_step;
static pid_t kgdb_sstep_pid;
/* to keep track of the CPU which is doing the single stepping*/
atomic_t kgdb_cpu_doing_single_step = ATOMIC_INIT(-1);
/*
* If you are debugging a problem where roundup (the collection of
* all other CPUs) is a problem [this should be extremely rare],
* then use the nokgdbroundup option to avoid roundup. In that case
* the other CPUs might interfere with your debugging context, so
* use this with care:
*/
kgdb: fix signedness mixmatches, add statics, add declaration to header Noticed by sparse: arch/x86/kernel/kgdb.c:556:15: warning: symbol 'kgdb_arch_pc' was not declared. Should it be static? kernel/kgdb.c:149:8: warning: symbol 'kgdb_do_roundup' was not declared. Should it be static? kernel/kgdb.c:193:22: warning: symbol 'kgdb_arch_pc' was not declared. Should it be static? kernel/kgdb.c:712:5: warning: symbol 'remove_all_break' was not declared. Should it be static? Related to kgdb_hex2long: arch/x86/kernel/kgdb.c:371:28: warning: incorrect type in argument 2 (different signedness) arch/x86/kernel/kgdb.c:371:28: expected long *long_val arch/x86/kernel/kgdb.c:371:28: got unsigned long *<noident> kernel/kgdb.c:469:27: warning: incorrect type in argument 2 (different signedness) kernel/kgdb.c:469:27: expected long *long_val kernel/kgdb.c:469:27: got unsigned long *<noident> kernel/kgdb.c:470:27: warning: incorrect type in argument 2 (different signedness) kernel/kgdb.c:470:27: expected long *long_val kernel/kgdb.c:470:27: got unsigned long *<noident> kernel/kgdb.c:894:27: warning: incorrect type in argument 2 (different signedness) kernel/kgdb.c:894:27: expected long *long_val kernel/kgdb.c:894:27: got unsigned long *<noident> kernel/kgdb.c:895:27: warning: incorrect type in argument 2 (different signedness) kernel/kgdb.c:895:27: expected long *long_val kernel/kgdb.c:895:27: got unsigned long *<noident> kernel/kgdb.c:1127:28: warning: incorrect type in argument 2 (different signedness) kernel/kgdb.c:1127:28: expected long *long_val kernel/kgdb.c:1127:28: got unsigned long *<noident> kernel/kgdb.c:1132:25: warning: incorrect type in argument 2 (different signedness) kernel/kgdb.c:1132:25: expected long *long_val kernel/kgdb.c:1132:25: got unsigned long *<noident> Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com> Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
2008-04-25 06:57:23 +09:00
static int kgdb_do_roundup = 1;
static int __init opt_nokgdbroundup(char *str)
{
kgdb_do_roundup = 0;
return 0;
}
early_param("nokgdbroundup", opt_nokgdbroundup);
/*
* Finally, some KGDB code :-)
*/
/*
* Weak aliases for breakpoint management,
* can be overriden by architectures when needed:
*/
int __weak kgdb_arch_set_breakpoint(struct kgdb_bkpt *bpt)
{
int err;
err = probe_kernel_read(bpt->saved_instr, (char *)bpt->bpt_addr,
BREAK_INSTR_SIZE);
if (err)
return err;
err = probe_kernel_write((char *)bpt->bpt_addr,
arch_kgdb_ops.gdb_bpt_instr, BREAK_INSTR_SIZE);
return err;
}
int __weak kgdb_arch_remove_breakpoint(struct kgdb_bkpt *bpt)
{
return probe_kernel_write((char *)bpt->bpt_addr,
(char *)bpt->saved_instr, BREAK_INSTR_SIZE);
}
int __weak kgdb_validate_break_address(unsigned long addr)
{
struct kgdb_bkpt tmp;
int err;
/* Validate setting the breakpoint and then removing it. If the
* remove fails, the kernel needs to emit a bad message because we
* are deep trouble not being able to put things back the way we
* found them.
*/
tmp.bpt_addr = addr;
err = kgdb_arch_set_breakpoint(&tmp);
if (err)
return err;
err = kgdb_arch_remove_breakpoint(&tmp);
if (err)
pr_err("Critical breakpoint error, kernel memory destroyed at: %lx\n",
addr);
return err;
}
unsigned long __weak kgdb_arch_pc(int exception, struct pt_regs *regs)
{
return instruction_pointer(regs);
}
int __weak kgdb_arch_init(void)
{
return 0;
}
int __weak kgdb_skipexception(int exception, struct pt_regs *regs)
{
return 0;
}
kgdb: Fix kgdb_roundup_cpus() for arches who used smp_call_function() When I had lockdep turned on and dropped into kgdb I got a nice splat on my system. Specifically it hit: DEBUG_LOCKS_WARN_ON(current->hardirq_context) Specifically it looked like this: sysrq: SysRq : DEBUG ------------[ cut here ]------------ DEBUG_LOCKS_WARN_ON(current->hardirq_context) WARNING: CPU: 0 PID: 0 at .../kernel/locking/lockdep.c:2875 lockdep_hardirqs_on+0xf0/0x160 CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.19.0 #27 pstate: 604003c9 (nZCv DAIF +PAN -UAO) pc : lockdep_hardirqs_on+0xf0/0x160 ... Call trace: lockdep_hardirqs_on+0xf0/0x160 trace_hardirqs_on+0x188/0x1ac kgdb_roundup_cpus+0x14/0x3c kgdb_cpu_enter+0x53c/0x5cc kgdb_handle_exception+0x180/0x1d4 kgdb_compiled_brk_fn+0x30/0x3c brk_handler+0x134/0x178 do_debug_exception+0xfc/0x178 el1_dbg+0x18/0x78 kgdb_breakpoint+0x34/0x58 sysrq_handle_dbg+0x54/0x5c __handle_sysrq+0x114/0x21c handle_sysrq+0x30/0x3c qcom_geni_serial_isr+0x2dc/0x30c ... ... irq event stamp: ...45 hardirqs last enabled at (...44): [...] __do_softirq+0xd8/0x4e4 hardirqs last disabled at (...45): [...] el1_irq+0x74/0x130 softirqs last enabled at (...42): [...] _local_bh_enable+0x2c/0x34 softirqs last disabled at (...43): [...] irq_exit+0xa8/0x100 ---[ end trace adf21f830c46e638 ]--- Looking closely at it, it seems like a really bad idea to be calling local_irq_enable() in kgdb_roundup_cpus(). If nothing else that seems like it could violate spinlock semantics and cause a deadlock. Instead, let's use a private csd alongside smp_call_function_single_async() to round up the other CPUs. Using smp_call_function_single_async() doesn't require interrupts to be enabled so we can remove the offending bit of code. In order to avoid duplicating this across all the architectures that use the default kgdb_roundup_cpus(), we'll add a "weak" implementation to debug_core.c. Looking at all the people who previously had copies of this code, there were a few variants. I've attempted to keep the variants working like they used to. Specifically: * For arch/arc we passed NULL to kgdb_nmicallback() instead of get_irq_regs(). * For arch/mips there was a bit of extra code around kgdb_nmicallback() NOTE: In this patch we will still get into trouble if we try to round up a CPU that failed to round up before. We'll try to round it up again and potentially hang when we try to grab the csd lock. That's not new behavior but we'll still try to do better in a future patch. Suggested-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Douglas Anderson <dianders@chromium.org> Cc: Vineet Gupta <vgupta@synopsys.com> Cc: Russell King <linux@armlinux.org.uk> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Richard Kuo <rkuo@codeaurora.org> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Paul Burton <paul.burton@mips.com> Cc: James Hogan <jhogan@kernel.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: Rich Felker <dalias@libc.org> Cc: "David S. Miller" <davem@davemloft.net> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Acked-by: Will Deacon <will.deacon@arm.com> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:26 +09:00
#ifdef CONFIG_SMP
/*
* Default (weak) implementation for kgdb_roundup_cpus
*/
static DEFINE_PER_CPU(call_single_data_t, kgdb_roundup_csd);
void __weak kgdb_call_nmi_hook(void *ignored)
{
/*
* NOTE: get_irq_regs() is supposed to get the registers from
* before the IPI interrupt happened and so is supposed to
* show where the processor was. In some situations it's
* possible we might be called without an IPI, so it might be
* safer to figure out how to make kgdb_breakpoint() work
* properly here.
*/
kgdb_nmicallback(raw_smp_processor_id(), get_irq_regs());
}
void __weak kgdb_roundup_cpus(void)
{
call_single_data_t *csd;
int this_cpu = raw_smp_processor_id();
int cpu;
kgdb: Don't round up a CPU that failed rounding up before If we're using the default implementation of kgdb_roundup_cpus() that uses smp_call_function_single_async() we can end up hanging kgdb_roundup_cpus() if we try to round up a CPU that failed to round up before. Specifically smp_call_function_single_async() will try to wait on the csd lock for the CPU that we're trying to round up. If the previous round up never finished then that lock could still be held and we'll just sit there hanging. There's not a lot of use trying to round up a CPU that failed to round up before. Let's keep a flag that indicates whether the CPU started but didn't finish to round up before. If we see that flag set then we'll skip the next round up. In general we have a few goals here: - We never want to end up calling smp_call_function_single_async() when the csd is still locked. This is accomplished because flush_smp_call_function_queue() unlocks the csd _before_ invoking the callback. That means that when kgdb_nmicallback() runs we know for sure the the csd is no longer locked. Thus when we set "rounding_up = false" we know for sure that the csd is unlocked. - If there are no timeouts rounding up we should never skip a round up. NOTE #1: In general trying to continue running after failing to round up CPUs doesn't appear to be supported in the debugger. When I simulate this I find that kdb reports "Catastrophic error detected" when I try to continue. I can overrule and continue anyway, but it should be noted that we may be entering the land of dragons here. Possibly the "Catastrophic error detected" was added _because_ of the future failure to round up, but even so this is an area of the code that hasn't been strongly tested. NOTE #2: I did a bit of testing before and after this change. I introduced a 10 second hang in the kernel while holding a spinlock that I could invoke on a certain CPU with 'taskset -c 3 cat /sys/...". Before this change if I did: - Invoke hang - Enter debugger - g (which warns about Catastrophic error, g again to go anyway) - g - Enter debugger ...I'd hang the rest of the 10 seconds without getting a debugger prompt. After this change I end up in the debugger the 2nd time after only 1 second with the standard warning about 'Timed out waiting for secondary CPUs.' I'll also note that once the CPU finished waiting I could actually debug it (aka "btc" worked) I won't promise that everything works perfectly if the errant CPU comes back at just the wrong time (like as we're entering or exiting the debugger) but it certainly seems like an improvement. NOTE #3: setting 'kgdb_info[cpu].rounding_up = false' is in kgdb_nmicallback() instead of kgdb_call_nmi_hook() because some implementations override kgdb_call_nmi_hook(). It shouldn't hurt to have it in kgdb_nmicallback() in any case. NOTE #4: this logic is really only needed because there is no API call like "smp_try_call_function_single_async()" or "smp_csd_is_locked()". If such an API existed then we'd use it instead, but it seemed a bit much to add an API like this just for kgdb. Signed-off-by: Douglas Anderson <dianders@chromium.org> Acked-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:27 +09:00
int ret;
kgdb: Fix kgdb_roundup_cpus() for arches who used smp_call_function() When I had lockdep turned on and dropped into kgdb I got a nice splat on my system. Specifically it hit: DEBUG_LOCKS_WARN_ON(current->hardirq_context) Specifically it looked like this: sysrq: SysRq : DEBUG ------------[ cut here ]------------ DEBUG_LOCKS_WARN_ON(current->hardirq_context) WARNING: CPU: 0 PID: 0 at .../kernel/locking/lockdep.c:2875 lockdep_hardirqs_on+0xf0/0x160 CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.19.0 #27 pstate: 604003c9 (nZCv DAIF +PAN -UAO) pc : lockdep_hardirqs_on+0xf0/0x160 ... Call trace: lockdep_hardirqs_on+0xf0/0x160 trace_hardirqs_on+0x188/0x1ac kgdb_roundup_cpus+0x14/0x3c kgdb_cpu_enter+0x53c/0x5cc kgdb_handle_exception+0x180/0x1d4 kgdb_compiled_brk_fn+0x30/0x3c brk_handler+0x134/0x178 do_debug_exception+0xfc/0x178 el1_dbg+0x18/0x78 kgdb_breakpoint+0x34/0x58 sysrq_handle_dbg+0x54/0x5c __handle_sysrq+0x114/0x21c handle_sysrq+0x30/0x3c qcom_geni_serial_isr+0x2dc/0x30c ... ... irq event stamp: ...45 hardirqs last enabled at (...44): [...] __do_softirq+0xd8/0x4e4 hardirqs last disabled at (...45): [...] el1_irq+0x74/0x130 softirqs last enabled at (...42): [...] _local_bh_enable+0x2c/0x34 softirqs last disabled at (...43): [...] irq_exit+0xa8/0x100 ---[ end trace adf21f830c46e638 ]--- Looking closely at it, it seems like a really bad idea to be calling local_irq_enable() in kgdb_roundup_cpus(). If nothing else that seems like it could violate spinlock semantics and cause a deadlock. Instead, let's use a private csd alongside smp_call_function_single_async() to round up the other CPUs. Using smp_call_function_single_async() doesn't require interrupts to be enabled so we can remove the offending bit of code. In order to avoid duplicating this across all the architectures that use the default kgdb_roundup_cpus(), we'll add a "weak" implementation to debug_core.c. Looking at all the people who previously had copies of this code, there were a few variants. I've attempted to keep the variants working like they used to. Specifically: * For arch/arc we passed NULL to kgdb_nmicallback() instead of get_irq_regs(). * For arch/mips there was a bit of extra code around kgdb_nmicallback() NOTE: In this patch we will still get into trouble if we try to round up a CPU that failed to round up before. We'll try to round it up again and potentially hang when we try to grab the csd lock. That's not new behavior but we'll still try to do better in a future patch. Suggested-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Douglas Anderson <dianders@chromium.org> Cc: Vineet Gupta <vgupta@synopsys.com> Cc: Russell King <linux@armlinux.org.uk> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Richard Kuo <rkuo@codeaurora.org> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Paul Burton <paul.burton@mips.com> Cc: James Hogan <jhogan@kernel.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: Rich Felker <dalias@libc.org> Cc: "David S. Miller" <davem@davemloft.net> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Acked-by: Will Deacon <will.deacon@arm.com> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:26 +09:00
for_each_online_cpu(cpu) {
/* No need to roundup ourselves */
if (cpu == this_cpu)
continue;
csd = &per_cpu(kgdb_roundup_csd, cpu);
kgdb: Don't round up a CPU that failed rounding up before If we're using the default implementation of kgdb_roundup_cpus() that uses smp_call_function_single_async() we can end up hanging kgdb_roundup_cpus() if we try to round up a CPU that failed to round up before. Specifically smp_call_function_single_async() will try to wait on the csd lock for the CPU that we're trying to round up. If the previous round up never finished then that lock could still be held and we'll just sit there hanging. There's not a lot of use trying to round up a CPU that failed to round up before. Let's keep a flag that indicates whether the CPU started but didn't finish to round up before. If we see that flag set then we'll skip the next round up. In general we have a few goals here: - We never want to end up calling smp_call_function_single_async() when the csd is still locked. This is accomplished because flush_smp_call_function_queue() unlocks the csd _before_ invoking the callback. That means that when kgdb_nmicallback() runs we know for sure the the csd is no longer locked. Thus when we set "rounding_up = false" we know for sure that the csd is unlocked. - If there are no timeouts rounding up we should never skip a round up. NOTE #1: In general trying to continue running after failing to round up CPUs doesn't appear to be supported in the debugger. When I simulate this I find that kdb reports "Catastrophic error detected" when I try to continue. I can overrule and continue anyway, but it should be noted that we may be entering the land of dragons here. Possibly the "Catastrophic error detected" was added _because_ of the future failure to round up, but even so this is an area of the code that hasn't been strongly tested. NOTE #2: I did a bit of testing before and after this change. I introduced a 10 second hang in the kernel while holding a spinlock that I could invoke on a certain CPU with 'taskset -c 3 cat /sys/...". Before this change if I did: - Invoke hang - Enter debugger - g (which warns about Catastrophic error, g again to go anyway) - g - Enter debugger ...I'd hang the rest of the 10 seconds without getting a debugger prompt. After this change I end up in the debugger the 2nd time after only 1 second with the standard warning about 'Timed out waiting for secondary CPUs.' I'll also note that once the CPU finished waiting I could actually debug it (aka "btc" worked) I won't promise that everything works perfectly if the errant CPU comes back at just the wrong time (like as we're entering or exiting the debugger) but it certainly seems like an improvement. NOTE #3: setting 'kgdb_info[cpu].rounding_up = false' is in kgdb_nmicallback() instead of kgdb_call_nmi_hook() because some implementations override kgdb_call_nmi_hook(). It shouldn't hurt to have it in kgdb_nmicallback() in any case. NOTE #4: this logic is really only needed because there is no API call like "smp_try_call_function_single_async()" or "smp_csd_is_locked()". If such an API existed then we'd use it instead, but it seemed a bit much to add an API like this just for kgdb. Signed-off-by: Douglas Anderson <dianders@chromium.org> Acked-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:27 +09:00
/*
* If it didn't round up last time, don't try again
* since smp_call_function_single_async() will block.
*
* If rounding_up is false then we know that the
* previous call must have at least started and that
* means smp_call_function_single_async() won't block.
*/
if (kgdb_info[cpu].rounding_up)
continue;
kgdb_info[cpu].rounding_up = true;
kgdb: Fix kgdb_roundup_cpus() for arches who used smp_call_function() When I had lockdep turned on and dropped into kgdb I got a nice splat on my system. Specifically it hit: DEBUG_LOCKS_WARN_ON(current->hardirq_context) Specifically it looked like this: sysrq: SysRq : DEBUG ------------[ cut here ]------------ DEBUG_LOCKS_WARN_ON(current->hardirq_context) WARNING: CPU: 0 PID: 0 at .../kernel/locking/lockdep.c:2875 lockdep_hardirqs_on+0xf0/0x160 CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.19.0 #27 pstate: 604003c9 (nZCv DAIF +PAN -UAO) pc : lockdep_hardirqs_on+0xf0/0x160 ... Call trace: lockdep_hardirqs_on+0xf0/0x160 trace_hardirqs_on+0x188/0x1ac kgdb_roundup_cpus+0x14/0x3c kgdb_cpu_enter+0x53c/0x5cc kgdb_handle_exception+0x180/0x1d4 kgdb_compiled_brk_fn+0x30/0x3c brk_handler+0x134/0x178 do_debug_exception+0xfc/0x178 el1_dbg+0x18/0x78 kgdb_breakpoint+0x34/0x58 sysrq_handle_dbg+0x54/0x5c __handle_sysrq+0x114/0x21c handle_sysrq+0x30/0x3c qcom_geni_serial_isr+0x2dc/0x30c ... ... irq event stamp: ...45 hardirqs last enabled at (...44): [...] __do_softirq+0xd8/0x4e4 hardirqs last disabled at (...45): [...] el1_irq+0x74/0x130 softirqs last enabled at (...42): [...] _local_bh_enable+0x2c/0x34 softirqs last disabled at (...43): [...] irq_exit+0xa8/0x100 ---[ end trace adf21f830c46e638 ]--- Looking closely at it, it seems like a really bad idea to be calling local_irq_enable() in kgdb_roundup_cpus(). If nothing else that seems like it could violate spinlock semantics and cause a deadlock. Instead, let's use a private csd alongside smp_call_function_single_async() to round up the other CPUs. Using smp_call_function_single_async() doesn't require interrupts to be enabled so we can remove the offending bit of code. In order to avoid duplicating this across all the architectures that use the default kgdb_roundup_cpus(), we'll add a "weak" implementation to debug_core.c. Looking at all the people who previously had copies of this code, there were a few variants. I've attempted to keep the variants working like they used to. Specifically: * For arch/arc we passed NULL to kgdb_nmicallback() instead of get_irq_regs(). * For arch/mips there was a bit of extra code around kgdb_nmicallback() NOTE: In this patch we will still get into trouble if we try to round up a CPU that failed to round up before. We'll try to round it up again and potentially hang when we try to grab the csd lock. That's not new behavior but we'll still try to do better in a future patch. Suggested-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Douglas Anderson <dianders@chromium.org> Cc: Vineet Gupta <vgupta@synopsys.com> Cc: Russell King <linux@armlinux.org.uk> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Richard Kuo <rkuo@codeaurora.org> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Paul Burton <paul.burton@mips.com> Cc: James Hogan <jhogan@kernel.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: Rich Felker <dalias@libc.org> Cc: "David S. Miller" <davem@davemloft.net> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Acked-by: Will Deacon <will.deacon@arm.com> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:26 +09:00
csd->func = kgdb_call_nmi_hook;
kgdb: Don't round up a CPU that failed rounding up before If we're using the default implementation of kgdb_roundup_cpus() that uses smp_call_function_single_async() we can end up hanging kgdb_roundup_cpus() if we try to round up a CPU that failed to round up before. Specifically smp_call_function_single_async() will try to wait on the csd lock for the CPU that we're trying to round up. If the previous round up never finished then that lock could still be held and we'll just sit there hanging. There's not a lot of use trying to round up a CPU that failed to round up before. Let's keep a flag that indicates whether the CPU started but didn't finish to round up before. If we see that flag set then we'll skip the next round up. In general we have a few goals here: - We never want to end up calling smp_call_function_single_async() when the csd is still locked. This is accomplished because flush_smp_call_function_queue() unlocks the csd _before_ invoking the callback. That means that when kgdb_nmicallback() runs we know for sure the the csd is no longer locked. Thus when we set "rounding_up = false" we know for sure that the csd is unlocked. - If there are no timeouts rounding up we should never skip a round up. NOTE #1: In general trying to continue running after failing to round up CPUs doesn't appear to be supported in the debugger. When I simulate this I find that kdb reports "Catastrophic error detected" when I try to continue. I can overrule and continue anyway, but it should be noted that we may be entering the land of dragons here. Possibly the "Catastrophic error detected" was added _because_ of the future failure to round up, but even so this is an area of the code that hasn't been strongly tested. NOTE #2: I did a bit of testing before and after this change. I introduced a 10 second hang in the kernel while holding a spinlock that I could invoke on a certain CPU with 'taskset -c 3 cat /sys/...". Before this change if I did: - Invoke hang - Enter debugger - g (which warns about Catastrophic error, g again to go anyway) - g - Enter debugger ...I'd hang the rest of the 10 seconds without getting a debugger prompt. After this change I end up in the debugger the 2nd time after only 1 second with the standard warning about 'Timed out waiting for secondary CPUs.' I'll also note that once the CPU finished waiting I could actually debug it (aka "btc" worked) I won't promise that everything works perfectly if the errant CPU comes back at just the wrong time (like as we're entering or exiting the debugger) but it certainly seems like an improvement. NOTE #3: setting 'kgdb_info[cpu].rounding_up = false' is in kgdb_nmicallback() instead of kgdb_call_nmi_hook() because some implementations override kgdb_call_nmi_hook(). It shouldn't hurt to have it in kgdb_nmicallback() in any case. NOTE #4: this logic is really only needed because there is no API call like "smp_try_call_function_single_async()" or "smp_csd_is_locked()". If such an API existed then we'd use it instead, but it seemed a bit much to add an API like this just for kgdb. Signed-off-by: Douglas Anderson <dianders@chromium.org> Acked-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:27 +09:00
ret = smp_call_function_single_async(cpu, csd);
if (ret)
kgdb_info[cpu].rounding_up = false;
kgdb: Fix kgdb_roundup_cpus() for arches who used smp_call_function() When I had lockdep turned on and dropped into kgdb I got a nice splat on my system. Specifically it hit: DEBUG_LOCKS_WARN_ON(current->hardirq_context) Specifically it looked like this: sysrq: SysRq : DEBUG ------------[ cut here ]------------ DEBUG_LOCKS_WARN_ON(current->hardirq_context) WARNING: CPU: 0 PID: 0 at .../kernel/locking/lockdep.c:2875 lockdep_hardirqs_on+0xf0/0x160 CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.19.0 #27 pstate: 604003c9 (nZCv DAIF +PAN -UAO) pc : lockdep_hardirqs_on+0xf0/0x160 ... Call trace: lockdep_hardirqs_on+0xf0/0x160 trace_hardirqs_on+0x188/0x1ac kgdb_roundup_cpus+0x14/0x3c kgdb_cpu_enter+0x53c/0x5cc kgdb_handle_exception+0x180/0x1d4 kgdb_compiled_brk_fn+0x30/0x3c brk_handler+0x134/0x178 do_debug_exception+0xfc/0x178 el1_dbg+0x18/0x78 kgdb_breakpoint+0x34/0x58 sysrq_handle_dbg+0x54/0x5c __handle_sysrq+0x114/0x21c handle_sysrq+0x30/0x3c qcom_geni_serial_isr+0x2dc/0x30c ... ... irq event stamp: ...45 hardirqs last enabled at (...44): [...] __do_softirq+0xd8/0x4e4 hardirqs last disabled at (...45): [...] el1_irq+0x74/0x130 softirqs last enabled at (...42): [...] _local_bh_enable+0x2c/0x34 softirqs last disabled at (...43): [...] irq_exit+0xa8/0x100 ---[ end trace adf21f830c46e638 ]--- Looking closely at it, it seems like a really bad idea to be calling local_irq_enable() in kgdb_roundup_cpus(). If nothing else that seems like it could violate spinlock semantics and cause a deadlock. Instead, let's use a private csd alongside smp_call_function_single_async() to round up the other CPUs. Using smp_call_function_single_async() doesn't require interrupts to be enabled so we can remove the offending bit of code. In order to avoid duplicating this across all the architectures that use the default kgdb_roundup_cpus(), we'll add a "weak" implementation to debug_core.c. Looking at all the people who previously had copies of this code, there were a few variants. I've attempted to keep the variants working like they used to. Specifically: * For arch/arc we passed NULL to kgdb_nmicallback() instead of get_irq_regs(). * For arch/mips there was a bit of extra code around kgdb_nmicallback() NOTE: In this patch we will still get into trouble if we try to round up a CPU that failed to round up before. We'll try to round it up again and potentially hang when we try to grab the csd lock. That's not new behavior but we'll still try to do better in a future patch. Suggested-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Douglas Anderson <dianders@chromium.org> Cc: Vineet Gupta <vgupta@synopsys.com> Cc: Russell King <linux@armlinux.org.uk> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Richard Kuo <rkuo@codeaurora.org> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Paul Burton <paul.burton@mips.com> Cc: James Hogan <jhogan@kernel.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: Rich Felker <dalias@libc.org> Cc: "David S. Miller" <davem@davemloft.net> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Acked-by: Will Deacon <will.deacon@arm.com> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:26 +09:00
}
}
#endif
/*
* Some architectures need cache flushes when we set/clear a
* breakpoint:
*/
static void kgdb_flush_swbreak_addr(unsigned long addr)
{
if (!CACHE_FLUSH_IS_SAFE)
return;
mm: per-thread vma caching This patch is a continuation of efforts trying to optimize find_vma(), avoiding potentially expensive rbtree walks to locate a vma upon faults. The original approach (https://lkml.org/lkml/2013/11/1/410), where the largest vma was also cached, ended up being too specific and random, thus further comparison with other approaches were needed. There are two things to consider when dealing with this, the cache hit rate and the latency of find_vma(). Improving the hit-rate does not necessarily translate in finding the vma any faster, as the overhead of any fancy caching schemes can be too high to consider. We currently cache the last used vma for the whole address space, which provides a nice optimization, reducing the total cycles in find_vma() by up to 250%, for workloads with good locality. On the other hand, this simple scheme is pretty much useless for workloads with poor locality. Analyzing ebizzy runs shows that, no matter how many threads are running, the mmap_cache hit rate is less than 2%, and in many situations below 1%. The proposed approach is to replace this scheme with a small per-thread cache, maximizing hit rates at a very low maintenance cost. Invalidations are performed by simply bumping up a 32-bit sequence number. The only expensive operation is in the rare case of a seq number overflow, where all caches that share the same address space are flushed. Upon a miss, the proposed replacement policy is based on the page number that contains the virtual address in question. Concretely, the following results are seen on an 80 core, 8 socket x86-64 box: 1) System bootup: Most programs are single threaded, so the per-thread scheme does improve ~50% hit rate by just adding a few more slots to the cache. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 50.61% | 19.90 | | patched | 73.45% | 13.58 | +----------------+----------+------------------+ 2) Kernel build: This one is already pretty good with the current approach as we're dealing with good locality. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 75.28% | 11.03 | | patched | 88.09% | 9.31 | +----------------+----------+------------------+ 3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 70.66% | 17.14 | | patched | 91.15% | 12.57 | +----------------+----------+------------------+ 4) Ebizzy: There's a fair amount of variation from run to run, but this approach always shows nearly perfect hit rates, while baseline is just about non-existent. The amounts of cycles can fluctuate between anywhere from ~60 to ~116 for the baseline scheme, but this approach reduces it considerably. For instance, with 80 threads: +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 1.06% | 91.54 | | patched | 99.97% | 14.18 | +----------------+----------+------------------+ [akpm@linux-foundation.org: fix nommu build, per Davidlohr] [akpm@linux-foundation.org: document vmacache_valid() logic] [akpm@linux-foundation.org: attempt to untangle header files] [akpm@linux-foundation.org: add vmacache_find() BUG_ON] [hughd@google.com: add vmacache_valid_mm() (from Oleg)] [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: adjust and enhance comments] Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Reviewed-by: Michel Lespinasse <walken@google.com> Cc: Oleg Nesterov <oleg@redhat.com> Tested-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 07:37:25 +09:00
if (current->mm) {
int i;
for (i = 0; i < VMACACHE_SIZE; i++) {
if (!current->vmacache.vmas[i])
mm: per-thread vma caching This patch is a continuation of efforts trying to optimize find_vma(), avoiding potentially expensive rbtree walks to locate a vma upon faults. The original approach (https://lkml.org/lkml/2013/11/1/410), where the largest vma was also cached, ended up being too specific and random, thus further comparison with other approaches were needed. There are two things to consider when dealing with this, the cache hit rate and the latency of find_vma(). Improving the hit-rate does not necessarily translate in finding the vma any faster, as the overhead of any fancy caching schemes can be too high to consider. We currently cache the last used vma for the whole address space, which provides a nice optimization, reducing the total cycles in find_vma() by up to 250%, for workloads with good locality. On the other hand, this simple scheme is pretty much useless for workloads with poor locality. Analyzing ebizzy runs shows that, no matter how many threads are running, the mmap_cache hit rate is less than 2%, and in many situations below 1%. The proposed approach is to replace this scheme with a small per-thread cache, maximizing hit rates at a very low maintenance cost. Invalidations are performed by simply bumping up a 32-bit sequence number. The only expensive operation is in the rare case of a seq number overflow, where all caches that share the same address space are flushed. Upon a miss, the proposed replacement policy is based on the page number that contains the virtual address in question. Concretely, the following results are seen on an 80 core, 8 socket x86-64 box: 1) System bootup: Most programs are single threaded, so the per-thread scheme does improve ~50% hit rate by just adding a few more slots to the cache. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 50.61% | 19.90 | | patched | 73.45% | 13.58 | +----------------+----------+------------------+ 2) Kernel build: This one is already pretty good with the current approach as we're dealing with good locality. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 75.28% | 11.03 | | patched | 88.09% | 9.31 | +----------------+----------+------------------+ 3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 70.66% | 17.14 | | patched | 91.15% | 12.57 | +----------------+----------+------------------+ 4) Ebizzy: There's a fair amount of variation from run to run, but this approach always shows nearly perfect hit rates, while baseline is just about non-existent. The amounts of cycles can fluctuate between anywhere from ~60 to ~116 for the baseline scheme, but this approach reduces it considerably. For instance, with 80 threads: +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 1.06% | 91.54 | | patched | 99.97% | 14.18 | +----------------+----------+------------------+ [akpm@linux-foundation.org: fix nommu build, per Davidlohr] [akpm@linux-foundation.org: document vmacache_valid() logic] [akpm@linux-foundation.org: attempt to untangle header files] [akpm@linux-foundation.org: add vmacache_find() BUG_ON] [hughd@google.com: add vmacache_valid_mm() (from Oleg)] [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: adjust and enhance comments] Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Reviewed-by: Michel Lespinasse <walken@google.com> Cc: Oleg Nesterov <oleg@redhat.com> Tested-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 07:37:25 +09:00
continue;
flush_cache_range(current->vmacache.vmas[i],
mm: per-thread vma caching This patch is a continuation of efforts trying to optimize find_vma(), avoiding potentially expensive rbtree walks to locate a vma upon faults. The original approach (https://lkml.org/lkml/2013/11/1/410), where the largest vma was also cached, ended up being too specific and random, thus further comparison with other approaches were needed. There are two things to consider when dealing with this, the cache hit rate and the latency of find_vma(). Improving the hit-rate does not necessarily translate in finding the vma any faster, as the overhead of any fancy caching schemes can be too high to consider. We currently cache the last used vma for the whole address space, which provides a nice optimization, reducing the total cycles in find_vma() by up to 250%, for workloads with good locality. On the other hand, this simple scheme is pretty much useless for workloads with poor locality. Analyzing ebizzy runs shows that, no matter how many threads are running, the mmap_cache hit rate is less than 2%, and in many situations below 1%. The proposed approach is to replace this scheme with a small per-thread cache, maximizing hit rates at a very low maintenance cost. Invalidations are performed by simply bumping up a 32-bit sequence number. The only expensive operation is in the rare case of a seq number overflow, where all caches that share the same address space are flushed. Upon a miss, the proposed replacement policy is based on the page number that contains the virtual address in question. Concretely, the following results are seen on an 80 core, 8 socket x86-64 box: 1) System bootup: Most programs are single threaded, so the per-thread scheme does improve ~50% hit rate by just adding a few more slots to the cache. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 50.61% | 19.90 | | patched | 73.45% | 13.58 | +----------------+----------+------------------+ 2) Kernel build: This one is already pretty good with the current approach as we're dealing with good locality. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 75.28% | 11.03 | | patched | 88.09% | 9.31 | +----------------+----------+------------------+ 3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 70.66% | 17.14 | | patched | 91.15% | 12.57 | +----------------+----------+------------------+ 4) Ebizzy: There's a fair amount of variation from run to run, but this approach always shows nearly perfect hit rates, while baseline is just about non-existent. The amounts of cycles can fluctuate between anywhere from ~60 to ~116 for the baseline scheme, but this approach reduces it considerably. For instance, with 80 threads: +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 1.06% | 91.54 | | patched | 99.97% | 14.18 | +----------------+----------+------------------+ [akpm@linux-foundation.org: fix nommu build, per Davidlohr] [akpm@linux-foundation.org: document vmacache_valid() logic] [akpm@linux-foundation.org: attempt to untangle header files] [akpm@linux-foundation.org: add vmacache_find() BUG_ON] [hughd@google.com: add vmacache_valid_mm() (from Oleg)] [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: adjust and enhance comments] Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Reviewed-by: Michel Lespinasse <walken@google.com> Cc: Oleg Nesterov <oleg@redhat.com> Tested-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 07:37:25 +09:00
addr, addr + BREAK_INSTR_SIZE);
}
}
mm: per-thread vma caching This patch is a continuation of efforts trying to optimize find_vma(), avoiding potentially expensive rbtree walks to locate a vma upon faults. The original approach (https://lkml.org/lkml/2013/11/1/410), where the largest vma was also cached, ended up being too specific and random, thus further comparison with other approaches were needed. There are two things to consider when dealing with this, the cache hit rate and the latency of find_vma(). Improving the hit-rate does not necessarily translate in finding the vma any faster, as the overhead of any fancy caching schemes can be too high to consider. We currently cache the last used vma for the whole address space, which provides a nice optimization, reducing the total cycles in find_vma() by up to 250%, for workloads with good locality. On the other hand, this simple scheme is pretty much useless for workloads with poor locality. Analyzing ebizzy runs shows that, no matter how many threads are running, the mmap_cache hit rate is less than 2%, and in many situations below 1%. The proposed approach is to replace this scheme with a small per-thread cache, maximizing hit rates at a very low maintenance cost. Invalidations are performed by simply bumping up a 32-bit sequence number. The only expensive operation is in the rare case of a seq number overflow, where all caches that share the same address space are flushed. Upon a miss, the proposed replacement policy is based on the page number that contains the virtual address in question. Concretely, the following results are seen on an 80 core, 8 socket x86-64 box: 1) System bootup: Most programs are single threaded, so the per-thread scheme does improve ~50% hit rate by just adding a few more slots to the cache. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 50.61% | 19.90 | | patched | 73.45% | 13.58 | +----------------+----------+------------------+ 2) Kernel build: This one is already pretty good with the current approach as we're dealing with good locality. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 75.28% | 11.03 | | patched | 88.09% | 9.31 | +----------------+----------+------------------+ 3) Oracle 11g Data Mining (4k pages): Similar to the kernel build workload. +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 70.66% | 17.14 | | patched | 91.15% | 12.57 | +----------------+----------+------------------+ 4) Ebizzy: There's a fair amount of variation from run to run, but this approach always shows nearly perfect hit rates, while baseline is just about non-existent. The amounts of cycles can fluctuate between anywhere from ~60 to ~116 for the baseline scheme, but this approach reduces it considerably. For instance, with 80 threads: +----------------+----------+------------------+ | caching scheme | hit-rate | cycles (billion) | +----------------+----------+------------------+ | baseline | 1.06% | 91.54 | | patched | 99.97% | 14.18 | +----------------+----------+------------------+ [akpm@linux-foundation.org: fix nommu build, per Davidlohr] [akpm@linux-foundation.org: document vmacache_valid() logic] [akpm@linux-foundation.org: attempt to untangle header files] [akpm@linux-foundation.org: add vmacache_find() BUG_ON] [hughd@google.com: add vmacache_valid_mm() (from Oleg)] [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: adjust and enhance comments] Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> Reviewed-by: Rik van Riel <riel@redhat.com> Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Reviewed-by: Michel Lespinasse <walken@google.com> Cc: Oleg Nesterov <oleg@redhat.com> Tested-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-04-08 07:37:25 +09:00
/* Force flush instruction cache if it was outside the mm */
flush_icache_range(addr, addr + BREAK_INSTR_SIZE);
}
/*
* SW breakpoint management:
*/
int dbg_activate_sw_breakpoints(void)
{
int error;
int ret = 0;
int i;
for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
if (kgdb_break[i].state != BP_SET)
continue;
error = kgdb_arch_set_breakpoint(&kgdb_break[i]);
if (error) {
ret = error;
pr_info("BP install failed: %lx\n",
kgdb_break[i].bpt_addr);
continue;
}
kgdb_flush_swbreak_addr(kgdb_break[i].bpt_addr);
kgdb_break[i].state = BP_ACTIVE;
}
return ret;
}
int dbg_set_sw_break(unsigned long addr)
{
int err = kgdb_validate_break_address(addr);
int breakno = -1;
int i;
if (err)
return err;
for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
if ((kgdb_break[i].state == BP_SET) &&
(kgdb_break[i].bpt_addr == addr))
return -EEXIST;
}
for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
if (kgdb_break[i].state == BP_REMOVED &&
kgdb_break[i].bpt_addr == addr) {
breakno = i;
break;
}
}
if (breakno == -1) {
for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
if (kgdb_break[i].state == BP_UNDEFINED) {
breakno = i;
break;
}
}
}
if (breakno == -1)
return -E2BIG;
kgdb_break[breakno].state = BP_SET;
kgdb_break[breakno].type = BP_BREAKPOINT;
kgdb_break[breakno].bpt_addr = addr;
return 0;
}
int dbg_deactivate_sw_breakpoints(void)
{
int error;
int ret = 0;
int i;
for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
if (kgdb_break[i].state != BP_ACTIVE)
continue;
error = kgdb_arch_remove_breakpoint(&kgdb_break[i]);
if (error) {
pr_info("BP remove failed: %lx\n",
kgdb_break[i].bpt_addr);
ret = error;
}
kgdb_flush_swbreak_addr(kgdb_break[i].bpt_addr);
kgdb_break[i].state = BP_SET;
}
return ret;
}
int dbg_remove_sw_break(unsigned long addr)
{
int i;
for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
if ((kgdb_break[i].state == BP_SET) &&
(kgdb_break[i].bpt_addr == addr)) {
kgdb_break[i].state = BP_REMOVED;
return 0;
}
}
return -ENOENT;
}
int kgdb_isremovedbreak(unsigned long addr)
{
int i;
for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
if ((kgdb_break[i].state == BP_REMOVED) &&
(kgdb_break[i].bpt_addr == addr))
return 1;
}
return 0;
}
int dbg_remove_all_break(void)
{
int error;
int i;
/* Clear memory breakpoints. */
for (i = 0; i < KGDB_MAX_BREAKPOINTS; i++) {
if (kgdb_break[i].state != BP_ACTIVE)
goto setundefined;
error = kgdb_arch_remove_breakpoint(&kgdb_break[i]);
if (error)
pr_err("breakpoint remove failed: %lx\n",
kgdb_break[i].bpt_addr);
setundefined:
kgdb_break[i].state = BP_UNDEFINED;
}
/* Clear hardware breakpoints. */
if (arch_kgdb_ops.remove_all_hw_break)
arch_kgdb_ops.remove_all_hw_break();
return 0;
}
/*
* Return true if there is a valid kgdb I/O module. Also if no
* debugger is attached a message can be printed to the console about
* waiting for the debugger to attach.
*
* The print_wait argument is only to be true when called from inside
* the core kgdb_handle_exception, because it will wait for the
* debugger to attach.
*/
static int kgdb_io_ready(int print_wait)
{
if (!dbg_io_ops)
return 0;
if (kgdb_connected)
return 1;
if (atomic_read(&kgdb_setting_breakpoint))
return 1;
if (print_wait) {
#ifdef CONFIG_KGDB_KDB
if (!dbg_kdb_mode)
pr_crit("waiting... or $3#33 for KDB\n");
#else
pr_crit("Waiting for remote debugger\n");
#endif
}
return 1;
}
static int kgdb_reenter_check(struct kgdb_state *ks)
{
unsigned long addr;
if (atomic_read(&kgdb_active) != raw_smp_processor_id())
return 0;
/* Panic on recursive debugger calls: */
exception_level++;
addr = kgdb_arch_pc(ks->ex_vector, ks->linux_regs);
dbg_deactivate_sw_breakpoints();
/*
* If the break point removed ok at the place exception
* occurred, try to recover and print a warning to the end
* user because the user planted a breakpoint in a place that
* KGDB needs in order to function.
*/
if (dbg_remove_sw_break(addr) == 0) {
exception_level = 0;
kgdb_skipexception(ks->ex_vector, ks->linux_regs);
dbg_activate_sw_breakpoints();
pr_crit("re-enter error: breakpoint removed %lx\n", addr);
WARN_ON_ONCE(1);
return 1;
}
dbg_remove_all_break();
kgdb_skipexception(ks->ex_vector, ks->linux_regs);
if (exception_level > 1) {
dump_stack();
kgdb_io_module_registered = false;
panic("Recursive entry to debugger");
}
pr_crit("re-enter exception: ALL breakpoints killed\n");
#ifdef CONFIG_KGDB_KDB
/* Allow kdb to debug itself one level */
return 0;
#endif
dump_stack();
panic("Recursive entry to debugger");
return 1;
}
static void dbg_touch_watchdogs(void)
{
touch_softlockup_watchdog_sync();
clocksource_touch_watchdog();
rcu_cpu_stall_reset();
}
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
static int kgdb_cpu_enter(struct kgdb_state *ks, struct pt_regs *regs,
int exception_state)
{
unsigned long flags;
int sstep_tries = 100;
int error;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
int cpu;
int trace_on = 0;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
int online_cpus = num_online_cpus();
u64 time_left;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
kgdb_info[ks->cpu].enter_kgdb++;
kgdb_info[ks->cpu].exception_state |= exception_state;
if (exception_state == DCPU_WANT_MASTER)
atomic_inc(&masters_in_kgdb);
else
atomic_inc(&slaves_in_kgdb);
if (arch_kgdb_ops.disable_hw_break)
arch_kgdb_ops.disable_hw_break(regs);
acquirelock:
kgdb: Avoid suspicious RCU usage warning [ Upstream commit 440ab9e10e2e6e5fd677473ee6f9e3af0f6904d6 ] At times when I'm using kgdb I see a splat on my console about suspicious RCU usage. I managed to come up with a case that could reproduce this that looked like this: WARNING: suspicious RCU usage 5.7.0-rc4+ #609 Not tainted ----------------------------- kernel/pid.c:395 find_task_by_pid_ns() needs rcu_read_lock() protection! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 1 3 locks held by swapper/0/1: #0: ffffff81b6b8e988 (&dev->mutex){....}-{3:3}, at: __device_attach+0x40/0x13c #1: ffffffd01109e9e8 (dbg_master_lock){....}-{2:2}, at: kgdb_cpu_enter+0x20c/0x7ac #2: ffffffd01109ea90 (dbg_slave_lock){....}-{2:2}, at: kgdb_cpu_enter+0x3ec/0x7ac stack backtrace: CPU: 7 PID: 1 Comm: swapper/0 Not tainted 5.7.0-rc4+ #609 Hardware name: Google Cheza (rev3+) (DT) Call trace: dump_backtrace+0x0/0x1b8 show_stack+0x1c/0x24 dump_stack+0xd4/0x134 lockdep_rcu_suspicious+0xf0/0x100 find_task_by_pid_ns+0x5c/0x80 getthread+0x8c/0xb0 gdb_serial_stub+0x9d4/0xd04 kgdb_cpu_enter+0x284/0x7ac kgdb_handle_exception+0x174/0x20c kgdb_brk_fn+0x24/0x30 call_break_hook+0x6c/0x7c brk_handler+0x20/0x5c do_debug_exception+0x1c8/0x22c el1_sync_handler+0x3c/0xe4 el1_sync+0x7c/0x100 rpmh_rsc_probe+0x38/0x420 platform_drv_probe+0x94/0xb4 really_probe+0x134/0x300 driver_probe_device+0x68/0x100 __device_attach_driver+0x90/0xa8 bus_for_each_drv+0x84/0xcc __device_attach+0xb4/0x13c device_initial_probe+0x18/0x20 bus_probe_device+0x38/0x98 device_add+0x38c/0x420 If I understand properly we should just be able to blanket kgdb under one big RCU read lock and the problem should go away. We'll add it to the beast-of-a-function known as kgdb_cpu_enter(). With this I no longer get any splats and things seem to work fine. Signed-off-by: Douglas Anderson <dianders@chromium.org> Link: https://lore.kernel.org/r/20200602154729.v2.1.I70e0d4fd46d5ed2aaf0c98a355e8e1b7a5bb7e4e@changeid Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Sasha Levin <sashal@kernel.org>
2020-06-03 07:47:39 +09:00
rcu_read_lock();
/*
* Interrupts will be restored by the 'trap return' code, except when
* single stepping.
*/
local_irq_save(flags);
cpu = ks->cpu;
kgdb_info[cpu].debuggerinfo = regs;
kgdb_info[cpu].task = current;
kgdb_info[cpu].ret_state = 0;
kgdb_info[cpu].irq_depth = hardirq_count() >> HARDIRQ_SHIFT;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
/* Make sure the above info reaches the primary CPU */
smp_mb();
if (exception_level == 1) {
if (raw_spin_trylock(&dbg_master_lock))
atomic_xchg(&kgdb_active, cpu);
goto cpu_master_loop;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
}
/*
* CPU will loop if it is a slave or request to become a kgdb
* master cpu and acquire the kgdb_active lock:
*/
while (1) {
cpu_loop:
if (kgdb_info[cpu].exception_state & DCPU_NEXT_MASTER) {
kgdb_info[cpu].exception_state &= ~DCPU_NEXT_MASTER;
goto cpu_master_loop;
} else if (kgdb_info[cpu].exception_state & DCPU_WANT_MASTER) {
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
if (raw_spin_trylock(&dbg_master_lock)) {
atomic_xchg(&kgdb_active, cpu);
break;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
}
} else if (kgdb_info[cpu].exception_state & DCPU_IS_SLAVE) {
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
if (!raw_spin_is_locked(&dbg_slave_lock))
goto return_normal;
} else {
return_normal:
/* Return to normal operation by executing any
* hw breakpoint fixup.
*/
if (arch_kgdb_ops.correct_hw_break)
arch_kgdb_ops.correct_hw_break();
if (trace_on)
tracing_on();
kgdb_info[cpu].debuggerinfo = NULL;
kgdb_info[cpu].task = NULL;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
kgdb_info[cpu].exception_state &=
~(DCPU_WANT_MASTER | DCPU_IS_SLAVE);
kgdb_info[cpu].enter_kgdb--;
smp_mb__before_atomic();
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
atomic_dec(&slaves_in_kgdb);
dbg_touch_watchdogs();
local_irq_restore(flags);
kgdb: Avoid suspicious RCU usage warning [ Upstream commit 440ab9e10e2e6e5fd677473ee6f9e3af0f6904d6 ] At times when I'm using kgdb I see a splat on my console about suspicious RCU usage. I managed to come up with a case that could reproduce this that looked like this: WARNING: suspicious RCU usage 5.7.0-rc4+ #609 Not tainted ----------------------------- kernel/pid.c:395 find_task_by_pid_ns() needs rcu_read_lock() protection! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 1 3 locks held by swapper/0/1: #0: ffffff81b6b8e988 (&dev->mutex){....}-{3:3}, at: __device_attach+0x40/0x13c #1: ffffffd01109e9e8 (dbg_master_lock){....}-{2:2}, at: kgdb_cpu_enter+0x20c/0x7ac #2: ffffffd01109ea90 (dbg_slave_lock){....}-{2:2}, at: kgdb_cpu_enter+0x3ec/0x7ac stack backtrace: CPU: 7 PID: 1 Comm: swapper/0 Not tainted 5.7.0-rc4+ #609 Hardware name: Google Cheza (rev3+) (DT) Call trace: dump_backtrace+0x0/0x1b8 show_stack+0x1c/0x24 dump_stack+0xd4/0x134 lockdep_rcu_suspicious+0xf0/0x100 find_task_by_pid_ns+0x5c/0x80 getthread+0x8c/0xb0 gdb_serial_stub+0x9d4/0xd04 kgdb_cpu_enter+0x284/0x7ac kgdb_handle_exception+0x174/0x20c kgdb_brk_fn+0x24/0x30 call_break_hook+0x6c/0x7c brk_handler+0x20/0x5c do_debug_exception+0x1c8/0x22c el1_sync_handler+0x3c/0xe4 el1_sync+0x7c/0x100 rpmh_rsc_probe+0x38/0x420 platform_drv_probe+0x94/0xb4 really_probe+0x134/0x300 driver_probe_device+0x68/0x100 __device_attach_driver+0x90/0xa8 bus_for_each_drv+0x84/0xcc __device_attach+0xb4/0x13c device_initial_probe+0x18/0x20 bus_probe_device+0x38/0x98 device_add+0x38c/0x420 If I understand properly we should just be able to blanket kgdb under one big RCU read lock and the problem should go away. We'll add it to the beast-of-a-function known as kgdb_cpu_enter(). With this I no longer get any splats and things seem to work fine. Signed-off-by: Douglas Anderson <dianders@chromium.org> Link: https://lore.kernel.org/r/20200602154729.v2.1.I70e0d4fd46d5ed2aaf0c98a355e8e1b7a5bb7e4e@changeid Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Sasha Levin <sashal@kernel.org>
2020-06-03 07:47:39 +09:00
rcu_read_unlock();
return 0;
}
cpu_relax();
}
/*
* For single stepping, try to only enter on the processor
* that was single stepping. To guard against a deadlock, the
* kernel will only try for the value of sstep_tries before
* giving up and continuing on.
*/
if (atomic_read(&kgdb_cpu_doing_single_step) != -1 &&
(kgdb_info[cpu].task &&
kgdb_info[cpu].task->pid != kgdb_sstep_pid) && --sstep_tries) {
atomic_set(&kgdb_active, -1);
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
raw_spin_unlock(&dbg_master_lock);
dbg_touch_watchdogs();
local_irq_restore(flags);
kgdb: Avoid suspicious RCU usage warning [ Upstream commit 440ab9e10e2e6e5fd677473ee6f9e3af0f6904d6 ] At times when I'm using kgdb I see a splat on my console about suspicious RCU usage. I managed to come up with a case that could reproduce this that looked like this: WARNING: suspicious RCU usage 5.7.0-rc4+ #609 Not tainted ----------------------------- kernel/pid.c:395 find_task_by_pid_ns() needs rcu_read_lock() protection! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 1 3 locks held by swapper/0/1: #0: ffffff81b6b8e988 (&dev->mutex){....}-{3:3}, at: __device_attach+0x40/0x13c #1: ffffffd01109e9e8 (dbg_master_lock){....}-{2:2}, at: kgdb_cpu_enter+0x20c/0x7ac #2: ffffffd01109ea90 (dbg_slave_lock){....}-{2:2}, at: kgdb_cpu_enter+0x3ec/0x7ac stack backtrace: CPU: 7 PID: 1 Comm: swapper/0 Not tainted 5.7.0-rc4+ #609 Hardware name: Google Cheza (rev3+) (DT) Call trace: dump_backtrace+0x0/0x1b8 show_stack+0x1c/0x24 dump_stack+0xd4/0x134 lockdep_rcu_suspicious+0xf0/0x100 find_task_by_pid_ns+0x5c/0x80 getthread+0x8c/0xb0 gdb_serial_stub+0x9d4/0xd04 kgdb_cpu_enter+0x284/0x7ac kgdb_handle_exception+0x174/0x20c kgdb_brk_fn+0x24/0x30 call_break_hook+0x6c/0x7c brk_handler+0x20/0x5c do_debug_exception+0x1c8/0x22c el1_sync_handler+0x3c/0xe4 el1_sync+0x7c/0x100 rpmh_rsc_probe+0x38/0x420 platform_drv_probe+0x94/0xb4 really_probe+0x134/0x300 driver_probe_device+0x68/0x100 __device_attach_driver+0x90/0xa8 bus_for_each_drv+0x84/0xcc __device_attach+0xb4/0x13c device_initial_probe+0x18/0x20 bus_probe_device+0x38/0x98 device_add+0x38c/0x420 If I understand properly we should just be able to blanket kgdb under one big RCU read lock and the problem should go away. We'll add it to the beast-of-a-function known as kgdb_cpu_enter(). With this I no longer get any splats and things seem to work fine. Signed-off-by: Douglas Anderson <dianders@chromium.org> Link: https://lore.kernel.org/r/20200602154729.v2.1.I70e0d4fd46d5ed2aaf0c98a355e8e1b7a5bb7e4e@changeid Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Sasha Levin <sashal@kernel.org>
2020-06-03 07:47:39 +09:00
rcu_read_unlock();
goto acquirelock;
}
if (!kgdb_io_ready(1)) {
kgdb_info[cpu].ret_state = 1;
goto kgdb_restore; /* No I/O connection, resume the system */
}
/*
* Don't enter if we have hit a removed breakpoint.
*/
if (kgdb_skipexception(ks->ex_vector, ks->linux_regs))
goto kgdb_restore;
kgdb: Disable WARN_CONSOLE_UNLOCKED for all kgdb [ Upstream commit 202164fbfa2b2ffa3e66b504e0f126ba9a745006 ] In commit 81eaadcae81b ("kgdboc: disable the console lock when in kgdb") we avoided the WARN_CONSOLE_UNLOCKED() yell when we were in kgdboc. That still works fine, but it turns out that we get a similar yell when using other I/O drivers. One example is the "I/O driver" for the kgdb test suite (kgdbts). When I enabled that I again got the same yells. Even though "kgdbts" doesn't actually interact with the user over the console, using it still causes kgdb to print to the consoles. That trips the same warning: con_is_visible+0x60/0x68 con_scroll+0x110/0x1b8 lf+0x4c/0xc8 vt_console_print+0x1b8/0x348 vkdb_printf+0x320/0x89c kdb_printf+0x68/0x90 kdb_main_loop+0x190/0x860 kdb_stub+0x2cc/0x3ec kgdb_cpu_enter+0x268/0x744 kgdb_handle_exception+0x1a4/0x200 kgdb_compiled_brk_fn+0x34/0x44 brk_handler+0x7c/0xb8 do_debug_exception+0x1b4/0x228 Let's increment/decrement the "ignore_console_lock_warning" variable all the time when we enter the debugger. This will allow us to later revert commit 81eaadcae81b ("kgdboc: disable the console lock when in kgdb"). Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Reviewed-by: Daniel Thompson <daniel.thompson@linaro.org> Link: https://lore.kernel.org/r/20200507130644.v4.1.Ied2b058357152ebcc8bf68edd6f20a11d98d7d4e@changeid Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Sasha Levin <sashal@kernel.org>
2020-05-08 05:08:39 +09:00
atomic_inc(&ignore_console_lock_warning);
/* Call the I/O driver's pre_exception routine */
if (dbg_io_ops->pre_exception)
dbg_io_ops->pre_exception();
/*
* Get the passive CPU lock which will hold all the non-primary
* CPU in a spin state while the debugger is active
*/
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
if (!kgdb_single_step)
raw_spin_lock(&dbg_slave_lock);
#ifdef CONFIG_SMP
/* If send_ready set, slaves are already waiting */
if (ks->send_ready)
atomic_set(ks->send_ready, 1);
/* Signal the other CPUs to enter kgdb_wait() */
else if ((!kgdb_single_step) && kgdb_do_roundup)
kgdb_roundup_cpus();
#endif
/*
* Wait for the other CPUs to be notified and be waiting for us:
*/
time_left = MSEC_PER_SEC;
while (kgdb_do_roundup && --time_left &&
(atomic_read(&masters_in_kgdb) + atomic_read(&slaves_in_kgdb)) !=
online_cpus)
udelay(1000);
if (!time_left)
pr_crit("Timed out waiting for secondary CPUs.\n");
/*
* At this point the primary processor is completely
* in the debugger and all secondary CPUs are quiescent
*/
dbg_deactivate_sw_breakpoints();
kgdb_single_step = 0;
kgdb_contthread = current;
exception_level = 0;
trace_on = tracing_is_on();
if (trace_on)
tracing_off();
while (1) {
cpu_master_loop:
if (dbg_kdb_mode) {
kgdb_connected = 1;
error = kdb_stub(ks);
if (error == -1)
continue;
kgdb_connected = 0;
} else {
error = gdb_serial_stub(ks);
}
if (error == DBG_PASS_EVENT) {
dbg_kdb_mode = !dbg_kdb_mode;
} else if (error == DBG_SWITCH_CPU_EVENT) {
kgdb_info[dbg_switch_cpu].exception_state |=
DCPU_NEXT_MASTER;
goto cpu_loop;
} else {
kgdb_info[cpu].ret_state = error;
break;
}
}
/* Call the I/O driver's post_exception routine */
if (dbg_io_ops->post_exception)
dbg_io_ops->post_exception();
kgdb: Disable WARN_CONSOLE_UNLOCKED for all kgdb [ Upstream commit 202164fbfa2b2ffa3e66b504e0f126ba9a745006 ] In commit 81eaadcae81b ("kgdboc: disable the console lock when in kgdb") we avoided the WARN_CONSOLE_UNLOCKED() yell when we were in kgdboc. That still works fine, but it turns out that we get a similar yell when using other I/O drivers. One example is the "I/O driver" for the kgdb test suite (kgdbts). When I enabled that I again got the same yells. Even though "kgdbts" doesn't actually interact with the user over the console, using it still causes kgdb to print to the consoles. That trips the same warning: con_is_visible+0x60/0x68 con_scroll+0x110/0x1b8 lf+0x4c/0xc8 vt_console_print+0x1b8/0x348 vkdb_printf+0x320/0x89c kdb_printf+0x68/0x90 kdb_main_loop+0x190/0x860 kdb_stub+0x2cc/0x3ec kgdb_cpu_enter+0x268/0x744 kgdb_handle_exception+0x1a4/0x200 kgdb_compiled_brk_fn+0x34/0x44 brk_handler+0x7c/0xb8 do_debug_exception+0x1b4/0x228 Let's increment/decrement the "ignore_console_lock_warning" variable all the time when we enter the debugger. This will allow us to later revert commit 81eaadcae81b ("kgdboc: disable the console lock when in kgdb"). Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Reviewed-by: Daniel Thompson <daniel.thompson@linaro.org> Link: https://lore.kernel.org/r/20200507130644.v4.1.Ied2b058357152ebcc8bf68edd6f20a11d98d7d4e@changeid Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Sasha Levin <sashal@kernel.org>
2020-05-08 05:08:39 +09:00
atomic_dec(&ignore_console_lock_warning);
if (!kgdb_single_step) {
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
raw_spin_unlock(&dbg_slave_lock);
/* Wait till all the CPUs have quit from the debugger. */
while (kgdb_do_roundup && atomic_read(&slaves_in_kgdb))
cpu_relax();
}
kgdb_restore:
if (atomic_read(&kgdb_cpu_doing_single_step) != -1) {
int sstep_cpu = atomic_read(&kgdb_cpu_doing_single_step);
if (kgdb_info[sstep_cpu].task)
kgdb_sstep_pid = kgdb_info[sstep_cpu].task->pid;
else
kgdb_sstep_pid = 0;
}
if (arch_kgdb_ops.correct_hw_break)
arch_kgdb_ops.correct_hw_break();
if (trace_on)
tracing_on();
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
kgdb_info[cpu].debuggerinfo = NULL;
kgdb_info[cpu].task = NULL;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
kgdb_info[cpu].exception_state &=
~(DCPU_WANT_MASTER | DCPU_IS_SLAVE);
kgdb_info[cpu].enter_kgdb--;
smp_mb__before_atomic();
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
atomic_dec(&masters_in_kgdb);
/* Free kgdb_active */
atomic_set(&kgdb_active, -1);
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
raw_spin_unlock(&dbg_master_lock);
dbg_touch_watchdogs();
local_irq_restore(flags);
kgdb: Avoid suspicious RCU usage warning [ Upstream commit 440ab9e10e2e6e5fd677473ee6f9e3af0f6904d6 ] At times when I'm using kgdb I see a splat on my console about suspicious RCU usage. I managed to come up with a case that could reproduce this that looked like this: WARNING: suspicious RCU usage 5.7.0-rc4+ #609 Not tainted ----------------------------- kernel/pid.c:395 find_task_by_pid_ns() needs rcu_read_lock() protection! other info that might help us debug this: rcu_scheduler_active = 2, debug_locks = 1 3 locks held by swapper/0/1: #0: ffffff81b6b8e988 (&dev->mutex){....}-{3:3}, at: __device_attach+0x40/0x13c #1: ffffffd01109e9e8 (dbg_master_lock){....}-{2:2}, at: kgdb_cpu_enter+0x20c/0x7ac #2: ffffffd01109ea90 (dbg_slave_lock){....}-{2:2}, at: kgdb_cpu_enter+0x3ec/0x7ac stack backtrace: CPU: 7 PID: 1 Comm: swapper/0 Not tainted 5.7.0-rc4+ #609 Hardware name: Google Cheza (rev3+) (DT) Call trace: dump_backtrace+0x0/0x1b8 show_stack+0x1c/0x24 dump_stack+0xd4/0x134 lockdep_rcu_suspicious+0xf0/0x100 find_task_by_pid_ns+0x5c/0x80 getthread+0x8c/0xb0 gdb_serial_stub+0x9d4/0xd04 kgdb_cpu_enter+0x284/0x7ac kgdb_handle_exception+0x174/0x20c kgdb_brk_fn+0x24/0x30 call_break_hook+0x6c/0x7c brk_handler+0x20/0x5c do_debug_exception+0x1c8/0x22c el1_sync_handler+0x3c/0xe4 el1_sync+0x7c/0x100 rpmh_rsc_probe+0x38/0x420 platform_drv_probe+0x94/0xb4 really_probe+0x134/0x300 driver_probe_device+0x68/0x100 __device_attach_driver+0x90/0xa8 bus_for_each_drv+0x84/0xcc __device_attach+0xb4/0x13c device_initial_probe+0x18/0x20 bus_probe_device+0x38/0x98 device_add+0x38c/0x420 If I understand properly we should just be able to blanket kgdb under one big RCU read lock and the problem should go away. We'll add it to the beast-of-a-function known as kgdb_cpu_enter(). With this I no longer get any splats and things seem to work fine. Signed-off-by: Douglas Anderson <dianders@chromium.org> Link: https://lore.kernel.org/r/20200602154729.v2.1.I70e0d4fd46d5ed2aaf0c98a355e8e1b7a5bb7e4e@changeid Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Sasha Levin <sashal@kernel.org>
2020-06-03 07:47:39 +09:00
rcu_read_unlock();
return kgdb_info[cpu].ret_state;
}
/*
* kgdb_handle_exception() - main entry point from a kernel exception
*
* Locking hierarchy:
* interface locks, if any (begin_session)
* kgdb lock (kgdb_active)
*/
int
kgdb_handle_exception(int evector, int signo, int ecode, struct pt_regs *regs)
{
struct kgdb_state kgdb_var;
struct kgdb_state *ks = &kgdb_var;
int ret = 0;
if (arch_kgdb_ops.enable_nmi)
arch_kgdb_ops.enable_nmi(0);
/*
* Avoid entering the debugger if we were triggered due to an oops
* but panic_timeout indicates the system should automatically
* reboot on panic. We don't want to get stuck waiting for input
* on such systems, especially if its "just" an oops.
*/
if (signo != SIGTRAP && panic_timeout)
return 1;
memset(ks, 0, sizeof(struct kgdb_state));
ks->cpu = raw_smp_processor_id();
ks->ex_vector = evector;
ks->signo = signo;
ks->err_code = ecode;
ks->linux_regs = regs;
if (kgdb_reenter_check(ks))
goto out; /* Ouch, double exception ! */
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
if (kgdb_info[ks->cpu].enter_kgdb != 0)
goto out;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
ret = kgdb_cpu_enter(ks, regs, DCPU_WANT_MASTER);
out:
if (arch_kgdb_ops.enable_nmi)
arch_kgdb_ops.enable_nmi(1);
return ret;
}
/*
* GDB places a breakpoint at this function to know dynamically loaded objects.
*/
static int module_event(struct notifier_block *self, unsigned long val,
void *data)
{
return 0;
}
static struct notifier_block dbg_module_load_nb = {
.notifier_call = module_event,
};
int kgdb_nmicallback(int cpu, void *regs)
{
#ifdef CONFIG_SMP
struct kgdb_state kgdb_var;
struct kgdb_state *ks = &kgdb_var;
kgdb: Don't round up a CPU that failed rounding up before If we're using the default implementation of kgdb_roundup_cpus() that uses smp_call_function_single_async() we can end up hanging kgdb_roundup_cpus() if we try to round up a CPU that failed to round up before. Specifically smp_call_function_single_async() will try to wait on the csd lock for the CPU that we're trying to round up. If the previous round up never finished then that lock could still be held and we'll just sit there hanging. There's not a lot of use trying to round up a CPU that failed to round up before. Let's keep a flag that indicates whether the CPU started but didn't finish to round up before. If we see that flag set then we'll skip the next round up. In general we have a few goals here: - We never want to end up calling smp_call_function_single_async() when the csd is still locked. This is accomplished because flush_smp_call_function_queue() unlocks the csd _before_ invoking the callback. That means that when kgdb_nmicallback() runs we know for sure the the csd is no longer locked. Thus when we set "rounding_up = false" we know for sure that the csd is unlocked. - If there are no timeouts rounding up we should never skip a round up. NOTE #1: In general trying to continue running after failing to round up CPUs doesn't appear to be supported in the debugger. When I simulate this I find that kdb reports "Catastrophic error detected" when I try to continue. I can overrule and continue anyway, but it should be noted that we may be entering the land of dragons here. Possibly the "Catastrophic error detected" was added _because_ of the future failure to round up, but even so this is an area of the code that hasn't been strongly tested. NOTE #2: I did a bit of testing before and after this change. I introduced a 10 second hang in the kernel while holding a spinlock that I could invoke on a certain CPU with 'taskset -c 3 cat /sys/...". Before this change if I did: - Invoke hang - Enter debugger - g (which warns about Catastrophic error, g again to go anyway) - g - Enter debugger ...I'd hang the rest of the 10 seconds without getting a debugger prompt. After this change I end up in the debugger the 2nd time after only 1 second with the standard warning about 'Timed out waiting for secondary CPUs.' I'll also note that once the CPU finished waiting I could actually debug it (aka "btc" worked) I won't promise that everything works perfectly if the errant CPU comes back at just the wrong time (like as we're entering or exiting the debugger) but it certainly seems like an improvement. NOTE #3: setting 'kgdb_info[cpu].rounding_up = false' is in kgdb_nmicallback() instead of kgdb_call_nmi_hook() because some implementations override kgdb_call_nmi_hook(). It shouldn't hurt to have it in kgdb_nmicallback() in any case. NOTE #4: this logic is really only needed because there is no API call like "smp_try_call_function_single_async()" or "smp_csd_is_locked()". If such an API existed then we'd use it instead, but it seemed a bit much to add an API like this just for kgdb. Signed-off-by: Douglas Anderson <dianders@chromium.org> Acked-by: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
2018-12-05 12:38:27 +09:00
kgdb_info[cpu].rounding_up = false;
memset(ks, 0, sizeof(struct kgdb_state));
ks->cpu = cpu;
ks->linux_regs = regs;
debug_core: refactor locking for master/slave cpus For quite some time there have been problems with memory barriers and various races with NMI on multi processor systems using the kernel debugger. The algorithm for entering the kernel debug core and resuming kernel execution was racy and had several known edge case problems with attempting to debug something on a heavily loaded system using breakpoints that are hit repeatedly and quickly. The prior "locking" design entry worked as follows: * The atomic counter kgdb_active was used with atomic exchange in order to elect a master cpu out of all the cpus that may have taken a debug exception. * The master cpu increments all elements of passive_cpu_wait[]. * The master cpu issues the round up cpus message. * Each "slave cpu" that enters the debug core increments its own element in cpu_in_kgdb[]. * Each "slave cpu" spins on passive_cpu_wait[] until it becomes 0. * The master cpu debugs the system. The new scheme removes the two arrays of atomic counters and replaces them with 2 single counters. One counter is used to count the number of cpus waiting to become a master cpu (because one or more hit an exception). The second counter is use to indicate how many cpus have entered as slave cpus. The new entry logic works as follows: * One or more cpus enters via kgdb_handle_exception() and increments the masters_in_kgdb. Each cpu attempts to get the spin lock called dbg_master_lock. * The master cpu sets kgdb_active to the current cpu. * The master cpu takes the spinlock dbg_slave_lock. * The master cpu asks to round up all the other cpus. * Each slave cpu that is not already in kgdb_handle_exception() will enter and increment slaves_in_kgdb. Each slave will now spin try_locking on dbg_slave_lock. * The master cpu waits for the sum of masters_in_kgdb and slaves_in_kgdb to be equal to the sum of the online cpus. * The master cpu debugs the system. In the new design the kgdb_active can only be changed while holding dbg_master_lock. Stress testing has not turned up any further entry/exit races that existed in the prior locking design. The prior locking design suffered from atomic variables not being truly atomic (in the capacity as used by kgdb) along with memory barrier races. Signed-off-by: Jason Wessel <jason.wessel@windriver.com> Acked-by: Dongdong Deng <dongdong.deng@windriver.com>
2010-05-21 22:46:00 +09:00
if (kgdb_info[ks->cpu].enter_kgdb == 0 &&
raw_spin_is_locked(&dbg_master_lock)) {
kgdb_cpu_enter(ks, regs, DCPU_IS_SLAVE);
return 0;
}
#endif
return 1;
}
int kgdb_nmicallin(int cpu, int trapnr, void *regs, int err_code,
atomic_t *send_ready)
{
#ifdef CONFIG_SMP
if (!kgdb_io_ready(0) || !send_ready)
return 1;
if (kgdb_info[cpu].enter_kgdb == 0) {
struct kgdb_state kgdb_var;
struct kgdb_state *ks = &kgdb_var;
memset(ks, 0, sizeof(struct kgdb_state));
ks->cpu = cpu;
ks->ex_vector = trapnr;
ks->signo = SIGTRAP;
ks->err_code = err_code;
ks->linux_regs = regs;
ks->send_ready = send_ready;
kgdb_cpu_enter(ks, regs, DCPU_WANT_MASTER);
return 0;
}
#endif
return 1;
}
static void kgdb_console_write(struct console *co, const char *s,
unsigned count)
{
unsigned long flags;
/* If we're debugging, or KGDB has not connected, don't try
* and print. */
if (!kgdb_connected || atomic_read(&kgdb_active) != -1 || dbg_kdb_mode)
return;
local_irq_save(flags);
gdbstub_msg_write(s, count);
local_irq_restore(flags);
}
static struct console kgdbcons = {
.name = "kgdb",
.write = kgdb_console_write,
.flags = CON_PRINTBUFFER | CON_ENABLED,
.index = -1,
};
static int __init opt_kgdb_con(char *str)
{
kgdb_use_con = 1;
if (kgdb_io_module_registered && !kgdb_con_registered) {
register_console(&kgdbcons);
kgdb_con_registered = 1;
}
return 0;
}
early_param("kgdbcon", opt_kgdb_con);
#ifdef CONFIG_MAGIC_SYSRQ
static void sysrq_handle_dbg(int key)
{
if (!dbg_io_ops) {
pr_crit("ERROR: No KGDB I/O module available\n");
return;
}
if (!kgdb_connected) {
#ifdef CONFIG_KGDB_KDB
if (!dbg_kdb_mode)
pr_crit("KGDB or $3#33 for KDB\n");
#else
pr_crit("Entering KGDB\n");
#endif
}
kgdb_breakpoint();
}
static struct sysrq_key_op sysrq_dbg_op = {
.handler = sysrq_handle_dbg,
.help_msg = "debug(g)",
.action_msg = "DEBUG",
};
#endif
kgdb: don't use a notifier to enter kgdb at panic; call directly Right now kgdb/kdb hooks up to debug panics by registering for the panic notifier. This works OK except that it means that kgdb/kdb gets called _after_ the CPUs in the system are taken offline. That means that if anything important was happening on those CPUs (like something that might have contributed to the panic) you can't debug them. Specifically I ran into a case where I got a panic because a task was "blocked for more than 120 seconds" which was detected on CPU 2. I nicely got shown stack traces in the kernel log for all CPUs including CPU 0, which was running 'PID: 111 Comm: kworker/0:1H' and was in the middle of __mmc_switch(). I then ended up at the kdb prompt where switched over to kgdb to try to look at local variables of the process on CPU 0. I found that I couldn't. Digging more, I found that I had no info on any tasks running on CPUs other than CPU 2 and that asking kdb for help showed me "Error: no saved data for this cpu". This was because all the CPUs were offline. Let's move the entry of kdb/kgdb to a direct call from panic() and stop using the generic notifier. Putting a direct call in allows us to order things more properly and it also doesn't seem like we're breaking any abstractions by calling into the debugger from the panic function. Daniel said: : This patch changes the way kdump and kgdb interact with each other. : However it would seem rather odd to have both tools simultaneously armed : and, even if they were, the user still has the option to use panic_timeout : to force a kdump to happen. Thus I think the change of order is : acceptable. Link: http://lkml.kernel.org/r/20190703170354.217312-1-dianders@chromium.org Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Daniel Thompson <daniel.thompson@linaro.org> Cc: Jason Wessel <jason.wessel@windriver.com> Cc: Kees Cook <keescook@chromium.org> Cc: Borislav Petkov <bp@suse.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Feng Tang <feng.tang@intel.com> Cc: YueHaibing <yuehaibing@huawei.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: "Steven Rostedt (VMware)" <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-26 08:47:45 +09:00
void kgdb_panic(const char *msg)
{
kgdb: don't use a notifier to enter kgdb at panic; call directly Right now kgdb/kdb hooks up to debug panics by registering for the panic notifier. This works OK except that it means that kgdb/kdb gets called _after_ the CPUs in the system are taken offline. That means that if anything important was happening on those CPUs (like something that might have contributed to the panic) you can't debug them. Specifically I ran into a case where I got a panic because a task was "blocked for more than 120 seconds" which was detected on CPU 2. I nicely got shown stack traces in the kernel log for all CPUs including CPU 0, which was running 'PID: 111 Comm: kworker/0:1H' and was in the middle of __mmc_switch(). I then ended up at the kdb prompt where switched over to kgdb to try to look at local variables of the process on CPU 0. I found that I couldn't. Digging more, I found that I had no info on any tasks running on CPUs other than CPU 2 and that asking kdb for help showed me "Error: no saved data for this cpu". This was because all the CPUs were offline. Let's move the entry of kdb/kgdb to a direct call from panic() and stop using the generic notifier. Putting a direct call in allows us to order things more properly and it also doesn't seem like we're breaking any abstractions by calling into the debugger from the panic function. Daniel said: : This patch changes the way kdump and kgdb interact with each other. : However it would seem rather odd to have both tools simultaneously armed : and, even if they were, the user still has the option to use panic_timeout : to force a kdump to happen. Thus I think the change of order is : acceptable. Link: http://lkml.kernel.org/r/20190703170354.217312-1-dianders@chromium.org Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Daniel Thompson <daniel.thompson@linaro.org> Cc: Jason Wessel <jason.wessel@windriver.com> Cc: Kees Cook <keescook@chromium.org> Cc: Borislav Petkov <bp@suse.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Feng Tang <feng.tang@intel.com> Cc: YueHaibing <yuehaibing@huawei.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: "Steven Rostedt (VMware)" <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-26 08:47:45 +09:00
if (!kgdb_io_module_registered)
return;
/*
kgdb: don't use a notifier to enter kgdb at panic; call directly Right now kgdb/kdb hooks up to debug panics by registering for the panic notifier. This works OK except that it means that kgdb/kdb gets called _after_ the CPUs in the system are taken offline. That means that if anything important was happening on those CPUs (like something that might have contributed to the panic) you can't debug them. Specifically I ran into a case where I got a panic because a task was "blocked for more than 120 seconds" which was detected on CPU 2. I nicely got shown stack traces in the kernel log for all CPUs including CPU 0, which was running 'PID: 111 Comm: kworker/0:1H' and was in the middle of __mmc_switch(). I then ended up at the kdb prompt where switched over to kgdb to try to look at local variables of the process on CPU 0. I found that I couldn't. Digging more, I found that I had no info on any tasks running on CPUs other than CPU 2 and that asking kdb for help showed me "Error: no saved data for this cpu". This was because all the CPUs were offline. Let's move the entry of kdb/kgdb to a direct call from panic() and stop using the generic notifier. Putting a direct call in allows us to order things more properly and it also doesn't seem like we're breaking any abstractions by calling into the debugger from the panic function. Daniel said: : This patch changes the way kdump and kgdb interact with each other. : However it would seem rather odd to have both tools simultaneously armed : and, even if they were, the user still has the option to use panic_timeout : to force a kdump to happen. Thus I think the change of order is : acceptable. Link: http://lkml.kernel.org/r/20190703170354.217312-1-dianders@chromium.org Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Daniel Thompson <daniel.thompson@linaro.org> Cc: Jason Wessel <jason.wessel@windriver.com> Cc: Kees Cook <keescook@chromium.org> Cc: Borislav Petkov <bp@suse.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Feng Tang <feng.tang@intel.com> Cc: YueHaibing <yuehaibing@huawei.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: "Steven Rostedt (VMware)" <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-26 08:47:45 +09:00
* We don't want to get stuck waiting for input from user if
* "panic_timeout" indicates the system should automatically
* reboot on panic.
*/
if (panic_timeout)
kgdb: don't use a notifier to enter kgdb at panic; call directly Right now kgdb/kdb hooks up to debug panics by registering for the panic notifier. This works OK except that it means that kgdb/kdb gets called _after_ the CPUs in the system are taken offline. That means that if anything important was happening on those CPUs (like something that might have contributed to the panic) you can't debug them. Specifically I ran into a case where I got a panic because a task was "blocked for more than 120 seconds" which was detected on CPU 2. I nicely got shown stack traces in the kernel log for all CPUs including CPU 0, which was running 'PID: 111 Comm: kworker/0:1H' and was in the middle of __mmc_switch(). I then ended up at the kdb prompt where switched over to kgdb to try to look at local variables of the process on CPU 0. I found that I couldn't. Digging more, I found that I had no info on any tasks running on CPUs other than CPU 2 and that asking kdb for help showed me "Error: no saved data for this cpu". This was because all the CPUs were offline. Let's move the entry of kdb/kgdb to a direct call from panic() and stop using the generic notifier. Putting a direct call in allows us to order things more properly and it also doesn't seem like we're breaking any abstractions by calling into the debugger from the panic function. Daniel said: : This patch changes the way kdump and kgdb interact with each other. : However it would seem rather odd to have both tools simultaneously armed : and, even if they were, the user still has the option to use panic_timeout : to force a kdump to happen. Thus I think the change of order is : acceptable. Link: http://lkml.kernel.org/r/20190703170354.217312-1-dianders@chromium.org Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Daniel Thompson <daniel.thompson@linaro.org> Cc: Jason Wessel <jason.wessel@windriver.com> Cc: Kees Cook <keescook@chromium.org> Cc: Borislav Petkov <bp@suse.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Feng Tang <feng.tang@intel.com> Cc: YueHaibing <yuehaibing@huawei.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: "Steven Rostedt (VMware)" <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-26 08:47:45 +09:00
return;
if (dbg_kdb_mode)
kgdb: don't use a notifier to enter kgdb at panic; call directly Right now kgdb/kdb hooks up to debug panics by registering for the panic notifier. This works OK except that it means that kgdb/kdb gets called _after_ the CPUs in the system are taken offline. That means that if anything important was happening on those CPUs (like something that might have contributed to the panic) you can't debug them. Specifically I ran into a case where I got a panic because a task was "blocked for more than 120 seconds" which was detected on CPU 2. I nicely got shown stack traces in the kernel log for all CPUs including CPU 0, which was running 'PID: 111 Comm: kworker/0:1H' and was in the middle of __mmc_switch(). I then ended up at the kdb prompt where switched over to kgdb to try to look at local variables of the process on CPU 0. I found that I couldn't. Digging more, I found that I had no info on any tasks running on CPUs other than CPU 2 and that asking kdb for help showed me "Error: no saved data for this cpu". This was because all the CPUs were offline. Let's move the entry of kdb/kgdb to a direct call from panic() and stop using the generic notifier. Putting a direct call in allows us to order things more properly and it also doesn't seem like we're breaking any abstractions by calling into the debugger from the panic function. Daniel said: : This patch changes the way kdump and kgdb interact with each other. : However it would seem rather odd to have both tools simultaneously armed : and, even if they were, the user still has the option to use panic_timeout : to force a kdump to happen. Thus I think the change of order is : acceptable. Link: http://lkml.kernel.org/r/20190703170354.217312-1-dianders@chromium.org Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Daniel Thompson <daniel.thompson@linaro.org> Cc: Jason Wessel <jason.wessel@windriver.com> Cc: Kees Cook <keescook@chromium.org> Cc: Borislav Petkov <bp@suse.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Feng Tang <feng.tang@intel.com> Cc: YueHaibing <yuehaibing@huawei.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: "Steven Rostedt (VMware)" <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-26 08:47:45 +09:00
kdb_printf("PANIC: %s\n", msg);
kgdb_breakpoint();
}
void __weak kgdb_arch_late(void)
{
}
void __init dbg_late_init(void)
{
dbg_is_early = false;
if (kgdb_io_module_registered)
kgdb_arch_late();
kdb_init(KDB_INIT_FULL);
}
static int
dbg_notify_reboot(struct notifier_block *this, unsigned long code, void *x)
{
/*
* Take the following action on reboot notify depending on value:
* 1 == Enter debugger
* 0 == [the default] detatch debug client
* -1 == Do nothing... and use this until the board resets
*/
switch (kgdbreboot) {
case 1:
kgdb_breakpoint();
case -1:
goto done;
}
if (!dbg_kdb_mode)
gdbstub_exit(code);
done:
return NOTIFY_DONE;
}
static struct notifier_block dbg_reboot_notifier = {
.notifier_call = dbg_notify_reboot,
.next = NULL,
.priority = INT_MAX,
};
static void kgdb_register_callbacks(void)
{
if (!kgdb_io_module_registered) {
kgdb_io_module_registered = 1;
kgdb_arch_init();
if (!dbg_is_early)
kgdb_arch_late();
register_module_notifier(&dbg_module_load_nb);
register_reboot_notifier(&dbg_reboot_notifier);
#ifdef CONFIG_MAGIC_SYSRQ
register_sysrq_key('g', &sysrq_dbg_op);
#endif
if (kgdb_use_con && !kgdb_con_registered) {
register_console(&kgdbcons);
kgdb_con_registered = 1;
}
}
}
static void kgdb_unregister_callbacks(void)
{
/*
kgdb: don't use a notifier to enter kgdb at panic; call directly Right now kgdb/kdb hooks up to debug panics by registering for the panic notifier. This works OK except that it means that kgdb/kdb gets called _after_ the CPUs in the system are taken offline. That means that if anything important was happening on those CPUs (like something that might have contributed to the panic) you can't debug them. Specifically I ran into a case where I got a panic because a task was "blocked for more than 120 seconds" which was detected on CPU 2. I nicely got shown stack traces in the kernel log for all CPUs including CPU 0, which was running 'PID: 111 Comm: kworker/0:1H' and was in the middle of __mmc_switch(). I then ended up at the kdb prompt where switched over to kgdb to try to look at local variables of the process on CPU 0. I found that I couldn't. Digging more, I found that I had no info on any tasks running on CPUs other than CPU 2 and that asking kdb for help showed me "Error: no saved data for this cpu". This was because all the CPUs were offline. Let's move the entry of kdb/kgdb to a direct call from panic() and stop using the generic notifier. Putting a direct call in allows us to order things more properly and it also doesn't seem like we're breaking any abstractions by calling into the debugger from the panic function. Daniel said: : This patch changes the way kdump and kgdb interact with each other. : However it would seem rather odd to have both tools simultaneously armed : and, even if they were, the user still has the option to use panic_timeout : to force a kdump to happen. Thus I think the change of order is : acceptable. Link: http://lkml.kernel.org/r/20190703170354.217312-1-dianders@chromium.org Signed-off-by: Douglas Anderson <dianders@chromium.org> Reviewed-by: Daniel Thompson <daniel.thompson@linaro.org> Cc: Jason Wessel <jason.wessel@windriver.com> Cc: Kees Cook <keescook@chromium.org> Cc: Borislav Petkov <bp@suse.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Feng Tang <feng.tang@intel.com> Cc: YueHaibing <yuehaibing@huawei.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: "Steven Rostedt (VMware)" <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-09-26 08:47:45 +09:00
* When this routine is called KGDB should unregister from
* handlers and clean up, making sure it is not handling any
* break exceptions at the time.
*/
if (kgdb_io_module_registered) {
kgdb_io_module_registered = 0;
unregister_reboot_notifier(&dbg_reboot_notifier);
unregister_module_notifier(&dbg_module_load_nb);
kgdb_arch_exit();
#ifdef CONFIG_MAGIC_SYSRQ
unregister_sysrq_key('g', &sysrq_dbg_op);
#endif
if (kgdb_con_registered) {
unregister_console(&kgdbcons);
kgdb_con_registered = 0;
}
}
}
/*
* There are times a tasklet needs to be used vs a compiled in
* break point so as to cause an exception outside a kgdb I/O module,
* such as is the case with kgdboe, where calling a breakpoint in the
* I/O driver itself would be fatal.
*/
static void kgdb_tasklet_bpt(unsigned long ing)
{
kgdb_breakpoint();
atomic_set(&kgdb_break_tasklet_var, 0);
}
static DECLARE_TASKLET(kgdb_tasklet_breakpoint, kgdb_tasklet_bpt, 0);
void kgdb_schedule_breakpoint(void)
{
if (atomic_read(&kgdb_break_tasklet_var) ||
atomic_read(&kgdb_active) != -1 ||
atomic_read(&kgdb_setting_breakpoint))
return;
atomic_inc(&kgdb_break_tasklet_var);
tasklet_schedule(&kgdb_tasklet_breakpoint);
}
EXPORT_SYMBOL_GPL(kgdb_schedule_breakpoint);
static void kgdb_initial_breakpoint(void)
{
kgdb_break_asap = 0;
pr_crit("Waiting for connection from remote gdb...\n");
kgdb_breakpoint();
}
/**
* kgdb_register_io_module - register KGDB IO module
* @new_dbg_io_ops: the io ops vector
*
* Register it with the KGDB core.
*/
int kgdb_register_io_module(struct kgdb_io *new_dbg_io_ops)
{
int err;
spin_lock(&kgdb_registration_lock);
if (dbg_io_ops) {
spin_unlock(&kgdb_registration_lock);
pr_err("Another I/O driver is already registered with KGDB\n");
return -EBUSY;
}
if (new_dbg_io_ops->init) {
err = new_dbg_io_ops->init();
if (err) {
spin_unlock(&kgdb_registration_lock);
return err;
}
}
dbg_io_ops = new_dbg_io_ops;
spin_unlock(&kgdb_registration_lock);
pr_info("Registered I/O driver %s\n", new_dbg_io_ops->name);
/* Arm KGDB now. */
kgdb_register_callbacks();
if (kgdb_break_asap)
kgdb_initial_breakpoint();
return 0;
}
EXPORT_SYMBOL_GPL(kgdb_register_io_module);
/**
* kkgdb_unregister_io_module - unregister KGDB IO module
* @old_dbg_io_ops: the io ops vector
*
* Unregister it with the KGDB core.
*/
void kgdb_unregister_io_module(struct kgdb_io *old_dbg_io_ops)
{
BUG_ON(kgdb_connected);
/*
* KGDB is no longer able to communicate out, so
* unregister our callbacks and reset state.
*/
kgdb_unregister_callbacks();
spin_lock(&kgdb_registration_lock);
WARN_ON_ONCE(dbg_io_ops != old_dbg_io_ops);
dbg_io_ops = NULL;
spin_unlock(&kgdb_registration_lock);
pr_info("Unregistered I/O driver %s, debugger disabled\n",
old_dbg_io_ops->name);
}
EXPORT_SYMBOL_GPL(kgdb_unregister_io_module);
int dbg_io_get_char(void)
{
int ret = dbg_io_ops->read_char();
if (ret == NO_POLL_CHAR)
return -1;
if (!dbg_kdb_mode)
return ret;
if (ret == 127)
return 8;
return ret;
}
/**
* kgdb_breakpoint - generate breakpoint exception
*
* This function will generate a breakpoint exception. It is used at the
* beginning of a program to sync up with a debugger and can be used
* otherwise as a quick means to stop program execution and "break" into
* the debugger.
*/
noinline void kgdb_breakpoint(void)
{
atomic_inc(&kgdb_setting_breakpoint);
wmb(); /* Sync point before breakpoint */
arch_kgdb_breakpoint();
wmb(); /* Sync point after breakpoint */
atomic_dec(&kgdb_setting_breakpoint);
}
EXPORT_SYMBOL_GPL(kgdb_breakpoint);
static int __init opt_kgdb_wait(char *str)
{
kgdb_break_asap = 1;
kdb_init(KDB_INIT_EARLY);
if (kgdb_io_module_registered)
kgdb_initial_breakpoint();
return 0;
}
early_param("kgdbwait", opt_kgdb_wait);