Recently we ran into a oom issue, kernel panic due to no killable process.
The dmesg shows huge unreclaimable slabs used almost 100% memory, but kdump doesn't capture vmcore due to some reason.
So, it may sound better to capture unreclaimable slab info in oom message when kernel panic to aid trouble shooting and cover the corner case.
Since kernel already panic, so capturing more information sounds worthy and doesn't bother normal oom killer.
With the patchset, tools/vm/slabinfo has a new option, "-U", to show unreclaimable slab only.
And, oom will print all non zero (num_objs * size != 0) unreclaimable slabs in oom killer message.
For details, please see the commit log for each commit.
Changelog v7 —> v8:
* Adopted Michal’s suggestion to dump unreclaim slab info when unreclaimable slabs amount > total user memory. Not only in oom panic path.
Changelog v6 -> v7:
* Added unreclaim_slabs_oom_ratio proc knob, unreclaimable slabs info will be dumped when unreclaimable slabs amount : all user memory > the ratio
Changelog v5 —> v6:
* Fixed a checkpatch.pl warning for patch #2
Changelog v4 —> v5:
* Solved the comments from David
* Build test SLABINFO = n
Changelog v3 —> v4:
* Solved the comments from David
* Added David’s Acked-by in patch 1
Changelog v2 —> v3:
* Show used size and total size of each kmem cache per David’s comment
Changelog v1 —> v2:
* Removed the original patch 1 (“mm: slab: output reclaimable flag in /proc/slabinfo”) since Christoph suggested it might break the compatibility and /proc/slabinfo is legacy
* Added Christoph’s Acked-by
* Removed acquiring slab_mutex per Tetsuo’s comment
Yang Shi (2):
tools: slabinfo: add "-U" option to show unreclaimable slabs only
mm: oom: show unreclaimable slab info when unreclaimable slabs > user memory
mm/oom_kill.c | 22 ++++++++++++++++++++++
mm/slab.h | 8 ++++++++
mm/slab_common.c | 29 +++++++++++++++++++++++++++++
tools/vm/slabinfo.c | 11 ++++++++++-
4 files changed, 69 insertions(+), 1 deletion(-)
Kernel may panic when oom happens without killable process sometimes it
is caused by huge unreclaimable slabs used by kernel.
Although kdump could help debug such problem, however, kdump is not
available on all architectures and it might be malfunction sometime.
And, since kernel already panic it is worthy capturing such information
in dmesg to aid touble shooting.
Print out unreclaimable slab info (used size and total size) which
actual memory usage is not zero (num_objs * size != 0) when
unreclaimable slabs amount is greater than total user memory (LRU
pages).
The output looks like:
Unreclaimable slab info:
Name Used Total
rpc_buffers 31KB 31KB
rpc_tasks 7KB 7KB
ebitmap_node 1964KB 1964KB
avtab_node 5024KB 5024KB
xfs_buf 1402KB 1402KB
xfs_ili 134KB 134KB
xfs_efi_item 115KB 115KB
xfs_efd_item 115KB 115KB
xfs_buf_item 134KB 134KB
xfs_log_item_desc 342KB 342KB
xfs_trans 1412KB 1412KB
xfs_ifork 212KB 212KB
Signed-off-by: Yang Shi <[email protected]>
---
mm/oom_kill.c | 22 ++++++++++++++++++++++
mm/slab.h | 8 ++++++++
mm/slab_common.c | 29 +++++++++++++++++++++++++++++
3 files changed, 59 insertions(+)
diff --git a/mm/oom_kill.c b/mm/oom_kill.c
index 99736e0..6d89397 100644
--- a/mm/oom_kill.c
+++ b/mm/oom_kill.c
@@ -43,6 +43,7 @@
#include <asm/tlb.h>
#include "internal.h"
+#include "slab.h"
#define CREATE_TRACE_POINTS
#include <trace/events/oom.h>
@@ -160,6 +161,25 @@ static bool oom_unkillable_task(struct task_struct *p,
return false;
}
+/*
+ * Print out unreclaimble slabs info when unreclaimable slabs amount is greater
+ * than all user memory (LRU pages)
+ */
+static bool is_dump_unreclaim_slabs(void)
+{
+ unsigned long nr_lru;
+
+ nr_lru = global_node_page_state(NR_ACTIVE_ANON) +
+ global_node_page_state(NR_INACTIVE_ANON) +
+ global_node_page_state(NR_ACTIVE_FILE) +
+ global_node_page_state(NR_INACTIVE_FILE) +
+ global_node_page_state(NR_ISOLATED_ANON) +
+ global_node_page_state(NR_ISOLATED_FILE) +
+ global_node_page_state(NR_UNEVICTABLE);
+
+ return (global_node_page_state(NR_SLAB_UNRECLAIMABLE) > nr_lru);
+}
+
/**
* oom_badness - heuristic function to determine which candidate task to kill
* @p: task struct of which task we should calculate
@@ -423,6 +443,8 @@ static void dump_header(struct oom_control *oc, struct task_struct *p)
mem_cgroup_print_oom_info(oc->memcg, p);
else
show_mem(SHOW_MEM_FILTER_NODES, oc->nodemask);
+ if (is_dump_unreclaim_slabs())
+ dump_unreclaimable_slab();
if (sysctl_oom_dump_tasks)
dump_tasks(oc->memcg, oc->nodemask);
}
diff --git a/mm/slab.h b/mm/slab.h
index 0733628..b0496d1 100644
--- a/mm/slab.h
+++ b/mm/slab.h
@@ -505,6 +505,14 @@ static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
void memcg_slab_stop(struct seq_file *m, void *p);
int memcg_slab_show(struct seq_file *m, void *p);
+#ifdef CONFIG_SLABINFO
+void dump_unreclaimable_slab(void);
+#else
+static inline void dump_unreclaimable_slab(void)
+{
+}
+#endif
+
void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr);
#ifdef CONFIG_SLAB_FREELIST_RANDOM
diff --git a/mm/slab_common.c b/mm/slab_common.c
index 904a83b..d08213d 100644
--- a/mm/slab_common.c
+++ b/mm/slab_common.c
@@ -1272,6 +1272,35 @@ static int slab_show(struct seq_file *m, void *p)
return 0;
}
+void dump_unreclaimable_slab(void)
+{
+ struct kmem_cache *s, *s2;
+ struct slabinfo sinfo;
+
+ pr_info("Unreclaimable slab info:\n");
+ pr_info("Name Used Total\n");
+
+ /*
+ * Here acquiring slab_mutex is unnecessary since we don't prefer to
+ * get sleep in oom path right before kernel panic, and avoid race
+ * condition.
+ * Since it is already oom, so there should be not any big allocation
+ * which could change the statistics significantly.
+ */
+ list_for_each_entry_safe(s, s2, &slab_caches, list) {
+ if (!is_root_cache(s) || (s->flags & SLAB_RECLAIM_ACCOUNT))
+ continue;
+
+ memset(&sinfo, 0, sizeof(sinfo));
+ get_slabinfo(s, &sinfo);
+
+ if (sinfo.num_objs > 0)
+ pr_info("%-17s %10luKB %10luKB\n", cache_name(s),
+ (sinfo.active_objs * s->size) / 1024,
+ (sinfo.num_objs * s->size) / 1024);
+ }
+}
+
#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
void *memcg_slab_start(struct seq_file *m, loff_t *pos)
{
--
1.8.3.1
Add "-U" option to show unreclaimable slabs only.
"-U" and "-S" together can tell us what unreclaimable slabs use the most
memory to help debug huge unreclaimable slabs issue.
Signed-off-by: Yang Shi <[email protected]>
Acked-by: Christoph Lameter <[email protected]>
Acked-by: David Rientjes <[email protected]>
---
tools/vm/slabinfo.c | 11 ++++++++++-
1 file changed, 10 insertions(+), 1 deletion(-)
diff --git a/tools/vm/slabinfo.c b/tools/vm/slabinfo.c
index b9d34b3..de8fa11 100644
--- a/tools/vm/slabinfo.c
+++ b/tools/vm/slabinfo.c
@@ -83,6 +83,7 @@ struct aliasinfo {
int sort_loss;
int extended_totals;
int show_bytes;
+int unreclaim_only;
/* Debug options */
int sanity;
@@ -132,6 +133,7 @@ static void usage(void)
"-L|--Loss Sort by loss\n"
"-X|--Xtotals Show extended summary information\n"
"-B|--Bytes Show size in bytes\n"
+ "-U|--Unreclaim Show unreclaimable slabs only\n"
"\nValid debug options (FZPUT may be combined)\n"
"a / A Switch on all debug options (=FZUP)\n"
"- Switch off all debug options\n"
@@ -568,6 +570,9 @@ static void slabcache(struct slabinfo *s)
if (strcmp(s->name, "*") == 0)
return;
+ if (unreclaim_only && s->reclaim_account)
+ return;
+
if (actual_slabs == 1) {
report(s);
return;
@@ -1346,6 +1351,7 @@ struct option opts[] = {
{ "Loss", no_argument, NULL, 'L'},
{ "Xtotals", no_argument, NULL, 'X'},
{ "Bytes", no_argument, NULL, 'B'},
+ { "Unreclaim", no_argument, NULL, 'U'},
{ NULL, 0, NULL, 0 }
};
@@ -1357,7 +1363,7 @@ int main(int argc, char *argv[])
page_size = getpagesize();
- while ((c = getopt_long(argc, argv, "aAd::Defhil1noprstvzTSN:LXB",
+ while ((c = getopt_long(argc, argv, "aAd::Defhil1noprstvzTSN:LXBU",
opts, NULL)) != -1)
switch (c) {
case '1':
@@ -1438,6 +1444,9 @@ int main(int argc, char *argv[])
case 'B':
show_bytes = 1;
break;
+ case 'U':
+ unreclaim_only = 1;
+ break;
default:
fatal("%s: Invalid option '%c'\n", argv[0], optopt);
--
1.8.3.1
On 2017/09/28 6:46, Yang Shi wrote:
> Changelog v7 —> v8:
> * Adopted Michal’s suggestion to dump unreclaim slab info when unreclaimable slabs amount > total user memory. Not only in oom panic path.
Holding slab_mutex inside dump_unreclaimable_slab() was refrained since V2
because there are
mutex_lock(&slab_mutex);
kmalloc(GFP_KERNEL);
mutex_unlock(&slab_mutex);
users. If we call dump_unreclaimable_slab() for non OOM panic path, aren't we
introducing a risk of crash (i.e. kernel panic) for regular OOM path?
We can try mutex_trylock() from dump_unreclaimable_slab() at best.
But it is still remaining unsafe, isn't it?
On 9/27/17 9:36 PM, Tetsuo Handa wrote:
> On 2017/09/28 6:46, Yang Shi wrote:
>> Changelog v7 —> v8:
>> * Adopted Michal’s suggestion to dump unreclaim slab info when unreclaimable slabs amount > total user memory. Not only in oom panic path.
>
> Holding slab_mutex inside dump_unreclaimable_slab() was refrained since V2
> because there are
>
> mutex_lock(&slab_mutex);
> kmalloc(GFP_KERNEL);
> mutex_unlock(&slab_mutex);
>
> users. If we call dump_unreclaimable_slab() for non OOM panic path, aren't we
> introducing a risk of crash (i.e. kernel panic) for regular OOM path?
I don't see the difference between regular oom path and oom path other
than calling panic() at last.
And, the slab dump may be called by panic path too, it is for both
regular and panic path.
Thanks,
Yang
>
> We can try mutex_trylock() from dump_unreclaimable_slab() at best.
> But it is still remaining unsafe, isn't it?
>
Yang Shi wrote:
> On 9/27/17 9:36 PM, Tetsuo Handa wrote:
> > On 2017/09/28 6:46, Yang Shi wrote:
> >> Changelog v7 -> v8:
> >> * Adopted Michal$B!G(Bs suggestion to dump unreclaim slab info when unreclaimable slabs amount > total user memory. Not only in oom panic path.
> >
> > Holding slab_mutex inside dump_unreclaimable_slab() was refrained since V2
> > because there are
> >
> > mutex_lock(&slab_mutex);
> > kmalloc(GFP_KERNEL);
> > mutex_unlock(&slab_mutex);
> >
> > users. If we call dump_unreclaimable_slab() for non OOM panic path, aren't we
> > introducing a risk of crash (i.e. kernel panic) for regular OOM path?
>
> I don't see the difference between regular oom path and oom path other
> than calling panic() at last.
>
> And, the slab dump may be called by panic path too, it is for both
> regular and panic path.
Calling a function that might cause kerneloops immediately before calling panic()
would be tolerable, for the kernel will panic after all. But calling a function
that might cause kerneloops when there is no plan to call panic() is a bug.
>
> Thanks,
> Yang
>
> >
> > We can try mutex_trylock() from dump_unreclaimable_slab() at best.
> > But it is still remaining unsafe, isn't it?
> >
>
On 9/28/17 12:57 PM, Tetsuo Handa wrote:
> Yang Shi wrote:
>> On 9/27/17 9:36 PM, Tetsuo Handa wrote:
>>> On 2017/09/28 6:46, Yang Shi wrote:
>>>> Changelog v7 -> v8:
>>>> * Adopted Michal$B!G(Bs suggestion to dump unreclaim slab info when unreclaimable slabs amount > total user memory. Not only in oom panic path.
>>>
>>> Holding slab_mutex inside dump_unreclaimable_slab() was refrained since V2
>>> because there are
>>>
>>> mutex_lock(&slab_mutex);
>>> kmalloc(GFP_KERNEL);
>>> mutex_unlock(&slab_mutex);
>>>
>>> users. If we call dump_unreclaimable_slab() for non OOM panic path, aren't we
>>> introducing a risk of crash (i.e. kernel panic) for regular OOM path?
>>
>> I don't see the difference between regular oom path and oom path other
>> than calling panic() at last.
>>
>> And, the slab dump may be called by panic path too, it is for both
>> regular and panic path.
>
> Calling a function that might cause kerneloops immediately before calling panic()
> would be tolerable, for the kernel will panic after all. But calling a function
> that might cause kerneloops when there is no plan to call panic() is a bug.
I got your point. slab_mutex is used to protect the list of all the
slabs, since we are already in oom, there should be not kmem cache
destroy happen during the list traverse. And, list_for_each_entry() has
been replaced to list_for_each_entry_safe() to make the traverse more
robust.
Thanks,
Yang
>
>>
>> Thanks,
>> Yang
>>
>>>
>>> We can try mutex_trylock() from dump_unreclaimable_slab() at best.
>>> But it is still remaining unsafe, isn't it?
>>>
>>
Yang Shi wrote:
> On 9/28/17 12:57 PM, Tetsuo Handa wrote:
> > Yang Shi wrote:
> >> On 9/27/17 9:36 PM, Tetsuo Handa wrote:
> >>> On 2017/09/28 6:46, Yang Shi wrote:
> >>>> Changelog v7 -> v8:
> >>>> * Adopted Michal$B!G(Bs suggestion to dump unreclaim slab info when unreclaimable slabs amount > total user memory. Not only in oom panic path.
> >>>
> >>> Holding slab_mutex inside dump_unreclaimable_slab() was refrained since V2
> >>> because there are
> >>>
> >>> mutex_lock(&slab_mutex);
> >>> kmalloc(GFP_KERNEL);
> >>> mutex_unlock(&slab_mutex);
> >>>
> >>> users. If we call dump_unreclaimable_slab() for non OOM panic path, aren't we
> >>> introducing a risk of crash (i.e. kernel panic) for regular OOM path?
> >>
> >> I don't see the difference between regular oom path and oom path other
> >> than calling panic() at last.
> >>
> >> And, the slab dump may be called by panic path too, it is for both
> >> regular and panic path.
> >
> > Calling a function that might cause kerneloops immediately before calling panic()
> > would be tolerable, for the kernel will panic after all. But calling a function
> > that might cause kerneloops when there is no plan to call panic() is a bug.
>
> I got your point. slab_mutex is used to protect the list of all the
> slabs, since we are already in oom, there should be not kmem cache
> destroy happen during the list traverse. And, list_for_each_entry() has
> been replaced to list_for_each_entry_safe() to make the traverse more
> robust.
I consider that OOM event and kmem chache destroy event can run concurrently
because slab_mutex is not held by OOM event (and unfortunately cannot be held
due to possibility of deadlock) in order to protect the list of all the slabs.
I don't think replacing list_for_each_entry() with list_for_each_entry_safe()
makes the traverse more robust, for list_for_each_entry_safe() does not defer
freeing of memory used by list element. Rather, replacing list_for_each_entry()
with list_for_each_entry_rcu() (and making relevant changes such as
rcu_read_lock()/rcu_read_unlock()/synchronize_rcu()) will make the traverse safe.
On 9/28/17 1:45 PM, Tetsuo Handa wrote:
> Yang Shi wrote:
>> On 9/28/17 12:57 PM, Tetsuo Handa wrote:
>>> Yang Shi wrote:
>>>> On 9/27/17 9:36 PM, Tetsuo Handa wrote:
>>>>> On 2017/09/28 6:46, Yang Shi wrote:
>>>>>> Changelog v7 -> v8:
>>>>>> * Adopted Michal$B!G(Bs suggestion to dump unreclaim slab info when unreclaimable slabs amount > total user memory. Not only in oom panic path.
>>>>>
>>>>> Holding slab_mutex inside dump_unreclaimable_slab() was refrained since V2
>>>>> because there are
>>>>>
>>>>> mutex_lock(&slab_mutex);
>>>>> kmalloc(GFP_KERNEL);
>>>>> mutex_unlock(&slab_mutex);
>>>>>
>>>>> users. If we call dump_unreclaimable_slab() for non OOM panic path, aren't we
>>>>> introducing a risk of crash (i.e. kernel panic) for regular OOM path?
>>>>
>>>> I don't see the difference between regular oom path and oom path other
>>>> than calling panic() at last.
>>>>
>>>> And, the slab dump may be called by panic path too, it is for both
>>>> regular and panic path.
>>>
>>> Calling a function that might cause kerneloops immediately before calling panic()
>>> would be tolerable, for the kernel will panic after all. But calling a function
>>> that might cause kerneloops when there is no plan to call panic() is a bug.
>>
>> I got your point. slab_mutex is used to protect the list of all the
>> slabs, since we are already in oom, there should be not kmem cache
>> destroy happen during the list traverse. And, list_for_each_entry() has
>> been replaced to list_for_each_entry_safe() to make the traverse more
>> robust.
>
> I consider that OOM event and kmem chache destroy event can run concurrently
> because slab_mutex is not held by OOM event (and unfortunately cannot be held
> due to possibility of deadlock) in order to protect the list of all the slabs.
>
> I don't think replacing list_for_each_entry() with list_for_each_entry_safe()
> makes the traverse more robust, for list_for_each_entry_safe() does not defer
> freeing of memory used by list element. Rather, replacing list_for_each_entry()
> with list_for_each_entry_rcu() (and making relevant changes such as
> rcu_read_lock()/rcu_read_unlock()/synchronize_rcu()) will make the traverse safe.
I'm not sure if rcu could satisfy this case. rcu just can protect
slab_caches_to_rcu_destroy list, which is used by SLAB_TYPESAFE_BY_RCU
slabs.
Yang
>
Yang Shi wrote:
> On 9/28/17 1:45 PM, Tetsuo Handa wrote:
> > Yang Shi wrote:
> >> On 9/28/17 12:57 PM, Tetsuo Handa wrote:
> >>> Yang Shi wrote:
> >>>> On 9/27/17 9:36 PM, Tetsuo Handa wrote:
> >>>>> On 2017/09/28 6:46, Yang Shi wrote:
> >>>>>> Changelog v7 -> v8:
> >>>>>> * Adopted Michal$B!G(Bs suggestion to dump unreclaim slab info when unreclaimable slabs amount > total user memory. Not only in oom panic path.
> >>>>>
> >>>>> Holding slab_mutex inside dump_unreclaimable_slab() was refrained since V2
> >>>>> because there are
> >>>>>
> >>>>> mutex_lock(&slab_mutex);
> >>>>> kmalloc(GFP_KERNEL);
> >>>>> mutex_unlock(&slab_mutex);
> >>>>>
> >>>>> users. If we call dump_unreclaimable_slab() for non OOM panic path, aren't we
> >>>>> introducing a risk of crash (i.e. kernel panic) for regular OOM path?
> >>>>
> >>>> I don't see the difference between regular oom path and oom path other
> >>>> than calling panic() at last.
> >>>>
> >>>> And, the slab dump may be called by panic path too, it is for both
> >>>> regular and panic path.
> >>>
> >>> Calling a function that might cause kerneloops immediately before calling panic()
> >>> would be tolerable, for the kernel will panic after all. But calling a function
> >>> that might cause kerneloops when there is no plan to call panic() is a bug.
> >>
> >> I got your point. slab_mutex is used to protect the list of all the
> >> slabs, since we are already in oom, there should be not kmem cache
> >> destroy happen during the list traverse. And, list_for_each_entry() has
> >> been replaced to list_for_each_entry_safe() to make the traverse more
> >> robust.
> >
> > I consider that OOM event and kmem chache destroy event can run concurrently
> > because slab_mutex is not held by OOM event (and unfortunately cannot be held
> > due to possibility of deadlock) in order to protect the list of all the slabs.
> >
> > I don't think replacing list_for_each_entry() with list_for_each_entry_safe()
> > makes the traverse more robust, for list_for_each_entry_safe() does not defer
> > freeing of memory used by list element. Rather, replacing list_for_each_entry()
> > with list_for_each_entry_rcu() (and making relevant changes such as
> > rcu_read_lock()/rcu_read_unlock()/synchronize_rcu()) will make the traverse safe.
>
> I'm not sure if rcu could satisfy this case. rcu just can protect
> slab_caches_to_rcu_destroy list, which is used by SLAB_TYPESAFE_BY_RCU
> slabs.
I'm not sure why you are talking about SLAB_TYPESAFE_BY_RCU.
What I meant is that
Upon registration:
// do initialize/setup stuff here
synchronize_rcu(); // <= for dump_unreclaimable_slab()
list_add_rcu(&kmem_cache->list, &slab_caches);
Upon unregistration:
list_del_rcu(&kmem_cache->list);
synchronize_rcu(); // <= for dump_unreclaimable_slab()
// do finalize/cleanup stuff here
then (if my understanding is correct)
rcu_read_lock();
list_for_each_entry_rcu(s, &slab_caches, list) {
if (!is_root_cache(s) || (s->flags & SLAB_RECLAIM_ACCOUNT))
continue;
memset(&sinfo, 0, sizeof(sinfo));
get_slabinfo(s, &sinfo);
if (sinfo.num_objs > 0)
pr_info("%-17s %10luKB %10luKB\n", cache_name(s),
(sinfo.active_objs * s->size) / 1024,
(sinfo.num_objs * s->size) / 1024);
}
rcu_read_unlock();
will make dump_unreclaimable_slab() safe.