[Intel-gfx] [PATCH v3 05/11] lib/ref_tracker: __ref_tracker_dir_print improve printing
Andrzej Hajda
andrzej.hajda at intel.com
Tue Feb 22 10:27:28 UTC 2022
On 22.02.2022 01:08, Eric Dumazet wrote:
> On Mon, Feb 21, 2022 at 3:26 PM Andrzej Hajda <andrzej.hajda at intel.com> wrote:
>> To improve readibility of ref_tracker printing following changes
> readability
>
>> have been performed:
>> - reports are printed per stack_handle - log is more compact,
>> - added display name for ref_tracker_dir,
>> - stack trace is printed indented, in the same printk call,
>> - total number of references is printed every time,
>> - print info about dropped references.
>>
>> Signed-off-by: Andrzej Hajda <andrzej.hajda at intel.com>
>> ---
>> include/linux/ref_tracker.h | 15 +++++--
>> lib/ref_tracker.c | 90 ++++++++++++++++++++++++++++++++-----
>> 2 files changed, 91 insertions(+), 14 deletions(-)
>>
>> diff --git a/include/linux/ref_tracker.h b/include/linux/ref_tracker.h
>> index 3e9e9df2a41f5..a2cf1f6309adb 100644
>> --- a/include/linux/ref_tracker.h
>> +++ b/include/linux/ref_tracker.h
>> @@ -17,12 +17,19 @@ struct ref_tracker_dir {
>> bool dead;
>> struct list_head list; /* List of active trackers */
>> struct list_head quarantine; /* List of dead trackers */
>> + char name[32];
>> #endif
>> };
>>
>> #ifdef CONFIG_REF_TRACKER
>> -static inline void ref_tracker_dir_init(struct ref_tracker_dir *dir,
>> - unsigned int quarantine_count)
>> +
>> +// Temporary allow two and three arguments, until consumers are converted
>> +#define ref_tracker_dir_init(_d, _q, args...) _ref_tracker_dir_init(_d, _q, ##args, #_d)
>> +#define _ref_tracker_dir_init(_d, _q, _n, ...) __ref_tracker_dir_init(_d, _q, _n)
>> +
>> +static inline void __ref_tracker_dir_init(struct ref_tracker_dir *dir,
>> + unsigned int quarantine_count,
>> + const char *name)
>> {
>> INIT_LIST_HEAD(&dir->list);
>> INIT_LIST_HEAD(&dir->quarantine);
>> @@ -31,6 +38,7 @@ static inline void ref_tracker_dir_init(struct ref_tracker_dir *dir,
>> dir->dead = false;
>> refcount_set(&dir->untracked, 1);
>> refcount_set(&dir->no_tracker, 1);
>> + strlcpy(dir->name, name, sizeof(dir->name));
>> stack_depot_init();
>> }
>>
>> @@ -51,7 +59,8 @@ int ref_tracker_free(struct ref_tracker_dir *dir,
>> #else /* CONFIG_REF_TRACKER */
>>
>> static inline void ref_tracker_dir_init(struct ref_tracker_dir *dir,
>> - unsigned int quarantine_count)
>> + unsigned int quarantine_count,
>> + ...)
>> {
>> }
>>
>> diff --git a/lib/ref_tracker.c b/lib/ref_tracker.c
>> index 5e9f90bbf771b..ab1253fde244e 100644
>> --- a/lib/ref_tracker.c
>> +++ b/lib/ref_tracker.c
>> @@ -1,11 +1,16 @@
>> // SPDX-License-Identifier: GPL-2.0-or-later
>> +
>> +#define pr_fmt(fmt) "ref_tracker: " fmt
>> +
>> #include <linux/export.h>
>> +#include <linux/list_sort.h>
>> #include <linux/ref_tracker.h>
>> #include <linux/slab.h>
>> #include <linux/stacktrace.h>
>> #include <linux/stackdepot.h>
>>
>> #define REF_TRACKER_STACK_ENTRIES 16
>> +#define STACK_BUF_SIZE 1024
>>
>> struct ref_tracker {
>> struct list_head head; /* anchor into dir->list or dir->quarantine */
>> @@ -14,24 +19,87 @@ struct ref_tracker {
>> depot_stack_handle_t free_stack_handle;
>> };
>>
>> -void __ref_tracker_dir_print(struct ref_tracker_dir *dir,
>> - unsigned int display_limit)
>> +struct ref_tracker_dir_stats {
>> + int total;
>> + int count;
>> + struct {
>> + depot_stack_handle_t stack_handle;
>> + unsigned int count;
>> + } stacks[];
>> +};
>> +
>> +static struct ref_tracker_dir_stats *
>> +ref_tracker_get_stats(struct ref_tracker_dir *dir, unsigned int limit)
>> {
>> + struct ref_tracker_dir_stats *stats;
>> struct ref_tracker *tracker;
>> - unsigned int i = 0;
>>
>> - lockdep_assert_held(&dir->lock);
>> + stats = kmalloc(struct_size(stats, stacks, limit),
>> + GFP_NOWAIT | __GFP_NOWARN);
> I would be more comfortable if the allocation was done by the caller,
> possibly using GFP_KERNEL and evenutally kvmalloc(),
> instead of under dir->lock ?
I though also about it, but decided to left this change to another patch
as the change can be substantial and could open another discussion.
I am not sure what you mean by 'caller' but it could be even external
user of the API:
1. alloc data for ref_tracker_dir_stats.
2. take locks, if necessary.
3. gather stats (ref_tracker_get_stats) atomically.
4. release taken locks.
5. print stats.
This way, allocation and printing would happen outside locks.
>
>
>> + if (!stats)
>> + return ERR_PTR(-ENOMEM);
>> + stats->total = 0;
>> + stats->count = 0;
>>
>> list_for_each_entry(tracker, &dir->list, head) {
>> - if (i < display_limit) {
>> - pr_err("leaked reference.\n");
>> - if (tracker->alloc_stack_handle)
>> - stack_depot_print(tracker->alloc_stack_handle);
>> - i++;
>> - } else {
>> - break;
>> + depot_stack_handle_t stack = tracker->alloc_stack_handle;
>> + int i;
>> +
>> + ++stats->total;
>> + for (i = 0; i < stats->count; ++i)
>> + if (stats->stacks[i].stack_handle == stack)
>> + break;
>> + if (i >= limit)
>> + continue;
>> + if (i >= stats->count) {
>> + stats->stacks[i].stack_handle = stack;
>> + stats->stacks[i].count = 0;
>> + ++stats->count;
>> }
>> + ++stats->stacks[i].count;
>> + }
>> +
>> + return stats;
>> +}
>> +
>> +void __ref_tracker_dir_print(struct ref_tracker_dir *dir,
>> + unsigned int display_limit)
>> +{
>> + struct ref_tracker_dir_stats *stats;
>> + unsigned int i = 0, skipped;
>> + depot_stack_handle_t stack;
>> + char *sbuf;
>> +
>> + lockdep_assert_held(&dir->lock);
>> +
>> + if (list_empty(&dir->list))
>> + return;
>> +
>> + stats = ref_tracker_get_stats(dir, display_limit);
>> + if (IS_ERR(stats)) {
>> + pr_err("%s@%pK: couldn't get stats, error %pe\n",
>> + dir->name, dir, stats);
>> + return;
>> }
>> +
>> + sbuf = kmalloc(STACK_BUF_SIZE, GFP_NOWAIT | __GFP_NOWARN);
> Same remark. These allocations are most probably going to happen from process
> context, I think GFP_KERNEL is more robust.
The problem is that in my scenario it can be called under spinlock, this
is why I want almost everywhere non-sleeping allocations.
>
> This is debugging infra, it would be sad if we give up at this point,
> after storing MB of traces :)
My approach was to avoid allocations if the system is short on memory -
better to keep it alive, and we still get the report, just without
stacktraces, one can print full stats later (for example via sysfs, or
trigger to dmesg) - big chances that the bug will be still there.
If you think that is no-go, alternatives I see:
- go back to GFP_ATOMIC,
- print stack directly, without using stack_depot_snprint,
- pre-allocate buffer.
Regards
Andrzej
>
>> +
>> + for (i = 0, skipped = stats->total; i < stats->count; ++i) {
>> + stack = stats->stacks[i].stack_handle;
>> + if (sbuf && !stack_depot_snprint(stack, sbuf, STACK_BUF_SIZE, 4))
>> + sbuf[0] = 0;
>> + pr_err("%s@%pK has %d/%d users at\n%s\n", dir->name, dir,
>> + stats->stacks[i].count, stats->total, sbuf);
>> + skipped -= stats->stacks[i].count;
>> + }
>> +
>> + if (skipped)
>> + pr_err("%s@%pK skipped reports about %d/%d users.\n",
>> + dir->name, dir, skipped, stats->total);
>> +
>> + kfree(sbuf);
>> +
>> + kfree(stats);
>> }
>> EXPORT_SYMBOL(__ref_tracker_dir_print);
>>
>> --
>> 2.25.1
>>
More information about the Intel-gfx
mailing list