Re: [PATCH v12] mm: slub: move sysfs slab alloc/free interfaces to debugfs

[Faiyaz Mohammed <faiyazm@codeaurora.org>]

On 6/16/2021 9:47 PM, Vlastimil Babka wrote:
> On 6/16/21 5:50 PM, Faiyaz Mohammed wrote:
>>
>>
>> On 6/16/2021 4:35 PM, Vlastimil Babka wrote:
>>> On 6/15/21 5:58 PM, Qian Cai wrote:
>>>>
>>>>
>>>> On 6/11/2021 3:03 PM, Faiyaz Mohammed wrote:
>>>>> alloc_calls and free_calls implementation in sysfs have two issues,
>>>>> one is PAGE_SIZE limitation of sysfs and other is it does not adhere
>>>>> to "one value per file" rule.
>>>>>
>>>>> To overcome this issues, move the alloc_calls and free_calls
>>>>> implementation to debugfs.
>>>>>
>>>>> Debugfs cache will be created if SLAB_STORE_USER flag is set.
>>>>>
>>>>> Rename the alloc_calls/free_calls to alloc_traces/free_traces,
>>>>> to be inline with what it does.
>>>>>
>>>>> Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
>>>>> Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
>>>>> Signed-off-by: Faiyaz Mohammed <faiyazm@codeaurora.org>
>>>>
>>>> Reverting this commit on today's linux-next fixed all leaks (hundreds) reported by kmemleak like below,
>>>>
>>>> unreferenced object 0xffff00091ae1b540 (size 64):
>>>>   comm "lsbug", pid 1607, jiffies 4294958291 (age 1476.340s)
>>>>   hex dump (first 32 bytes):
>>>>     02 00 00 00 00 00 00 00 6b 6b 6b 6b 6b 6b 6b 6b  ........kkkkkkkk
>>>>     6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b  kkkkkkkkkkkkkkkk
>>>>   backtrace:
>>>>     [<ffff8000106b06b8>] slab_post_alloc_hook+0xa0/0x418
>>>>     [<ffff8000106b5c7c>] kmem_cache_alloc_trace+0x1e4/0x378
>>>>     [<ffff8000106b5e40>] slab_debugfs_start+0x30/0x50
>>>>     slab_debugfs_start at /usr/src/linux-next/mm/slub.c:5831
>>>>     [<ffff8000107b3dbc>] seq_read_iter+0x214/0xd50
>>>>     [<ffff8000107b4b84>] seq_read+0x28c/0x418
>>>>     [<ffff8000109560b4>] full_proxy_read+0xdc/0x148
>>>>     [<ffff800010738f24>] vfs_read+0x104/0x340
>>>>     [<ffff800010739ee0>] ksys_read+0xf8/0x1e0
>>>>     [<ffff80001073a03c>] __arm64_sys_read+0x74/0xa8
>>>>     [<ffff8000100358d4>] invoke_syscall.constprop.0+0xdc/0x1d8
>>>>     [<ffff800010035ab4>] do_el0_svc+0xe4/0x298
>>>>     [<ffff800011138528>] el0_svc+0x20/0x30
>>>>     [<ffff800011138b08>] el0t_64_sync_handler+0xb0/0xb8
>>>>     [<ffff80001001259c>] el0t_64_sync+0x178/0x17c
>>>>
>>>
>>> I think the problem is here:
>>>
>>>>> +static void slab_debugfs_stop(struct seq_file *seq, void *v)
>>>>> +{
>>>>> +	kfree(v);
>>>>> +}
>>>>> +
>>>>> +static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
>>>>> +{
>>>>> +	loff_t *spos = v;
>>>>> +	struct loc_track *t = seq->private;
>>>>> +
>>>>> +	if (*ppos < t->count) {
>>>>> +		*ppos = ++*spos;
>>>>> +		return spos;
>>>>> +	}
>>>>> +	*ppos = ++*spos;
>>>>> +	return NULL;
>>>>> +}
>>>
>>> If we return NULL, then NULL is passed to slab_debugfs_stop and thus we don't
>>> kfree ppos. kfree(NULL) is silently ignored.
>>>
>> I think yes, if NULL passed to kfree, it simply do return.
>>> I think as we have private struct loc_track, we can add a pos field there and
>>> avoid the kmaloc/kfree altogether.
>>>
>> Hmm, yes we can add pos field "or" we can use argument "v" mean we can
>> update v with pos in ->next() and use in ->show() to avoid the leak
>> (kmalloc/kfree).
> 
> Can you explain the "or" part more. It's exactly what we already do, no?I am thinking if we simplly do ppos return from slab_debugfs_start() and
in slab_debugfs_next() assign ppos to "v", update it and return if
records are there. something like below (approach 1):
...
static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
{
...
        v = ppos;
        if (*ppos < t->count) {
                 ++*ppos;
                return v;
        }

        ++*ppos;
        return NULL;
}
...
static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
{
        return ppos;
}
...

> "v" as you said. The problem is, if next(); returns NULL, then stop() gets the
> NULL as "v". It's just what I see in the code of seq_read_iter() and traverse()
> in fs/seq_file.c. I don't see another way to say there are no more records to
> print - only to return NULL in next().
> Ah, ok so we could maybe do the kfree() in next() then before returning NULL,
> which is the last moment we have the pointer. But really, if we already have a
> loc_track in private, why kmalloc an additional loff_t.
> 
Yes, we can do kfree() before returning NULL, but better to add ppos in
lock_track. (approach 2)

> Anyway it seems to me also that
> Documentation/filesystems/seq_file.rst should be updated, as the kfree() in
> stop() is exactly what it suggests, and it doesn't show how next() indicates
> that there are no more records by returning NULL, and what to do about kfree() then.

Can you please suggest me which approach would be good to avoid the
leak?. I will update in next patch version.