Serge E. Hallyn wrote:
> Quoting Oren Laadan (orenl@xxxxxxxxxxx):
>>
>> Serge E. Hallyn wrote:
>>> Quoting Matt Helsley (matthltc@xxxxxxxxxx):
>>>>> @@ -401,6 +409,9 @@ char *ckpt_generate_fmt(struct ckpt_ctx *ctx, char *fmt)
>>>>> case 'E':
>>>>> len += sprintf(format+len, "[%s]", "err %d");
>>>>> break;
>>>>> + case 'C': /* count of bytes read/written to checkpoint image */
>>>>> + len += sprintf(format+len, "[%s]", "pos %d");
>>>>> + break;
>>>> Instead we could always output ckpt->total and then we wouldn't need %(C). I
>>>> suspect it's such a useful piece of information that it'll be repeated
>>>> in many/all format strings eventually.
>>> Yes, likewise %(T). If that's what we want to do.
>> I agree. For the cases when there is not task, can put "none"
>>
>>> Should we discuss here what we want an entry to look like? For both
>>> ckpt_write_err (to the checkpoint image) and ckpt_error()?
>>>
>> Yes please !
>
> Actually %T isn't the current task, right, so it shouldn't always be prepended?
> It actually is only meaningful during checkpoint_task(), collect_objs(), and
> __tree_count_tasks?
>
> Ok, so how about:
>
> 1. ckpt_write_err() always also calls ckpt_error() (which in turn calls
> ckpt_debug). Avoid duplication which exists in several places
> right now.
> 2. We always prepend:
>
> [current->pid]:[ctx->root_pid]:[ctx->active_pid]:[ctx->errno][ctx->total]
>
> The %(X) expansions if specified come whereever they are in the fmt
> string (which is what's happening now with my patchset).
So somewhere should set ctx->errno during a checkpoint.
I suppose active_pid is for restart, but it's redundant isn't it ?
(it's always active_pid) - is it the different between top-level pid-ns
and "current" pid-ns ?
Instead of writing root_pid repeatedly, why not write sometime at the
beginning some "global" info about the checkpoint/restart ? (e.g.
the root_pid ...)
>
> Kind of long, but again this is for ckpt_error and ckpt_write_err, not for all
> ckpt_debugs().
>
>>>>> case 'O':
>>>>> len += sprintf(format+len, "[%s]", "obj %d");
>>>>> break;
>>>>> @@ -435,6 +446,51 @@ char *ckpt_generate_fmt(struct ckpt_ctx *ctx, char *fmt)
>>>>> return format;
>>>>> }
>>>>>
>>>>> +void ckpt_log_error(struct ckpt_ctx *ctx, char *fmt, ...)
>>>>> +{
>>>>> + mm_segment_t fs;
>>>>> + struct file *file;
>>>>> + int count;
>>>>> + va_list ap, aq, az;
>>>>> + char *format;
>>>>> + char buf[200], *bufp = buf;
>>>> I believe this buffer is too big for a kernel stack -- especially
>>>> for ckpt_log_error() which might be invoked "deep" in
>>>> the kernel stack.
>>> 200 bytes? Well, I guess I can try with 50 which still may often be
>>> enough.
>> How about using a dedicated buffer on @ctx for that ?
>
> I was going to do that originally, but then thought back to your
> comments about parallel checkpoint, and didn't feel like also adding
> a spinlock.
We _will_ have some sort of locking when doing a parallel checkpoint.
So when we get there either use that lock, or (what I believe is more
likely) create a per-checkpointer sub-data structure (a-la per-cpu).
Oren.
_______________________________________________
Containers mailing list
Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linux-foundation.org/mailman/listinfo/containers
