[Devel] Re: [RFC v14][PATCH 31/54] powerpc: checkpoint/restart implementation

[Oren Laadan]

Nathan Lynch wrote:
> Hello Oren,
> 
> Oren Laadan <orenl at cs.columbia.edu> writes:
> 
>> From: Nathan Lynch <ntl at pobox.com>
>>
>> Support for checkpointing and restarting GPRs, FPU state, DABR, and
>> Altivec state.
> 
> ...
> 
>> Signed-off-by: Nathan Lynch <ntl at pobox.com>
> 
> ...
> 
>> +/* dump the cpu state and registers of a given task */
>> +int checkpoint_write_cpu(struct ckpt_ctx *ctx, struct task_struct *t)
>> +{
>> +	struct ckpt_hdr_cpu *cpu_hdr;
>> +	int rc;
>> +
>> +	rc = -ENOMEM;
>> +	cpu_hdr = ckpt_hdr_get(ctx, sizeof(*cpu_hdr), CKPT_HDR_CPU);
> 
> This won't build (should be ckpt_hdr_get_type?).
> 
> I didn't write this code (I used kzalloc).
> 
> In the code I did write, I deliberately preferred the slab allocator to
> the checkpoint-specific APIs.  I do not see the advantage of using an
> arbitrarily fixed size special allocation stack that is prone to
> overflow or, worse, data corruption if someone improperly interleaves
> their gets and puts.
> 

There is a reason I insist on it: I plan to optimize c/r app downtime
by buffering data in the kernel while apps are frozen, and write-back
the output after the resume execution. To do this efficiently without
extra data copy you need something smarted than just kmalloc/kfree().

Since some API and implementation will be used later, it makes sense
to me to enforce at API requirement already. IOW, I want the code to
use ckpt_hdr_get(), ckpt_hdr_put() and ckpt_hdr_get_type().

How they are implemented, now, doesn't really matter. Your point about
using kmalloc() is correct, in particular at this stage of development.

So here's what I'll do:  I'll keep the interface requirement, and
change the implementation behind to use kmalloc/kfree().

Oren.

_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers