[PATCH] x86/mce: Dynamically size space for machine check records

[Naik, Avadhut]

Hi,

On 2/28/2024 17:14, Tony Luck wrote:
> Systems with a large number of CPUs may generate a large
> number of machine check records when things go seriously
> wrong. But Linux has a fixed buffer that can only capture
> a few dozen errors.
> 
> Allocate space based on the number of CPUs (with a minimum
> value based on the historical fixed buffer that could store
> 80 records).
> 
> Signed-off-by: Tony Luck <tony.luck@xxxxxxxxx>
> ---
> 
> Discussion earlier concluded with the realization that it is
> safe to dynamically allocate the mce_evt_pool at boot time.
> So here's a patch to do that. Scaling algorithm here is a
> simple linear "4 records per possible CPU" with a minimum
> of 80 to match the legacy behavior. I'm open to other
> suggestions.
> 
> Note that I threw in a "+1" to the return from ilog2() when
> calling gen_pool_create(). From reading code, and running
> some tests, it appears that the min_alloc_order argument
> needs to be large enough to allocate one of the mce_evt_llist
> structures.
> 
> Some other gen_pool users in Linux may also need this "+1".
> 
>  arch/x86/kernel/cpu/mce/genpool.c | 22 ++++++++++++++++------
>  1 file changed, 16 insertions(+), 6 deletions(-)
> 
> diff --git a/arch/x86/kernel/cpu/mce/genpool.c b/arch/x86/kernel/cpu/mce/genpool.c
> index fbe8b61c3413..a1f0a8f29cf5 100644
> --- a/arch/x86/kernel/cpu/mce/genpool.c
> +++ b/arch/x86/kernel/cpu/mce/genpool.c
> @@ -16,14 +16,13 @@
>   * used to save error information organized in a lock-less list.
>   *
>   * This memory pool is only to be used to save MCE records in MCE context.
> - * MCE events are rare, so a fixed size memory pool should be enough. Use
> - * 2 pages to save MCE events for now (~80 MCE records at most).
> + * MCE events are rare, so a fixed size memory pool should be enough.
> + * Allocate on a sliding scale based on number of CPUs.
>   */
> -#define MCE_POOLSZ	(2 * PAGE_SIZE)
> +#define MCE_MIN_ENTRIES	80
>  
>  static struct gen_pool *mce_evt_pool;
>  static LLIST_HEAD(mce_event_llist);
> -static char gen_pool_buf[MCE_POOLSZ];
>  
>  /*
>   * Compare the record "t" with each of the records on list "l" to see if
> @@ -118,14 +117,25 @@ int mce_gen_pool_add(struct mce *mce)
>  
>  static int mce_gen_pool_create(void)
>  {
> +	int mce_numrecords, mce_poolsz;
>  	struct gen_pool *tmpp;
>  	int ret = -ENOMEM;
> +	void *mce_pool;
> +	int order;
>  
> -	tmpp = gen_pool_create(ilog2(sizeof(struct mce_evt_llist)), -1);
> +	order = ilog2(sizeof(struct mce_evt_llist)) + 1;
> +	tmpp = gen_pool_create(order, -1);
>  	if (!tmpp)
>  		goto out;
>  
> -	ret = gen_pool_add(tmpp, (unsigned long)gen_pool_buf, MCE_POOLSZ, -1);
> +	mce_numrecords = max(80, num_possible_cpus() * 4);

Per Boris's below suggestion, shouldn't this be:
	mce_numrecords = max(80, num_possible_cpus() * 16);

>> 	min(4*PAGE_SIZE, num_possible_cpus() * PAGE_SIZE);
> 
> max() ofc.
> 
>> There's a sane minimum and one page pro logical CPU should be fine on
>> pretty much every configuration...

4 MCE records per CPU equates to 1024 bytes, considering the genpool intrinsic
behavior you explained in the other subthread.

Apart from this, tested the patch on a couple of AMD systems. Didn't observe any
issues.

> +	mce_poolsz = mce_numrecords * (1 << order);
> +	mce_pool = kmalloc(mce_poolsz, GFP_KERNEL);
> +	if (!mce_pool) {
> +		gen_pool_destroy(tmpp);
> +		goto out;
> +	}
> +	ret = gen_pool_add(tmpp, (unsigned long)mce_pool, mce_poolsz, -1);
>  	if (ret) {
>  		gen_pool_destroy(tmpp);
>  		goto out;

-- 
Thanks,
Avadhut Naik