Re: [PATCH v5 09/16] kexec: enable KHO support for memory preservation

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Apr 03, 2025 at 03:50:04PM +0000, Pratyush Yadav wrote:

> The patch currently has a limitation where it does not free any of the
> empty tables after a unpreserve operation. But Changyuan's patch also
> doesn't do it so at least it is not any worse off.

We do we even have unpreserve? Just discard the entire KHO operation
in a bulk.

> When working on this patch, I realized that kho_mem_deserialize() is
> currently _very_ slow. It takes over 2 seconds to make memblock
> reservations for 48 GiB of 0-order pages. I suppose this can later be
> optimized by teaching memblock_free_all() to skip preserved pages
> instead of making memblock reservations.

Yes, this was my prior point of not having actual data to know what
the actual hot spots are.. This saves a few ms on an operation that
takes over 2 seconds :)

> +typedef unsigned long khomem_desc_t;

This should be more like:

union {
      void *table;
      phys_addr_t table_phys;
};

Since we are not using the low bits right now and it is alot cheaper
to convert from va to phys only once during the final step. __va is
not exactly fast.

> +#define PTRS_PER_LEVEL		(PAGE_SIZE / sizeof(unsigned long))
> +#define KHOMEM_L1_BITS		(PAGE_SIZE * BITS_PER_BYTE)
> +#define KHOMEM_L1_MASK		((1 << ilog2(KHOMEM_L1_BITS)) - 1)
> +#define KHOMEM_L1_SHIFT		(PAGE_SHIFT)
> +#define KHOMEM_L2_SHIFT		(KHOMEM_L1_SHIFT + ilog2(KHOMEM_L1_BITS))
> +#define KHOMEM_L3_SHIFT		(KHOMEM_L2_SHIFT + ilog2(PTRS_PER_LEVEL))
> +#define KHOMEM_L4_SHIFT		(KHOMEM_L3_SHIFT + ilog2(PTRS_PER_LEVEL))
> +#define KHOMEM_PFN_MASK		PAGE_MASK

This all works better if you just use GENMASK and FIELD_GET

> +static int __khomem_table_alloc(khomem_desc_t *desc)
> +{
> +	if (khomem_desc_none(*desc)) {

Needs READ_ONCE

> +struct kho_mem_track {
> +	/* Points to L4 KHOMEM descriptor, each order gets its own table. */
> +	struct xarray orders;
> +};

I think it would be easy to add a 5th level and just use bits 63:57 as
a 6 bit order. Then you don't need all this stuff either.

> +int kho_preserve_folio(struct folio *folio)
> +{
> +	unsigned long pfn = folio_pfn(folio);
> +	unsigned int order = folio_order(folio);
> +	int err;
> +
> +	if (!kho_enable)
> +		return -EOPNOTSUPP;
> +
> +	down_read(&kho_out.tree_lock);

This lock still needs to go away

> +static void kho_mem_serialize(void)
> +{
> +	struct kho_mem_track *tracker = &kho_mem_track;
> +	khomem_desc_t *desc;
> +	unsigned long order;
> +
> +	xa_for_each(&tracker->orders, order, desc) {
> +		if (WARN_ON(order >= NR_PAGE_ORDERS))
> +			break;
> +		kho_out.mem_tables[order] = *desc;

Missing the virt_to_phys?

> +	nr_tables = min_t(unsigned int, len / sizeof(*tables), NR_PAGE_ORDERS);
> +	for (order = 0; order < nr_tables; order++)
> +		khomem_walk_preserved((khomem_desc_t *)&tables[order], order,

Missing phys_to_virt

Please dont' remove the KHOSER stuff, and do use it with proper
structs and types. It is part of keeping this stuff understandable.

Jason