[PATCH v5 2/6] kho: support incremental entry publication in kho_block

Pasha Tatashin pasha.tatashin at soleen.com
Wed Nov 11 07:15:36 PST 2026


On Wed Nov 11, 2026 at 6:18 AM EST, Mike Rapoport wrote:
> Hi,
>
>> To allow LUO sessions to be frozen and committed into the outgoing KHO
>> subtree from userspace ahead of reboot, kho_block needs to support
>> publishing entries one by one into a live serialized block chain while
>> new blocks may still be preallocated or shrunk for uncommitted sessions.
>> 
>> Currently, kho_block assumes outgoing block sets are populated in a
>> single batch during reboot() before the chain is exposed: preallocated
>> blocks are linked onto the chain immediately even while empty
>> (ser->count == 0), and reserving a slot increments ser->count before the
>> entry is populated.
>> 
>> Address this by keeping preallocated outgoing blocks unlinked in
>> kho_block_add(), linking blocks lazily only as entries are written. Add
>> kho_block_set_it_append() to copy each entry into its slot before
>> publishing the updated ser->count and next-block link via
>> smp_store_release(), paired with smp_load_acquire() in
>> kho_block_set_it_read_entry().
>> 
>> Signed-off-by: Pasha Tatashin <pasha.tatashin at soleen.com>
>> ---
>>  include/linux/kho_block.h     |  1 +
>>  kernel/liveupdate/kho_block.c | 86 +++++++++++++++++++++++++++++++----
>>  2 files changed, 77 insertions(+), 10 deletions(-)
>>
>> diff --git a/include/linux/kho_block.h b/include/linux/kho_block.h
>> index 93a7cc2be5f51..a9e696ef3e08e 100644
>> --- a/include/linux/kho_block.h
>> +++ b/include/linux/kho_block.h
>> @@ -100,6 +100,7 @@ static inline bool kho_block_set_is_empty(struct kho_block_set *bs)
>>  
>>  void kho_block_set_it_init(struct kho_block_set_it *it, struct kho_block_set *bs);
>>  void *kho_block_set_it_reserve_entry(struct kho_block_set_it *it);
>> +void *kho_block_set_it_append(struct kho_block_set_it *it, const void *src);
>>  void *kho_block_set_it_read_entry(struct kho_block_set_it *it);
>>  void *kho_block_set_it_prev(struct kho_block_set_it *it);
>>  
>> diff --git a/kernel/liveupdate/kho_block.c b/kernel/liveupdate/kho_block.c
>> index 0d2a342ef4228..2813c6fd593c4 100644
>> --- a/kernel/liveupdate/kho_block.c
>> +++ b/kernel/liveupdate/kho_block.c
>> @@ -16,6 +16,12 @@
>>   * array of serialized entries. Multiple blocks are linked together via a
>>   * physical pointer in the header, forming a linked list that can be easily
>>   * traversed in both the current and the next kernel.
>> + *
>> + * Outgoing blocks preallocated via kho_block_set_grow() remain unlinked until
>> + * entries are written into them. Callers publishing entries incrementally into
>> + * a live KHO subtree can use kho_block_set_it_append(), which populates each
>> + * entry before exposing the updated block count or next-block link via
>> + * release-store ordering.
>
>
>>   */
>>  
>>  #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
>> @@ -93,9 +99,7 @@ static int kho_block_add(struct kho_block_set *bs,
>>  	list_add_tail(&block->list, &bs->blocks);
>>  	bs->nblocks++;
>>  
>> -	if (last)
>> -		last->ser->next = virt_to_phys(ser);
>> -	else
>> +	if (!last)
>>  		bs->head_pa = virt_to_phys(ser);
>
>
>>  
>>  	return 0;
>> @@ -129,14 +133,15 @@ static void kho_block_set_shrink_one(struct kho_block_set *bs)
>>  	last = list_last_entry(&bs->blocks, struct kho_block, list);
>>  	list_del(&last->list);
>>  	bs->nblocks--;
>> -	kho_block_free_ser(bs, last->ser);
>> -	kfree(last);
>>  
>>  	new_last = list_last_entry_or_null(&bs->blocks, struct kho_block, list);
>>  	if (new_last)
>> -		new_last->ser->next = 0;
>> +		WRITE_ONCE(new_last->ser->next, 0);
>
> Why this is needed? Docs explicitly state that "Synchronization and
> locking are the responsibility of the caller".

This was done as prep for the kho-ro changes: when an uncommitted 
session closes and shrinks an unused trailing block, new_last may 
already be published and read concurrently via kho-ro. I will move it 
into a separate patch in the kho-ro series to clearly explain the 
intent.

>
>>  	else
>>  		bs->head_pa = 0;
>> +
>> +	kho_block_free_ser(bs, last->ser);
>> +	kfree(last);
>
> This change seems unnecessary.

Leftover from the earlier pop support (unlinking last->ser before 
unpreserving/freeing it). Will drop this change in v6 as well; I have 
since simplified the logic so adding to KHO is always a one-way trip.

>>  }
>>  
>>  /**
>> @@ -144,9 +149,9 @@ static void kho_block_set_shrink_one(struct kho_block_set *bs)
>>   * @bs:    The block set.
>>   * @count: The target number of valid entries to accommodate.
>>   *
>> - * Dynamically preallocates and links preserved memory blocks if the target
>> - * entry count exceeds the current total capacity of the set, ensuring they
>> - * are available during serialization/deserialization.
>> + * Dynamically preallocates preserved memory blocks if the target entry count
>> + * exceeds the current total capacity of the set, ensuring they are available
>> + * during serialization/deserialization.
>>   *
>>   * Context: Caller must hold a lock protecting the block set.
>>   * Return: 0 on success, or a negative errno on failure.
>> @@ -318,6 +323,7 @@ void kho_block_set_clear(struct kho_block_set *bs)
>>  	struct kho_block *block;
>>  
>>  	list_for_each_entry(block, &bs->blocks, list) {
>> +		block->ser->next = 0;
>>  		block->ser->count = 0;
>>  		memset(block->ser + 1, 0, KHO_BLOCK_SIZE - sizeof(*block->ser));
>>  	}
>> @@ -348,6 +354,7 @@ void kho_block_set_it_init(struct kho_block_set_it *it, struct kho_block_set *bs
>>   */
>>  void *kho_block_set_it_reserve_entry(struct kho_block_set_it *it)
>>  {
>> +	struct kho_block *prev = NULL;
>>  	void *entry;
>>  
>>  	if (!it->block)
>> @@ -356,15 +363,73 @@ void *kho_block_set_it_reserve_entry(struct kho_block_set_it *it)
>>  	if (it->i == it->bs->count_per_block) {
>>  		if (list_is_last(&it->block->list, &it->bs->blocks))
>>  			return NULL;
>> +		prev = it->block;
>>  		it->block = list_next_entry(it->block, list);
>>  		it->i = 0;
>>  	}
>>  
>> +	if (!it->i)
>> +		it->block->ser->next = 0;
>> +
>>  	entry = kho_block_entry(it, it->i++);
>>  	it->block->ser->count = it->i;
>> +	if (prev)
>> +		prev->ser->next = virt_to_phys(it->block->ser);
>
> kernel-doc should mention that this function now links blocks.
>
> And I believe there should be a paragraph about the new flow in the DOC:
> section.

+1, I will update the kernel-doc.

>
>> +
>>  	return entry;
>>  }
>>  
>> +/**
>> + * kho_block_set_it_append - Write and publish an entry at the next slot.
>> + * @it:  The block iterator.
>> + * @src: Entry data of size @it->bs->entry_size to copy into the slot.
>> + *
>> + * Copies @src into the next available slot before updating the block's entry
>> + * count or linking a newly occupied block to the previous block, so concurrent
>> + * readers never observe an uninitialized slot or an empty block.
>> + *
>> + * Return: A pointer to the written entry slot, or NULL if the block set's
>> + * capacity is fully exhausted.
>> + */
>> +void *kho_block_set_it_append(struct kho_block_set_it *it, const void *src)
>> +{
>> +	struct kho_block *prev = NULL;
>> +	void *dst;
>> +
>> +	if (!it->block)
>> +		return NULL;
>> +
>> +	if (it->i == it->bs->count_per_block) {
>> +		if (list_is_last(&it->block->list, &it->bs->blocks))
>> +			return NULL;
>> +		prev = it->block;
>> +		it->block = list_next_entry(it->block, list);
>> +		it->i = 0;
>> +	}
>> +
>> +	if (!it->i)
>> +		WRITE_ONCE(it->block->ser->next, 0);
>> +
>> +	dst = kho_block_entry(it, it->i++);
>> +	memcpy(dst, src, it->bs->entry_size);
>> +
>> +	/*
>> +	 * Ensure the serialized entry is visible before exposing the
>> +	 * incremented count to concurrent readers.
>> +	 */
>> +	smp_store_release(&it->block->ser->count, it->i);
>> +	if (prev) {
>> +		/*
>> +		 * Ensure the new block's first entry and header are visible
>> +		 * before linking prev->ser->next for concurrent readers.
>> +		 */
>> +		smp_store_release(&prev->ser->next,
>> +				  virt_to_phys(it->block->ser));
>
> Until now the locking model was "the caller is responsible to hold a
> lock".
>
> Using WRITE_ONCE() and smp_store_release() instead of plain assignments
> implies that it is not anymore, or at least not always.
>
> This should be a well documented separate change.

Agreed. I wrote this patch as preparation for both early freeze and 
kho-ro: mutations to struct kho_block_set itself still require caller 
locking, while the WRITE_ONCE() and release/acquire ordering allow a 
concurrent kho-ro reader to walk the live serialized physical block 
chain while the host continues committing entries under its lock.

In v6, I will remove the kho-ro parts, and submit them together with the 
next iteration of that series.

-- 
Pasha




More information about the kexec mailing list