On 11/18/24 16:55, Mikulas Patocka wrote:
>
>
> On Mon, 18 Nov 2024, Meir Elisha wrote:
>
>>> diff --git a/drivers/md/persistent-data/dm-transaction-manager.c b/drivers/md/persistent-data/dm-transaction-manager.c
>>> index c7ba4e6cbbc7..8d486b1e6693 100644
>>> --- a/drivers/md/persistent-data/dm-transaction-manager.c
>>> +++ b/drivers/md/persistent-data/dm-transaction-manager.c
>>> @@ -10,6 +10,7 @@
>>> #include "dm-space-map-metadata.h"
>>> #include "dm-persistent-data-internal.h"
>>>
>>> +#include <linux/moduleparam.h>
>>> #include <linux/export.h>
>>> #include <linux/mutex.h>
>>> #include <linux/hash.h>
>>> @@ -84,8 +85,35 @@ struct shadow_info {
>>> /*
>>> * It would be nice if we scaled with the size of transaction.
>>> */
>>> -#define DM_HASH_SIZE 256
>>> -#define DM_HASH_MASK (DM_HASH_SIZE - 1)
>>> +static uint tm_hash_table_size = 256;
>>> +
>>> +static int param_set_hash_size(const char *val, const struct kernel_param *kp)
>>> +{
>>> + unsigned int num;
>>> + int ret;
>>> +
>>> + if (!val)
>>> + return -EINVAL;
>>> +
>>> + ret = kstrtouint(val, 0, &num);
>>> + if (ret)
>>> + return ret;
>>> +
>>> + /* Hash size must be a power of 2 */
>>> + if (!(num && !(num & (num - 1))))
>>> + return -EINVAL;
>>> +
>>> + *((unsigned int *)kp->arg) = num;
>>> + return 0;
>>> +}
>>> +
>>> +static const struct kernel_param_ops tm_hash_table_size_ops = {
>>> + .set = param_set_hash_size,
>>> + .get = param_get_uint
>>> +};
>>> +
>>> +module_param_cb(tm_hash_table_size, &tm_hash_table_size_ops, &tm_hash_table_size, 0644);
>>> +MODULE_PARM_DESC(tm_hash_table_size, "transaction manager hash size");
>>>
>>> struct dm_transaction_manager {
>>> int is_clone;
>>> @@ -95,8 +123,9 @@ struct dm_transaction_manager {
>>> struct dm_space_map *sm;
>>>
>>> spinlock_t lock;
>>> - struct hlist_head buckets[DM_HASH_SIZE];
>>> -
>>> + struct hlist_head *buckets;
>>> + uint hash_size;
>>> + uint hash_mask;
>>> struct prefetch_set prefetches;
>>> };
>>>
>>> @@ -105,7 +134,7 @@ struct dm_transaction_manager {
>>> static int is_shadow(struct dm_transaction_manager *tm, dm_block_t b)
>>> {
>>> int r = 0;
>>> - unsigned int bucket = dm_hash_block(b, DM_HASH_MASK);
>>> + unsigned int bucket = dm_hash_block(b, tm->hash_mask);
>>> struct shadow_info *si;
>>>
>>> spin_lock(&tm->lock);
>>> @@ -131,7 +160,7 @@ static void insert_shadow(struct dm_transaction_manager *tm, dm_block_t b)
>>> si = kmalloc(sizeof(*si), GFP_NOIO);
>>> if (si) {
>>> si->where = b;
>>> - bucket = dm_hash_block(b, DM_HASH_MASK);
>>> + bucket = dm_hash_block(b, tm->hash_mask);
>>> spin_lock(&tm->lock);
>>> hlist_add_head(&si->hlist, tm->buckets + bucket);
>>> spin_unlock(&tm->lock);
>>> @@ -146,7 +175,7 @@ static void wipe_shadow_table(struct dm_transaction_manager *tm)
>>> int i;
>>>
>>> spin_lock(&tm->lock);
>>> - for (i = 0; i < DM_HASH_SIZE; i++) {
>>> + for (i = 0; i < tm->hash_size; i++) {
>>> bucket = tm->buckets + i;
>>> hlist_for_each_entry_safe(si, tmp, bucket, hlist)
>>> kfree(si);
>>> @@ -169,13 +198,19 @@ static struct dm_transaction_manager *dm_tm_create(struct dm_block_manager *bm,
>>> if (!tm)
>>> return ERR_PTR(-ENOMEM);
>>>
>>> + tm->hash_size = tm_hash_table_size;
>>> + tm->buckets = kmalloc_array(tm->hash_size, sizeof(*tm->buckets), GFP_KERNEL);
>>> + if (!tm->buckets)
>>> + return ERR_PTR(-ENOMEM);
>>> +
>>> tm->is_clone = 0;
>>> tm->real = NULL;
>>> tm->bm = bm;
>>> tm->sm = sm;
>>> + tm->hash_mask = tm->hash_size - 1;
>>>
>>> spin_lock_init(&tm->lock);
>>> - for (i = 0; i < DM_HASH_SIZE; i++)
>>> + for (i = 0; i < tm->hash_size; i++)
>>> INIT_HLIST_HEAD(tm->buckets + i);
>>>
>>> prefetch_init(&tm->prefetches);
>>
>> Hi all,
>>
>> I wanted to follow up on this patch.
>> Could you please review it when you have a chance?
>> If you need any additional information or have concerns, let me know.
>> Thanks for your time.
>
> Hi
>
> I discussed it with Joe. We concluded that the patch is not much useful,
> because few users will know about this option and few users will change it
> if they hit bad performance.
>
> Joe will be reworking the code so that it will use rb-trees instead of
> linked lists when there will be hash collision, this could improve
> performance without adding a tunable variable.
>
> Mikulas
>
First of all, thanks for the response.
I get that probably not too many users will use the module_param
but it has a significant performance impact on our use case.
Any general idea when the rb-trees solution should be submitted?
