On Thu, Oct 17, 2024 at 11:41 PM Kanchana P Sridhar
<kanchana.p.sridhar@xxxxxxxxx> wrote:
>
> Intel IAA hardware acceleration can be used effectively to improve the
> zswap_store() performance of large folios by batching multiple pages in a
> folio to be compressed in parallel by IAA. Hence, to build compress batching
> of zswap large folio stores using IAA, we need to be able to submit a batch
> of compress jobs from zswap to the hardware to compress in parallel if the
> iaa_crypto "async" mode is used.
>
> The IAA compress batching paradigm works as follows:
>
> 1) Submit N crypto_acomp_compress() jobs using N requests.
> 2) Use the iaa_crypto driver async poll() method to check for the jobs
> to complete.
> 3) There are no ordering constraints implied by submission, hence we
> could loop through the requests and process any job that has
> completed.
> 4) This would repeat until all jobs have completed with success/error
> status.
>
> To facilitate this, we need to provide for multiple acomp_reqs in
> "struct crypto_acomp_ctx", each representing a distinct compress
> job. Likewise, there needs to be a distinct destination buffer
> corresponding to each acomp_req.
>
> If CONFIG_ZSWAP_STORE_BATCHING_ENABLED is enabled, this patch will set the
> SWAP_CRYPTO_SUB_BATCH_SIZE constant to 8UL. This implies each per-cpu
> crypto_acomp_ctx associated with the zswap_pool can submit up to 8
> acomp_reqs at a time to accomplish parallel compressions.
>
> If IAA is not present and/or CONFIG_ZSWAP_STORE_BATCHING_ENABLED is not
> set, SWAP_CRYPTO_SUB_BATCH_SIZE will be set to 1UL.
>
> On an Intel Sapphire Rapids server, each socket has 4 IAA, each of which
> has 2 compress engines and 8 decompress engines. Experiments modeling a
> contended system with say 72 processes running under a cgroup with a fixed
> memory-limit, have shown that there is a significant performance
> improvement with dispatching compress jobs from all cores to all the
> IAA devices on the socket. Hence, SWAP_CRYPTO_SUB_BATCH_SIZE is set to
> 8 to maximize compression throughput if IAA is available.
>
> The definition of "struct crypto_acomp_ctx" is modified to make the
> req/buffer be arrays of size SWAP_CRYPTO_SUB_BATCH_SIZE. Thus, the
> added memory footprint cost of this per-cpu structure for batching is
> incurred only for platforms that have Intel IAA.
>
> Suggested-by: Ying Huang <ying.huang@xxxxxxxxx>
> Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@xxxxxxxxx>
Does this really need to be done in zswap? Why can't zswap submit a
single compression request with the supported number of pages, and
have the driver handle it as it sees fit?
> ---
> mm/swap.h | 11 ++++++
> mm/zswap.c | 104 ++++++++++++++++++++++++++++++++++-------------------
> 2 files changed, 78 insertions(+), 37 deletions(-)
>
> diff --git a/mm/swap.h b/mm/swap.h
> index ad2f121de970..566616c971d4 100644
> --- a/mm/swap.h
> +++ b/mm/swap.h
> @@ -8,6 +8,17 @@ struct mempolicy;
> #include <linux/swapops.h> /* for swp_offset */
> #include <linux/blk_types.h> /* for bio_end_io_t */
>
> +/*
> + * For IAA compression batching:
> + * Maximum number of IAA acomp compress requests that will be processed
> + * in a sub-batch.
> + */
> +#if defined(CONFIG_ZSWAP_STORE_BATCHING_ENABLED)
> +#define SWAP_CRYPTO_SUB_BATCH_SIZE 8UL
> +#else
> +#define SWAP_CRYPTO_SUB_BATCH_SIZE 1UL
> +#endif
> +
> /* linux/mm/page_io.c */
> int sio_pool_init(void);
> struct swap_iocb;
> diff --git a/mm/zswap.c b/mm/zswap.c
> index 4893302d8c34..579869d1bdf6 100644
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -152,9 +152,9 @@ bool zswap_never_enabled(void)
>
> struct crypto_acomp_ctx {
> struct crypto_acomp *acomp;
> - struct acomp_req *req;
> + struct acomp_req *req[SWAP_CRYPTO_SUB_BATCH_SIZE];
> + u8 *buffer[SWAP_CRYPTO_SUB_BATCH_SIZE];
> struct crypto_wait wait;
> - u8 *buffer;
> struct mutex mutex;
> bool is_sleepable;
> };
> @@ -832,49 +832,64 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
> struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
> struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
> struct crypto_acomp *acomp;
> - struct acomp_req *req;
> int ret;
> + int i, j;
>
> mutex_init(&acomp_ctx->mutex);
>
> - acomp_ctx->buffer = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
> - if (!acomp_ctx->buffer)
> - return -ENOMEM;
> -
> acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
> if (IS_ERR(acomp)) {
> pr_err("could not alloc crypto acomp %s : %ld\n",
> pool->tfm_name, PTR_ERR(acomp));
> - ret = PTR_ERR(acomp);
> - goto acomp_fail;
> + return PTR_ERR(acomp);
> }
> acomp_ctx->acomp = acomp;
> acomp_ctx->is_sleepable = acomp_is_async(acomp);
>
> - req = acomp_request_alloc(acomp_ctx->acomp);
> - if (!req) {
> - pr_err("could not alloc crypto acomp_request %s\n",
> - pool->tfm_name);
> - ret = -ENOMEM;
> - goto req_fail;
> + for (i = 0; i < SWAP_CRYPTO_SUB_BATCH_SIZE; ++i) {
> + acomp_ctx->buffer[i] = kmalloc_node(PAGE_SIZE * 2,
> + GFP_KERNEL, cpu_to_node(cpu));
> + if (!acomp_ctx->buffer[i]) {
> + for (j = 0; j < i; ++j)
> + kfree(acomp_ctx->buffer[j]);
> + ret = -ENOMEM;
> + goto buf_fail;
> + }
> + }
> +
> + for (i = 0; i < SWAP_CRYPTO_SUB_BATCH_SIZE; ++i) {
> + acomp_ctx->req[i] = acomp_request_alloc(acomp_ctx->acomp);
> + if (!acomp_ctx->req[i]) {
> + pr_err("could not alloc crypto acomp_request req[%d] %s\n",
> + i, pool->tfm_name);
> + for (j = 0; j < i; ++j)
> + acomp_request_free(acomp_ctx->req[j]);
> + ret = -ENOMEM;
> + goto req_fail;
> + }
> }
> - acomp_ctx->req = req;
>
> + /*
> + * The crypto_wait is used only in fully synchronous, i.e., with scomp
> + * or non-poll mode of acomp, hence there is only one "wait" per
> + * acomp_ctx, with callback set to req[0].
> + */
> crypto_init_wait(&acomp_ctx->wait);
> /*
> * if the backend of acomp is async zip, crypto_req_done() will wakeup
> * crypto_wait_req(); if the backend of acomp is scomp, the callback
> * won't be called, crypto_wait_req() will return without blocking.
> */
> - acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
> + acomp_request_set_callback(acomp_ctx->req[0], CRYPTO_TFM_REQ_MAY_BACKLOG,
> crypto_req_done, &acomp_ctx->wait);
>
> return 0;
>
> req_fail:
> + for (i = 0; i < SWAP_CRYPTO_SUB_BATCH_SIZE; ++i)
> + kfree(acomp_ctx->buffer[i]);
> +buf_fail:
> crypto_free_acomp(acomp_ctx->acomp);
> -acomp_fail:
> - kfree(acomp_ctx->buffer);
> return ret;
> }
>
> @@ -884,11 +899,17 @@ static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
> struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
>
> if (!IS_ERR_OR_NULL(acomp_ctx)) {
> - if (!IS_ERR_OR_NULL(acomp_ctx->req))
> - acomp_request_free(acomp_ctx->req);
> + int i;
> +
> + for (i = 0; i < SWAP_CRYPTO_SUB_BATCH_SIZE; ++i)
> + if (!IS_ERR_OR_NULL(acomp_ctx->req[i]))
> + acomp_request_free(acomp_ctx->req[i]);
> +
> + for (i = 0; i < SWAP_CRYPTO_SUB_BATCH_SIZE; ++i)
> + kfree(acomp_ctx->buffer[i]);
> +
> if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> crypto_free_acomp(acomp_ctx->acomp);
> - kfree(acomp_ctx->buffer);
> }
>
> return 0;
> @@ -911,7 +932,7 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
>
> mutex_lock(&acomp_ctx->mutex);
>
> - dst = acomp_ctx->buffer;
> + dst = acomp_ctx->buffer[0];
> sg_init_table(&input, 1);
> sg_set_page(&input, page, PAGE_SIZE, 0);
>
> @@ -921,7 +942,7 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
> * giving the dst buffer with enough length to avoid buffer overflow.
> */
> sg_init_one(&output, dst, PAGE_SIZE * 2);
> - acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
> + acomp_request_set_params(acomp_ctx->req[0], &input, &output, PAGE_SIZE, dlen);
>
> /*
> * If the crypto_acomp provides an asynchronous poll() interface,
> @@ -940,19 +961,20 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
> * parallel.
> */
> if (acomp_ctx->acomp->poll) {
> - comp_ret = crypto_acomp_compress(acomp_ctx->req);
> + comp_ret = crypto_acomp_compress(acomp_ctx->req[0]);
> if (comp_ret == -EINPROGRESS) {
> do {
> - comp_ret = crypto_acomp_poll(acomp_ctx->req);
> + comp_ret = crypto_acomp_poll(acomp_ctx->req[0]);
> if (comp_ret && comp_ret != -EAGAIN)
> break;
> } while (comp_ret);
> }
> } else {
> - comp_ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
> + comp_ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req[0]),
> + &acomp_ctx->wait);
> }
>
> - dlen = acomp_ctx->req->dlen;
> + dlen = acomp_ctx->req[0]->dlen;
> if (comp_ret)
> goto unlock;
>
> @@ -1006,31 +1028,39 @@ static void zswap_decompress(struct zswap_entry *entry, struct folio *folio)
> */
> if ((acomp_ctx->is_sleepable && !zpool_can_sleep_mapped(zpool)) ||
> !virt_addr_valid(src)) {
> - memcpy(acomp_ctx->buffer, src, entry->length);
> - src = acomp_ctx->buffer;
> + memcpy(acomp_ctx->buffer[0], src, entry->length);
> + src = acomp_ctx->buffer[0];
> zpool_unmap_handle(zpool, entry->handle);
> }
>
> sg_init_one(&input, src, entry->length);
> sg_init_table(&output, 1);
> sg_set_folio(&output, folio, PAGE_SIZE, 0);
> - acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, PAGE_SIZE);
> + acomp_request_set_params(acomp_ctx->req[0], &input, &output,
> + entry->length, PAGE_SIZE);
> if (acomp_ctx->acomp->poll) {
> - ret = crypto_acomp_decompress(acomp_ctx->req);
> + ret = crypto_acomp_decompress(acomp_ctx->req[0]);
> if (ret == -EINPROGRESS) {
> do {
> - ret = crypto_acomp_poll(acomp_ctx->req);
> + ret = crypto_acomp_poll(acomp_ctx->req[0]);
> BUG_ON(ret && ret != -EAGAIN);
> } while (ret);
> }
> } else {
> - BUG_ON(crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait));
> + BUG_ON(crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req[0]),
> + &acomp_ctx->wait));
> }
> - BUG_ON(acomp_ctx->req->dlen != PAGE_SIZE);
> - mutex_unlock(&acomp_ctx->mutex);
> + BUG_ON(acomp_ctx->req[0]->dlen != PAGE_SIZE);
>
> - if (src != acomp_ctx->buffer)
> + if (src != acomp_ctx->buffer[0])
> zpool_unmap_handle(zpool, entry->handle);
> +
> + /*
> + * It is safer to unlock the mutex after the check for
> + * "src != acomp_ctx->buffer[0]" so that the value of "src"
> + * does not change.
> + */
> + mutex_unlock(&acomp_ctx->mutex);
> }
>
> /*********************************
> --
> 2.27.0
>
