Merge: dm-crypt: limit the size of encryption requests

MR: https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/4779

JIRA: https://issues.redhat.com/browse/RHEL-49548  
Tested: cryptsetup tests and the bug reproducer  
Upstream Status: kernel/git/device-mapper/linux-dm.git  
  
commit 0d815e3  
Author: Mikulas Patocka <mpatocka@redhat.com>  
Date:   Wed Jul 3 15:00:29 2024 +0200  
  
    dm-crypt: limit the size of encryption requests  
  
    There was a performance regression reported where dm-crypt would perform  
    worse on new kernels than on old kernels. The reason is that the old  
    kernels split the bios to NVMe request size (that is usually 65536 or  
    131072 bytes) and the new kernels pass the big bios through dm-crypt and  
    split them underneath.  
  
    If a big 1MiB bio is passed to dm-crypt, dm-crypt processes it on a single  
    core without parallelization and this is what causes the performance  
    degradation.  
  
    This commit introduces new tunable variables  
    /sys/module/dm_crypt/parameters/max_read_size and  
    /sys/module/dm_crypt/parameters/max_write_size that specify the maximum  
    bio size for dm-crypt. Bios larger than this value are split, so that  
    they can be encrypted in parallel by multiple cores. If these variables  
    are '0', a default 131072 is used.  
  
    Splitting bios may cause performance regressions in other workloads - if  
    this happens, the user should increase the value in max_read_size and  
    max_write_size variables.  
  
    max_read_size:  
    128k    2399MiB/s  
    256k    2368MiB/s  
    512k    1986MiB/s  
    1024    1790MiB/s  
  
    max_write_size:  
    128k    1712MiB/s  
    256k    1651MiB/s  
    512k    1537MiB/s  
    1024k   1332MiB/s  
  
    Note that if you run dm-crypt inside a virtual machine, you may need to do  
    "echo numa >/sys/module/workqueue/parameters/default_affinity_scope" to  
    improve performance.  
  
    Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>  
    Tested-by: Laurence Oberman <loberman@redhat.com>  
  
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>

Approved-by: Mikuláš Patočka <mpatocka@redhat.com>
Approved-by: John B. Wyatt IV <jwyatt@redhat.com>
Approved-by: CKI KWF Bot <cki-ci-bot+kwf-gitlab-com@redhat.com>

Merged-by: Lucas Zampieri <lzampier@redhat.com>

Author: zampierilucas

Documentation/admin-guide/device-mapper/dm-crypt.rst

@@ -155,6 +155,17 @@ iv_large_sectors
    The <iv_offset> must be multiple of <sector_size> (in 512 bytes units)
    if this flag is specified.
 
+
+Module parameters::
+max_read_size
+max_write_size
+   Maximum size of read or write requests. When a request larger than this size
+   is received, dm-crypt will split the request. The splitting improves
+   concurrency (the split requests could be encrypted in parallel by multiple
+   cores), but it also causes overhead. The user should tune these parameters to
+   fit the actual workload.
+
+
 Example scripts
 ===============
 LUKS (Linux Unified Key Setup) is now the preferred way to set up disk

drivers/md/dm-crypt.c

@@ -238,6 +238,31 @@ static unsigned int dm_crypt_clients_n;
 static volatile unsigned long dm_crypt_pages_per_client;
 #define DM_CRYPT_MEMORY_PERCENT			2
 #define DM_CRYPT_MIN_PAGES_PER_CLIENT		(BIO_MAX_VECS * 16)
+#define DM_CRYPT_DEFAULT_MAX_READ_SIZE		131072
+#define DM_CRYPT_DEFAULT_MAX_WRITE_SIZE		131072
+
+static unsigned int max_read_size = 0;
+module_param(max_read_size, uint, 0644);
+MODULE_PARM_DESC(max_read_size, "Maximum size of a read request");
+static unsigned int max_write_size = 0;
+module_param(max_write_size, uint, 0644);
+MODULE_PARM_DESC(max_write_size, "Maximum size of a write request");
+static unsigned get_max_request_size(struct crypt_config *cc, bool wrt)
+{
+	unsigned val, sector_align;
+	val = !wrt ? READ_ONCE(max_read_size) : READ_ONCE(max_write_size);
+	if (likely(!val))
+		val = !wrt ? DM_CRYPT_DEFAULT_MAX_READ_SIZE : DM_CRYPT_DEFAULT_MAX_WRITE_SIZE;
+	if (wrt || cc->on_disk_tag_size) {
+		if (unlikely(val > BIO_MAX_VECS << PAGE_SHIFT))
+			val = BIO_MAX_VECS << PAGE_SHIFT;
+	}
+	sector_align = max(bdev_logical_block_size(cc->dev->bdev), (unsigned)cc->sector_size);
+	val = round_down(val, sector_align);
+	if (unlikely(!val))
+		val = sector_align;
+	return val >> SECTOR_SHIFT;
+}
 
 static void crypt_endio(struct bio *clone);
 static void kcryptd_queue_crypt(struct dm_crypt_io *io);
@@ -3435,6 +3460,7 @@ static int crypt_map(struct dm_target *ti, struct bio *bio)
 {
 	struct dm_crypt_io *io;
 	struct crypt_config *cc = ti->private;
+	unsigned max_sectors;
 
 	/*
 	 * If bio is REQ_PREFLUSH or REQ_OP_DISCARD, just bypass crypt queues.
@@ -3453,9 +3479,9 @@ static int crypt_map(struct dm_target *ti, struct bio *bio)
 	/*
 	 * Check if bio is too large, split as needed.
 	 */
-	if (unlikely(bio->bi_iter.bi_size > (BIO_MAX_VECS << PAGE_SHIFT)) &&
-	    (bio_data_dir(bio) == WRITE || cc->on_disk_tag_size))
-		dm_accept_partial_bio(bio, ((BIO_MAX_VECS << PAGE_SHIFT) >> SECTOR_SHIFT));
+	max_sectors = get_max_request_size(cc, bio_data_dir(bio) == WRITE);
+	if (unlikely(bio_sectors(bio) > max_sectors))
+		dm_accept_partial_bio(bio, max_sectors);
 
 	/*
 	 * Ensure that bio is a multiple of internal sector encryption size