bcache: introduce bcache sysfs entries for ioprio-based bypass/writeback hints

Add sysfs entries to support to hint for bypass/writeback by the ioprio
assigned to the bio.  If the bio is unassigned, use current's io-context
ioprio for cache writeback or bypass (configured per-process with
`ionice`).

Having idle IOs bypass the cache can increase performance elsewhere
since you probably don't care about their performance.  In addition,
this prevents idle IOs from promoting into (polluting) your cache and
evicting blocks that are more important elsewhere.

If you really nead the performance at the expense of SSD wearout,
then configure ioprio_writeback and set your `ionice` appropriately.

For example:
	echo 2,7 > /sys/block/bcache0/bcache/ioprio_bypass
	echo 2,0 > /sys/block/bcache0/bcache/ioprio_writeback

See the documentation commit for details.

Signed-off-by: Eric Wheeler <[email protected]>
Acked-by: Kent Overstreet <[email protected]>
Tested-by: Kai Krakow <[email protected]>
Cc: [email protected]

Author: Eric Wheeler

drivers/md/bcache/bcache.h

@@ -392,6 +392,9 @@ struct cached_dev {
 	unsigned int		error_limit;
 	unsigned int		offline_seconds;
 
+	unsigned short		ioprio_writeback;
+	unsigned short		ioprio_bypass;
+
 	char			backing_dev_name[BDEVNAME_SIZE];
 };
 

drivers/md/bcache/request.c

@@ -379,6 +379,8 @@ static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
 	unsigned int sectors, congested;
 	struct task_struct *task = current;
 	struct io *i;
+	struct io_context *ioc;
+	unsigned short ioprio;
 
 	if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
 	    c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
@@ -406,6 +408,28 @@ static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
 			goto skip;
 	}
 
+	/* If the ioprio already exists on the bio, use that.  We assume that
+	 * the upper layer properly assigned the calling process's ioprio to
+	 * the bio being passed to bcache. Otherwise, use current's ioc. */
+	ioprio = bio_prio(bio);
+	if (!ioprio_valid(ioprio)) {
+		ioc = get_task_io_context(current, GFP_NOIO, NUMA_NO_NODE);
+		if (ioc) {
+			if (ioprio_valid(ioc->ioprio))
+				ioprio = ioc->ioprio;
+			put_io_context(ioc);
+			ioc = NULL;
+		}
+	}
+
+	/* If process ioprio is lower-or-equal to dc->ioprio_bypass, then
+	 * hint for bypass. Note that a lower-priority IO class+value
+	 * has a greater numeric value. */
+	if (ioprio_valid(ioprio) && ioprio_valid(dc->ioprio_writeback)
+		&& ioprio >= dc->ioprio_bypass) {
+		goto skip;
+	}
+
 	if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
 	    bio_sectors(bio) & (c->sb.block_size - 1)) {
 		pr_debug("skipping unaligned io");

drivers/md/bcache/sysfs.c

@@ -144,6 +144,9 @@ rw_attribute(copy_gc_enabled);
 rw_attribute(gc_after_writeback);
 rw_attribute(size);
 
+rw_attribute(ioprio_writeback);
+rw_attribute(ioprio_bypass);
+
 static ssize_t bch_snprint_string_list(char *buf,
 				       size_t size,
 				       const char * const list[],
@@ -272,6 +275,16 @@ SHOW(__bch_cached_dev)
 		return strlen(buf);
 	}
 
+	if (attr == &sysfs_ioprio_bypass)
+		return snprintf(buf, PAGE_SIZE-1, "%d,%ld\n",
+			IOPRIO_PRIO_CLASS(dc->ioprio_bypass),
+			IOPRIO_PRIO_DATA(dc->ioprio_bypass));
+
+	if (attr == &sysfs_ioprio_writeback)
+		return snprintf(buf, PAGE_SIZE-1, "%d,%ld\n",
+			IOPRIO_PRIO_CLASS(dc->ioprio_writeback),
+			IOPRIO_PRIO_DATA(dc->ioprio_writeback));
+
 #undef var
 	return 0;
 }
@@ -284,6 +297,10 @@ STORE(__cached_dev)
 	ssize_t v;
 	struct cache_set *c;
 	struct kobj_uevent_env *env;
+	unsigned ioprio_class = 0; /* invalid initial ioprio values */
+	unsigned ioprio_level = IOPRIO_BE_NR;
+	unsigned short *ioprio_hint = NULL;
+	char *ioprio_type = NULL;
 
 	/* no user space access if system is rebooting */
 	if (bcache_is_reboot)
@@ -430,6 +447,57 @@ STORE(__cached_dev)
 	if (attr == &sysfs_stop)
 		bcache_device_stop(&dc->disk);
 
+	/* ioprio hinting: we use ioprio_hint to reduce duplicate printk verbiage */
+	if (attr == &sysfs_ioprio_writeback) {
+		ioprio_hint = &dc->ioprio_writeback;
+		ioprio_type = "writeback";
+	}
+
+	if (attr == &sysfs_ioprio_bypass) {
+		ioprio_hint = &dc->ioprio_bypass;
+		ioprio_type = "bypass";
+	}
+
+	if (ioprio_hint != NULL)
+	{
+		if (sscanf(buf, "%u,%u", &ioprio_class, &ioprio_level) != 2
+			|| ioprio_class > IOPRIO_CLASS_IDLE
+			|| ioprio_level >= IOPRIO_BE_NR) {
+			pr_err("ioprio_%s invalid, expecting: (class,level) but parsed (%u,%u); ignored.",
+				ioprio_type,
+				ioprio_class, ioprio_level);
+			return size;
+		}
+
+		/* Use the maximum(/minimum) value in the class shift space to make integer
+		  comparison correct for ioprio_writeback(/ioprio_bypass) for IOPRIO_CLASS_IDLE.
+		  This is necessary because there are no ioprio levels for the idle class. */
+		if (ioprio_class == IOPRIO_CLASS_IDLE) {
+			if (ioprio_hint == &dc->ioprio_writeback)
+				ioprio_level = IOPRIO_PRIO_MASK;
+			else
+				/* Same, but 0 for bypass (inverted vs. writeback) */
+				ioprio_level = 0;
+		}
+
+		*ioprio_hint = IOPRIO_PRIO_VALUE(ioprio_class, ioprio_level);
+
+		if (!ioprio_valid(*ioprio_hint))
+			pr_info("disabled ioprio_%s hints.", ioprio_type);
+		else
+			pr_info("set hint for cache %s with priority %s: (class,level) = (%u,%u)",
+				ioprio_type,
+				( ioprio_hint == &dc->ioprio_writeback ? "at-or-above" : "at-or-below" ),
+				ioprio_class, ioprio_level);
+
+		if (ioprio_valid(dc->ioprio_writeback)
+			&& ioprio_valid(dc->ioprio_bypass)
+			&& dc->ioprio_writeback >= dc->ioprio_bypass)
+			pr_warning(
+				"warning: ioprio_writeback hint is neither disabled nor higher priority than the bypass hint; "
+				"will always writeback!");
+	}
+
 	return size;
 }
 
@@ -517,6 +585,8 @@ static struct attribute *bch_cached_dev_files[] = {
 #endif
 	&sysfs_backing_dev_name,
 	&sysfs_backing_dev_uuid,
+	&sysfs_ioprio_bypass,
+	&sysfs_ioprio_writeback,
 	NULL
 };
 KTYPE(bch_cached_dev);

drivers/md/bcache/writeback.c

@@ -821,6 +821,16 @@ void bch_cached_dev_writeback_init(struct cached_dev *dc)
 	dc->writeback_rate_p_term_inverse = 40;
 	dc->writeback_rate_i_term_inverse = 10000;
 
+	/*
+	 * These defaults provide the best SSD life by enabling bypass
+	 * for priorities at-or-below BE-7. This also provides better
+	 * performance (cache hits) by preventing (near-)idle processes from
+	 * polluting the cache working set.  Only set ioprio_writeback if
+	 * you really need it: it will wear out your SSD sooner.
+	 */
+	dc->ioprio_writeback = IOPRIO_PRIO_VALUE(0, 0);
+	dc->ioprio_bypass    = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_BE, (IOPRIO_BE_NR-1));
+
 	WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
 	INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
 }

drivers/md/bcache/writeback.h

@@ -65,6 +65,8 @@ static inline bool should_writeback(struct cached_dev *dc, struct bio *bio,
 				    unsigned int cache_mode, bool would_skip)
 {
 	unsigned int in_use = dc->disk.c->gc_stats.in_use;
+	struct io_context *ioc;
+	unsigned short ioprio;
 
 	if (cache_mode != CACHE_MODE_WRITEBACK ||
 	    test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
@@ -82,6 +84,28 @@ static inline bool should_writeback(struct cached_dev *dc, struct bio *bio,
 	if (would_skip)
 		return false;
 
+	/* If the ioprio already exists on the bio, use that.  We assume that
+	 * the upper layer properly assigned the calling process's ioprio to
+	 * the bio being passed to bcache. Otherwise, use current's ioc. */
+	ioprio = bio_prio(bio);
+	if (!ioprio_valid(ioprio)) {
+		ioc = get_task_io_context(current, GFP_NOIO, NUMA_NO_NODE);
+		if (ioc) {
+			if (ioprio_valid(ioc->ioprio))
+				ioprio = ioc->ioprio;
+			put_io_context(ioc);
+			ioc = NULL;
+		}
+	}
+
+	/* If process ioprio is higher-or-equal to dc->ioprio_writeback, then
+	 * hint for writeback. Note that a higher-priority IO class+value
+	 * has a lesser numeric value. */
+	if (ioprio_valid(ioprio) && ioprio_valid(dc->ioprio_writeback)
+		&& ioprio <= dc->ioprio_writeback) {
+		return true;
+	}
+
 	return (op_is_sync(bio->bi_opf) ||
 		bio->bi_opf & (REQ_META|REQ_PRIO) ||
 		in_use <= bch_cutoff_writeback);