Skip to contentSkip to conversation

blk-throttle: verify format of bandwidth limit and detect overflows

PATCH In review Konstantin Khlebnikov 2 participants 3 messages Feb 27, 2019 – Mar 5, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 blk-throttle: verify format of bandwidth limit and detect overflows

Unlike to memory cgroup blkio throttler does not support value suffixes.

It silently ignores everything after last digit. For example this command will set rate limit 1 byte per second rather than 1 megabyte per second:

# echo "7:0 1M" > blkio.throttle.read_bps_device
# cat blkio.throttle.read_bps_device
7:0 1

Cgroup2 interface has the same flaw:

# echo "7:0 rbps=1M" > io.max
# cat io.max
7:0 rbps=1 wbps=max riops=max wiops=max

Also sscanf does not care much about overflows.

This patch uses modern function kstrtou64 for parsing. It rejects trailing garbage and detects integer overflows.

Also this patch handles iops limit overflows for cgroup-v1 in the same as
cgroup-v2: limits >= UINT_MAX becomes unlimited.

Fixes 2ee867dcfa2e ("blkcg: implement interface for the unified hierarchy")
---
 block/blk-throttle.c |   40 ++++++++++++++++++++++++----------------
 1 file changed, 24 insertions(+), 16 deletions(-)
block/blk-throttle.c +24 -16
index 1b97a73d2fb1..ce00060b3a48 100644
@@ -1433,8 +1433,10 @@ static ssize_t tg_set_conf(struct kernfs_open_file *of,
 	if (ret)
 		return ret;
 
-	ret = -EINVAL;
-	if (sscanf(ctx.body, "%llu", &v) != 1)
+	/* remove trailing spaces, kstrto* are strict about them */
+	ctx.body = strim(ctx.body);
+	ret = kstrtou64(ctx.body, 10, &v);
+	if (ret)
 		goto out_finish;
 	if (!v)
 		v = U64_MAX;
@@ -1444,7 +1446,8 @@ static ssize_t tg_set_conf(struct kernfs_open_file *of,
 	if (is_u64)
 		*(u64 *)((void *)tg + of_cft(of)->private) = v;
 	else
-		*(unsigned int *)((void *)tg + of_cft(of)->private) = v;
+		*(unsigned int *)((void *)tg + of_cft(of)->private) =
+							min_t(u64, v, UINT_MAX);
 
 	tg_conf_updated(tg, false);
 	ret = 0;
@@ -1609,23 +1612,25 @@ static ssize_t tg_set_limit(struct kernfs_open_file *of,
 	idle_time = tg->idletime_threshold_conf;
 	latency_time = tg->latency_target_conf;
 	while (true) {
-		char tok[27];	/* wiops=18446744073709551616 */
-		char *p;
-		u64 val = U64_MAX;
-		int len;
-
-		if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
-			break;
-		if (tok[0] == '\0')
-			break;
-		ctx.body += len;
+		char tok[8];	/* "latency" */
+		char buf[21];	/* U64_MAX */
+		int end;
+		u64 val;
 
 		ret = -EINVAL;
-		p = tok;
-		strsep(&p, "=");
-		if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
+		if (sscanf(ctx.body, "%7[^=]=%20s %n", tok, buf, &end) != 2)
 			goto out_finish;
 
+		/* skip this field and trailing spaces */
+		ctx.body += end;
+
+		ret = kstrtou64(buf, 10, &val);
+		if (ret) {
+			if (strcmp(buf, "max"))
+				goto out_finish;
+			val = U64_MAX;
+		}
+
 		ret = -ERANGE;
 		if (!val)
 			goto out_finish;
@@ -1645,6 +1650,9 @@ static ssize_t tg_set_limit(struct kernfs_open_file *of,
 			latency_time = val;
 		else
 			goto out_finish;
+
+		if (!*ctx.body)
+			break;
 	}
 
 	tg->bps_conf[READ][index] = v[0];
maintainer
#1 lore raw

Hello, Konstantin.

Quoted 17 lines · Konstantin Khlebnikov

Unlike to memory cgroup blkio throttler does not support value suffixes.

It silently ignores everything after last digit. For example this command will set rate limit 1 byte per second rather than 1 megabyte per second:

# echo "7:0 1M" > blkio.throttle.read_bps_device
# cat blkio.throttle.read_bps_device
7:0 1

Cgroup2 interface has the same flaw:

# echo "7:0 rbps=1M" > io.max
# cat io.max
7:0 rbps=1 wbps=max riops=max wiops=max

Also sscanf does not care much about overflows.

Quoted 8 lines

This patch uses modern function kstrtou64 for parsing. It rejects trailing garbage and detects integer overflows.

Also this patch handles iops limit overflows for cgroup-v1 in the same as
cgroup-v2: limits >= UINT_MAX becomes unlimited.

Fixes 2ee867dcfa2e ("blkcg: implement interface for the unified hierarchy")

So, I'd much rather keep the parsing implementations simple. Unless there's a correctness problem (you mentioned overflowing, how would that happen?), the simpler the better. I don't think the kernel needs to be in the business of strict input verification here.

Thanks.

Signature · 1 line
tejun
#2 lore raw
Quoted 26 lines · Tejun Heo

Hello, Konstantin.

Quoted 23 lines · Konstantin Khlebnikov

Unlike to memory cgroup blkio throttler does not support value suffixes.

It silently ignores everything after last digit. For example this command will set rate limit 1 byte per second rather than 1 megabyte per second:

# echo "7:0 1M" > blkio.throttle.read_bps_device
# cat blkio.throttle.read_bps_device
7:0 1

Cgroup2 interface has the same flaw:

# echo "7:0 rbps=1M" > io.max
# cat io.max
7:0 rbps=1 wbps=max riops=max wiops=max

Also sscanf does not care much about overflows.

This patch uses modern function kstrtou64 for parsing. It rejects trailing garbage and detects integer overflows.

Also this patch handles iops limit overflows for cgroup-v1 in the same as
cgroup-v2: limits >= UINT_MAX becomes unlimited.

Quoted 7 lines
Quoted 2 lines
Fixes 2ee867dcfa2e ("blkcg: implement interface for the unified hierarchy")

So, I'd much rather keep the parsing implementations simple. Unless there's a correctness problem (you mentioned overflowing, how would that happen?), the simpler the better. I don't think the kernel needs to be in the business of strict input verification here.

This is inconsistency between controllers. IO controller should either accept size suffixes or reject them. Ignoring garbage is not an option. To be honest, I've successfully shot in the foot by bps limit "1M".

Integer overflows are happens when userspace software uses floats for data representation and calculations. This way values bigger than 2^32-1 or 2^64-1 could appear easily in text representation.

Almost all integer knobs in sysctl and sysfs use modern kstrtoXX functions with strict validation of format and overflows.

Quoted 3 lines

Thanks.