Skip to contentSkip to conversation

btrfs: sysfs: Use scnprintf() for avoiding potential buffer overflow

PATCH In review Takashi Iwai 3 participants 6 messages Mar 11, 2020 – Mar 20, 2020 Btrfs file system

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 btrfs: sysfs: Use scnprintf() for avoiding potential buffer overflow

Since snprintf() returns the would-be-output size instead of the actual output size, the succeeding calls may go beyond the given buffer limit. Fix it by replacing with scnprintf().

---
 fs/btrfs/sysfs.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
fs/btrfs/sysfs.c +3 -3
index 93cf76118a04..d3dc069789a5 100644
@@ -310,12 +310,12 @@ static ssize_t supported_checksums_show(struct kobject *kobj,
 		 * This "trick" only works as long as 'enum btrfs_csum_type' has
 		 * no holes in it
 		 */
-		ret += snprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
+		ret += scnprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
 				(i == 0 ? "" : " "), btrfs_super_csum_name(i));
 
 	}
 
-	ret += snprintf(buf + ret, PAGE_SIZE - ret, "\n");
+	ret += scnprintf(buf + ret, PAGE_SIZE - ret, "\n");
 	return ret;
 }
 BTRFS_ATTR(static_feature, supported_checksums, supported_checksums_show);
@@ -992,7 +992,7 @@ char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags)
 			continue;
 
 		name = btrfs_feature_attrs[set][i].kobj_attr.attr.name;
-		len += snprintf(str + len, bufsize - len, "%s%s",
+		len += scnprintf(str + len, bufsize - len, "%s%s",
 				len ? "," : "", name);
 	}
 
Signature · 1 line
2.16.4
#2 lore raw
In reply to Takashi Iwai
Quoted 3 lines · Takashi Iwai

Since snprintf() returns the would-be-output size instead of the actual output size, the succeeding calls may go beyond the given buffer limit. Fix it by replacing with scnprintf().

Is this a mechanical conversion or is there actually a potential overflow in the code?

Quoted 10 lines
---
 fs/btrfs/sysfs.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
fs/btrfs/sysfs.c +0 -0
index 93cf76118a04..d3dc069789a5 100644

fs/btrfs/sysfs.c static ssize_t supported_checksums_show(struct kobject *kobj, · line 314

Quoted 6 lines
fs/btrfs/sysfs.c +1 -1
@@ -310,12 +310,12 @@ static ssize_t supported_checksums_show(struct kobject *kobj,
 		 * This "trick" only works as long as 'enum btrfs_csum_type' has
 		 * no holes in it
 		 */
-		ret += snprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
+		ret += scnprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
 				(i == 0 ? "" : " "), btrfs_super_csum_name(i));

Loop count is a constant, each iteration filling with two %s of constant length, buffer size is PAGE_SIZE.

Quoted 9 lines
fs/btrfs/sysfs.c +1 -1
 	}
 
-	ret += snprintf(buf + ret, PAGE_SIZE - ret, "\n");
+	ret += scnprintf(buf + ret, PAGE_SIZE - ret, "\n");
 	return ret;
 }
 BTRFS_ATTR(static_feature, supported_checksums, supported_checksums_show);

fs/btrfs/sysfs.c char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags) · line 996

Quoted 6 lines
fs/btrfs/sysfs.c +1 -1
@@ -992,7 +992,7 @@ char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags)
 			continue;
 
 		name = btrfs_feature_attrs[set][i].kobj_attr.attr.name;
-		len += snprintf(str + len, bufsize - len, "%s%s",
+		len += scnprintf(str + len, bufsize - len, "%s%s",
 				len ? "," : "", name);

Similar, compile-time constant for number of loops, filling with strings of bounded length.

If the patch is a precaution, then ok, but I don't see what it's trying to fix.

#3 lore raw
Quoted 8 lines · David Sterba
Quoted 3 lines · Takashi Iwai

Since snprintf() returns the would-be-output size instead of the actual output size, the succeeding calls may go beyond the given buffer limit. Fix it by replacing with scnprintf().

Is this a mechanical conversion or is there actually a potential overflow in the code?

It's rather a result of pattern matching.

Quoted 5 lines
Quoted 5 lines
---
 fs/btrfs/sysfs.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
Quoted 14 lines

fs/btrfs/sysfs.c static ssize_t supported_checksums_show(struct kobject *kobj, · line 314

Quoted 11 lines
fs/btrfs/sysfs.c +1 -1
index 93cf76118a04..d3dc069789a5 100644
@@ -310,12 +310,12 @@ static ssize_t supported_checksums_show(struct kobject *kobj,
 		 * This "trick" only works as long as 'enum btrfs_csum_type' has
 		 * no holes in it
 		 */
-		ret += snprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
+		ret += scnprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
 				(i == 0 ? "" : " "), btrfs_super_csum_name(i));

Loop count is a constant, each iteration filling with two %s of constant length, buffer size is PAGE_SIZE.

Yes, it's likely OK with the current code, but then snprintf() usage is utterly bogus.

Quoted 18 lines
Quoted 8 lines
fs/btrfs/sysfs.c +1 -1
 	}
 
-	ret += snprintf(buf + ret, PAGE_SIZE - ret, "\n");
+	ret += scnprintf(buf + ret, PAGE_SIZE - ret, "\n");
 	return ret;
 }
 BTRFS_ATTR(static_feature, supported_checksums, supported_checksums_show);

fs/btrfs/sysfs.c char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags) · line 996

Quoted 6 lines
fs/btrfs/sysfs.c +1 -1
@@ -992,7 +992,7 @@ char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags)
 			continue;
 
 		name = btrfs_feature_attrs[set][i].kobj_attr.attr.name;
-		len += snprintf(str + len, bufsize - len, "%s%s",
+		len += scnprintf(str + len, bufsize - len, "%s%s",
 				len ? "," : "", name);

Similar, compile-time constant for number of loops, filling with strings of bounded length.

Quoted 2 lines

If the patch is a precaution, then ok, but I don't see what it's trying to fix.

Take it rather a precaution, yes.

The problem is that the usage like

	pos += snprintf(buf + pos, len - pos, ...);

to append strings is already wrong per design unless it has a return value check right after each call. It might work if the string really doesn't go over the limit; but then it makes no sense to use snprintf(), you can use the plain sprintf().

thanks,

Takashi

#4 lore raw
Quoted 33 lines · Takashi Iwai
Quoted 8 lines · David Sterba
Quoted 3 lines · Takashi Iwai

Since snprintf() returns the would-be-output size instead of the actual output size, the succeeding calls may go beyond the given buffer limit. Fix it by replacing with scnprintf().

Is this a mechanical conversion or is there actually a potential overflow in the code?

It's rather a result of pattern matching.

Quoted 5 lines
Quoted 5 lines
---
 fs/btrfs/sysfs.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
Quoted 14 lines

fs/btrfs/sysfs.c static ssize_t supported_checksums_show(struct kobject *kobj, · line 314

Quoted 11 lines
fs/btrfs/sysfs.c +1 -1
index 93cf76118a04..d3dc069789a5 100644
@@ -310,12 +310,12 @@ static ssize_t supported_checksums_show(struct kobject *kobj,
 		 * This "trick" only works as long as 'enum btrfs_csum_type' has
 		 * no holes in it
 		 */
-		ret += snprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
+		ret += scnprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
 				(i == 0 ? "" : " "), btrfs_super_csum_name(i));

Loop count is a constant, each iteration filling with two %s of constant length, buffer size is PAGE_SIZE.

Quoted 2 lines

Yes, it's likely OK with the current code, but then snprintf() usage is utterly bogus.

I'm not sure what exactly are you calling bogus. We want to keep the code maintainable, the snprintf could be replaced by

	strcpy(buf, "crc32c xxhash64 sha256 blake2b\n");

yes, but now we have 2 places with hardcoded values. What I consided a good practice is to have one definition like

	static const struct btrfs_csums {
		u16             size;
		const char      *name;
		const char      *driver;
	} btrfs_csums[] = {
		[BTRFS_CSUM_TYPE_CRC32] = { .size = 4, .name = "crc32c" },
		[BTRFS_CSUM_TYPE_XXHASH] = { .size = 8, .name = "xxhash64" },
		[BTRFS_CSUM_TYPE_SHA256] = { .size = 32, .name = "sha256" },
		[BTRFS_CSUM_TYPE_BLAKE2] = { .size = 32, .name = "blake2b",
					     .driver = "blake2b-256" },
	};

and the extract what's needed.

Quoted 16 lines
Quoted 11 lines

fs/btrfs/sysfs.c char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags) · line 996

Quoted 7 lines
fs/btrfs/sysfs.c +1 -1
@@ -992,7 +992,7 @@ char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags)
 			continue;
 
 		name = btrfs_feature_attrs[set][i].kobj_attr.attr.name;
-		len += snprintf(str + len, bufsize - len, "%s%s",
+		len += scnprintf(str + len, bufsize - len, "%s%s",
 				len ? "," : "", name);

Similar, compile-time constant for number of loops, filling with strings of bounded length.

Quoted 2 lines

If the patch is a precaution, then ok, but I don't see what it's trying to fix.

Take it rather a precaution, yes.

Quoted 8 lines

The problem is that the usage like

	pos += snprintf(buf + pos, len - pos, ...);

to append strings is already wrong per design unless it has a return value check right after each call. It might work if the string really doesn't go over the limit; but then it makes no sense to use snprintf(), you can use the plain sprintf().

I'm afraid that when we use snprintf, somebody comes that it's unsafe because that's what some code scanning tool said that, without looking at the context of use. The code can simply use strcat and be fine, but that I don't want to encourage to be used, when code is copied to similar functions.

I'm fine with scnprintf as this should make everybody happy, while there would be effectively no change, just that the changelog should be worded accordingly.

#5 lore raw
Quoted 35 lines · David Sterba
Quoted 33 lines · Takashi Iwai
Quoted 8 lines · David Sterba
Quoted 3 lines · Takashi Iwai

Since snprintf() returns the would-be-output size instead of the actual output size, the succeeding calls may go beyond the given buffer limit. Fix it by replacing with scnprintf().

Is this a mechanical conversion or is there actually a potential overflow in the code?

It's rather a result of pattern matching.

Quoted 19 lines
Quoted 10 lines
---
 fs/btrfs/sysfs.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
fs/btrfs/sysfs.c +0 -0
index 93cf76118a04..d3dc069789a5 100644

fs/btrfs/sysfs.c static ssize_t supported_checksums_show(struct kobject *kobj, · line 314

Quoted 6 lines
fs/btrfs/sysfs.c +1 -1
@@ -310,12 +310,12 @@ static ssize_t supported_checksums_show(struct kobject *kobj,
 		 * This "trick" only works as long as 'enum btrfs_csum_type' has
 		 * no holes in it
 		 */
-		ret += snprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
+		ret += scnprintf(buf + ret, PAGE_SIZE - ret, "%s%s",
 				(i == 0 ? "" : " "), btrfs_super_csum_name(i));

Loop count is a constant, each iteration filling with two %s of constant length, buffer size is PAGE_SIZE.

Quoted 4 lines
Quoted 2 lines

Yes, it's likely OK with the current code, but then snprintf() usage is utterly bogus.

I'm not sure what exactly are you calling bogus.

Heh, this shows the exact problem of snprintf() :)

Actually, the snprintf() usage in btrfs/sysfs.c is harmful because it looks as if it were safer than plain sprintf() although it doesn't protect correctly.

For example, try to run the code below (better to compile with -O option):

-- 8< --
#include <stdio.h>

#define SIZE 128
int main()

{
	char buf[SIZE];
	int i, len = 0;

	for (i = 0; i < 1000; i++)
		len += snprintf(buf + len, SIZE - len, "%d", i);
	printf("%d, %s\n", len, buf);
	return 0;
}

-- 8< --

The code looks as if correct, limiting the buffer to the given size; but it'll lead to either a segfault or a totally bogus output.

And when you compare the above with the code in btrfs/sysfs.c, you'll see the same logic, and see why snprintf() usage there is harmful.

Quoted 25 lines
Quoted 16 lines
Quoted 11 lines

fs/btrfs/sysfs.c char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags) · line 996

Quoted 7 lines
fs/btrfs/sysfs.c +1 -1
@@ -992,7 +992,7 @@ char *btrfs_printable_features(enum btrfs_feature_set set, u64 flags)
 			continue;
 
 		name = btrfs_feature_attrs[set][i].kobj_attr.attr.name;
-		len += snprintf(str + len, bufsize - len, "%s%s",
+		len += scnprintf(str + len, bufsize - len, "%s%s",
 				len ? "," : "", name);

Similar, compile-time constant for number of loops, filling with strings of bounded length.

Quoted 2 lines

If the patch is a precaution, then ok, but I don't see what it's trying to fix.

Take it rather a precaution, yes.

Quoted 8 lines

The problem is that the usage like

	pos += snprintf(buf + pos, len - pos, ...);

to append strings is already wrong per design unless it has a return value check right after each call. It might work if the string really doesn't go over the limit; but then it makes no sense to use snprintf(), you can use the plain sprintf().

Quoted 5 lines

I'm afraid that when we use snprintf, somebody comes that it's unsafe because that's what some code scanning tool said that, without looking at the context of use. The code can simply use strcat and be fine, but that I don't want to encourage to be used, when code is copied to similar functions.

That's why scnprintf() was proposed in a decade ago. It's definitely less confusing than snprintf() and leads to more expected results.

Quoted 3 lines

I'm fine with scnprintf as this should make everybody happy, while there would be effectively no change, just that the changelog should be worded accordingly.

I'm fine for rephrasing the changelog if you agree with applying the code change. I've done similarly rephrasing for some other patches already.

thanks,

Takashi