Skip to contentSkip to conversation

Bug#763917: mdadm: rounding errors in human_size()

DISCUSSION Michael Tokarev 2 participants 2 messages Dec 5, 2014 – Dec 18, 2014 1 message not in the archive MD RAID

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw

Neil, will you take the patch in this bugreport for the next version?

http://bugs.debian.org/763917.

Thanks,

/mjt

Quoted 60 lines · On Fri, 3 Oct 2014 20:24:54 +0200 Jan Echternach <[email protected]> wrote:

Package: mdadm
Version: 3.3.2-1
Severity: minor
Tags: patch

While setting up a new system, I noticed an incorrect value in the output of mdadm --examine:

Avail Dev Size : 2095080 (1023.16 MiB 1072.68 MB)

The 1023.16 MiB were quite irritating because the underlying partition has a size of exactly 1023 MiB.

The number of sectors seems plausible: 2095080 sectors * 512 bytes/sector are 1022.988 MiB or 1072.681 MB. I looked into the code and found an inaccuracy in human_size() and human_size_brief(). The formula used for the MiB value is essentially

  long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;

but (1LL<<20) / 200LL is not an integer. It's rounded down and cMiB becomes too large. The quick fix would have been multiplying by 200 before dividing by 1<<20, but that might cause integer overflows in the GiB case.

The following patch uses a more complicated formula that computes the fractional portion separately from the integer portion. It also changes some longs to long longs to eliminate a different cause of integer overflows.

util.c +7 -5
@@ -671,15 +671,17 @@
 	if (bytes < 5000*1024)
 		buf[0] = 0;
 	else if (bytes < 2*1024LL*1024LL*1024LL) {
-		long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;
+		long cMiB = bytes / (1LL<<20) * 100
+			    + ((bytes % (1LL<<20)) * 200 / (1LL<<20) +1) /2;
 		long cMB  = (bytes / ( 1000000LL / 200LL ) +1) /2;
 		snprintf(buf, sizeof(buf), " (%ld.%02ld MiB %ld.%02ld MB)",
 			cMiB/100 , cMiB % 100,
 			cMB/100, cMB % 100);
 	} else {
-		long cGiB = (bytes / ( (1LL<<30) / 200LL ) +1) /2;
-		long cGB  = (bytes / (1000000000LL/200LL ) +1) /2;
-		snprintf(buf, sizeof(buf), " (%ld.%02ld GiB %ld.%02ld GB)",
+		long long cGiB = bytes / (1LL<<30) * 100
+				 + ((bytes % (1LL<<30)) * 200 / (1LL<<30) +1) /2;
+		long long cGB  = (bytes / (1000000000LL/200LL ) +1) /2;
+		snprintf(buf, sizeof(buf), " (%lld.%02lld GiB %lld.%02lld GB)",
 			cGiB/100 , cGiB % 100,
 			cGB/100, cGB % 100);
 	}
@@ -706,12 +708,14 @@
 		buf[0] = 0;
 	else if (prefix == IEC) {
 		if (bytes < 2*1024LL*1024LL*1024LL) {
-			long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;
+			long cMiB = bytes / (1LL<<20) * 100
#2 lore raw
Quoted 33 lines · On Fri, 05 Dec 2014 13:28:00 +0300 Michael Tokarev <[email protected]> wrote:

Neil, will you take the patch in this bugreport for the next version?

http://bugs.debian.org/763917.

Thanks,

/mjt

Quoted 24 lines · On Fri, 3 Oct 2014 20:24:54 +0200 Jan Echternach <[email protected]> wrote:

Package: mdadm
Version: 3.3.2-1
Severity: minor
Tags: patch

While setting up a new system, I noticed an incorrect value in the output of mdadm --examine:

Avail Dev Size : 2095080 (1023.16 MiB 1072.68 MB)

The 1023.16 MiB were quite irritating because the underlying partition has a size of exactly 1023 MiB.

The number of sectors seems plausible: 2095080 sectors * 512 bytes/sector are 1022.988 MiB or 1072.681 MB. I looked into the code and found an inaccuracy in human_size() and human_size_brief(). The formula used for the MiB value is essentially

  long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;

but (1LL<<20) / 200LL is not an integer. It's rounded down and cMiB becomes too large. The quick fix would have been multiplying by 200 before dividing by 1<<20, but that might cause integer overflows in the GiB case.

Given that we are doing 64bit arithmetic, we would need about 56bits of bytes for there to be a rounding problem. That's 64 petabytes.

I decide to just do the simple transformation. If we get arrays close to petabytes I would want to make other changes, like reporting the number of terabytes for larger arrays.

Thanks,
NeilBrown

Quoted 41 lines
Quoted 31 lines

The following patch uses a more complicated formula that computes the fractional portion separately from the integer portion. It also changes some longs to long longs to eliminate a different cause of integer overflows.

util.c +6 -4
@@ -671,15 +671,17 @@
 	if (bytes < 5000*1024)
 		buf[0] = 0;
 	else if (bytes < 2*1024LL*1024LL*1024LL) {
-		long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;
+		long cMiB = bytes / (1LL<<20) * 100
+			    + ((bytes % (1LL<<20)) * 200 / (1LL<<20) +1) /2;
 		long cMB  = (bytes / ( 1000000LL / 200LL ) +1) /2;
 		snprintf(buf, sizeof(buf), " (%ld.%02ld MiB %ld.%02ld MB)",
 			cMiB/100 , cMiB % 100,
 			cMB/100, cMB % 100);
 	} else {
-		long cGiB = (bytes / ( (1LL<<30) / 200LL ) +1) /2;
-		long cGB  = (bytes / (1000000000LL/200LL ) +1) /2;
-		snprintf(buf, sizeof(buf), " (%ld.%02ld GiB %ld.%02ld GB)",
+		long long cGiB = bytes / (1LL<<30) * 100
+				 + ((bytes % (1LL<<30)) * 200 / (1LL<<30) +1) /2;
+		long long cGB  = (bytes / (1000000000LL/200LL ) +1) /2;
+		snprintf(buf, sizeof(buf), " (%lld.%02lld GiB %lld.%02lld GB)",
 			cGiB/100 , cGiB % 100,
 			cGB/100, cGB % 100);
 	}

util.c line 711

Quoted 5 lines
util.c +1 -1
@@ -706,12 +708,14 @@
 		buf[0] = 0;
 	else if (prefix == IEC) {
 		if (bytes < 2*1024LL*1024LL*1024LL) {
-			long cMiB = (bytes / ( (1LL<<20) / 200LL ) +1) /2;
+			long cMiB = bytes / (1LL<<20) * 100

_______________________________________________
pkg-mdadm-devel mailing list
[email protected]
http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-mdadm-devel