Bug 2089316

This bug was initially created as a copy of Bug #1913569

I am copying this bug because: 
The same bug appears in the RHEL-9 version of tar package.


Description of problem:

This is a follow up bug to the bug raised previously - Bug 1913566 - tar silently is padding with zeros (for Read error at byte...), should say it is Padding with zeros
https://bugzilla.redhat.com/show_bug.cgi?id=1913566

While using the reproducing steps to test that tar is writing 0's only until the file size is reached (as file header is already written) on Read error and  tar is not writing infinitely.

During the reproducer steps provided by customer we see that sometimes tar report "Read I/O errors" as "File shrank, padding with 0" , we see in the step(2) provided.

Version-Release number of selected component (if applicable):
# rpm -q tar libtar
tar-1.30-5.el8.x86_64
libtar-1.2.20-15.el8.x86_64


How reproducible:
For me it was around 50%
The below script may not give you read error on every try. I was getting error for (read error + file shrank) on every alternate try.

Steps to Reproduce:
=================
The following steps provided by customer I can reproduce the issue. It setup a device with dmsetup using error target.


#!/bin/bash
# Reproducer "tardust"
#
# When "tar create" reads a file there are several shortcomings when it hits read error
#
# 1) When read() returns 0 bytes due to read error, then this happens
# read(4, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# write(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# read(4, 0x563adef7b000, 3584) = -1 EIO (Input/output error)
# write(2, "tar: ", 5tar: ) = 5
# write(2, "/mntx/testfile: Read error at by"..., 70/mntx/testfile: Read error at byte 260653056, while reading 3584 bytes) = 70
# write(2, ": Input/output error", 20: Input/output error) = 20
# write(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# write(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# write(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# Actual behaviour: it prints a message about "Read error", but it conceals the information it will pad the output with zeros
# Expected behaviour: it should also print the information "padding with zero"
# 2) There is a 2nd shortcoming about tar not differentiate between "read error" and "file shrinkage"
# That means when it sees a short read due to read error, it does not report read error.
# It looks like this:
# read(4, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# write(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# read(4, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 2560          <<< HERE
# write(2, "tar: ", 5tar: ) = 5
# write(2, "/mntx/testfile: File shrank by 5"..., 65/mntx/testfile: File shrank by 53927936 bytes; padding with zeros) = 65
# write(2, "\n", 1
# ) = 1
# write(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# write(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# Summary: A read error is not reported here. At least it now says "padding with zeros"
# Expected behaviour: it should report a read error, so the user knows what it going on.
#
# 3) Side-Note:
# The blocking factor is applied to the output. When reading a file, all reads are misaligned by 512 bytes.
# This is because it writes a 512 header for every archived file.
# That means the first read from file is 512bytes too short:
# Running with tar-blocking-factor=7
# fstat(1, {st_mode=S_IFREG|0644, st_size=17827, ...}) = 0
# write(1, "/mntx/testfile\n", 15/mntx/testfile
# ) = 15
# read(4, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3072) = 3072 #1st read 512bytes too short
# write(3, "mntx/testfile\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# read(4, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
# write(3, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3584) = 3584
#
# 4) Reproducer overview:
# - Create a 500MB testimage, then create a testfile1 in the image
# - Use losetup/dmsetup with the "dust" target type
# - you can inject IO errors at specified block number in "dust"
# - You must hit a 4K boundary to see EIO, so use tar-blocking-factor=7 and
# - vary the bad blocknumber to find the case (1)
echo Step 1 Create disk image
dd if=/dev/zero of=/tmp/testimage bs=1M count=500 || exit
echo Step 2 Create XFS in image
mkfs.xfs /tmp/testimage || exit
echo Step 3 Use losetup so the file can be used a block device
losetup /dev/loop1 /tmp/testimage || exit
losetup
echo Step 6 Now create the testfile, this will have read error injected later
mkdir /mntx
mount /dev/loop1 /mntx || exit
dd if=/dev/zero of=/mntx/testfile bs=1M count=300 || exit
umount /mntx
echo Step7 Now iterating through bad blocks
echo As result, there are strace output file a1000 ... a1040
for i in `seq 1000 1 1040`
do
echo
echo Badblock $i
let ERR=i
let ERR1=i+1
let NUMSECTOR2=1024000-ERR1
#echo ERR1 is $ERR1
#echo NUMSECTOR2 is $NUMSECTOR2
dmsetup create tardust <<EOF
0 $ERR linear /dev/loop1 0
$ERR 1 error 
$ERR1 $NUMSECTOR2 linear /dev/loop1 $ERR1
EOF
#dmsetup ls
#dmsetup status
#dmsetup table
mount /dev/mapper/tardust /mntx || exit
strace tar cvbf 7 /tmp/tardust.tar /mntx/testfile >&/tmp/a$i
umount /mntx 
dmsetup remove tardust
grep -e error -e shrank /tmp/a$i
done
echo "Done: inspect the strace output file for error behaviour (grep error ; Look at last read()-call )"
losetup -d /dev/loop1

=================

Actual results:
- When tar hits a disk read error when reading file from disk and creating an archive, then it prints "file shrank"
- This happens in most cases (due to tar-block-size / disk-block-size / read-shift-by-512-bytes interaction)
- I provided a reproducer which shows under which circumstances it correctly prints "Read error at byte…"

Expected results:
- When there is a read error, THEN tar shall report a read error
- When there is a read error, THEN tar shall NOT report a "file shrank"

Additional info:

See bug Bug 1913566 - tar silently is padding with zeros (for Read error at byte...), should say it is Padding with zeros
https://bugzilla.redhat.com/show_bug.cgi?id=1913566