Sunday, March 26, 2017

How to prevent Wine from auto-running .exe files in Fedora

After installing Wine on Fedora 25 (for a reason that isn't terribly important here) I've noticed that I can run Window executables straight from the bash command line:

$ file -b ~/.wine/drive_c/Program\ Files/Internet\ Explorer/iexplore.exe
PE32+ executable (GUI) x86-64, for MS Windows
$ !$

That iexplore.exe file is clearly not a ELF, yet the kernel successfully executes it.

Since Windows ecosystems is known for its total absence of malware & ransomware; & Wine, in turn, is known for a military grade, bulletproof sandbox (i.e., it provides no protection whatsoever), I find the notion of auto-running very difficult to reconcile w/ common sence.

How does it work

If a kernel was compiled w/ CONFIG_BINFMT_MISC option, execve(2) (I'm simplifying here a little) obtains an ability to delegate the execution of unknown binaries to external processes. The subsystem that manages the format → interpreter association table is called binfmt_misc.

binfmt_misc maintains its ephemeral database in the procfs. At the boot time you mount /proc/sys/fs/binfmt_misc directory & feed /proc/sys/fs/binfmt_misc/register file w/ specially crafted text lines to create association entries (or rules as systemd docs like to call them).

E.g., in the good old days before systemd, we would have put

echo :win:M:0:MZ::/usr/bin/wine: > /proc/sys/fs/binfmt_misc/register

somewhere in /etc/rc.local.

Such a command creates /proc/sys/fs/binfmt_misc/win file:

enabled
interpreter /usr/bin/wine
flags:
offset 0
magic 4d5a

where the offset & the magic 4d5a (MZ) means the 1st 2 bytes of a typical Windows executable:

$ hexdump -C -n10 iexplore.exe
00000000  4d 5a 40 00 01 00 00 00  06 00                    |MZ@.......|
0000000a

We can be even more fancier & make an extension → interpreter association, e.g.:

# echo :xv:E::gif::/usr/bin/xv: > /proc/sys/fs/binfmt_misc/register
$ chmod +x slave-ship-daily-schedules.gif
$ ./!$

creates .gif → xv entry, & starts /usr/bin/xv slave-ship-daily-schedules.gif under the hood when we try to "run" the .gif image.

To delete all the entries, we write -1 to /proc/sys/fs/binfmt_misc/status file or if we want to delete only a particular entry:

# echo -1 > /proc/sys/fs/binfmt_misc/xv

The systemd way

systemd doesn't allow us to mingle w/ binfmt_misc subsystem directly. We ought to write the text lines in the same format binfmt_misc undestands but put them in special .conf files, where a separate program systemd-binfmt expects to find them.

If we provide an rpm for our app, we should put my-app.conf file in /usr/lib/binfmt.d/ directory during the installation & run

/usr/lib/systemd/systemd-binfmt my-app.conf

in the post-install hook.

The algo systemd-binfmt uses is fairly straightforward. When run w/o any command line args, it deletes all the existing entries & recreates the ones specified in the .conf files. If provided w/ the name of the .conf file (w/o a directory prefix!), it scans the file to add new entries.

An excerpt from src/binfmt/binfmt.c (commit 1539a65):

if (argc > optind) {
  int i;

  for (i = optind; i < argc; i++) {
    k = apply_file(argv[i], false);
    if (k < 0 && r == 0)
      r = k;
  }
} else {
  _cleanup_strv_free_ char **files = NULL;
  char **f;

  r = conf_files_list_nulstr(&files, ".conf", NULL, conf_file_dirs);
  if (r < 0) {
    log_error_errno(r, "Failed to enumerate binfmt.d files: %m");
    goto finish;
  }

  /* Flush out all rules */
  write_string_file("/proc/sys/fs/binfmt_misc/status", "-1", 0);

  STRV_FOREACH(f, files) {
    k = apply_file(*f, true);
    if (k < 0 && r == 0)
      r = k;
  }
}

It sounds fine, until we remember that a typical user (a) doesn't know much about binfmt-misc kernel module, (b) doesn't know anything about the standalone systemd-binfmt program, for he uses systemd-binfmt.service unit, as in

# systemctl restart systemd-binfmt

That unit (/usr/lib/systemd/system/systemd-binfmt.service) incorporates a handful of preconditions. The relevant ones:

ConditionDirectoryNotEmpty=|/lib/binfmt.d
ConditionDirectoryNotEmpty=|/usr/lib/binfmt.d
ConditionDirectoryNotEmpty=|/usr/local/lib/binfmt.d
ConditionDirectoryNotEmpty=|/etc/binfmt.d
ConditionDirectoryNotEmpty=|/run/binfmt.d

When we install Wine, it makes 2 entries in the systemd-binfmt DB. If we (a) remove the offending package wine-systemd that contains the evil /usr/lib/binfmt.d/wine.conf file & (b) dutifully restart systemd-binfmt service--no binfmt-misc association entries gets removed!

# rpm -qf /lib/binfmt.d/wine.conf
wine-systemd-2.3-1.fc25.noarch
# rpm --nodeps -e wine-systemd
# systemctl restart systemd-binfmt
# ls -l /proc/sys/fs/binfmt_misc/
total 0K
--w------- 1 root root 0 Mar 24 20:49 register
-rw-r--r-- 1 root root 0 Mar 25 20:48 status
-rw-r--r-- 1 root root 0 Mar 25 20:58 windows
-rw-r--r-- 1 root root 0 Mar 25 20:58 windowsPE

because systemd will forbid systemd-binfmt to run at all, because all the conf directories are empty. The remedy is to run systemd-binfmt by hand:

# /usr/lib/systemd/systemd-binfmt
# ls -l /proc/sys/fs/binfmt_misc/
total 0K
--w------- 1 root root 0 Mar 24 20:49 register
-rw-r--r-- 1 root root 0 Mar 25 20:48 status

You should not manually delete the rpm w/ the evil .conf file for dnf will reinstall wine-systemd package during the next Wine update. The recommended systemd-style solution is to create a symlink in /etc/binfmt.d/ to /dev/null named wine.conf & then restart systemd-binfmt service:

# ln -s /dev/null /etc/binfmt.d/wine.conf
# systemctl restart systemd-binfmt

Saturday, March 4, 2017

Bloatware comes when nobody's lookin'

If you write a Chrome extension for distribution outside of Google Webstore, you will inevitably need to pack the extension into a .crx file. It can be done either using the Chrome UI (the "Pack extension..." button) or "manually" via creating a .zip file & prepending a specially crafted header to it.

Obviously, the 2nd variant can be easily automated. If you have several extensions to maintain, a part of your build system that's responsible for the .crx generation could be extracted to some kind of a "plugin" that can be shared between all the extensions you maintain. In my case, the "plugin" consists of 2 files: a small sh script & a tiny makefile (that can be safely included into another makefile).

I was thinking about uploading those 2 files to npm (& hesitating a little for if a package wouldn't contain any JS code soever, should it be on npm?) but then decided to check the registry first.

The npm registry indeed contains multiple packages for dealing w/ .crx files. One of the most popular is called "crx" (at the time of writing it had 192 ★ on Github). Glancing through its code I failed not to notice the unfortunate difference between the classic Unix approach to such a problem & the JavaScript one. Perhaps, the latter is everything you may do ironically, unless you're a grand, hardcore troll.

Before I explain myself further, let's digress for a bit to explain what .crx files are & how to create them.

CRX Package Format

In the ideal world, Alice would package her extension (a set of .js & .json files) into a .zip file, would upload the file to her http://cs.a-nice-university.edu/~alice/ page & would tell her dear friend Bob about it.

In reality, the extension must be protected from tampering. If Eve, the evil sysadmin of cs.a-nice-university.edu zone, will entertain an idea of mingling w/ Alice's extension via adding a hot Bitcoin miner to it, Bob should be able to detect that before the extension gets intalled.

One way of doing it is to generate a pair of public & private keys. You sign the extension w/ your private key & users check the downloaded .zip w/ you public key. The trouble is there is no standard way to "sign" a .zip file for neither its metadata supports such a thing as an embedded signature, not any of "archive managers" would know what to do w/ such an upgraded metadata.

This is where .crx files come in: they contain a copy of an RSA public key + an encrypted SHA1 of a .zip archive in question. This metadata simply gets prepended to the original .zip file.

http://gromnitsky.users.sourceforge.net/articles/crx.svg

E.g., Alice, after testing her Chrome extension, zips it to a file:

$ touch manifest.json
$ zip foo.zip !$

Next, she generates a private 1024-bit RSA key:

$ openssl genrsa 1024 > private.pem

Then prepends the aforementioned block to foo.zip. As it's a multiple step process, she writes it down in zip2crx script (this is a modified version of the script from developer.chrome.com):

#!/bin/sh

[ $# -ne 2 ] && {
    echo "Usage: `basename $0` file.zip private_key" 1>&2
    exit 1
}

zip=$1
key=$2

name=${zip%.zip}
crx="$name.crx"
pub="$name.pub"
sig="$name.sig"
zip="$name.zip"
trap 'rm -f "$pub" "$sig"' EXIT

# signature
openssl sha1 -sha1 -binary -sign "$key" < "$zip" > "$sig"

# public key
openssl rsa -pubout -outform DER < "$key" > "$pub" 2>/dev/null

# Take "abcdefgh" and return it as "ghefcdab"
byte_swap() {
    echo "${1:6:2}${1:4:2}${1:2:2}${1:0:2}"
}

crmagic_hex="4372 3234" # Cr24
version_hex="0200 0000" # 2
pub_len_hex=$(byte_swap $(printf '%08x\n' $(stat -c %s "$pub")))
sig_len_hex=$(byte_swap $(printf '%08x\n' $(stat -c %s "$sig")))

(
    echo "$crmagic_hex $version_hex $pub_len_hex $sig_len_hex" | xxd -r -p
    cat "$pub" "$sig" "$zip"
) > "$crx"

Her script signs the .zip, automatically derives a public key from private.pem, combines together all the necessary data for a .crx file consumer (like the length of the private key, &c) & concatenates the obtained header w/ the .zip

$ ./zip2crx foo.zip private.pem
$ file foo.crx
foo.crx: Google Chrome extension, version 2

The resulting .crx is ready to be used in Chrome. She uploads it to her web page & sends to Bob (via email) her public key (possibly encrypted w/ his GPG public key, but we won't get into that).

Bob, having obtained Alice's public key, downloads foo.crx, extracts from it the embedded public key, the signature & foo.zip & checks (w/ Alice's public key) the validity of foo.zip.

(The whole verification process is a little too convoluted to present it here, but if you're interested, download this script & run it against the provided key:

$ ./crx2zip foo.crx alice.pub
RSA key                 1024-bit
Total header size:      306 bytes
Public key:             alice.pub
Signature status:       Verified OK

where alice.pub would be a public key in the DER-format.)

If Eve, the evil sysadmin, did indeed modify foo.crx in any way, Bob is able to detect that, for despite that Eve can do all the same operations as Bob did, she doesn't have Alice's private key & thus is unable to re-sign the tampered foo.zip properly in an undetectable way. All she can hope for is that Bob, being a lazybones, won't bother to do the necessary checks before installing foo.crx into his browser.

The example is somewhat contrived, for if Alice decides to upload foo.zip to Google Webstore instead, by this virtue she exempts herself from managing the crypto keys. The Webstore makes the key pair for her, does all the checks of all the sub-sequential updates of foo.zip & generates the correct .crx. The final extension is being delivered to Bob via HTTPS, thus leaving Eve, the evil sysadmin, out of luck.

The JavaScript way

Back to our findings from the npm registry.

crx package has dual nature: it's a library & a CLI util. The distinguishing feature of this program is that "It is written purely in JavaScript and does not require OpenSSL!".

The author states it as if the mere act of depending on OpenSSL code is somehow inconvenient or morally repugnant. I must say he's not alone in his view. I may sympatise for the overly cautious approach in the case of maintaining a big farm of cloud services like Amazon does, but I cannot reconcile w/ this stance in the case of a local developer machine.

A user of a .crx file doesn't have to interfere w/ OpenSSL, it's solely the developer's job to create the proper .crx. To find a developer machine that doesn't have OpenSSL intalled already is a challenge that only a few of us can achieve. (Certainly, we may be so bold; we may think of some poor souls who have to use old versions of Windows but even they can always download Cygwin.)

Then there is the size. Our zip2crx example is 37 lines long. Anybody can read it (even indolent Bob) & in a minute understand what the script is doing. crx package on the other hand

$ find node_modules/crx/{bin,src} -type f | xargs wc -l
  150 node_modules/crx/bin/crx.js
  294 node_modules/crx/src/crx.js
   47 node_modules/crx/src/resolver.js
  491 total

is ~13 times bigger.

The package also provides some kind of an artisan build system surrogate. It "packs" the source code into different places depending on its command line options. It generates RSA keys.

I know that it's quite fashionable in the JavaScript world to write a replacement for Make every year, but it still amazes me why would anyone do the job in the courageous crx package way, instead of reusing available, well-tested tools on your machine.

The Unix way

Everything except zip2crx is already here. Suppose you have foobar extension:

foobar/
├── Makefile
├── src/
│   ├── hello.js
│   └── manifest.json
└── zip2crx*

src directory contains the source code for the extension. Makefile is a primitive 18-lines long set of shell instructions:

pkg.name := foobar
out := _build
src := $(shell find src -type f)

mkdir = @mkdir -p $(dir $@)     # a canned recipe

.PHONY: crx
crx: $(out)/$(pkg.name).crx

$(out)/$(pkg.name).zip: $(src)
        $(mkdir)
        cd $(dir $<) && zip -qr $(CURDIR)/$@ *

%.crx: %.zip private.pem
        ./zip2crx $< private.pem

private.pem:
        openssl genrsa 2048 > $@

If you type make it'll generate in _build/ directory foobar.zip & foobar.crx. If you don't have an RSA key, Make will generate it for you too.

Why would you need JavaScript for that?

I'll conclude by quoting Doug McIlroy:

"A first engineering question to ask is: how often is one likely to have to do this exact task? Not at all often, I contend. It is plausible, though, that similar, but not identical, problems might arise. A wise engineering solution would produce--or better, exploit--reusable parts."