Proper host name validation was not being performed for certificates
kept in the local keystore. If an attacker could convince a user to
accept and store an attacker's certificate, then that certificate
could be used for MITM attacks, giving the attacker access to all
connections to all servers in all accounts in K-9.
This commit changes how the certificates are stored. Previously, an
entire certificate chain was stored for a server (and any of those
certificates in the chain were available for validating signatures on
certificates received when connecting). Now just the single
certificate for the server is stored.
This commit changes how locally stored certificates are retrieved.
They can only be retrieved using the host:port that the user
configured for the server.
This also fixes issue 1326. Users can now use different certificates
for different servers on the same host (listening to different ports).
The above changes mean that users might have to re-accept certificates
that they had previously accepted and are still using (but only if the
certificate's Subject doesn't match the host that they are connecting
to).
This commit modifies AccountSetupBasics so that it now calls
AccountSetupCheckSettings twice -- once for checking the incoming
settings and once for the outgoing settings. Otherwise, an exception
could occur while checking incoming settings, the user could say
continue (or the user could accept a certificate key), and the
outgoing settings would not be checked. This also helps with
determining if a certificate exception was for the incoming or
outgoing server, which is needed if the user decides to add the
certificate to the keystore.
The TrustedSocketFactory, which provides goodies like better cipher suites and
TLSv1.2, was only being used for tunnelled connections. Use it for STARTTLS
connections as well.
As Georg Lukas wrote in his blog post about how Android handles TLS
handshake (http://op-co.de/blog/posts/android_ssl_downgrade/), an
explicit order of cipher suites and TLS versions must be supplied to
avoid having the weak (presumably broken) RC4 cipher at the top of the
preference list.
This commit adds the list included in the blog post to every TLS socket
creation, including IMAP, POP3 and SMTP, see Wireshark screenshots done
during testing at http://vsza.hu/k9mail-tls-hardening/
This builds upon the efforts started 2 commits back where \r\n is used for
all message text and \n is only used when the text is inside an
EolConvertingEditText widget.
The problem:
Configure the account (just an example -- problems can occur in other
configurations as well):
Message Format: HTML
Reply quoting style: Prefix
Quote message when replying: yes
Reply after quoted text: yes
Reply to a message that has a large quantity (20+) of \r\n scattered in
the body of its HTML version (not an unusual scenario).
Add a reply. Save the message as a draft. Go back & open the draft
again. A fatal IndexOutOfBoundsException occurs.
The cause:
When the draft was saved, the X-K9mail-Identity header was computed and
added to the message, then the text of the message was processed with
MimeUtility.fixDraftTextBody, replacing all occurrences of \r\n with \n in
the quoted message before being saved in LocalStore, thus invalidating the
X-K9mail-Identity header.
The fix:
Remove MimeUtility.fixDraftTextBody and implement
MessageCompose$EolConvertingEditText instead. Any message text placed in
an EolConvertingEditText widget is assured to have \n line endings. Any
message text extracted from an EolConvertingEditText widget is assured to
have \r\n line endings. The X-K9mail-Identity header will always be
computed correctly.
Issues thought to be related: 4782, 5010, 5634, 5725
As noted in some of the referenced issues, errors didn't always result in
a fatal exception, but instead with mixed up text.
Ref: commit f9a35aeaee
Fixes assignment problems for emails sent by some issue tracking
systems, which send out mails with a fixed mail address on behalf of
different people.
Update LocalStore code to handle the newly introduced temporary files
for attachments
Conflicts:
res/values/strings.xml
src/com/fsck/k9/activity/MessageCompose.java
Android allows other apps to access protected content of an app without requesting the
necessary permission when the app returns an Intent with FLAG_GRANT_READ_URI_PERMISSION.
This regularly happens as a result of ACTION_GET_CONTENT, i.e. what we use to pick content
to be attached to a message. Accessing that content only works while the receiving activity
is running. Afterwards accessing the content throws a SecurityException because of the
missing permission.
This commit changes K-9 Mail's behavior to copy the content to a temporary file in K-9's
cache directory while the activity is still running.
Fixes issue 4847, 5821
This also fixes bugs related to the fact that K-9 Mail didn't save a copy of attached content
in the message database.
Fixes issue 1187, 3330, 4930
Don't convert the content-type to lower case in
MimeMessage.getContentType. The content-type may have optional parameters
that are case sensitive (boundary, name).
In removing the lower-case conversion from getContentType, a review was
made for inappropriate case-sensitive comparisons which use data obtained
with getContentType. The only ones found were in isMimeType in both
Message and MimeBodyPart.
Case-sensitive instances of isMimeType were made case-insensitive. Also,
isMimeType was moved from Message to MimeMessage for symmetry with
MimeBodyPart (MimeMessage & MimeBodyPart are similar and contain a good
bit of duplication such as this).
The unit test required fixing now that the case of the boundary text is
preserved.
References:
Commits 2c5186 and dc4002 added the toLowerCase to getContentType in
MimeMessage & MimeBodyPart (Issue 94).
Later, commit 50cd60 removed the toLowerCase addition from MimeBodyPart
(Issue 1289).
Fix the unit test to match.
All line endings in the unit test are now the same.
(Just for consistency. Not a big deal, since such problems are fixed when
the messages are run through EOLConvertingOutputStream.)
The preceding commit resulted in attachments of type message/rfc822 being
sent with 8bit encoding even when the SMTP server did not support
8BITMIME. This commit assures that messages will be converted to 7bit
when necessary.
A new interface CompositeBody was created that extends Body, and classes
Message and Multipart were changed from implementing Body to
CompositeBody. Additional classes BinaryTempFileMessageBody and
LocalAttachmentMessageBody were created (by extending BinaryTempFileBody
and LocalAttachmentBody, respectively), and they too implement
CompositeBody.
A CompositeBody is a Body containing a composite-type that can contain
subparts that may require recursive processing when converting from 8bit
to 7bit. The Part to which a CompositeBody belongs is only permitted to
use 8bit or 7bit encoding for the CompositeBody.
Previously, a Message was created so that it was 7bit clean by default
(even though that meant base64 encoding all attachments, including
messages). Then, if the SMTP server supported 8BITMIME,
Message.setEncoding("8bit") was called so that bodies of type TextBody
would been transmitted using 8bit encoding rather than quoted-printable.
Now, messages are created with 8bit encoding by default. Then, if the
SMTP server does not support 8BITMIME, Message.setUsing7bitTransport is
called to recursively convert the message and its subparts to 7bit. The
method setUsing7bitTransport was added to the interfaces Part and
CompositeBody.
setEncoding no longer iterates over parts in Multipart. That task belongs
to setUsing7bitTransport, which may in turn call setEncoding on the parts.
MimeUtility.getEncodingforType was created as a helper function for
choosing a default encoding that should be used for a given MIME type when
an attachment is added to a message (either while composing or when
retrieving from LocalStore).
setEncoding was implemented in MimeBodyPart to assure that the encoding
set in the Part's headers was the same as set for the Part's Body. (The
method already existed in MimeMessage, which has similarities with
MimeBodyPart.)
MimeMessage.parse(InputStream in, boolean recurse) was implemented so that
the parser could be told to recursively process nested messages read from
the InputStream, thus giving access to all subparts at any level that may
need to be converted from 8bit to 7bit.
The problem: Receive a message with an attachment of type message/rfc822
and forward it. When the message is sent, K-9 Mail uses base64 encoding
for the attachment. (Alternatively, you could compose a new message and
add such an attachment from a file using a filing-picking app, but that is
not 100% effective because the app may not choose the correct
message/rfc822 MIME type for the attachment.)
Such encoding is prohibited per RFC 2046 (5.2.1) and RFC 2045 (6.4). Only
8bit or 7bit encoding is permitted for attachments of type message/rfc822.
Thunderbird refuses to decode such attachments. All that is shown is the
base64 encoded body.
This commit implements LocalAttachmentBody.setEncoding. If an attachment
to a newly composed message is itself a message, then setEncoding("8bit")
is called, otherwise setEncoding("base64") is called for the attachment.
Similar behavior occurs when an attachment is retrieved from LocalStore.
The setEncoding method was added to the Body interface, since all
implementations of Body now declare the method.
The problem here differs from that in the preceding commit: Here, the
encoding problem occurs on sending, not on receipt. Here, the entire
message (headers and body) is base64 encoded, not just the body. Here,
the headers correctly identify the encoding used; it's just that the RFC
does not permit such encoding of attached messages. The problem here
could in fact occur in combination with the preceding problem.
Issue 5734 exemplifies the problem: receive a message with an attachment
of type message/rfc822 that doesn't use base64 encoding for the body of
the attached message. K-9 Mail incorrectly stores the attached message
locally with its original headers but using base64 encoding for the body.
A discrepancy thus exists between what the headers say about the encoding
of the body versus the actual encoding used. This is obvious when
attempting to view the attachment (either by using a compatible message
viewer available on the device or by saving the attachment to a file and
viewing the file contents).
The process: When a message with an attached sub-message is received,
Message.parse puts the attachment in a new MimeMessage with the
attachment's body in a BinaryTempFileBody. LocalFolder.saveAttachment
then calls Message.writeTo (which later calls BinaryTempFileBody.writeTo)
to place the entire attachment (headers and body) in a new file that will
become a LocalAttachmentBody. Until now, BinaryTempFileBody.writeTo
could only save the message body using base64 encoding.
This commit implements BinaryTempFileBody.setEncoding and assures that the
body is written out with the same encoding that was found in its headers.
Currently, K-9 Mail detects if an SMTP server supports 8BITMIME (RFC
6152), and if so, TextBody parts are sent with content-transfer-ecoding =
8bit. Otherwise, they are sent using quoted-printable.
This adds the required "BODY=8BITMIME" parameter to the MAIL command when
sending messages to servers that support 8BITMIME.
If you attempted to use SSL to connect to a server that speaks
STARTTLS, you should get an SSL protocol error. Instead, you
were likely to get an "Unrecognized Certificate" error that shows
you an unrelated certificate chain and asks you to accept it or
reject it. Neither action would work because the actual problem
had nothing to do with certificates. The unrelated certificate
chain that popped up had been statically stored when validating
a prior connection to a different server.
With this patch, certificate chains are no longer stored statically
when validating server connections.
Issue 5886 is an example of a user experiencing this problem.
Previously the app crashed when upgrading the database failed. Now we
reset the database version and run the upgrade code again (recreating
all tables).
This reverts commit bbdec62e37.
Aside from being the incorrect solution for fixing the problem
described in pull request 211, the patch generates 'Dead code'
warnings inside the if(){} statements on lines 46 and 47.
The correct fix for the problem was already implemented in commit
5678786c97.
Although the logcat in the pull request was generated after the fix,
line numbers in the log indicate that it was based on an outdated
version of MimeUtility.java from before the fix.
This requires another database schema change. With this change messages
at the root of a thread reference themselves in the 'threads' table,
i.e. 'root' contains the value of 'id' for these messages. It makes
selecting all messages in a thread much simpler.
Previously, <html>, <head>, & <body> tags were
attached to messages before they were stored locally.
But now that the <head> element also needs to include
a <meta> element (for proper MessageWebView display),
it seems unecesary to store all these tags with each
message.
Now the tags are no longer stored with the messages. Instead,
MessageWebView applies the tags before displaying the message.
This also eliminates the need to upgrade an older
message database where all the old messages would have
otherwise needed to be wrapped with the new tags.
Database updates can be surprisingly slow. This lead to slow updates of
the user interface which in turn made working with K-9 Mail not as fun
as it should be. This commit hopefully changes that.
The commit that introduced those notifications also introduced a rather
... interesting design pattern: The CertificateValidationException
notified the user of its pure existance - it's no longer a 'message'
only, but defines policy. As this is more than unusual, replace this
pattern by the MessagingController treating
CertificateValidationException specially when accessing remote folders.
Also make clear which account failed when constructing the notification.
With this fix, a CertPathValidatorException or CertificateException will
create a "Certificate error: Check your server settings" notification
in the status bar. When the user clicks on the notification, they are
taken to the appropriate server settings screen where they can review their
settings and can accept a different server certificate.
LocalMessage already has a content preview in it; reuse that.
Remove unneeded MimeMessage#getPreview() method now that we don't need
to generate a preview anymore.
This changes the interface to MessagingController and the way flags are
updated in the database. Now messages aren't changed one by one but in
batches of 500. This should give better performance, but breaks the
unread and flagged count. I'm not very sad about this, because now we
can move towards only displaying the number of unread/flagged messages
in the local database.
