most content is in markdown files in website/, some is taken out of the repo
README and rfc/index.txt. a Go file generates html. static files are kept in a
separate repo due to size.
the http basic auth we had was very simple to reason about, and to implement.
but it has a major downside:
there is no way to logout, browsers keep sending credentials. ideally, browsers
themselves would show a button to stop sending credentials.
a related downside: the http auth mechanism doesn't indicate for which server
paths the credentials are.
another downside: the original password is sent to the server with each
request. though sending original passwords to web servers seems to be
considered normal.
our new approach uses session cookies, along with csrf values when we can. the
sessions are server-side managed, automatically extended on each use. this
makes it easy to invalidate sessions and keeps the frontend simpler (than with
long- vs short-term sessions and refreshing). the cookies are httponly,
samesite=strict, scoped to the path of the web interface. cookies are set
"secure" when set over https. the cookie is set by a successful call to Login.
a call to Logout invalidates a session. changing a password invalidates all
sessions for a user, but keeps the session with which the password was changed
alive. the csrf value is also random, and associated with the session cookie.
the csrf must be sent as header for api calls, or as parameter for direct form
posts (where we cannot set a custom header). rest-like calls made directly by
the browser, e.g. for images, don't have a csrf protection. the csrf value is
returned by the Login api call and stored in localstorage.
api calls without credentials return code "user:noAuth", and with bad
credentials return "user:badAuth". the api client recognizes this and triggers
a login. after a login, all auth-failed api calls are automatically retried.
only for "user:badAuth" is an error message displayed in the login form (e.g.
session expired).
in an ideal world, browsers would take care of most session management. a
server would indicate authentication is needed (like http basic auth), and the
browsers uses trusted ui to request credentials for the server & path. the
browser could use safer mechanism than sending original passwords to the
server, such as scram, along with a standard way to create sessions. for now,
web developers have to do authentication themselves: from showing the login
prompt, ensuring the right session/csrf cookies/localstorage/headers/etc are
sent with each request.
webauthn is a newer way to do authentication, perhaps we'll implement it in the
future. though hardware tokens aren't an attractive option for many users, and
it may be overkill as long as we still do old-fashioned authentication in smtp
& imap where passwords can be sent to the server.
for issue #58
we were already accepting, processing and displaying incoming tls reports. now
we start tracking TLS connection and security-policy-related errors for
outgoing message deliveries as well. we send reports once a day, to the
reporting addresses specified in TLSRPT records (rua) of a policy domain. these
reports are about MTA-STS policies and/or DANE policies, and about
STARTTLS-related failures.
sending reports is enabled by default, but can be disabled through setting
NoOutgoingTLSReports in mox.conf.
only at the end of the implementation process came the realization that the
TLSRPT policy domain for DANE (MX) hosts are separate from the TLSRPT policy
for the recipient domain, and that MTA-STS and DANE TLS/policy results are
typically delivered in separate reports. so MX hosts need their own TLSRPT
policies.
config for the per-host TLSRPT policy should be added to mox.conf for existing
installs, in field HostTLSRPT. it is automatically configured by quickstart for
new installs. with a HostTLSRPT config, the "dns records" and "dns check" admin
pages now suggest the per-host TLSRPT record. by creating that record, you're
requesting TLS reports about your MX host.
gathering all the TLS/policy results is somewhat tricky. the tentacles go
throughout the code. the positive result is that the TLS/policy-related code
had to be cleaned up a bit. for example, the smtpclient TLS modes now reflect
reality better, with independent settings about whether PKIX and/or DANE
verification has to be done, and/or whether verification errors have to be
ignored (e.g. for tls-required: no header). also, cached mtasts policies of
mode "none" are now cleaned up once the MTA-STS DNS record goes away.
getting mox to compile required changing code in only a few places where
package "syscall" was used: for accessing file access times and for umask
handling. an open problem is how to start a process as an unprivileged user on
windows. that's why "mox serve" isn't implemented yet. and just finding a way
to implement it now may not be good enough in the near future: we may want to
starting using a more complete privilege separation approach, with a process
handling sensitive tasks (handling private keys, authentication), where we may
want to pass file descriptors between processes. how would that work on
windows?
anyway, getting mox to compile for windows doesn't mean it works properly on
windows. the largest issue: mox would normally open a file, rename or remove
it, and finally close it. this happens during message delivery. that doesn't
work on windows, the rename/remove would fail because the file is still open.
so this commit swaps many "remove" and "close" calls. renames are a longer
story: message delivery had two ways to deliver: with "consuming" the
(temporary) message file (which would rename it to its final destination), and
without consuming (by hardlinking the file, falling back to copying). the last
delivery to a recipient of a message (and the only one in the common case of a
single recipient) would consume the message, and the earlier recipients would
not. during delivery, the already open message file was used, to parse the
message. we still want to use that open message file, and the caller now stays
responsible for closing it, but we no longer try to rename (consume) the file.
we always hardlink (or copy) during delivery (this works on windows), and the
caller is responsible for closing and removing (in that order) the original
temporary file. this does cost one syscall more. but it makes the delivery code
(responsibilities) a bit simpler.
there is one more obvious issue: the file system path separator. mox already
used the "filepath" package to join paths in many places, but not everywhere.
and it still used strings with slashes for local file access. with this commit,
the code now uses filepath.FromSlash for path strings with slashes, uses
"filepath" in a few more places where it previously didn't. also switches from
"filepath" to regular "path" package when handling mailbox names in a few
places, because those always use forward slashes, regardless of local file
system conventions. windows can handle forward slashes when opening files, so
test code that passes path strings with forward slashes straight to go stdlib
file i/o functions are left unchanged to reduce code churn. the regular
non-test code, or test code that uses path strings in places other than
standard i/o functions, does have the paths converted for consistent paths
(otherwise we would end up with paths with mixed forward/backward slashes in
log messages).
windows cannot dup a listening socket. for "mox localserve", it isn't
important, and we can work around the issue. the current approach for "mox
serve" (forking a process and passing file descriptors of listening sockets on
"privileged" ports) won't work on windows. perhaps it isn't needed on windows,
and any user can listen on "privileged" ports? that would be welcome.
on windows, os.Open cannot open a directory, so we cannot call Sync on it after
message delivery. a cursory internet search indicates that directories cannot
be synced on windows. the story is probably much more nuanced than that, with
long deep technical details/discussions/disagreement/confusion, like on unix.
for "mox localserve" we can get away with making syncdir a no-op.
so users can understand what is going on. e.g. a CAA record that doesn't allow
a CA to sign a certificate. previously, the error message would just be "no
viable challenge type found", which doesn't help the user.
the vendored dns resolver code is a copy of the go stdlib dns resolver, with
awareness of the "authentic data" (i.e. dnssec secure) added, as well as support
for enhanced dns errors, and looking up tlsa records (for dane). ideally it
would be upstreamed, but the chances seem slim.
dnssec-awareness is added to all packages, e.g. spf, dkim, dmarc, iprev. their
dnssec status is added to the Received message headers for incoming email.
but the main reason to add dnssec was for implementing dane. with dane, the
verification of tls certificates can be done through certificates/public keys
published in dns (in the tlsa records). this only makes sense (is trustworthy)
if those dns records can be verified to be authentic.
mox now applies dane to delivering messages over smtp. mox already implemented
mta-sts for webpki/pkix-verification of certificates against the (large) pool
of CA's, and still enforces those policies when present. but it now also checks
for dane records, and will verify those if present. if dane and mta-sts are
both absent, the regular opportunistic tls with starttls is still done. and the
fallback to plaintext is also still done.
mox also makes it easy to setup dane for incoming deliveries, so other servers
can deliver with dane tls certificate verification. the quickstart now
generates private keys that are used when requesting certificates with acme.
the private keys are pre-generated because they must be static and known during
setup, because their public keys must be published in tlsa records in dns.
autocert would generate private keys on its own, so had to be forked to add the
option to provide the private key when requesting a new certificate. hopefully
upstream will accept the change and we can drop the fork.
with this change, using the quickstart to setup a new mox instance, the checks
at internet.nl result in a 100% score, provided the domain is dnssec-signed and
the network doesn't have any issues.
it was far down on the roadmap, but implemented earlier, because it's
interesting, and to help prepare for a jmap implementation. for jmap we need to
implement more client-like functionality than with just imap. internal data
structures need to change. jmap has lots of other requirements, so it's already
a big project. by implementing a webmail now, some of the required data
structure changes become clear and can be made now, so the later jmap
implementation can do things similarly to the webmail code. the webmail
frontend and webmail are written together, making their interface/api much
smaller and simpler than jmap.
one of the internal changes is that we now keep track of per-mailbox
total/unread/unseen/deleted message counts and mailbox sizes. keeping this
data consistent after any change to the stored messages (through the code base)
is tricky, so mox now has a consistency check that verifies the counts are
correct, which runs only during tests, each time an internal account reference
is closed. we have a few more internal "changes" that are propagated for the
webmail frontend (that imap doesn't have a way to propagate on a connection),
like changes to the special-use flags on mailboxes, and used keywords in a
mailbox. more changes that will be required have revealed themselves while
implementing the webmail, and will be implemented next.
the webmail user interface is modeled after the mail clients i use or have
used: thunderbird, macos mail, mutt; and webmails i normally only use for
testing: gmail, proton, yahoo, outlook. a somewhat technical user is assumed,
but still the goal is to make this webmail client easy to use for everyone. the
user interface looks like most other mail clients: a list of mailboxes, a
search bar, a message list view, and message details. there is a top/bottom and
a left/right layout for the list/message view, default is automatic based on
screen size. the panes can be resized by the user. buttons for actions are just
text, not icons. clicking a button briefly shows the shortcut for the action in
the bottom right, helping with learning to operate quickly. any text that is
underdotted has a title attribute that causes more information to be displayed,
e.g. what a button does or a field is about. to highlight potential phishing
attempts, any text (anywhere in the webclient) that switches unicode "blocks"
(a rough approximation to (language) scripts) within a word is underlined
orange. multiple messages can be selected with familiar ui interaction:
clicking while holding control and/or shift keys. keyboard navigation works
with arrows/page up/down and home/end keys, and also with a few basic vi-like
keys for list/message navigation. we prefer showing the text instead of
html (with inlined images only) version of a message. html messages are shown
in an iframe served from an endpoint with CSP headers to prevent dangerous
resources (scripts, external images) from being loaded. the html is also
sanitized, with javascript removed. a user can choose to load external
resources (e.g. images for tracking purposes).
the frontend is just (strict) typescript, no external frameworks. all
incoming/outgoing data is typechecked, both the api request parameters and
response types, and the data coming in over SSE. the types and checking code
are generated with sherpats, which uses the api definitions generated by
sherpadoc based on the Go code. so types from the backend are automatically
propagated to the frontend. since there is no framework to automatically
propagate properties and rerender components, changes coming in over the SSE
connection are propagated explicitly with regular function calls. the ui is
separated into "views", each with a "root" dom element that is added to the
visible document. these views have additional functions for getting changes
propagated, often resulting in the view updating its (internal) ui state (dom).
we keep the frontend compilation simple, it's just a few typescript files that
get compiled (combined and types stripped) into a single js file, no additional
runtime code needed or complicated build processes used. the webmail is served
is served from a compressed, cachable html file that includes style and the
javascript, currently just over 225kb uncompressed, under 60kb compressed (not
minified, including comments). we include the generated js files in the
repository, to keep Go's easily buildable self-contained binaries.
authentication is basic http, as with the account and admin pages. most data
comes in over one long-term SSE connection to the backend. api requests signal
which mailbox/search/messages are requested over the SSE connection. fetching
individual messages, and making changes, are done through api calls. the
operations are similar to imap, so some code has been moved from package
imapserver to package store. the future jmap implementation will benefit from
these changes too. more functionality will probably be moved to the store
package in the future.
the quickstart enables webmail on the internal listener by default (for new
installs). users can enable it on the public listener if they want to. mox
localserve enables it too. to enable webmail on existing installs, add settings
like the following to the listeners in mox.conf, similar to AccountHTTP(S):
WebmailHTTP:
Enabled: true
WebmailHTTPS:
Enabled: true
special thanks to liesbeth, gerben, andrii for early user feedback.
there is plenty still to do, see the list at the top of webmail/webmail.ts.
feedback welcome as always.
the mailbox select/examine responses now return all flags used in a mailbox in
the FLAGS response. and indicate in the PERMANENTFLAGS response that clients
can set new keywords. we store these values on the new Message.Keywords field.
system/well-known flags are still in Message.Flags, so we're recognizing those
and handling them separately.
the imap store command handles the new flags. as does the append command, and
the search command.
we store keywords in a mailbox when a message in that mailbox gets the keyword.
we don't automatically remove the keywords from a mailbox. there is currently
no way at all to remove a keyword from a mailbox.
the import commands now handle non-system/well-known keywords too, when
importing from mbox/maildir.
jmap requires keyword support, so best to get it out of the way now.
before, we used heuristics to decide when to train/untrain a message as junk or
nonjunk: the message had to be seen, be in certain mailboxes. then if a message
was marked as junk, it was junk. and otherwise it was nonjunk. this wasn't good
enough: you may want to keep some messages around as neither junk or nonjunk.
and that wasn't possible.
ideally, we would just look at the imap $Junk and $NotJunk flags. the problem
is that mail clients don't set these flags, or don't make it easy. thunderbird
can set the flags based on its own bayesian filter. it has a shortcut for
marking Junk and moving it to the junk folder (good), but the counterpart of
notjunk only marks a message as notjunk without showing in the UI that it was
marked as notjunk. there is also no "move and mark as notjunk" mechanism. e.g.
"archive" does not mark a message as notjunk. ios mail and mutt don't appear to
have any way to see or change the $Junk and $NotJunk flags.
what email clients do have is the ability to move messages to other
mailboxes/folders. so mox now has a mechanism that allows you to configure
mailboxes that automatically set $Junk or $NotJunk (or clear both) when a
message is moved/copied/delivered to that folder. e.g. a mailbox called junk or
spam or rejects marks its messags as junk. inbox, postmaster, dmarc, tlsrpt,
neutral* mark their messages as neither junk or notjunk. other folders mark
their messages as notjunk. e.g. list/*, archive. this functionality is
optional, but enabled with the quickstart and for new accounts.
also, mox now keeps track of the previous training of a message and will only
untrain/train if needed. before, there probably have been duplicate or missing
(un)trainings.
this also includes a new subcommand "retrain" to recreate the junkfilter for an
account. you should run it after updating to this version. and you should
probably also modify your account config to include the AutomaticJunkFlags.
