rootvector2 [Thu, 21 May 2026 12:56:17 +0000 (18:26 +0530)]
lib-dict-backend: cdb - Fix use-after-realloc of returned key pointer
cdb_dict_next() reuses ctx->buffer for both the key and the value:
the key is placed at offset 0 of the buffer, and a pointer to it is
returned via *key_r. cdb_dict_iterate() then writes that pointer back
to *key_r in the caller's slot and only afterwards appends the value
into the same buffer with buffer_append_space_unsafe(). If the total
size (key_len + 1 + value_len + 1) exceeds the buffer's current
allocation (initially 256), the append triggers a realloc on the
default_pool, which moves the backing memory and frees the previous
allocation. The key pointer saved in *key_r is now dangling, and the
caller's next dereference is a heap-use-after-free.
The fix sets *key_r from ctx->buffer->data after the value has been
appended, so it always reflects the buffer's current (post-realloc)
base address. The key is at offset 0, so reading the buffer's data
pointer is sufficient.
Reproduced as a standalone case with AddressSanitizer, which reports
heap-use-after-free on the saved key pointer after the value append
forces the buffer to grow past its initial 256-byte allocation.
Paul Howarth [Thu, 26 Jun 2025 09:47:35 +0000 (10:47 +0100)]
m4: Fix for systems with signed 32-bit time_t
The code and tests have some hard-coded values based on the size of
time_t but they don't take into account whether or not time_t is a
signed value.
On systems such as i686 with time_t as a signed 32-bit integer, the
maximum value is the same as an unsigned 31-bit integer. Tweak the
configure test to treat 32-bit signed as 31 bits.
overflow the maximum value of 2147483647, causing the test to fail.
This can be fixed by using a "long long" (which is used for intmax_t)
instead, e.g.
Aki Tuomi [Wed, 20 May 2026 12:21:01 +0000 (15:21 +0300)]
lib: panic on nearest_power() overflow with the offending value
Replace the i_assert() with an i_panic() that reports the input
value, making overflow failures easier to diagnose. Add unit tests
for the boundary input and a fatal test for the overflow case.
Aki Tuomi [Wed, 13 May 2026 15:33:08 +0000 (15:33 +0000)]
quota: Fix maildir quota drop to zero after IMAP MOVE/REPLACE
When mailbox_move() or IMAP REPLACE is used with the maildir quota
driver, the source mail expunge is double-counted, causing the quota
to drop to zero.
quota_try_alloc() called quota_alloc_with_size(ctx, size, expunged_size),
which subtracted expunged_size from the destination transaction's
bytes_used. At the same time mail_expunge() on the source/expunged mail
fired quota_mail_expunge(), and the expunge was committed independently
via qbox->expunge_qt in sync_notify. Both deltas commit to the same
quota root, netting -size when size == expunged_size.
Co-Authored: Aleksei Fedorov <aleksei.fedorov@webpros.com>
Fixes issues with false positives on certain phrase searches, as searches
on tokenized pieces are really intended to be used as a limiter on the
full phrase searching as opposed to an indication of full phrase match
on those messages.
Example 1: Exact phrase no-match (the canonical case)
Message body: "abcde fghij"
Search: BODY "abcd fghi"
Before: MATCHES (false positive — tokenized "abcd*" AND "fghi*" found)
After: NO MATCH (correct — exact phrase "abcd fghi" not in body)
Example 2: Exact phrase that should match
Message body: "the quick brown fox"
Search: BODY "quick brown"
Before: MATCHES (worked, but for wrong reason — tokenized terms matched)
After: MATCHES (correct — exact phrase "quick brown" verified by core)
Example 3: Words exist but not as a phrase
Message body: "brown fox quick"
Search: BODY "quick brown"
Before: MATCHES (false positive — both words present, just not adjacent)
After: NO MATCH (correct — "quick brown" doesn't appear as a substring)
Example 4: Single-word split (e.g., dotted word)
Message body: "foo.bar value"
Search: TEXT "foo.bar"
Before: NO MATCH (broken — phrase logic incorrectly applied, original
string "foo.bar" was skipped during token expansion)
After: MATCHES (correct — single word without spaces is not treated
as a phrase; "foo.bar" is indexed and matched directly)
Example 5: Multi-word search with one non-phrase word
Message body: "hello world"
Search: BODY hello "world peace"
Before: "hello" matched; "world peace" could produce false positives
from tokenized "world*" AND "peace*"
After: "hello" matches; "world peace" correctly requires exact phrase
Phrase arguments, and the tokenized children are now labeled into
three categories:
- This is the full phrase
- This is a tokenized word of a phrase, and it is the first word
- This is a tokenized word of a phrase
Before recent changes, the mailbox_list_escape_name_params() leading-'~' rule
fired on every hierarchy part, so vname "parent/~child" was written to
disk as "parent/\7echild". The new code applies the rule only on the
first part, encoding the same vname as "parent/~child". On-disk
directories written by older Dovecot versions therefore become
inaccessible after upgrade: the lookup constructs the new-format
storage name and stat()s a path that no longer matches.
Timo Sirainen [Thu, 30 Apr 2026 14:41:33 +0000 (14:41 +0000)]
lib-storage: mailbox_name_hdr_encode/decode - Distinguish INBOX from <ns prefix>/INBOX
The box-name index header stored the storage name with per-part unescape,
which collapses <escape>49NBOX to "INBOX". On rebuild from box-name headers,
this made <ns prefix>/INBOX indistinguishable from the regular INBOX, so
list index rebuild restored <ns prefix>/INBOX as INBOX.
Append a backwards compatible flag byte to the box-name header after two
trailing NUL bytes, with MAILBOX_NAME_HDR_FLAG_INBOX_INBOX bit signaling
that the on-disk "INBOX" name actually refers to <ns prefix>/INBOX. The
flag byte is only emitted when at least one bit is set, so old headers
(no flag byte) round-trip unchanged. The decoder strips trailing NUL
padding added by lib-index, then detects the flag suffix unambiguously:
two consecutive NULs at the end never occur in the old format because
the encoder splits on the hierarchy separator and real storage names
contain no empty hierarchy parts.
Old Dovecot versions reading a new-format header will parse the flag
byte as a one-byte trailing hierarchy component. This is acceptable
because old Dovecot never wrote <ns prefix>/INBOX in the first place.
Timo Sirainen [Thu, 30 Apr 2026 09:22:10 +0000 (09:22 +0000)]
imapc: Document round-trip requirements for invalid mUTF-7 names
Note in the imapc-list.c top-of-file comment that the four name
forms (remote_name, storage_name, vname, fs_name) round-trip
losslessly even for invalid mUTF-7 remote names, but only when
mailbox_list_visible_escape_char is set. With it unset and
mailbox_list_utf8=no, vname falls through to the raw mUTF-7 form
and SELECT round-trips would re-encode literal '&' as "&-",
breaking access to those mailboxes.
This documents the existing behavior; no code change.
Timo Sirainen [Thu, 30 Apr 2026 09:06:07 +0000 (09:06 +0000)]
lib-storage: Add tests for mailbox_list_escape_name_params() first_part
Cover the leading-'~' rule (escaped only when first_part=TRUE), the
per-part dirstart escapes for "." and "..", separator conversion,
and a round-trip property test against
mailbox_list_unescape_name_params() for inputs that look like parts
vs full vnames.
The leading '~' check used to fire whenever the input started with '~'.
That assumed callers always passed the full vname. Since
mailbox_list_default_get_storage_name() and
mailbox_name_hdr_decode_storage_name() now split into hierarchy parts
before escaping, the '~' check would wrongly match the start of any
part - e.g. "foo/~bar" would have its second component escaped as
"+7ebar" even though '~' in the middle of a path has no special
meaning.
Add a first_part parameter. Only when first_part=TRUE do we apply the
leading-'~' escape. Callers operating on parts pass first_part=FALSE
for non-leading parts.
For mailbox_list_default_get_storage_name() also require prefix[0] ==
'\\0' so that when the INBOX/INBOX prefix has been prepended to
storage_name, the first split-part is no longer the very start of the
escaped output.
Timo Sirainen [Thu, 30 Apr 2026 09:03:08 +0000 (09:03 +0000)]
lib-storage: mailbox_list_unescape_name_params() - Drop unused ns_prefix parameter
Symmetric to the previous escape_name_params() cleanup. The three
callers (mailbox_list_default_get_vname_part(),
imapc_list_storage_to_remote_name(), imapc_list_fs_to_storage_name())
either pass an empty string or operate on a single hierarchy part
that won't start with the namespace prefix.
Timo Sirainen [Thu, 30 Apr 2026 09:01:58 +0000 (09:01 +0000)]
lib-storage: mailbox_list_escape_name_params() - Drop unused ns_prefix parameter
All callers either pass an empty string or pass list->ns->prefix
operating on input where the prefix has already been stripped earlier
(by mailbox_list_vname_prepare() or by splitting into hierarchy parts).
The str_begins(vname, ns_prefix, ...) match was effectively dead.
Drop the parameter. Callers updated: mailbox_list_default_get_storage_name(),
mailbox_name_hdr_decode_storage_name(), imapc_list_remote_to_storage_name(),
imapc_list_storage_to_fs_name().
Timo Sirainen [Mon, 26 Jan 2026 22:09:24 +0000 (00:09 +0200)]
lib-storage: Fix handling escaped mailbox names in mailbox list index
Code paths using mailbox_list_index_node_get_path() weren't escaping the
mailbox name, breaking some of the functionality for such mailboxes.
This includes:
* Auto-deleting \Noselect leaf mailboxes
* Notifying changes to mailbox (IMAP NOTIFY)
* Storing mailbox GUID in dovecot.list.index
Timo Sirainen [Tue, 5 May 2026 21:15:22 +0000 (21:15 +0000)]
login-common: Emit proxy_session_finished and new _started event when client is redirected
When LOGIN_PROXY_FAILURE_TYPE_AUTH_REDIRECT was handled, the disconnect
event chain never produced an error_code=proxy_dest_redirected entry.
Fire a new proxy_session_finished event with this error_code and start
a new proxy_session_started for the new destination.
Timo Sirainen [Tue, 5 May 2026 21:14:04 +0000 (21:14 +0000)]
login-proxy: Set destination event fields in login_proxy_set_destination()
Move dest_host, dest_ip and dest_port event_add_*() calls from
login_proxy_new() into login_proxy_set_destination(). This ensures the
fields stay in sync.
Timo Sirainen [Wed, 6 May 2026 15:19:16 +0000 (15:19 +0000)]
imap-login: Reject duplicate MULTIPLEX from backend
If a buggy backend IMAP server sends "* MULTIPLEX 0" twice, the second
parse call re-entered login_proxy_multiplex_input_start() and crashed
on the multiplex_orig_input==NULL assert. Detect the duplicate and
fail the proxy connection with a protocol error instead.
Returns whether login_proxy_multiplex_input_start() has already been
called on the proxy. Lets callers detect a duplicate MULTIPLEX
notification from a buggy backend instead of triggering the
multiplex_orig_input==NULL assert.
fts-flatcurve: Clean up FTS directory on mailbox deletion
This code is only needed for Maildir. dbox handles FTS cleanup
already because it can just recursively delete the entire index
directory natively. Maildir (for safety reasons) can only delete
directories that appear in the "is_internal_name" list.
Fred Morcos [Thu, 30 Apr 2026 14:44:27 +0000 (16:44 +0200)]
auth: Fix stale cache entry blocking auth after password change in worker path
When passdb cache verification runs on an auth worker, a successful
prior login plus a subsequent password change leaves a stale cached
password. The non-worker code path already handles this: if the
cached password mismatches and the entry's last_success flag (or
neg_expired) indicates the cache is likely stale, fall back to a
fresh passdb lookup. The worker callback was missing this handling,
so users would fail to authenticate until the cache entry expired.
After the worker returns a PASSWORD_MISMATCH, re-lookup the cache
node and apply the same passdb_cache_use_password_mismatch() rule
to decide whether to invoke the fallback. The node pointer cannot
be carried across the worker round-trip (see auth_cache_node
lifetime), so the lookup is redone here. orig_cache_result is
preserved so the re-lookup does not clobber the request's recorded
cache result.
If the re-lookup misses because the entry was evicted while the
worker was busy, fall back unconditionally instead of delivering
the mismatch as-is: a password change would otherwise be reported
as a mismatch instead of being retried, and the cached evidence
that would have driven the stale-cache rule no longer exists, so
an extra fresh lookup is the right behavior.
Fred Morcos [Thu, 30 Apr 2026 15:34:55 +0000 (17:34 +0200)]
auth: Convert passdb_cache_verify_plain() to continuation-passing style
Change passdb_cache_verify_plain() from a bool-returning function with
an out-param to one that always invokes a continuation: either it
finishes the verification itself (cache hit, negative entry, etc.)
or it invokes the supplied fallback callback (cache miss, password
mismatch with stale-cache eligibility, no cache configured).
Callers no longer need to handle the result themselves; the helper
decides and dispatches. This unifies how cache hits and misses are
handed back into the auth flow and is preparation for fixing the
worker path's stale-cache handling.
Add auth_request_verify_plain_internal_failure() as the fallback
callback for the INTERNAL_FAILURE/expired-cache path in
auth_request_verify_plain_passdb_callback(). The existing
auth_request_verify_plain_passdb() helper is reused as the fallback
for the initial verify path.
Move the post-cache-check passdb dispatch (state transition,
wanted_credentials_scheme reset, blocking/direct iface call) out of
auth_request_default_verify_plain_continue() into a static helper. Pure
refactoring with no behavior change.
Fred Morcos [Thu, 30 Apr 2026 15:31:47 +0000 (17:31 +0200)]
auth: Rename auth_request_verify_plain_callback() to auth_request_verify_plain_passdb_callback()
Pure rename. The callback is invoked from passdb verification paths
(both blocking-worker and direct), so the new name better describes
where it sits in the call chain and disambiguates it from
auth_request_verify_plain_callback_finish().
Timo Sirainen [Fri, 24 Apr 2026 19:52:41 +0000 (19:52 +0000)]
lib-event: Add event_get_global_root()
Returns the bottom-most (oldest pushed) event on the global event
stack, or the current global event if the stack is empty. Lets
plugins annotate the outermost active caller's event from within a
nested global event context.
Timo Sirainen [Fri, 24 Apr 2026 19:42:55 +0000 (19:42 +0000)]
imap: Add [THROTTLED] response code to tagged replies
When a plugin marks cmd->global_event with a positive integer
"throttled_any" field during command execution, the tagged response
is prefixed with [THROTTLED] resp-code (skipped when the reply text
already carries a resp-code per RFC 3501) and gets a
"(throttled N.NNN secs)" annotation before the timing stats. The
annotation is always added when throttled, so clients that cannot
act on the resp-code still see throttling in the reply text.
The header value reaching fts_build_mail_header() was already RFC 2047
decoded by message_decoder, so passing it to message_address_parse()
meant decoded display name characters were reinterpreted as RFC 5322
syntax. For example:
decodes to "Testuser (Example)" and "(Example)" was then silently dropped as
an RFC 5322 comment. Similarly, display names containing characters
like '[' made parse_name_addr() fail, leaving neither the display name
nor the email address indexed.
Enable MESSAGE_DECODER_FLAG_RAW_ADDRESS_HEADERS so that address headers
arrive with encoded-words intact, parse the raw value, write the
normalised address, and then decode the result to UTF-8 before handing
it to the FTS backend.
Timo Sirainen [Fri, 17 Apr 2026 19:32:23 +0000 (19:32 +0000)]
lib-mail: message-decoder - Add flag to skip RFC 2047 decoding for address headers
Add MESSAGE_DECODER_FLAG_RAW_ADDRESS_HEADERS. When set, address headers
(From, To, Cc, etc.) are passed through with their raw bytes intact
instead of having encoded-words decoded to UTF-8. This lets callers
feed the raw value to message_address_parse() so that decoded display
name characters like '(' or '[' are not reinterpreted as RFC 5322
comments or other specials during address parsing.
Timo Sirainen [Wed, 22 Apr 2026 12:44:24 +0000 (15:44 +0300)]
imap-acl: Fail if ACL identifier is invalid
Reject invalid identifiers early in imap_acl_identifier_parse() using
acl_id_is_valid(). This prevents CR/LF injection and rejects identifiers
that are too long, contain control characters or are not valid UTF-8.
Timo Sirainen [Thu, 23 Apr 2026 12:35:17 +0000 (12:35 +0000)]
login-common: Translate haproxy ssl_client_cert into VALID_CLIENT_CERT
When haproxy terminates TLS for the end client and reports via the
PP2_TYPE_SSL TLV that a client certificate was presented and
successfully verified, propagate this to the auth request as
AUTH_REQUEST_FLAG_VALID_CLIENT_CERT. Without this, settings such as
auth_ssl_require_client_cert cannot be used with plaintext haproxy
listeners, even though haproxy has performed certificate verification
on the dovecot operator's behalf.
Add a haproxy_ssl_client_cert field to struct client that mirrors the
master_service_connection_haproxy.ssl_client_cert field, and use it
alongside the existing ssl_iostream check in client_get_auth_flags().
Previously the parser ignored both the verify field and the
PP2_CLIENT_CERT_CONN / PP2_CLIENT_CERT_SESS bits in the client byte
of the PP2_TYPE_SSL TLV, and unconditionally copied the client
certificate common name into the connection. This meant that a
common name reported by HAProxy was trusted even when HAProxy itself
had reported a verification failure (verify != 0).
Set ssl_client_cert from the cert presence bits, and only trust the
common name when a client certificate was presented and verified
successfully. When verify is non-zero, the certificate sub-TLVs are
ignored.
Timo Sirainen [Thu, 23 Apr 2026 10:43:28 +0000 (10:43 +0000)]
lib-master: haproxy - Read PP2 SSL TLV verify field as big-endian
The PROXY protocol v2 specification does not explicitly state the
endianness of the verify field in the PP2_TYPE_SSL TLV, but the
reference HAProxy implementation transmits it in network byte order
via htonl().
This doesn't really matter, because the field is checked either as
0 or non-0 in the next commit as well.
Timo Sirainen [Wed, 29 Apr 2026 15:13:02 +0000 (18:13 +0300)]
lib: file_create_locked() - Make sure locked file isn't a symlink
Since this is used for files that are auto-created, there's likely no reason
for it to support the file being a symlink. So its behavior is now safer by
not allowing locking symlinks.
Timo Sirainen [Mon, 27 Apr 2026 22:12:50 +0000 (22:12 +0000)]
anvil: Introduce struct anvil_command for array dispatch
Register all anvil commands in an array and dispatch them with a
for-loop. The connection type allowed for each command is now declared
in the array entry, and the type check happens in the generic
dispatcher rather than in each command handler.
Timo Sirainen [Mon, 27 Apr 2026 22:10:52 +0000 (22:10 +0000)]
anvil: Make connection type explicit per command
Each command handler now explicitly checks that the connection type is
allowed for that command. This replaces the implicit 'AUTH_PENALTY only
allows PENALTY-*' check in the dispatcher and the special-case KILL
check.
Timo Sirainen [Mon, 27 Apr 2026 22:31:10 +0000 (22:31 +0000)]
login-common, lmtp: Switch LOOKUP queries to anvil-connect-limit socket
LOOKUP commands are now only accepted on connections of type
ANVIL_CONNECTION_TYPE_CONNECT_LIMIT. Connect to the dedicated
anvil-connect-limit socket instead of the generic anvil admin socket.
Timo Sirainen [Mon, 27 Apr 2026 22:04:53 +0000 (22:04 +0000)]
anvil: Rename ANVIL_CONNECTION_TYPE_MASTER to ANVIL_CONNECTION_TYPE_SHARED_FIFO
This connection type is the shared FIFO from the master process that all
services use to send CONNECT, DISCONNECT and KILL commands. Rename
accordingly.