# -*- coding: utf-8 -*-

# Copyright © Cloud Linux GmbH & Cloud Linux Software, Inc 2010-2021 All Rights Reserved
#
# Licensed under CLOUD LINUX LICENSE AGREEMENT
# http://cloudlinux.com/docs/LICENSE.TXT

"""
This module contains contains classes implementing SSA Agent behaviour
"""

import atexit
import json
import logging
import pwd
import re
import socket as socket_module
import struct
import time
from concurrent.futures import ThreadPoolExecutor
from threading import BoundedSemaphore, Lock, current_thread

from clcommon.cpapi import (
    cpusers,
    domain_owner,
    get_domains_php_info,
    get_main_username_by_uid,
)

from .internal.constants import agent_sock
from .internal.exceptions import SSAError
from .internal.utils import canonical_stored_domain, create_socket
from .modules.processor import RequestProcessor

# Maximum number of concurrent worker threads for handling requests.
# Limits memory usage on high-traffic servers.
MAX_WORKERS = 50

# Upper bound on bytes accepted from a single connection.
# The PHP extension's JSON payload is always under 1 KB; 8 KB gives
# ample headroom while preventing unbounded reads.
MAX_MSG_SIZE = 8192

# Pre-auth read budget: seconds a peer has to finish sending its payload
# before the worker+admission slot it holds is reclaimed. The whole read in
# handle() runs BEFORE any sender validation, so on the world-writable socket
# an idle/drip-feeding peer can pin a worker for this long with O(1) effort
# (slow-read / Slowloris primitive). The per-UID admission cap bounds how many
# such slots one UID holds concurrently, but not how long each is held, so this
# budget is kept tight. The legitimate PHP client writes its <1 KB payload in
# one shot and closes (see src/processors.c: ssa_process_request), so the read
# returns on EOF within milliseconds; the client's own SO_SNDTIMEO/SO_RCVTIMEO
# is SSA_IO_TIMEOUT_SEC=2, so a well-behaved sender never needs more than 2s.
SOCKET_READ_TIMEOUT = 2

# Largest value storable in a SQLite signed 64-bit INTEGER column. Metric
# values above this raise OverflowError on bulk insert and would poison the
# shared flush buffer, so they must be rejected during input validation.
_MAX_SQLITE_INT = 2**63 - 1

# Bounded admission control. ThreadPoolExecutor.submit() enqueues into an
# UNBOUNDED queue.SimpleQueue, so a local connect-flood on the world-writable
# socket would back up accepted-but-unhandled connections without limit and
# grow the agent's memory toward MemoryMax=1G (OOM/restart of the telemetry
# daemon). Cap in-flight + queued work at the worker count plus a small queue:
# past that, the accept loop closes the connection instead of queueing it.
MAX_PENDING = MAX_WORKERS
MAX_INFLIGHT = MAX_WORKERS + MAX_PENDING

# Per-UID admission cap. The socket is world-writable by design, so without
# this a single local UID can hold every MAX_INFLIGHT slot before any sender
# validation runs (the peer UID is only checked inside handle(), after a worker
# and slot are already committed), starving every other tenant's telemetry —
# a cross-principal denial of monitoring. Cap one peer UID at half the worker
# pool so no single UID can monopolise the executor; the remainder always stays
# available to other UIDs. Over-cap connections are dropped (best-effort
# telemetry). Privileged service peers (_PRIVILEGED_SERVICE_UIDS — root
# collectors, the LiteSpeed/Apache master) have no single-tenant identity and
# are exempt.
MAX_INFLIGHT_PER_UID = MAX_WORKERS // 2

# Sentinel default for handle()'s admission-release argument. It distinguishes
# "this connection went through the accept loop and holds an admission slot to
# release" (admitted_uid is an int or None) from "handle() was invoked directly
# without going through listen() and holds no slot" (admitted_uid is
# _NO_ADMISSION), so the finally block never over-releases the semaphore.
_NO_ADMISSION = object()


class SimpleAgent:
    """
    SSA Simple Agent class
    """

    def __init__(self):
        self.logger = logging.getLogger('agent')
        self.request_processor = RequestProcessor()
        self.executor = ThreadPoolExecutor(max_workers=MAX_WORKERS)
        # Counts in-flight + queued connections; bounds the executor's
        # otherwise-unbounded work queue (admission control).
        self._admission = BoundedSemaphore(MAX_INFLIGHT)
        # Per-UID in-flight counters (peer UID -> count), guarded by a lock:
        # incremented by the accept loop at admission and decremented by
        # handle()'s finally when a worker completes the connection. Bounds how
        # many admission slots any single UID can hold (MAX_INFLIGHT_PER_UID).
        self._per_uid_lock = Lock()
        self._per_uid_inflight = {}
        # Shared-web service UIDs (nobody/apache) resolved from pwd on first use
        # and cached for the process lifetime; None until resolved. Used to gate
        # the shared-handler owner-mismatch exemption in _authorize_sender.
        self._shared_web_uids_cache = None
        # Rate-limiting state for repetitive per-peer decision logs (accept/
        # reject), keyed by peer UID and guarded by its own lock so one local
        # peer cannot flood the agent log.
        self._log_rate_lock = Lock()
        self._log_rate_state = {}
        atexit.register(self._shutdown)
        # start serving incoming connections
        self.listen()

    def _shutdown(self):
        """Gracefully shutdown the thread pool executor."""
        self.executor.shutdown(wait=False)

    def _try_admit_uid(self, peer_uid: int) -> bool:
        """
        Reserve a per-UID admission slot for peer_uid. Privileged service
        peers (_PRIVILEGED_SERVICE_UIDS — root collectors / the LiteSpeed/
        Apache master) have no single-tenant identity and legitimately serve
        every domain, so they are exempt. Returns False when peer_uid already
        holds MAX_INFLIGHT_PER_UID in-flight connections, so one local UID
        cannot monopolise the executor and deny other tenants' telemetry.
        """
        if peer_uid in self._PRIVILEGED_SERVICE_UIDS:
            return True
        with self._per_uid_lock:
            if self._per_uid_inflight.get(peer_uid, 0) >= MAX_INFLIGHT_PER_UID:
                return False
            self._per_uid_inflight[peer_uid] = self._per_uid_inflight.get(peer_uid, 0) + 1
        return True

    def _release_uid(self, peer_uid: int) -> None:
        """Release a per-UID admission slot reserved by _try_admit_uid."""
        if peer_uid in self._PRIVILEGED_SERVICE_UIDS:
            return
        with self._per_uid_lock:
            remaining = self._per_uid_inflight.get(peer_uid, 0) - 1
            if remaining > 0:
                self._per_uid_inflight[peer_uid] = remaining
            else:
                self._per_uid_inflight.pop(peer_uid, None)

    def _release_admission(self, peer_uid) -> None:
        """Release the global admission slot and, when the connection was
        attributed to a UID, its per-UID slot."""
        self._admission.release()
        if peer_uid is not None:
            self._release_uid(peer_uid)

    # Usernames under which shared/DSO PHP runs as the SHARED web UID rather
    # than the per-account UID (cPanel/LiteSpeed DSO 'nobody', Apache-module
    # 'apache'). Resolved to numeric UIDs once, lazily, from pwd.
    _SHARED_WEB_USERNAMES = ('nobody', 'apache')

    def _shared_web_uids(self) -> frozenset:
        """
        Numeric UIDs of the shared web service accounts (nobody/apache),
        resolved from pwd and cached for the process lifetime. Usernames absent
        from the passwd database are skipped (KeyError tolerated), so this is
        robust on panels/OSes that ship only one of them. The shared-handler
        owner-mismatch exemption in _authorize_sender applies ONLY when the peer
        UID is one of these — a per-user tenant peer (its own account UID) can
        never claim it, closing the cross-tenant spoof where one tenant reports
        another tenant's dso/mod_php/module domain.
        """
        if self._shared_web_uids_cache is None:
            uids = set()
            for name in self._SHARED_WEB_USERNAMES:
                try:
                    uids.add(pwd.getpwnam(name).pw_uid)
                except KeyError:
                    pass
            self._shared_web_uids_cache = frozenset(uids)
        return self._shared_web_uids_cache

    # Repetitive per-peer decision logs (accept/reject) are rate-limited so a
    # single local peer cannot flood the agent log: at most _LOG_RATE_MAX lines
    # per key per _LOG_RATE_WINDOW seconds, then a one-line suppression summary
    # when the window next rolls over. Error/exception logs are NEVER routed
    # through this — only the per-request accept/reject lines are.
    _LOG_RATE_WINDOW = 60
    _LOG_RATE_MAX = 20

    def _log_rate_limited(self, key, level: str, msg: str, *args) -> None:
        """
        Emit ``self.logger.<level>(msg, *args)`` unless *key* (a peer UID) has
        already produced _LOG_RATE_MAX lines in the current _LOG_RATE_WINDOW;
        suppressed lines are counted and reported in a single summary line when
        the next window opens. Thread-safe (guarded by _log_rate_lock), matching
        the module's lock discipline.
        """
        now = time.monotonic()
        emit = False
        summary = 0
        with self._log_rate_lock:
            state = self._log_rate_state.get(key)
            if state is None or now - state['start'] >= self._LOG_RATE_WINDOW:
                summary = state['suppressed'] if state else 0
                self._log_rate_state[key] = {'start': now, 'emitted': 1, 'suppressed': 0}
                emit = True
            elif state['emitted'] < self._LOG_RATE_MAX:
                state['emitted'] += 1
                emit = True
            else:
                state['suppressed'] += 1
        if summary:
            getattr(self.logger, level)(
                '[%s] suppressed %d repetitive log line(s) for peer key %r', current_thread().name, summary, key
            )
        if emit:
            getattr(self.logger, level)(msg, *args)

    def listen(self) -> None:
        """
        Start listening socket
        """
        _socket = create_socket(agent_sock)
        while True:
            connection, address = _socket.accept()
            if not self._admission.acquire(blocking=False):
                # Pool + bounded queue saturated: refuse rather than queue
                # unbounded work. The peer (best-effort telemetry) just drops
                # this sample; the agent's memory stays bounded under a flood.
                self.logger.warning('[Admission] pool saturated (%d in-flight); dropping connection', MAX_INFLIGHT)
                connection.close()
                continue
            # Resolve the kernel-trusted peer UID cheaply (SO_PEERCRED, no
            # payload read) and enforce a per-UID cap BEFORE a worker is
            # committed: sender validation only happens later inside handle(),
            # so without this one local UID could hold every slot with idle
            # connections and deny all other tenants' telemetry. A connection
            # whose credentials cannot be read is not dropped here (it cannot
            # be attributed to a UID); handle() closes it on the same failure.
            try:
                peer_uid = self._get_peer_uid(connection)
            except (OSError, struct.error) as e:
                self.logger.debug('[Admission] peer credentials unreadable, skipping per-UID cap: %s', str(e))
                peer_uid = None
            if peer_uid is not None and not self._try_admit_uid(peer_uid):
                self.logger.warning(
                    '[Admission] per-UID cap (%d) reached for UID=%d; dropping connection',
                    MAX_INFLIGHT_PER_UID,
                    peer_uid,
                )
                connection.close()
                self._admission.release()
                continue
            try:
                # Pass peer_uid so handle() releases exactly the admission
                # slot(s) reserved above (global + per-UID) in its finally,
                # once per handled connection.
                self.executor.submit(self.handle, connection, peer_uid)
            except RuntimeError as e:
                # CPython's ThreadPoolExecutor.submit() enqueues the work
                # item BEFORE calling _adjust_thread_count(); a RuntimeError
                # from Thread.start() (e.g. the cgroup pids controller
                # rejecting a new thread because TasksMax was hit,
                # CLPRO-3118) therefore leaves the work item in the queue,
                # and existing workers will drain it. Do NOT close the
                # connection here — that would race with a worker that may
                # yet process it and turn a real burst-induced failure into
                # spurious "Bad file descriptor" errors. Do NOT release the
                # admission slot here either: the slot is released by handle()'s
                # finally when a worker actually runs the queued item. If
                # thread-starts keep failing and no worker ever runs it, the
                # slot stays held and the gate correctly REFUSES new connections
                # (bounded, fail-safe) rather than letting the executor's queue
                # grow past the admission bound. Killing the agent would only
                # trigger a restart loop, so just log and keep accepting.
                self.logger.error('[ThreadPool] submit raised (work item queued for existing workers): %s', str(e))
                continue
            self.logger.debug('[ThreadPool] Submitted task')

    # Fields that the PHP extension sends (see dump.c: ssa_agent_dump)
    _REQUIRED_FIELDS = frozenset(
        {'timestamp', 'url', 'duration', 'hitting_limits', 'throttled_time', 'io_throttled_time', 'wordpress'}
    )

    # 8190 matches Apache's default LimitRequestLine, which is the effective
    # upper bound on URLs reaching the PHP extension in practice.
    _MAX_URL_LENGTH = 8190
    _URL_RE = re.compile(r'^https?\??://[^\x00-\x1f\s<>"{}|\\^`\[\]]+\Z')

    @staticmethod
    def _get_peer_uid(connection: socket_module.socket) -> int:
        """
        Get the UID of the peer process using SO_PEERCRED.
        :param connection: socket object
        :return: UID of the connecting process
        """
        cred = connection.getsockopt(socket_module.SOL_SOCKET, socket_module.SO_PEERCRED, struct.calcsize('3i'))
        _pid, uid, _gid = struct.unpack('3i', cred)
        return uid

    @classmethod
    def _validate_input(cls, data: dict) -> bool:
        """
        Validate that input data contains exactly the expected metric fields
        with the correct value types. The PHP extension always sends all 7
        fields (see dump.c), so we require an exact key match and enforce
        the types produced by the C formatter to reject both malformed and
        spoofed payloads.
        """
        if not isinstance(data, dict) or not data:
            return False
        if set(data.keys()) != cls._REQUIRED_FIELDS:
            return False
        if not isinstance(data['timestamp'], str) or not data['timestamp'].isascii() or not data['timestamp'].isdigit():
            return False
        if not isinstance(data['url'], str) or not data['url']:
            return False
        if len(data['url']) > cls._MAX_URL_LENGTH:
            return False
        if not cls._URL_RE.match(data['url']):
            return False
        if (
            isinstance(data['duration'], bool)
            or not isinstance(data['duration'], int)
            or not 0 <= data['duration'] <= _MAX_SQLITE_INT
        ):
            return False
        if not isinstance(data['hitting_limits'], bool):
            return False
        if (
            isinstance(data['throttled_time'], bool)
            or not isinstance(data['throttled_time'], int)
            or not 0 <= data['throttled_time'] <= _MAX_SQLITE_INT
        ):
            return False
        if (
            isinstance(data['io_throttled_time'], bool)
            or not isinstance(data['io_throttled_time'], int)
            or not 0 <= data['io_throttled_time'] <= _MAX_SQLITE_INT
        ):
            return False
        if not isinstance(data['wordpress'], bool):
            return False
        return True

    # Peer UIDs that legitimately serve every tenant and therefore have no
    # single per-tenant identity to enforce: uid 0 (root) — the LiteSpeed/
    # Apache master and root-owned telemetry collectors. Such a peer is
    # trusted for any domain; every OTHER non-tenant peer (a shared DSO
    # worker such as nobody/apache, or an unmappable UID) is still subjected
    # to the cross-tenant owner check below.
    _PRIVILEGED_SERVICE_UIDS = frozenset({0})

    # PHP handlers under which PHP runs as the SHARED web UID (nobody/apache)
    # rather than the per-account UID. Enumerated across every clcommon.cpapi
    # backend: cPanel 'dso' (WHM php_get_handlers current_handler; mod_php),
    # DirectAdmin 'mod_php' (custombuild php{N}_mode), and Plesk 'module' (the
    # Apache-module handler — plesk.get_domains_php_info passes the raw <type>
    # through verbatim, so PHP runs as the shared `apache` UID). For a domain on
    # one of these a shared-UID peer (nobody/apache) IS the legitimate server,
    # so an owner-UID mismatch is expected and must NOT be treated as a spoof.
    # Every other handler (suphp/cgi/lsapi/fpm/fastcgi/php-fpm/x-httpd-*lsphp)
    # runs PHP as the per-user account UID.
    _SHARED_UID_HANDLERS = frozenset({'dso', 'mod_php', 'module'})

    def _authorize_sender(self, peer_uid: int, url: str) -> bool:
        """
        Authenticate the sender against the domain it reports for.

        The socket is world-writable by design (PHP workers run under
        arbitrary tenant UIDs), so the peer UID is the only sender identity
        available. The sender is classified by panel hosting-account
        membership, NOT by a numeric UID threshold: a UID_MIN cutoff is wrong
        panel/OS-dependently — DirectAdmin's `webapps` (uid 1001) and `admin`
        (uid 1000) sit at/above UID_MIN yet own no per-tenant identity, and
        cPanel/LiteSpeed on EL8/EL9 runs DSO/non-suEXEC PHP as `nobody`
        (uid 65534, not 99), also ≥ UID_MIN. A threshold would false-reject
        all of their legitimate metrics.

        Instead the peer's UID is resolved to a username via the panel
        (get_main_username_by_uid) and checked against the panel's
        authoritative tenant set (cpusers(), via _panel_tenant_users).
        A peer whose UID is a privileged root-owned service
        (_PRIVILEGED_SERVICE_UIDS — uid 0: the LiteSpeed/Apache master,
        root-run collectors) has no per-tenant identity to enforce and is
        trusted for any domain. Every OTHER peer — a hosting account, a
        shared DSO worker (apache, nobody=65534), a service account
        (webapps), or a UID that maps to no panel user — is subjected to a
        POSITIVE cross-tenant mismatch check: the reported domain's owner
        resolves to a system UID different from the peer UID. A mismatch is NOT
        rejected unconditionally, because it is also the legitimate shape of
        DSO/mod_php self-telemetry: on a DSO domain PHP runs as the shared web
        UID (nobody) while the domain's OWNER is the account, so owner_uid !=
        peer_uid is expected. The mismatch is therefore disambiguated by the
        REPORTED domain's own PHP handler (get_domains_php_info keyed on the
        reported domain, see _reported_domain_allows_shared_peer): handlers are
        PER-DOMAIN, so get_domains_php_info maps every domain — main, addon, sub
        and alias — to its own handler_type. A shared/DSO handler ('dso'/
        'mod_php'/'module') on the reported domain means the shared-UID peer
        legitimately serves it -> accept (DSO self-telemetry preserved); a
        per-user handler (suphp/cgi/lsapi/fpm/…) means PHP runs as the account
        UID, so a shared/service peer reporting it is a spoof -> reject. Keying
        on the REPORTED domain rather than the account's main domain is required:
        a mixed account can serve a shared handler on its main domain yet a
        per-user handler on a secondary/addon domain, so classifying by the main
        domain would admit a spoof on the account's per-user secondary domains.
        This closes the mixed-config spoof where a shared/service UID reports
        metrics for a per-user tenant's domain it does not own, without dropping
        legitimate DSO metrics. When ownership
        cannot be established (domain_owner raises, returns nothing, or names a
        user absent from the passwd database — panel DB/file unavailable, domain
        mid-provisioning, parked/alias domains) we fail OPEN (accept + warn).
        This fail-open-on-unresolvable is a deliberate product choice:
        dropping such payloads buys no security over the confirmed spoof
        (which needs a resolvable owner that differs) while silently losing
        legitimate metrics. If cpusers() itself fails the tenant set degrades
        to empty, but the owner check still runs for non-privileged peers, so
        a positive cross-tenant mismatch on a per-user-handler domain is still
        rejected.

        Ownership is compared by UID, NOT by username string (CLPRO-3231).
        The peer identity is the kernel-trusted SO_PEERCRED UID; with per-user
        PHP (lsphp / suEXEC / PHP-FPM-per-user) that UID is the domain owner's
        UID by construction, so mapping the reported domain's owner to a UID
        (pwd.getpwnam) and comparing UIDs can never false-reject a site's own
        traffic. Comparing two independently-resolved name strings —
        getpwuid(peer_uid).pw_name versus the owner recorded in the panel's
        domain DB — is NOT safe: the live passwd DB and the panel DB can spell
        the same account differently (case, account rename, duplicate/alias
        passwd entry, reseller naming), which false-rejected entire legitimate
        sites and flooded Sentry once 0.4-28 shipped the string comparison.

        The shared-handler mismatch exemption is bound to a VERIFIED shared-web
        peer UID: on a positive owner-UID mismatch the reported-domain handler
        check runs only when peer_uid is itself a shared-web service UID
        (nobody/apache, resolved by _shared_web_uids); every other peer — a
        per-user tenant reporting a domain it does not own, even a DSO one — is
        REJECTED. This closes the cross-tenant spoof where a legitimate per-user
        tenant reports telemetry for another tenant's dso/mod_php/module domain.

        Residual limit: fail-open on a mismatch is reserved for the case where
        the OWNER cannot be resolved (domain_owner returns nothing / raises, or
        the owner is not in passwd) — no distinct per-tenant kernel identity to
        bind to, and the threat model here is per-user PHP. Once the owner IS
        resolved to a distinct account, classification keys on the REPORTED
        domain's own handler and errs toward REJECT: a reported domain that is
        absent from an available map, or whose handler is None/unknown, is
        rejected, so an unknown handler no longer fails open. When the panel
        handler API itself is unavailable (get_domains_php_info raises — e.g.
        InterWorx/ispmanager where it is unsupported, or a transient panel-DB
        outage), the classifier now fails CLOSED: a shared-web peer's
        owner-mismatch self-telemetry is DROPPED (rejected) rather than risk a
        spoof, since a resolved distinct owner plus an unavailable classifier
        must not authorize. This narrow residual FLIP is documented in
        docs/design/ssa-agent-socket.md.

        Lookups (cpusers() and domain_owner) are performed FRESH on every
        payload — there is no cache. The previous owner/tenant-set memo was
        removed: it guarded a security decision and produced repeated
        staleness/race findings. Fresh lookups are obviously correct and
        acceptable here because the panel layer (clcommon.cpapi) memoizes
        internally (the DirectAdmin backend caches its domain DB), and this
        runs on a background daemon (50-worker pool), not the PHP request hot
        path.

        The owner lookup MUST key on the SAME domain string the processor
        stores, or a crafted URL can authorize as an unresolvable string
        (fail-open) yet be stored under a victim's domain. To guarantee that
        parity by construction, the lookup domain is derived from
        canonical_stored_domain() — the single source of truth for
        RequestProcessor's RequestResult.domain (netloc with every 'www.'
        substring removed) — and then normalized to the panel's lookup form:
        userinfo (user:pass@) and :port that survive in the netloc are
        dropped, the trailing root-label dot is stripped, and the host is
        lowercased (the panel keys ownership on the bare lowercase hostname).
        Because this normalization is a pure function of the stored string,
        every URL that stores as a given victim domain — 'www.' prefix or
        suffix insertion, trailing dot, uppercase, embedded port — resolves to
        the same owner lookup and is forced through the cross-tenant check.
        Deriving from hostname or stripping the trailing dot BEFORE removing
        'www.' broke this parity for the trailing-dot suffix case (e.g.
        victim.comwww.) — see bugbot 95cac397, CLPRO-3190, !45 note 557376.
        (url_split's substring replace can also rewrite domains that merely
        contain 'www.' mid-name; that attribution quirk is pre-existing and
        shared deliberately — parity is the security property.)
        """
        peer_user = self._peer_panel_user(peer_uid)
        if (
            peer_user is None or peer_user not in self._panel_tenant_users()
        ) and peer_uid in self._PRIVILEGED_SERVICE_UIDS:
            # Privileged root-owned service peer: no per-tenant identity to
            # enforce, trusted for any domain. Every other non-tenant or
            # unmappable peer falls through to the cross-tenant owner check
            # below, so a shared/service UID cannot report a per-user
            # tenant's domain it does not own.
            return True
        domain = self._owner_lookup_domain(url)
        owner = self._resolve_domain_owner(domain, peer_user)
        if not owner:
            return True
        try:
            owner_uid = pwd.getpwnam(owner).pw_uid
        except KeyError:
            self.logger.warning(
                '[%s] Owner %r of %r not in passwd (peer UID=%d), accepting',
                current_thread().name,
                owner,
                domain,
                peer_uid,
            )
            return True
        if owner_uid == peer_uid:
            return True
        # Positive owner-UID mismatch. The shared-handler exemption applies ONLY
        # when the PEER is itself a shared web service UID (nobody/apache) — the
        # UID under which shared/DSO PHP actually runs. A per-user tenant peer
        # (its OWN account UID, present in cpusers) can never legitimately serve
        # another tenant's domain, so it is REJECTED here even if that domain
        # runs a shared handler: this closes the cross-tenant spoof where tenant
        # A reports telemetry for tenant B's dso/mod_php/module domain. Only a
        # verified shared-web peer falls through to the per-reported-domain
        # handler check (which keys on the REPORTED domain's own handler, not the
        # owner account's main domain, so a mixed account's per-user secondary
        # domain cannot be spoofed), preserving legitimate DSO self-telemetry.
        if peer_uid in self._shared_web_uids():
            return self._reported_domain_allows_shared_peer(domain)
        return False

    def _reported_domain_allows_shared_peer(self, domain: str) -> bool:
        """
        True if the REPORTED domain serves PHP as a shared web UID, judged by
        that domain's OWN handler in get_domains_php_info. Handlers are
        per-domain: get_domains_php_info maps every domain — main, addon, sub and
        alias — to its own handler_type (a mixed account can serve a shared
        handler on its main domain and a per-user handler on a secondary), so the
        reported domain must be classified directly.

        This method is reached ONLY for a peer that is itself a verified shared-
        web service UID (nobody/apache) on a positive owner-UID mismatch — the
        caller (_authorize_sender) gates it on that. Returns True (accept) only
        when the reported domain's handler is a shared-UID handler
        (_SHARED_UID_HANDLERS). A per-user handler, a None/unknown handler, or a
        reported domain absent from an available map returns False (reject): the
        owner is a resolved DISTINCT per-user account, so even a shared-web peer
        reporting it is a spoof.

        Fails CLOSED (False, reject) when the panel cannot yield the handler map
        at all — get_domains_php_info() raises (e.g. InterWorx/ispmanager where
        it is unsupported, or the panel DB is momentarily down). RESIDUAL FLIP
        (documented in docs/design/ssa-agent-socket.md): previously this branch
        returned True, so a shared-UID self-report was accepted on API-unavailable
        panels and a spoof was possible there. It now DROPS (rejects) a shared-web
        peer's owner-mismatch self-telemetry on those backends rather than risk a
        spoof — with the peer-UID gate above, this branch is reached only for a
        shared-web peer whose reported domain resolves to a distinct owner, so a
        resolved distinct owner plus an unavailable classifier must not authorize.
        """
        try:
            info = get_domains_php_info()
        except Exception as e:
            self.logger.warning(
                '[%s] Panel handler API unavailable for %r, rejecting (fail closed): %s',
                current_thread().name,
                domain,
                str(e),
            )
            return False
        handler = (info.get(domain) or {}).get('handler_type')
        return handler in self._SHARED_UID_HANDLERS

    @staticmethod
    def _owner_lookup_domain(url: str) -> str:
        """
        Panel owner-lookup key for url, derived from the SAME string the
        processor stores (canonical_stored_domain), so authorization and
        storage cannot diverge. Strips userinfo and :port that survive in the
        netloc, removes the trailing root-label dot, lowercases, then removes
        'www.' once more: the netloc-level strip is case-sensitive, so an
        uppercase 'WWW.' prefix only collapses after lowercasing — this keeps
        authorization at least as strict as the prior hostname-based path
        while preserving stored-domain parity for every clean-apex variant.
        """
        host = canonical_stored_domain(url)
        if '@' in host:
            host = host.rsplit('@', 1)[1]
        if host.startswith('[') and ']' in host:
            # IPv6 literal: keep [..] intact, drop any :port after it
            host = host[: host.index(']') + 1]
        else:
            left, sep, right = host.rpartition(':')
            if sep and right.isdigit():
                host = left
        return host.rstrip('.').lower().replace('www.', '')

    @staticmethod
    def _peer_panel_user(peer_uid: int):
        """
        Resolve the peer UID to its panel main username, or None when it
        cannot be mapped (no such panel user). An unmappable UID is not a
        tenant and is trusted (fail open).
        """
        try:
            return get_main_username_by_uid(peer_uid)
        except Exception:
            return None

    def _panel_tenant_users(self) -> frozenset:
        """Fresh frozenset(cpusers()). On failure -> empty frozenset, so every
        peer is treated as a non-tenant and trusted (fail open), consistent with
        the unresolvable-owner / nopanel behaviour."""
        try:
            return frozenset(cpusers())
        except Exception as e:
            self.logger.warning(
                '[%s] Could not resolve panel hosting accounts, treating every peer as non-tenant (fail open): %s',
                current_thread().name,
                str(e),
            )
            return frozenset()

    def _resolve_domain_owner(self, domain: str, peer_user: str):
        """
        Resolve the owner username of a bare hostname via the panel API.
        Returns the owner username, or None for every case where ownership
        cannot be established (the fail-open cases documented in
        _authorize_sender).
        """
        try:
            owner = domain_owner(domain)
        except Exception as e:
            self.logger.warning(
                '[%s] Could not resolve owner of %r for user=%r, accepting: %s',
                current_thread().name,
                domain,
                peer_user,
                str(e),
            )
            return None
        if not owner:
            self.logger.warning('[%s] No owner for %r (user=%r), accepting', current_thread().name, domain, peer_user)
            return None
        return owner

    def handle(self, connection: socket_module.socket, admitted_uid=_NO_ADMISSION) -> None:
        """
        Handle incoming connection.

        :param connection: socket object usable to send and receive data on the
            connection.
        :param admitted_uid: the peer UID this connection was admitted under in
            the accept loop (an int; or None when its peer credentials were
            unreadable there but a global admission slot was still taken). The
            admission slot(s) reserved in listen() are released here in the
            finally block — exactly once per handled connection — so a
            queued-but-not-yet-run work item keeps its slot until a worker
            actually runs handle() (F-02 admission accounting). The
            _NO_ADMISSION sentinel means handle() was invoked directly (unit
            tests) without going through the accept loop, so there is no
            admission slot to release.
        """
        try:
            try:
                peer_uid = self._get_peer_uid(connection)
            except (OSError, struct.error) as e:
                self.logger.error('[%s] Failed to get peer credentials: %s', current_thread().name, str(e))
                connection.close()
                return

            try:
                input_data = self.read_input(connection)
                if not self._validate_input(input_data):
                    self._log_rate_limited(
                        peer_uid,
                        'warning',
                        '[%s] Rejected invalid payload from UID=%d: keys=%s',
                        current_thread().name,
                        peer_uid,
                        sorted(input_data.keys()) if isinstance(input_data, dict) else type(input_data).__name__,
                    )
                    return
                if not self._authorize_sender(peer_uid, input_data['url']):
                    self._log_rate_limited(
                        peer_uid,
                        'warning',
                        '[%s] Rejected cross-tenant payload from UID=%d for url=%r',
                        current_thread().name,
                        peer_uid,
                        input_data['url'],
                    )
                    return
                # The sustained-insert / on-disk storage vector is bounded
                # OUTSIDE this dispatch, not by an insert quota here:
                # RETENTION_TIME_DAYS=1 plus the daily cleanup_old_data + VACUUM
                # (py/ssa/db.py) cap on-disk growth, and RequestProcessor's
                # per-sender queue fairness cap (PER_SENDER_QUEUE_SHARE, keyed on
                # this same peer UID) bounds the in-memory backlog. So the only
                # per-request abuse surface left here is log volume, which the
                # rate-limited accept/reject lines below address.
                self.request_processor.handle(input_data, peer_uid)
                self._log_rate_limited(
                    peer_uid, 'info', '[%s] Accepted payload from UID %d', current_thread().name, peer_uid
                )
            except socket_module.timeout as e:
                self.logger.warning(
                    '[%s] Connection timed out (peer UID=%d): %s', current_thread().name, peer_uid, str(e)
                )
            except (SSAError, json.JSONDecodeError, ValueError) as e:
                self.logger.error('Handled exception in [%s]: %s', current_thread().name, str(e))
            except Exception as e:
                self.logger.exception('Unexpected exception in [%s]: %s', current_thread().name, str(e))
            finally:
                connection.close()
        finally:
            if admitted_uid is not _NO_ADMISSION:
                self._release_admission(admitted_uid)

    def read_input(self, connection: socket_module.socket) -> dict:
        """
        Read the pre-auth payload from *connection* under a TOTAL wall-clock
        deadline of SOCKET_READ_TIMEOUT seconds, then return the decoded JSON.

        connection.settimeout() is only a PER-recv inactivity timeout, so a peer
        that drips one byte just before each timeout could otherwise pin the
        worker (and the admission slot it holds) indefinitely — a slow-read /
        Slowloris primitive on the world-writable socket. The whole read is
        therefore bounded by an absolute monotonic deadline: before every recv
        the remaining budget is computed and installed as the socket timeout,
        and a non-positive remainder raises socket.timeout. At most MAX_MSG_SIZE
        bytes are read; an empty recv (EOF) ends the read. A legitimate client
        sends its <1 KB payload in one recv well within the budget, so this is
        transparent to well-behaved senders. Decoding / json.loads behaviour is
        unchanged from the previous makefile-based read.
        """
        deadline = time.monotonic() + SOCKET_READ_TIMEOUT
        chunks = []
        total = 0
        while total < MAX_MSG_SIZE:
            remaining = deadline - time.monotonic()
            if remaining <= 0:
                raise socket_module.timeout('pre-auth read budget exceeded')
            connection.settimeout(remaining)
            chunk = connection.recv(MAX_MSG_SIZE - total)
            if not chunk:
                break
            chunks.append(chunk)
            total += len(chunk)
        data = b''.join(chunks).decode(errors='ignore')
        self.logger.info('[%s] I received %i bytes', current_thread().name, len(data.encode()))
        self.logger.debug('[%s] payload: %s', current_thread().name, data)
        if data:
            return json.loads(data.strip())
        return {}