path: root/eclass/ehooker.eclass

# Copyright 1999-2013 Gentoo Foundation
# Distributed under the terms of the GNU General Public License v2
# $Header: $

# @ECLASS: ehooker.eclass 
# @MAINTAINER:
# Greg Turner <gmt@be-evil.net>
# @BLURB: Simple hooking framework consumable by other ebuilds/eclasses
# @DESCRIPTION:
# ehooker.eclass Provides a very simple api for exposing and consuming hookable
# events in ebuilds and eclasses.  This can be a handy way for eclasses to
# effect polymorphic behavior or to offer callbacks to ebuilds.
#
# Hooks are identified by arbitrary names.  The names should have no
# spaces in them and should not contain characters likely to cause
# interactions with shell globbing.  There's no need to push your luck :)
#
# Hook names share a global namespace, so it's a good practice to prefix groups
# hooks with something consistent and memorable.  The name of an eclass, for
# example, seems like a sensible prefix.
#
# No action is required to expose a hook to consumers; anyone can register
# to listen to any hook by name, and any hook may be fired, without any
# declarative prerequisite.  This unfortunately makes it pretty easy to
# miss your callbacks due to typos.  Fixing that is kinda hard without
# placing limitations on the framework, and probably not worth it.  Instead,
# it is easy to debug hook processing by setting ECLASS_DEBUG_OUTPUT=on.
# 
# Hook listener's effects on the bash environment can be very aggressively
# debugged using the EHOOKER_DEBUG_HOOKS variable -- just remember to turn
# this off once you're done debugging, as it's non-portable (to non-portage
# package managers) and damned expensive.
#
# To consume a hook, all that's necessary is to call the ehook function.
# Hook consumers are tracked in a global bash variable (exactly which depends
# on the version of bash used -- it is the associative array _EHOOKER_ADATA for 
# bash v4 and up, and the indexed array _EHOOKER_IDATA for earlier bash versions).
#
# Hooks should probably be avoided in error-handling code-paths, as this
# could result in situations where you never get around to failing due to bugs
# in other people's code.
#
# Must-execute framework-level cleanup code-paths also probably should not be
# in hook-listeners.  It's simply too easy for someone to install hooks on top
# of yours and cause some kind of problem where you never get your hook notification.
#
# The ehooks framework enforces no particular policy with respect
# to the flow of information between the invoking function and the
# hook-listeners.  Only the name of the hook is provided as an argument to
# the hook-listeners, and ehook_fire returns zero if and only if all invoked
# listeners returned zero.  Aside from that, you're on your own.
#
# However, in non-parallel modes, the global and local bash lexical environment,
# including exported environment variables, is preserved, and no sub-shells are
# employed in the hook-listener invocation codepath.  Therefore, bash variables
# and environment variables may be used as a means of coordinating the flow of
# information both into and out of hook-listener functions from hook-firing functions,
# according to any design pattern you can dream up.
#
# Here is an example illustrating how the author mostly tends to use the framework:
#
# @EXAMPLE:
# ebuild hook consumer code:
#
# @CODE
# src_configure() {
#     # sign up for hook
#     ehook whizbang-pre_src_compile my_perabi_pre_config
#
#     # call potentially hook-firing code
#     whizbang-multi-abi_src_configure
# }
#
# my_perabi_pre_config() {
#     export LIBX11_PATH="/usr/$(get_libdir)"
# }
# @CODE
#
# @EXAMPLE:
# eclass hook publisher code:
#
# @CODE
# whizbang-multi-abi_src_configure() {
#     multilib_foreach_abi _whizbang_perabi_econf "$@"
# }
#
# _whizbang_perabi_econf() {
#     declare -a econf_args=("$@")
#     ehook_fire whizbang-pre_src_compile &&
#         econf "${econf_args[@]}"a
#     ehook_fire whizbang-post_src_compile --uninterruptible
# }
# @CODE

# @ECLASS-VARIABLE: EHOOKER_DEBUG_HOOKS
# @DEFAULT_UNSET
# @DESCRIPTION:
# If set to an array of hook names, detailed diagnostic
# informational messages are displayed during listener invocation
# describing changes to the envoironment effected by the
# hook listener.
#
# Does not work for parallel hooks.  This will not work for
# non-portage package managers and contains nonportable, non-future-proof
# code.  For diagnostic purposes only.

if [[ ${___ECLASS_ONCE_EHOOKER} != "recur -_+^+_- spank" ]] ; then
___ECLASS_ONCE_EHOOKER="recur -_+^+_- spank"

inherit multiprocessing

declare -g -a _EHOOKER_HOOKS_FIRING=( )
declare -g -i _EHOOKER_BVMAJOR=${BASH_VERSION%%.*}
# do we have bash 4.x or newer (and hence the ability to
# use associative arrays)?
_EHOOKER_BASH4() {
	if (( _EHOOKER_BVMAJOR >= 4 )) ; then
		return 0
	else
		return 1
	fi
}
# This is the registration datastore.  Just to be sure we don't get
# mixed up somehow, use different names for the bash4/+ associative
# and bash<4 indexed versions.  The associative version is expected
# to perform considerably better in the face of massive numbers of
# registrations and listeners (although it's far from clear that
# scenario will ever occur, and it's not a ton of work anyhow).
if _EHOOKER_BASH4 ; then
	declare -g -A _EHOOKER_ADATA=( )
else
	declare -g -a _EHOOKER_IDATA=( )
fi

# some internal functions to manipulate the ehook registration data
# <hook> -> _EHOOKER_GET_HOOK_LISTENERS_RESULT="<space-separated-list-of-listeners>"
_EHOOKER_GET_HOOK_LISTENERS() {
	declare -g _EHOOKER_GET_HOOK_LISTENERS_RESULT=""
	local item
	if _EHOOKER_BASH4 ; then
		_EHOOKER_GET_HOOK_LISTENERS_RESULT="${_EHOOKER_ADATA["${1}"]}"
		return 0
	else
		for item in "${_EHOOKER_IDATA[@]}" ; do
			if [[ ${item%% *} == ${1} ]] ; then
				_EHOOKER_GET_HOOK_LISTENERS_RESULT="${item#* }"
				return 0
			fi
		done
	fi
	return 0
}

# <hook> [<listener(s)>] (no data returned)
_EHOOKER_SET_HOOK_LISTENERS() {
	local hook="$1"
	local listeners="${*:2}"
	if _EHOOKER_BASH4 ; then
		[[ -n ${listeners} ]] || { unset -v _EHOOKER_ADATA["${hook}"] ; return 0 ; }
		_EHOOKER_ADATA["${hook}"]="${listeners}"
	else
		declare -i idx
		for idx in "${!_EHOOKER_IDATA[@]}" ; do
			[[ ${_EHOOKER_IDATA[${idx}]%% *} == ${hook} ]] || continue
			# we found an entry.  if the caller requested an empty list then
			# we ca simply delete it.  Otherwise, replace the old list with
			# the new one
			[[ -n ${listeners} ]] || { unset -v "_EHOOKER_IDATA[${idx}]" ; return 0 ; }
			_EHOOKER_IDATA[${idx}]="${hook} ${listeners}"
			return 0
		done
		# no entry exists yet, so we only need to make one if the caller
		# requested a non-empty list of listener functions
		[[ -n ${listeners} ]] && _EHOOKER_IDATA+=("${hook} ${listeners}")
	fi
	return 0
}

# @FUNCTION: ehook_is_firing
# @USAGE: <hook-name>
# @DESCRIPTION:
# There is no provision by ehooker.eclass for "hook recursion."
#
# In liberally configured userspace environments, infinite recursion
# could sometimes lead to system-wide resource-starvation-induced
# paralysis, so ehooker.eclass implements a simple and effective
# countermeasure: recursion is "impossible" within a given bash process
# environment.  No hook may fire while ehook_fire is already servicing
# that same hook's listeners somewhere lower down in the function call
# chain. 
#
# However, in multi-process environments, no guarantee exists that the same
# hook will not fire simultaneously in different processes.  Furthermore,
# the mechanisms prohibiting recursion are not designed to be resilient
# against deliberate subversion -- only to prevent mistakes by well-meaning users.
#
# The policy is enforced by calling die any time a recursive firing request
# is made to ehook_fire().  In order to avoid such death dynamically, ehook_is_firing
# may be used to test whether the immediate execution context is already servicing
# a hook firing request for the hook in question.
#
# ehook_is_firing returns a nonzero- ("true") result if the hook indicated
# is currently being processed (fired) by ehook_fire and/or hook-listener
# functions, and a zero- ("false") result when the indicated hook is
# currently idle.
ehook_is_firing() {
	local hot_hook
	for hot_hook in "${_EHOOKER_HOOKS_FIRING[@]}" ; do
		[[ ${hot_hook} == $1 ]] && return 0
	done
	return 1
}

# @FUNCTION: ehook_fire
# @USAGE: <hook-name> [-u | --uninterruptible] | [-p | --parallel]
# @DESCRIPTION:
# Causes hook-listener functions that have registered for the
# named hook events to be invoked, either one-by-one, or when the
# --parallel option is used, in parallel.
#
# When not parallelizing, hooks are fired in reverse-order of
# registration.  No such guarantee can exist in the parallelized case,
# for reasons that should be obvious.
#
# By default, the first hook consumer that returns a nonzero value from
# its hook function will immediately interrupt processing and cause
# a nonzero value to be returned from ehook_fire.
#
# If the --uninterruptible option is provided, then all registered hook
# consumers will be invoked regardless of their return values, and
# ehook_fire will return the logical AND of all their return values.
#
# Regardless, ehook_fire returns zero if and only if all listeners
# invoked return zero (or if there are no listeners).
# If any hook returns a nonzero value, the function returns 1.
#
# The --parallel option implies the --uninterruptible option.
# 
# If an attempt is made to fire a hook while that hook is already firing,
# ehook_fire will die.
#
# Years of bitter experince have led most experienced ebuild and eclass
# developers to the counter-intuitive conclusion that ameliorative
# error handling is undesirable in ebuild code, as this leads to
# non-reproducibility.
#
# That stated, if you don't want your ebuild to die due to recursive
# hook invocation, you can check ehook_is_firing before invocation
# and take some other action if the hook event is in-progress.
ehook_fire() {
	debug-print-function ${FUNCNAME} "$@"

	[[ "$#" -eq 0 ]] && die "ehook_fire invoked with no arguments"

	# N.B....!
	#
	# Although a given hook may not recursively fire itself,
	# recursion through this /code path/ is completely
	# legitimate and supported, as hook-listener functions
	# responding to hook "a" are free to fire hook "b."
	#
	# Therefore we must:
	#
	#   o meticulously ensure all scratch-variables
	#     are declared as local variables.... but really,
	#     better yet, don't use scratch-variables at all --
	#     use the positional parameters and relegate
	#     any scratchwork to sub-functions which can't
	#     muck up the lexical environment we bequeath to
	#     the listener functions.
	#
	#   o architect any interactions with global variables
	#     in a manner resilient to recursive invocation
	#
	# Note: in bash, "local" variables bleed right through function
	# invocations as if they were global -- basically almost nothing
	# distinguishes local from global variables until they go out
	# of scope.  This is why we go to such pains to avoid creating
	# local variables below: such variables would flow into the
	# hook-listener functions we invoked, scribbling over any like-named variable
	# the hook-firing function intended to provide to listeners.  Although
	# we could make this a non-issue by using high-bandwith variable names,
	# this would only clutter up the environment even more and we'd still
	# have to worry about recursion-safety at every turn.

	[[ -n "$1" ]] || die "empty hook name not allowed"

	_EHOOKER_GET_HOOK_LISTENERS "$1"

	# no point screwing around any further if nobody's listening
	[[ -n "${_EHOOKER_GET_HOOK_LISTENERS_RESULT}" ]] || { 
		debug-print "${FUNCNAME}: nobody listening for \"$1\""
		return 0
	}

	ehook_is_firing "$1" && \
		die "Recursive invocation of hook \"$1\""

	_EHOOKER_HOOKS_FIRING=( "$1" "${_EHOOKER_HOOKS_FIRING[@]}" )

	# boil down the other arguments to something more managable
	_ehook_fire_normalize_args() {
		local result="d" arg
		for arg in "${@}" ; do
			case "${arg}" in
				-u|--uninterruptible) 
					case "${result}" in
						p|u) continue ;; # already uninterruptible
						d) result="u" ;;
						*) result="e:wtf???" ;;
					esac
					;;
				-p|--parallel)
					case "${result}" in
						p|d|u) result="p" ;; 
						*) result="e:wtf???" ;;
					esac
					;;
				*) result="e:${arg}" ; break ;;
			esac
		done
		echo "${result}"
		return 0
	}

	# new positional parameters:
	# $1 - hook name
	# $2 - one of:
	#        o "d" - default: interruptible invocation
	#        o "u" - uninterruptible: nonparallel, ordered, uninterruptible invocation
	#        o "p" - parallel: parallel, and implicitly uninterruptible, invocation
	#        o "e:<bad argument>" - some bad argument was provided, here it is
	# $3 - one or more listener functions we want to invoke
	#
	# These are the arguments we pass to _ehooker_fire_away, which does all the
	# heavy lifting.
	#
	# NB: it's critically important that we stop using _EHOOKER_GET_HOOK_LISTENERS_RESULT
	# before any recursion occurs, as it's global and likely to be changed by listener code
	set -- "$1" "$(_ehook_fire_normalize_args "${@:2}" )" ${_EHOOKER_GET_HOOK_LISTENERS_RESULT}

	[[ $2 == e* ]] && \
		die "ehook_fire called with illegal argument \"${2#e:}\""

	debug-print "${FUNCNAME}: firing notifications for \"$1\" event to listeners: ${_EHOOKER_GET_HOOK_LISTENERS_RESULT}"
	# can't hurt just to tidy this up...
	unset _EHOOKER_GET_HOOK_LISTENERS_RESULT

	# ready. set....
	_ehooker_fire_away "$@"

	# the hook processing is complete so now it's fine to use some variables.
	local _EHOOKER_HOOK_FIRE_RESULT=$?

	# filter $1 back out of the _EHOOKER_HOOKS_FIRING array
	# pretty sure it will always be at the beginning and we don't
	# need to go to the trouble, but if the semantics changed later
	# this should be more robust... plus I'm really not sure.
	_EHOOKER_HOOKS_FIRING=(
		$(
			for hot_hook in "${_EHOOKER_HOOKS_FIRING[@]}" ; do
				[[ ${hot_hook} != $1 ]] && echo "${hot_hook}"
			done
		)
	)

	debug-print "${FUNCNAME}: notifications for hook \"$1\" complete (result-code: ${_EHOOKER_HOOK_FIRE_RESULT})"
	return ${_EHOOKER_HOOK_FIRE_RESULT}
}

# <hookname> -> true(0) - debug | false(1) - dont debug
_ehooker_hook_is_debuggee() {
	[[ -z ${EHOOKER_DEBUG_HOOKS+1} ]] && return 1
	local ahook
	if [[ $(declare -p EHOOKER_DEBUG_HOOKS) == 'declare -'[aA]* ]] ; then
		for ahook in ${EHOOKER_DEBUG_HOOKS[@]} ; do
			[[ ${ahook} == $1 ]] && return 0
		done
	else
		ewarn "EHOOKER_DEBUG_HOOKS is not an array.  Treating as string."
		for ahook in ${EHOOKER_DEBUG_HOOKS} ; do
			[[ ${ahook} == $1 ]] && return 0
		done
	fi
	return 1
}

# args: <hook-name> <mode: d|u|p> <function-name> [<debug-env-file-name>]
# invoke the listener and handle debug mode processing if indicated.
# The fourth parameter is "fake" -- that is, one cant really pass it in;
# it is conditionally set in code, if we decide to activate debugging.
_ehooker_invoke_listener() {
	debug-print-function ${FUNCNAME} "$@"

	if _ehooker_hook_is_debuggee $1 ; then
		if [[ $2 == p ]] ; then
			ewarn "Unable to debug ${1} -> ${3} hook invocation (PID=$$): parallelized"
		elif ! type md5sum &> /dev/null ; then
			ewarn "Unable to debug ${1} -> ${3} hook invocation (PID=$$): missing md5sum executable"
		else
			set -- "$@" "${T}/ehook_debug_env_$(mktemp XXXX)_$$"
		fi
	fi

	# the number of parameters now tells us whether we are debugged or not
	if [[ $# -eq 4 ]] ; then
		( source "${PORTAGE_BIN_PATH}"/save-ebuild-env.sh && __save_ebuild_env | \
			grep -v '^declare\s-a\sBASH_LINENO' > "${4}.old" )
	fi

	if ! declare -F "$3" >/dev/null 2>&1 ; then
		die "Listener \"$3\" for ehook \"$1\" does not appear to exist."
	fi

	# and... invoke!
	$3 "$1"

	# OK, invocation is complete.  No lexical scope pollution threat remains, but we
	# still don't want to make local variables, as they would clutter up debugging output.
	# Yet we need at least to check for debug mode before returning.  So we need to
	# store $? somehow.  The best solution is unchanged -- positional parameter abuse.
	#
	# The new positional parameters will be:
	# <return-code> <hook-name> <mode: d|u|p> <function-name> [<debug-env-file-name>]
	set -- "$?" "$@"

	# after the above we must check for five parameters to test for debug mode :)
	if [[ $# -eq 5 ]] ; then
		# yes, this is completely evil and wrong.  consider whether or not this
		# is of sufficient utility to leave in...
		( source "${PORTAGE_BIN_PATH}"/save-ebuild-env.sh && __save_ebuild_env | \
			grep -v '^declare\s-a\sBASH_LINENO' > "${5}.new" )
		md5sum "${5}.old" | sed 's:\.old$:.new:' | {
			if md5sum --quiet --status -c ; then
				einfo
				einfo "environment not modified during execution of function"
				einfo "${4} while processing ehook: ${2}"
				einfo
			else
				einfo
				einfo "environment modified during execution of function"
				einfo "${4} while processing ehook: ${2}"
				einfo
				einfo "$(diff -du "${5}.old" "${5}.new")"
				einfo
			fi
		}
		case ${EAPI:-0} in
			4|5) nonfatal rm -f "${T}"/${5}.{old,new} ;;
			*) rm -f "${T}"/${5}.{old,new} ;;
		esac
	fi

	return $1
}

# ehook_fire helper function:
# args: <hookname> <mode: d|u|p> <list of hook-listners to invoke>
_ehooker_fire_away() {
	[[ $2 == p ]] && multijob_init
	while [[ $# -gt 2 ]] ; do
		case $2 in
			d)
				_ehooker_invoke_listener "${@:1:3}" || return 1
				;;
			u)
				if ! _ehooker_invoke_listener "${@:1:3}" ; then
					# kind of a hack -- we know we want to return
					# 1 no matter what, now, so just recurse and
					# then return it.  Argument #3 should disappear,
					# same as if we were iterating.
					_ehooker_fire_away "$1" u "${@:4}"
					return 1
				fi
				;;
			p)
				multijob_child_init _ehooker_invoke_listener "${@:1:3}"
				;;
		esac
		# make argument #3 disappear
		set -- "$1" "$2" "${@:4}"
	done
	if [[ $2 == p ]] ; then
		# multijob_finish will return with the correct result-code
		multijob_finish
	else
		# we get here via "d" and "u" mode when nobody failed, so...
		return 0
	fi
}

# @FUNCTION: ehook
# @USAGE: <hook-name> <hook-listener-function>
# @DESCRIPTION:
# Registers a new listener for the named hook.  New regsitrations
# go to the top of the list of listeners so if the hook is fired
# unparallelized, the most recently registered listener gets the
# chance to abort processing first.
#
# The provided hook-listener function need not exist until the hook
# is invoked.  However, at that time, if it isn't defined,
# ehook_fire dies.
#
# Hook listener functions may be registered to more than one hook.
# This could be useful, i.e., for logging.
#
# Hook listener functions recieve a single argument, containing the
# name of the hook that triggered their execution.
#
# If your hook listener function returns a nonzero value, this will
# be communicated to the hook-firing function, and, at the option of
# the hook-firing function, may also prevent additional hook-listening
# functions from recieving notification.
#
# Note that manipulation of hook registration using the ehook and
# eunhook functions is a fatal error if the hook is already in the process
# of being fired.  If this is unacceptable then check ehook_is_firing before
# invoking these functions.
ehook() {
	[[ "$#" -ne 2 ]] && die "ehook called with invalid number of arguments."
	local hookname="$1"
	local hookfunc="$2"
	[[ -n "${hookname}" ]] || die "ehook called with empty hook name"
	[[ -n "${hookfunc}" ]] || die "ehook called with empty listener argument"
	# [[ -n "$( declare -F "${hookfunc}" )" ]] || \
	#	die "ehook called with non-function listener argument \"${hookfunc}\""
	case "${hookname}${hookfunc}" in
		*\ *) die "ehook called with illegal space-containing argument" ;;
	esac
	ehook_is_firing "${hookname}" && \
		die "ehook \"${hookname}\" already firing during ehook invocation."
	
	_EHOOKER_GET_HOOK_LISTENERS "${hookname}"
	local new_listeners=$( echo ${hookfunc} $(
		for listener in $_EHOOKER_GET_HOOK_LISTENERS_RESULT ; do
			[[ ${listener} != ${hookfunc} ]] && echo "${listener}"
		done
	) )
	# figure out if the request is a noop.  If so, do nothing.  Otherwise
	# enforce the rule about hook-listener registration manipulation during
	# hook firing.
	if [[ ${new_listeners} != ${_EHOOKER_GET_HOOK_LISTENERS_RESULT} ]] ; then
		ehook_is_firing "${hookname}" && \
			die "ehook \"${hookname}\" already firing during ehook invocation."
		_EHOOKER_SET_HOOK_LISTENERS "${hookname}" ${new_listeners}
	fi
	return 0
}

# @FUNCTION: eunhook
# @USAGE: <hook-name> <hook-listener-function>
# @DESCRIPTION:
# reverses any active registration of <hook-listener-function> to <hook-name>
# A noop if no such registration exists.
#
# Note that manipulation of hook registration using the ehook and
# eunhook functions is a fatal error if the hook is already in the process
# of being fired.  If this is unacceptable then check ehook_is_firing before
# invoking these functions.
eunhook() {
	[[ "$#" -ne 2 ]] && die "ehook called with invalid number of arguments."
	local hookname="$1"
	local hookfunc="$2"
	[[ -n ${hookname} ]] || die "ehook called with empty hook name"
	[[ -n ${hookfunc} ]] || die "ehook called with empty listener argument"
	case "${hookname}${hookfunc}" in
		*\ *) die "ehook called with illegal space-containing argument" ;;
	esac

	_EHOOKER_GET_HOOK_LISTENERS "${hookname}"
	local new_listeners=$( echo $(
		for listener in $_EHOOKER_GET_HOOK_LISTENERS_RESULT ; do
			[[ ${listener} != ${hookfunc} ]] && echo "${listener}"
		done
	) )
	# figure out if the request is a noop.  If so, do nothing.  Otherwise
	# enforce the rule about hook-listener registration manipulation during
	# hook firing.
	if [[ ${new_listeners} != ${_EHOOKER_GET_HOOK_LISTENERS_RESULT} ]] ; then
		ehook_is_firing "${hookname}" && \
			die "ehook \"${hookname}\" already firing during eunhook invocation."
		_EHOOKER_SET_HOOK_LISTENERS "${hookname}" ${new_listeners}
	fi
	return 0
}

fi # ___ECLASS_ONCE_EHOOKER