30206523120

Committed 26 Jul 2026 02:39PM UTC coverage: 85.11% (+0.04%) from 85.072%

Build # 30206523120

Build Type

push

github

Committed by

Tieske

Commit Message

fix(semaphore): stop leaking resources to canceled queue waiters

Same root cause as the copas.lock deadlock (#199): semaphore:give()
deducted a queued waiter's requested resources from the pool and then
called copas.wakeup() without checking whether it actually resumed
the coroutine. If that waiter had been canceled externally (eg. via
copas.removethread()/future:cancel()) while queued, the resources
were permanently subtracted from the pool even though the canceled
coroutine never runs to consume or return them - silently shrinking
the semaphore's effective capacity forever, and starving any real
waiter still queued behind it.

give() now only deducts the resources once copas.wakeup() confirms
the waiter actually resumed; a stale/canceled entry is dropped without
touching the count, and the loop continues to the next waiter in line.

Coverage Stats

18 of 18 new or added lines in 2 files covered. (100.0%)

127 existing lines in 5 files now uncovered.

1429 of 1679 relevant lines covered (85.11%)

67776.83 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

97.25

/src/copas/semaphore.lua

local copas = require("copas")

local coroutine_running = coroutine.running
if _VERSION=="Lua 5.1" and not jit then     -- obsolete: only for Lua 5.1 compatibility
  coroutine_running = require("coxpcall").running
end

local DEFAULT_TIMEOUT = 10

local semaphore = {}
semaphore.__index = semaphore


-- registry, semaphore indexed by the coroutines using them.
local registry = setmetatable({}, { __mode="kv" })


-- create a new semaphore
-- @param max maximum number of resources the semaphore can hold (this maximum does NOT include resources that have been given but not yet returned).
-- @param start (optional, default 0) the initial resources available
-- @param seconds (optional, default 10) default semaphore timeout in seconds, or `math.huge` to have no timeout.
function semaphore.new(max, start, seconds)
  local timeout = tonumber(seconds or DEFAULT_TIMEOUT) or -1
  if timeout < 0 then
    error("expected timeout (2nd argument) to be a number greater than or equal to 0, got: " .. tostring(seconds), 2)
  end
  if type(max) ~= "number" or max < 1 then
    error("expected max resources (1st argument) to be a number greater than 0, got: " .. tostring(max), 2)
  end

  local self = setmetatable({
      count = start or 0,
      max = max,
      timeout = timeout,
      q_tip = 1,    -- position of next entry waiting
      q_tail = 1,   -- position where next one will be inserted
      queue = {},
      to_flags = setmetatable({}, { __mode = "k" }), -- timeout flags indexed by coroutine
    }, semaphore)

  return self
end


do
  local destroyed_func = function()
    return nil, "destroyed"
  end

  local destroyed_semaphore_mt = {
    __index = function()
      return destroyed_func
    end
  }

  -- destroy a semaphore.
  -- Releases all waiting threads with `nil+"destroyed"`
  function semaphore:destroy()
    self:give(math.huge)
    self.destroyed = true
    setmetatable(self, destroyed_semaphore_mt)
    return true
  end
end


-- Gives resources.
-- @param given (optional, default 1) number of resources to return. If more
-- than the maximum are returned then it will be capped at the maximum and
-- error "too many" will be returned.
function semaphore:give(given)
  local err
  given = given or 1
  local count = self.count + given
  --print("now at",count, ", after +"..given)
  if count > self.max then
    count = self.max
    err = "too many"
  end

  while self.q_tip < self.q_tail do
    local i = self.q_tip
    local nxt = self.queue[i] -- there can be holes, so nxt might be nil
    if not nxt then
      self.q_tip = i + 1
    elseif not copas.issleeping(nxt.co) then
      -- stale entry; 'nxt.co' was canceled externally (eg. via
      -- copas.removethread/future:cancel) and will never resume to
      -- consume or give back resources. Drop it without touching 'count',
      -- and regardless of how much it had requested, so it doesn't block
      -- waiters behind it that request less than it did.
      self.queue[i] = nil
      self.to_flags[nxt.co] = nil
      self.q_tip = i + 1
      nxt.co = nil
    elseif count >= nxt.requested then
      -- release it
      self.queue[i] = nil
      self.to_flags[nxt.co] = nil
      count = count - nxt.requested
      self.q_tip = i + 1
      copas.wakeup(nxt.co)
      nxt.co = nil
    else
      break -- we ran out of resources
    end
  end

  if self.q_tip == self.q_tail then  -- reset queue
    self.queue = {}
    self.q_tip = 1
    self.q_tail = 1
  end

  self.count = count
  if err then
    return nil, err
  end
  return true
end



local function timeout_handler(co)
  local self = registry[co]
  --print("checking timeout ", co)
  if not self then
    return
  end

  for i = self.q_tip, self.q_tail do
    local item = self.queue[i]
    if item and co == item.co then
      self.queue[i] = nil
      self.to_flags[co] = true
      --print("marked timeout ", co)
      copas.wakeup(co)
      return
    end
  end
  -- nothing to do here...
end


-- Requests resources from the semaphore.
-- Waits if there are not enough resources available before returning.
-- @param requested (optional, default 1) the number of resources requested
-- @param timeout (optional, defaults to semaphore timeout) timeout in
-- seconds. If 0 it will either succeed or return immediately with error "timeout".
-- If `math.huge` it will wait forever.
-- @return true, or nil+"destroyed"
function semaphore:take(requested, timeout)
  requested = requested or 1
  if self.q_tail == 1 and self.count >= requested then
    -- nobody is waiting before us, and there is enough in store
    self.count = self.count - requested
    return true
  end

  if requested > self.max then
    return nil, "too many"
  end

  local to = timeout or self.timeout
  if to == 0 then
    return nil, "timeout"
  end

  -- get in line
  local co = coroutine_running()
  self.to_flags[co] = nil
  registry[co] = self
  copas.timeout(to, timeout_handler)

  self.queue[self.q_tail] = {
    co = co,
    requested = requested,
    --timeout = nil, -- flag indicating timeout
  }
  self.q_tail = self.q_tail + 1

  copas.pauseforever() -- block until woken
  registry[co] = nil

  if self.to_flags[co] then
    -- a timeout happened
    self.to_flags[co] = nil
    return nil, "timeout"
  end

  copas.timeout(0)

  if self.destroyed then
    return nil, "destroyed"
  end

  return true
end

-- returns current available resources
function semaphore:get_count()
  return self.count
end

-- returns total shortage for requested resources
function semaphore:get_wait()
  local wait = 0
  for i = self.q_tip, self.q_tail - 1 do
    wait = wait + ((self.queue[i] or {}).requested or 0)
  end
  return wait - self.count
end


return semaphore

1	local copas = require("copas")	162✔
2
3	local coroutine_running = coroutine.running	162✔
4	if _VERSION=="Lua 5.1" and not jit then -- obsolete: only for Lua 5.1 compatibility	162✔
UNCOV 5	coroutine_running = require("coxpcall").running	27✔
6	end
7
8	local DEFAULT_TIMEOUT = 10	162✔
9
10	local semaphore = {}	162✔
11	semaphore.__index = semaphore	162✔
12
13
14	-- registry, semaphore indexed by the coroutines using them.
15	local registry = setmetatable({}, { __mode="kv" })	162✔
16
17
18	-- create a new semaphore
19	-- @param max maximum number of resources the semaphore can hold (this maximum does NOT include resources that have been given but not yet returned).
20	-- @param start (optional, default 0) the initial resources available
21	-- @param seconds (optional, default 10) default semaphore timeout in seconds, or `math.huge` to have no timeout.
22	function semaphore.new(max, start, seconds)	162✔
23	local timeout = tonumber(seconds or DEFAULT_TIMEOUT) or -1	506✔
24	if timeout < 0 then	506✔
25	error("expected timeout (2nd argument) to be a number greater than or equal to 0, got: " .. tostring(seconds), 2)	×
26	end
27	if type(max) ~= "number" or max < 1 then	506✔
28	error("expected max resources (1st argument) to be a number greater than 0, got: " .. tostring(max), 2)	×
29	end
30
31	local self = setmetatable({	1,012✔
32	count = start or 0,	506✔
33	max = max,	506✔
34	timeout = timeout,	506✔
35	q_tip = 1, -- position of next entry waiting	423✔
36	q_tail = 1, -- position where next one will be inserted	423✔
37	queue = {},	506✔
38	to_flags = setmetatable({}, { __mode = "k" }), -- timeout flags indexed by coroutine	506✔
39	}, semaphore)	506✔
40
41	return self	506✔
42	end
43
44
45	do
46	local destroyed_func = function()
47	return nil, "destroyed"	24✔
48	end
49
50	local destroyed_semaphore_mt = {	162✔
51	__index = function()
52	return destroyed_func	24✔
53	end
54	}
55
56	-- destroy a semaphore.
57	-- Releases all waiting threads with `nil+"destroyed"`
58	function semaphore:destroy()	162✔
59	self:give(math.huge)	386✔
60	self.destroyed = true	386✔
61	setmetatable(self, destroyed_semaphore_mt)	386✔
62	return true	386✔
63	end
64	end
65
66
67	-- Gives resources.
68	-- @param given (optional, default 1) number of resources to return. If more
69	-- than the maximum are returned then it will be capped at the maximum and
70	-- error "too many" will be returned.
71	function semaphore:give(given)	162✔
72	local err
73	given = given or 1	638✔
74	local count = self.count + given	638✔
75	--print("now at",count, ", after +"..given)
76	if count > self.max then	638✔
77	count = self.max	392✔
78	err = "too many"	392✔
79	end
80
81	while self.q_tip < self.q_tail do	860✔
82	local i = self.q_tip	228✔
83	local nxt = self.queue[i] -- there can be holes, so nxt might be nil	228✔
84	if not nxt then	228✔
85	self.q_tip = i + 1	12✔
86	elseif not copas.issleeping(nxt.co) then	252✔
87	-- stale entry; 'nxt.co' was canceled externally (eg. via
88	-- copas.removethread/future:cancel) and will never resume to
89	-- consume or give back resources. Drop it without touching 'count',
90	-- and regardless of how much it had requested, so it doesn't block
91	-- waiters behind it that request less than it did.
92	self.queue[i] = nil	18✔
93	self.to_flags[nxt.co] = nil	18✔
94	self.q_tip = i + 1	18✔
95	nxt.co = nil	18✔
96	elseif count >= nxt.requested then	198✔
97	-- release it
98	self.queue[i] = nil	192✔
99	self.to_flags[nxt.co] = nil	192✔
100	count = count - nxt.requested	192✔
101	self.q_tip = i + 1	192✔
102	copas.wakeup(nxt.co)	192✔
103	nxt.co = nil	192✔
104	else
105	break -- we ran out of resources
106	end
107	end
108
109	if self.q_tip == self.q_tail then -- reset queue	638✔
110	self.queue = {}	632✔
111	self.q_tip = 1	632✔
112	self.q_tail = 1	632✔
113	end
114
115	self.count = count	638✔
116	if err then	638✔
117	return nil, err	392✔
118	end
119	return true	246✔
120	end
121
122
123
124	local function timeout_handler(co)
125	local self = registry[co]	12✔
126	--print("checking timeout ", co)
127	if not self then	12✔
128	return	×
129	end
130
131	for i = self.q_tip, self.q_tail do	12✔
132	local item = self.queue[i]	12✔
133	if item and co == item.co then	12✔
134	self.queue[i] = nil	12✔
135	self.to_flags[co] = true	12✔
136	--print("marked timeout ", co)
137	copas.wakeup(co)	12✔
138	return	12✔
139	end
140	end
141	-- nothing to do here...
142	end
143
144
145	-- Requests resources from the semaphore.
146	-- Waits if there are not enough resources available before returning.
147	-- @param requested (optional, default 1) the number of resources requested
148	-- @param timeout (optional, defaults to semaphore timeout) timeout in
149	-- seconds. If 0 it will either succeed or return immediately with error "timeout".
150	-- If `math.huge` it will wait forever.
151	-- @return true, or nil+"destroyed"
152	function semaphore:take(requested, timeout)	162✔
153	requested = requested or 1	348✔
154	if self.q_tail == 1 and self.count >= requested then	348✔
155	-- nobody is waiting before us, and there is enough in store
156	self.count = self.count - requested	108✔
157	return true	108✔
158	end
159
160	if requested > self.max then	240✔
161	return nil, "too many"	6✔
162	end
163
164	local to = timeout or self.timeout	234✔
165	if to == 0 then	234✔
166	return nil, "timeout"	12✔
167	end
168
169	-- get in line
170	local co = coroutine_running()	222✔
171	self.to_flags[co] = nil	222✔
172	registry[co] = self	222✔
173	copas.timeout(to, timeout_handler)	222✔
174
175	self.queue[self.q_tail] = {	222✔
176	co = co,	222✔
177	requested = requested,	222✔
178	--timeout = nil, -- flag indicating timeout
179	}	222✔
180	self.q_tail = self.q_tail + 1	222✔
181
182	copas.pauseforever() -- block until woken	222✔
183	registry[co] = nil	204✔
184
185	if self.to_flags[co] then	204✔
186	-- a timeout happened
187	self.to_flags[co] = nil	12✔
188	return nil, "timeout"	12✔
189	end
190
191	copas.timeout(0)	192✔
192
193	if self.destroyed then	192✔
194	return nil, "destroyed"	54✔
195	end
196
197	return true	138✔
198	end
199
200	-- returns current available resources
201	function semaphore:get_count()	162✔
202	return self.count	78✔
203	end
204
205	-- returns total shortage for requested resources
206	function semaphore:get_wait()	162✔
207	local wait = 0	72✔
208	for i = self.q_tip, self.q_tail - 1 do	156✔
209	wait = wait + ((self.queue[i] or {}).requested or 0)	84✔
210	end
211	return wait - self.count	72✔
212	end
213
214
215	return semaphore	162✔

lunarmodules / copas / 30206523120

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous