Environment: queue master (07fd732), Tarantool 3.8.0 (also the same code in 1.5.0). memtx.
Summary
fifottl (and therefore limfifottl) creates its stop channel as an unbuffered fiber.channel() (fifottl.lua#L200), but the TTL fiber only reads it in the box.info.ro == true branch (fifottl.lua#L171-L177). So:
tube.raw:stop() called while the instance is RW blocks the caller forever.
fifottl has no method.drop, so tube:drop() never calls stop() at all: the space is dropped and the TTL fiber leaks, one per dropped tube. After the instance becomes RO the leaked fibers sit in the sync_chan:get(0.1) poll loop forever.
utubettl uses fiber.channel(1) so its stop() returns, but the fiber keeps running in RW; utubettl's drop() then makes it crash on the dropped space (attempt to index field 'watch' (a nil value)).
Repro
local fiber = require('fiber')
box.cfg{}
local queue = require('queue')
-- 1. stop() blocks while RW
local tube = queue.create_tube('t', 'fifottl')
local done = false
fiber.create(function() tube.raw:stop(); done = true end)
fiber.sleep(1)
print('fifottl raw:stop() returned within 1s:', done) -- false (utubettl: true)
-- 2. drop() leaks the fiber
local function ttl_fibers()
local n = 0
for _, f in pairs(fiber.info()) do if f.name == 'fifottl' then n = n + 1 end end
return n
end
local before = ttl_fibers()
for i = 1, 3 do queue.create_tube('leak_' .. i, 'fifottl'):drop() end
fiber.sleep(0.5)
print('fifottl fibers leaked by 3 drops:', ttl_fibers() - before) -- 3
Output on master:
fifottl raw:stop() returned within 1s: false
fifottl fibers leaked by 3 drops: 3
Expected
stop() returns in both RW and RO mode and the fiber actually terminates; drop() stops the fiber before dropping the space (as utubettl tries to do), and does not leave the fiber running against a dropped space.
Suggested fix
Check a stop flag (or the channel with a zero timeout) in the RW branch of the loop as well, make stop() wait for the fiber to finish before clearing self.fiber, and add method.drop to fifottl that calls stop() first. fifottl should also use a buffered channel like utubettl so stop() cannot block the state machine fiber.
Environment: queue master (07fd732), Tarantool 3.8.0 (also the same code in 1.5.0). memtx.
Summary
fifottl(and thereforelimfifottl) creates its stop channel as an unbufferedfiber.channel()(fifottl.lua#L200), but the TTL fiber only reads it in thebox.info.ro == truebranch (fifottl.lua#L171-L177). So:tube.raw:stop()called while the instance is RW blocks the caller forever.fifottlhas nomethod.drop, sotube:drop()never callsstop()at all: the space is dropped and the TTL fiber leaks, one per dropped tube. After the instance becomes RO the leaked fibers sit in thesync_chan:get(0.1)poll loop forever.utubettlusesfiber.channel(1)so itsstop()returns, but the fiber keeps running in RW;utubettl'sdrop()then makes it crash on the dropped space (attempt to index field 'watch' (a nil value)).Repro
Output on master:
Expected
stop()returns in both RW and RO mode and the fiber actually terminates;drop()stops the fiber before dropping the space (asutubettltries to do), and does not leave the fiber running against a dropped space.Suggested fix
Check a stop flag (or the channel with a zero timeout) in the RW branch of the loop as well, make
stop()wait for the fiber to finish before clearingself.fiber, and addmethod.droptofifottlthat callsstop()first.fifottlshould also use a buffered channel likeutubettlsostop()cannot block the state machine fiber.