[Tarantool-patches] [PATCH 1/1] fio: close unused descriptors automatically

Georgy Kirichenko georgy at tarantool.org
Sun Feb 9 21:34:38 MSK 2020


Hi, thanks for the patch.

Unfortunately your patch is not correct because fiber.create
yields (which could be fixed by using fiber.new) and raises an error what is not 
allowed inside gc callbacks.

>From my point of view the possible approach is a dedicated fiber with
a list of pending callbacks.

WBR

On Sunday, 9 February 2020 20:37:09 MSK Vladislav Shpilevoy wrote:
> Fio.open() returned a file descriptor, which was not closed
> automatically after all its links were nullified. In other words,
> GC didn't close the descriptor.
> 
> This was not really useful, because after fio.open() an exception
> may appear, and user needed to workaround this to manually call
> fio_object:close(). Also this was not consistent with io.open().
> 
> Now fio.open() object closes the descriptor automatically when
> GCed.
> 
> Closes #4727
> 
> @TarantoolBot document
> Title: fio descriptor is closed automatically by GC
> 
> fio.open() returns a descriptor which can be closed manually by
> calling :close() method, or it will be closed automatically, when
> it has no references, and GC deletes it.
> 
> :close() method existed always, auto GC was added just now.
> 
> ---
> 
> @ChangeLog
> - fio descriptors are closed on garbage collection (gh-4727).
> 
>  src/lua/fio.lua       | 33 ++++++++++++---
>  test/app/fio.result   | 93 +++++++++++++++++++++++++++++++++++++++++++
>  test/app/fio.test.lua | 58 +++++++++++++++++++++++++++
>  3 files changed, 178 insertions(+), 6 deletions(-)
> 
> diff --git a/src/lua/fio.lua b/src/lua/fio.lua
> index 4692e1026..41901904c 100644
> --- a/src/lua/fio.lua
> +++ b/src/lua/fio.lua
> @@ -141,10 +141,12 @@ end
> 
>  fio_methods.close = function(self)
>      local res, err = internal.close(self.fh)
> -    self.fh = -1
>      if err ~= nil then
>          return false, err
>      end
> +    ffi.gc(self._gc, nil)
> +    self._gc = nil
> +    self.fh = -1
>      return res
>  end
> 
> @@ -160,7 +162,23 @@ fio_methods.stat = function(self)
>      return internal.fstat(self.fh)
>  end
> 
> -local fio_mt = { __index = fio_methods }
> +local fio_mt = {
> +    __index = fio_methods,
> +    __serialize = function(obj)
> +        return {fh = obj.fh}
> +    end,
> +}
> +
> +local function fio_wrap(fh)
> +    return setmetatable({
> +        fh = fh,
> +        _gc = ffi.gc(ffi.new('char[1]'), function()
> +            -- FFI GC can't yield. Internal.close() yields.
> +            -- Collect the garbage later, in a separate fiber.
> +            fiber.create(internal.close, fh)
> +        end)
> +    }, fio_mt)
> +end
> 
>  fio.open = function(path, flags, mode)
>      local iflag = 0
> @@ -202,10 +220,13 @@ fio.open = function(path, flags, mode)
>      if err ~= nil then
>          return nil, err
>      end
> -
> -    fh = { fh = fh }
> -    setmetatable(fh, fio_mt)
> -    return fh
> +    local ok, res = pcall(fio_wrap, fh)
> +    if not ok then
> +        internal.close(fh)
> +        -- This is either OOM or bad syntax, both require throw.
> +        return error(res)
> +    end
> +    return res
>  end
> 
>  fio.pathjoin = function(...)
> diff --git a/test/app/fio.result b/test/app/fio.result
> index f83c43f44..6345ac22e 100644
> --- a/test/app/fio.result
> +++ b/test/app/fio.result
> @@ -1456,3 +1456,96 @@ fio.mktree('/dev/null/dir')
>  - false
>  - 'Error creating directory /dev/null: File exists'
>  ...
> +--
> +-- gh-4727: fio handler GC.
> +--
> +flags = {'O_CREAT', 'O_RDWR'}
> +---
> +...
> +mode = {'S_IRWXU'}
> +---
> +...
> +filename = 'test4727.txt'
> +---
> +...
> +fh1 = nil
> +---
> +...
> +fh2 = nil
> +---
> +...
> +-- Idea of the test is that according to the Open Group standard,
> +-- open() always returns the smallest available descriptor. This
> +-- means, that in 'open() + close() + open()' the second open()
> +-- should return the same value as the first call, if no other
> +-- threads/fibers managed to interfere. Because of the
> +-- interference the sequence may need to be called multiple times
> +-- to catch a couple of equal descriptors.
> +-- GC function of a fio object creates a new fiber. Give it         \
> +-- time to execute.                                                 \
> +test_run:wait_cond(function()                                       \
> +    collectgarbage('collect')                                       \
> +    local f = fio.open(filename, flags, mode)                       \
> +    fh1 = f.fh                                                      \
> +    f = nil                                                         \
> +    collectgarbage('collect')                                       \
> +    fiber.yield()                                                   \
> +    f = fio.open(filename, flags, mode)                             \
> +    fh2 = f.fh                                                      \
> +    f = nil                                                         \
> +    collectgarbage('collect')                                       \
> +    fiber.yield()                                                   \
> +    return fh1 == fh2                                               \
> +end) or {fh1, fh2}
> +---
> +- true
> +...
> +-- Ensure, that GC does not break anything after explicit close.
> +-- Idea of the test is the same as in the previous test, but now
> +-- the second descriptor is used for something. If GC of the first
> +-- fio object is called even after close(), it would close the
> +-- same descriptor, already used by the second fio object. And it
> +-- won't be able to write anything. Or will write, but to a
> +-- totally different descriptor created by some other
> +-- fiber/thread. This is why read() is called on the same file
> +-- afterwards.
> +f = nil
> +---
> +...
> +test_run:wait_cond(function()                                       \
> +    f = fio.open(filename, flags, mode)                             \
> +    fh1 = f.fh                                                      \
> +    f:close()                                                       \
> +    f = fio.open(filename, flags, mode)                             \
> +    fh2 = f.fh                                                      \
> +    return fh1 == fh2                                               \
> +end)
> +---
> +- true
> +...
> +collectgarbage('collect')
> +---
> +- 0
> +...
> +fiber.yield()
> +---
> +...
> +f:write('test')
> +---
> +- true
> +...
> +f:close()
> +---
> +- true
> +...
> +f = fio.open(filename, flags, mode)
> +---
> +...
> +f:read()
> +---
> +- test
> +...
> +f:close()
> +---
> +- true
> +...
> diff --git a/test/app/fio.test.lua b/test/app/fio.test.lua
> index 56c957d8a..c726bade6 100644
> --- a/test/app/fio.test.lua
> +++ b/test/app/fio.test.lua
> @@ -474,3 +474,61 @@ test_run:cmd("clear filter")
>  --
>  fio.mktree('/dev/null')
>  fio.mktree('/dev/null/dir')
> +
> +--
> +-- gh-4727: fio handler GC.
> +--
> +flags = {'O_CREAT', 'O_RDWR'}
> +mode = {'S_IRWXU'}
> +filename = 'test4727.txt'
> +fh1 = nil
> +fh2 = nil
> +-- Idea of the test is that according to the Open Group standard,
> +-- open() always returns the smallest available descriptor. This
> +-- means, that in 'open() + close() + open()' the second open()
> +-- should return the same value as the first call, if no other
> +-- threads/fibers managed to interfere. Because of the
> +-- interference the sequence may need to be called multiple times
> +-- to catch a couple of equal descriptors.
> +test_run:wait_cond(function()                                       \
> +    collectgarbage('collect')                                       \
> +    local f = fio.open(filename, flags, mode)                       \
> +    fh1 = f.fh                                                      \
> +    f = nil                                                         \
> +    collectgarbage('collect')                                       \
> +-- GC function of a fio object creates a new fiber. Give it         \
> +-- time to execute.                                                 \
> +    fiber.yield()                                                   \
> +    f = fio.open(filename, flags, mode)                             \
> +    fh2 = f.fh                                                      \
> +    f = nil                                                         \
> +    collectgarbage('collect')                                       \
> +    fiber.yield()                                                   \
> +    return fh1 == fh2                                               \
> +end) or {fh1, fh2}
> +
> +-- Ensure, that GC does not break anything after explicit close.
> +-- Idea of the test is the same as in the previous test, but now
> +-- the second descriptor is used for something. If GC of the first
> +-- fio object is called even after close(), it would close the
> +-- same descriptor, already used by the second fio object. And it
> +-- won't be able to write anything. Or will write, but to a
> +-- totally different descriptor created by some other
> +-- fiber/thread. This is why read() is called on the same file
> +-- afterwards.
> +f = nil
> +test_run:wait_cond(function()                                       \
> +    f = fio.open(filename, flags, mode)                             \
> +    fh1 = f.fh                                                      \
> +    f:close()                                                       \
> +    f = fio.open(filename, flags, mode)                             \
> +    fh2 = f.fh                                                      \
> +    return fh1 == fh2                                               \
> +end)
> +collectgarbage('collect')
> +fiber.yield()
> +f:write('test')
> +f:close()
> +f = fio.open(filename, flags, mode)
> +f:read()
> +f:close()

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 488 bytes
Desc: This is a digitally signed message part.
URL: <https://lists.tarantool.org/pipermail/tarantool-patches/attachments/20200209/6facfb48/attachment.sig>


More information about the Tarantool-patches mailing list