diff options
| -rw-r--r-- | erts/emulator/internal_doc/CarrierMigration.md | 50 | ||||
| -rw-r--r-- | lib/kernel/src/inet.erl | 30 | ||||
| -rw-r--r-- | lib/kernel/test/inet_SUITE.erl | 32 | ||||
| -rw-r--r-- | lib/ssl/src/ssl.appup.src | 4 |
4 files changed, 71 insertions, 45 deletions
diff --git a/erts/emulator/internal_doc/CarrierMigration.md b/erts/emulator/internal_doc/CarrierMigration.md index 7afdb70aef..2a9594db25 100644 --- a/erts/emulator/internal_doc/CarrierMigration.md +++ b/erts/emulator/internal_doc/CarrierMigration.md @@ -16,12 +16,12 @@ When a carrier is empty, i.e. contains only one large free block, it is deallocated. Since multiblock carriers can contain both allocated blocks and free blocks at the same time, an allocator instance might be stuck with a large amount of poorly utilized carriers if the memory -load decrease. After a peak in memory usage it is expected that not -all memory can be returned since the blocks still allocated is likely +load decreases. After a peak in memory usage it is expected that not +all memory can be returned since the blocks still allocated are likely to be dispersed over multiple carriers. Such poorly utilized carriers -can usually be reused if the memory load increase again. However, +can usually be reused if the memory load increases again. However, since each scheduler thread manages its own set of allocator -instances, and memory load is not necessarily connected to CPU load we +instances, and memory load is not necessarily correlated to CPU load, we might get into a situation where there are lots of poorly utilized multiblock carriers on some allocator instances while we need to allocate new multiblock carriers on other allocator instances. In @@ -50,13 +50,13 @@ the allocator instance manages. Free blocks in one specific carrier can be referred to from potentially every other carrier that is managed, and the amount of such references can be huge. That is, the work of removing the free blocks of such a carrier from the search -tree will be huge. One way of solving this could be to not migrate +tree will be huge. One way of solving this could be not to migrate carriers that contain lots of free blocks, but this would prevent us -from migrating carriers that potentially needs to be migrated in order +from migrating carriers that potentially need to be migrated in order to solve the problem we set out to solve. By using one data structure of free blocks in each carrier and an -allocator instance wide data structure of carriers managed by the +allocator instance-wide data structure of carriers managed by the allocator instance, the work needed in order to remove and add carriers can be kept to a minimum. When migration of carriers is enabled on a specific allocator type, we require that an allocation @@ -76,9 +76,9 @@ through a pool of carriers. In order for a carrier migration to complete, one scheduler needs to move the carrier into the pool, and another scheduler needs to take the carrier out of the pool. -The pool is implemented as a lock free, circular, double linked, +The pool is implemented as a lock-free, circular, double linked, list. The list contains a sentinel which is used as the starting point -when inserting to, or fetching from the pool. Carriers in the pool are +when inserting to, or fetching from, the pool. Carriers in the pool are elements in this list. The list can be modified by all scheduler threads @@ -108,19 +108,19 @@ all search operations need to read the content of the sentinel. If we were to modify the sentinel, the cache line containing the sentinel would unnecessarily be bounced between processors. -The `prev`, and `next` fields in the elements of the list contains the +The `prev` and `next` fields in the elements of the list contain the value of the pointer, a modification marker, and a deleted marker. Memory operations on these fields are done using atomic memory operations. When a thread has set the modification marker in a field, no-one except the thread that set the marker is allowed to modify the -field. If multiple modification markers needs to be set, we always +field. If multiple modification markers need to be set, we always begin with `next` fields followed by `prev` fields in the order following the actual pointers. This guarantees that no deadlocks will occur. When a carrier is being removed from a pool, we mark it with a thread progress value that needs to be reached before we are allowed to -modify the `next`, and `prev` fields. That is, until we reach this +modify the `next` and `prev` fields. That is, until we reach this thread progress we are not allowed to insert the carrier into the pool again, and we are not allowed to deallocate the carrier. This ensures that threads inspecting the pool always will be able to traverse the @@ -130,12 +130,12 @@ threads may have references to it via the pool. ### Migration ### -There exist one pool for each allocator type enabling migration of +There exists one pool for each allocator type enabling migration of carriers between scheduler specific allocator instances of the same allocator type. Each allocator instance keeps track of the current utilization of its -multiblock carriers. When the utilization falls below the "abandon +multiblock carriers. When the total utilization falls below the "abandon carrier utilization limit" it starts to inspect the utilization of the current carrier when deallocations are made. If also the utilization of the carrier falls below the "abandon carrier utilization limit" it @@ -163,7 +163,7 @@ If a carrier in the pool becomes empty, it will be withdrawn from the pool. All carriers that become empty are also always passed to its **owning** allocator instance for deallocation using the delayed dealloc functionality. Since carriers this way always will be -deallocated by the owner, that allocated the carrier, the +deallocated by the owner that allocated the carrier, the underlying functionality of allocating and deallocating carriers can remain simple and doesn't have to bother about multiple threads. In a NUMA system we will also not mix carriers originating from multiple @@ -185,10 +185,10 @@ In short: To harbor real time characteristics, searching the pool is limited. We only inspect a limited number of carriers. If none of those carriers had a free block large enough to satisfy the allocation -request, the search will fail. A carrier in the pool can also be busy, +request, the search will fail. A carrier in the pool can also be busy if another thread is currently doing block deallocation work on the carrier. A busy carrier will also be skipped by the search as it can -not satisfy the request. The pool is lock free and we do not want to +not satisfy the request. The pool is lock-free and we do not want to block, waiting for the other thread to finish. #### Before OTP 17.4 #### @@ -212,9 +212,9 @@ may later be inserted into the pool due to bad utilization. If the frequency of insertions into the pool is higher than successful fetching from the pool, memory will eventually get exhausted. -This "bad" state, consist of a cluster of small and/or highly +This "bad" state consists of a cluster of small and/or highly fragmented carriers located at the sentinel in the pool. The largest free -block in such a "bad" carrier is rather small, making it not able to satisfy +block in such a "bad" carrier is rather small, making it unable to satisfy most allocation requests. As the search always started at the sentinel, any such "bad" carriers that had been left in the pool would eventually cluster together at the sentinel. All searches first @@ -229,7 +229,7 @@ allocator itself and later had been abandoned to the pool. If none of our own abandoned carrier would do, then the search continues into the pool, as before, to look for carriers created by other allocators. However, if we have at least one abandoned carrier of our -own, that could not satisfy the request, we can use that as entry point +own that could not satisfy the request, we can use that as entry point into the pool. The result is that we prefer carriers created by the thread itself, @@ -237,15 +237,15 @@ which is good for NUMA performance. And we get more entry points when searching the pool, which will ease contention and clustering. To do the first search among own carriers, every allocator instance -has two new lists; `pooled_list` and `traitor_list`. These lists are only -accessed by the allocator itself and they only contain the allocators +has two new lists: `pooled_list` and `traitor_list`. These lists are only +accessed by the allocator itself and they only contain the allocator's own carriers. When an owned carrier is abandoned and put in the pool, it is also linked into `pooled_list`. When we search our `pooled_list` and find a carrier that is no longer in the pool, we move that carrier from `pooled_list` to `traitor_list` as it is now employed by another allocator. If searching `pooled_list` fails, we also do a limited search of `traitor_list`. When finding an abandoned -carrier in `traitor_list` it is either employed, or moved back to +carrier in `traitor_list` it is either employed or moved back to `pooled_list` if it could not satisfy the allocation request. When searching `pooled_list` and `traitor_list` we always start at the @@ -256,14 +256,14 @@ allocator instance, they need no thread synchronization at all. Furthermore, the search for own carriers that are scheduled for deallocation is now done as the last search option. The idea is -that it is better to reuse a poorly utilized carrier, than to +that it is better to reuse a poorly utilized carrier than to resurrect an empty carrier that was just about to be released back to the OS. ### Result ### The use of this strategy of abandoning carriers with poor utilization -and reusing these in allocator instances with an increased carrier +and reusing them in allocator instances with an increased carrier demand is extremely effective and completely eliminates the problems that otherwise sometimes occurred when CPU load dropped while memory load did not. diff --git a/lib/kernel/src/inet.erl b/lib/kernel/src/inet.erl index d17da2d329..43bab8bcf0 100644 --- a/lib/kernel/src/inet.erl +++ b/lib/kernel/src/inet.erl @@ -695,9 +695,9 @@ connect_options(Opts, Family) -> Error -> Error end. -con_opt([{raw,A,B,C}|Opts],R,As) -> +con_opt([{raw,A,B,C}|Opts],#connect_opts{} = R,As) -> con_opt([{raw,{A,B,C}}|Opts],R,As); -con_opt([Opt | Opts], R, As) -> +con_opt([Opt | Opts], #connect_opts{} = R, As) -> case Opt of {ip,IP} -> con_opt(Opts, R#connect_opts { ifaddr = IP }, As); {ifaddr,IP} -> con_opt(Opts, R#connect_opts { ifaddr = IP }, As); @@ -722,10 +722,10 @@ con_opt([Opt | Opts], R, As) -> {Name,Val} when is_atom(Name) -> con_add(Name, Val, R, Opts, As); _ -> {error, badarg} end; -con_opt([], R, _) -> +con_opt([], #connect_opts{} = R, _) -> {ok, R}. -con_add(Name, Val, R, Opts, AllOpts) -> +con_add(Name, Val, #connect_opts{} = R, Opts, AllOpts) -> case add_opt(Name, Val, R#connect_opts.opts, AllOpts) of {ok, SOpts} -> con_opt(Opts, R#connect_opts { opts = SOpts }, AllOpts); @@ -763,9 +763,9 @@ listen_options(Opts, Family) -> Error -> Error end. -list_opt([{raw,A,B,C}|Opts], R, As) -> +list_opt([{raw,A,B,C}|Opts], #listen_opts{} = R, As) -> list_opt([{raw,{A,B,C}}|Opts], R, As); -list_opt([Opt | Opts], R, As) -> +list_opt([Opt | Opts], #listen_opts{} = R, As) -> case Opt of {ip,IP} -> list_opt(Opts, R#listen_opts { ifaddr = IP }, As); {ifaddr,IP} -> list_opt(Opts, R#listen_opts { ifaddr = IP }, As); @@ -791,10 +791,10 @@ list_opt([Opt | Opts], R, As) -> {Name,Val} when is_atom(Name) -> list_add(Name, Val, R, Opts, As); _ -> {error, badarg} end; -list_opt([], R, _SockOpts) -> +list_opt([], #listen_opts{} = R, _SockOpts) -> {ok, R}. -list_add(Name, Val, R, Opts, As) -> +list_add(Name, Val, #listen_opts{} = R, Opts, As) -> case add_opt(Name, Val, R#listen_opts.opts, As) of {ok, SOpts} -> list_opt(Opts, R#listen_opts { opts = SOpts }, As); @@ -821,9 +821,9 @@ udp_options(Opts, Family) -> Error -> Error end. -udp_opt([{raw,A,B,C}|Opts], R, As) -> +udp_opt([{raw,A,B,C}|Opts], #udp_opts{} = R, As) -> udp_opt([{raw,{A,B,C}}|Opts], R, As); -udp_opt([Opt | Opts], R, As) -> +udp_opt([Opt | Opts], #udp_opts{} = R, As) -> case Opt of {ip,IP} -> udp_opt(Opts, R#udp_opts { ifaddr = IP }, As); {ifaddr,IP} -> udp_opt(Opts, R#udp_opts { ifaddr = IP }, As); @@ -838,7 +838,7 @@ udp_opt([Opt | Opts], R, As) -> BinNS = filename2binary(NS), case prim_inet:is_sockopt_val(netns, BinNS) of true -> - list_opt(Opts, R#udp_opts { fd = [{netns,BinNS}] }, As); + udp_opt(Opts, R#udp_opts { fd = [{netns,BinNS}] }, As); false -> {error, badarg} end; @@ -848,10 +848,10 @@ udp_opt([Opt | Opts], R, As) -> {Name,Val} when is_atom(Name) -> udp_add(Name, Val, R, Opts, As); _ -> {error, badarg} end; -udp_opt([], R, _SockOpts) -> +udp_opt([], #udp_opts{} = R, _SockOpts) -> {ok, R}. -udp_add(Name, Val, R, Opts, As) -> +udp_add(Name, Val, #udp_opts{} = R, Opts, As) -> case add_opt(Name, Val, R#udp_opts.opts, As) of {ok, SOpts} -> udp_opt(Opts, R#udp_opts { opts = SOpts }, As); @@ -895,7 +895,7 @@ sctp_options(Opts, Mod) -> Error -> Error end. -sctp_opt([Opt|Opts], Mod, R, As) -> +sctp_opt([Opt|Opts], Mod, #sctp_opts{} = R, As) -> case Opt of {ip,IP} -> sctp_opt_ifaddr(Opts, Mod, R, As, IP); @@ -938,7 +938,7 @@ sctp_opt([], _Mod, #sctp_opts{ifaddr=IfAddr}=R, _SockOpts) -> {ok, R} end. -sctp_opt(Opts, Mod, R, As, Name, Val) -> +sctp_opt(Opts, Mod, #sctp_opts{} = R, As, Name, Val) -> case add_opt(Name, Val, R#sctp_opts.opts, As) of {ok,SocketOpts} -> sctp_opt(Opts, Mod, R#sctp_opts{opts=SocketOpts}, As); diff --git a/lib/kernel/test/inet_SUITE.erl b/lib/kernel/test/inet_SUITE.erl index ed43749cc0..d45dfc2173 100644 --- a/lib/kernel/test/inet_SUITE.erl +++ b/lib/kernel/test/inet_SUITE.erl @@ -1,7 +1,7 @@ %% %% %CopyrightBegin% %% -%% Copyright Ericsson AB 1997-2013. All Rights Reserved. +%% Copyright Ericsson AB 1997-2014. All Rights Reserved. %% %% The contents of this file are subject to the Erlang Public License, %% Version 1.1, (the "License"); you may not use this file except in @@ -38,7 +38,7 @@ gethostnative_debug_level/0, gethostnative_debug_level/1, getif/1, getif_ifr_name_overflow/1,getservbyname_overflow/1, getifaddrs/1, - parse_strict_address/1, simple_netns/1]). + parse_strict_address/1, simple_netns/1, simple_netns_open/1]). -export([get_hosts/1, get_ipv6_hosts/1, parse_hosts/1, parse_address/1, kill_gethost/0, parallell_gethost/0, test_netns/0]). @@ -53,7 +53,7 @@ all() -> t_gethostnative, gethostnative_parallell, cname_loop, gethostnative_debug_level, gethostnative_soft_restart, getif, getif_ifr_name_overflow, getservbyname_overflow, - getifaddrs, parse_strict_address, simple_netns]. + getifaddrs, parse_strict_address, simple_netns, simple_netns_open]. groups() -> [{parse, [], [parse_hosts, parse_address]}]. @@ -1128,6 +1128,32 @@ jog_netns_opt(S) -> ok. +simple_netns_open(Config) when is_list(Config) -> + case gen_udp:open(0, [binary,{netns,"/"},inet]) of + {ok,U} -> + ok = gen_udp:close(U); + {error,E1} when E1 =:= einval; E1 =:= eperm -> + ok + end, + case gen_tcp:listen(0, [binary,{netns,"/"},inet]) of + {ok,T} -> + ok = gen_tcp:close(T); + {error,E2} when E2 =:= einval; E2 =:= eperm -> + ok + end, + try gen_sctp:open(0, [binary,{netns,"/"},inet]) of + {ok,S} -> + ok = gen_sctp:close(S); + {error,E3} + when E3 =:= einval; E3 =:= eperm; E3 =:= eprotonosupport -> + ok + catch + error:badarg -> + %% Some older platforms does not allow netns for sctp + ok + end. + + %% Manual test to be run outside test_server in an emulator %% started by root, in a machine with setns() support... test_netns() -> diff --git a/lib/ssl/src/ssl.appup.src b/lib/ssl/src/ssl.appup.src index b713f86c1e..7986722094 100644 --- a/lib/ssl/src/ssl.appup.src +++ b/lib/ssl/src/ssl.appup.src @@ -1,13 +1,13 @@ %% -*- erlang -*- {"%VSN%", [ - {<<"5\\.3\\.[1-4]($|\\..*)">>, [{restart_application, ssl}]}, + {<<"5\\.3\\.[1-7]($|\\..*)">>, [{restart_application, ssl}]}, {<<"5\\.[0-2]($|\\..*)">>, [{restart_application, ssl}]}, {<<"4\\..*">>, [{restart_application, ssl}]}, {<<"3\\..*">>, [{restart_application, ssl}]} ], [ - {<<"5\\.3\\.[1-4]($|\\..*)">>, [{restart_application, ssl}]}, + {<<"5\\.3\\.[1-7]($|\\..*)">>, [{restart_application, ssl}]}, {<<"5\\.[0-2]($|\\..*)">>, [{restart_application, ssl}]}, {<<"4\\..*">>, [{restart_application, ssl}]}, {<<"3\\..*">>, [{restart_application, ssl}]} |