linux-morello December 2024

linux-morello@op-lists.linaro.org

7 participants
3 discussions

[PATCH v4 0/5] netfilter: Create new type for kernel pointers.

by Joshua Lant

Hi Kevin, Sorry I know I said that I would get a spare moment to do this work over a month ago, but the University of Manchester's elves are very busy this time of year! Here are the changes you requested... Please let me know if there are issues building. I still don't much understand why that last series had a problem. Changes from v3: - [XX/05] Modified the use of __nf_kptr_t in the xtables plugin structs to use a union, with the original struct as a member. This trick allows for removal of the heavy casting in the kernel which was required in the earlier version. - [XX/05] Squashed many of the commits (those from the xtables plugin header files) into a single commit, since each individual commit now makes far fewer changes. Testing: - Tested with purecap iptables tests (nftables only), passing 65/68 tests. Those which fail are expected to fail at this point, due primarily to improperly written test cases, or missing versions of userspace tooling. Joshua Lant (5): netfilter: Create new type for kernel pointers. x_tables.h: pointers to unions in uapi struct xt plugins: pointers to unions in uapi struct ebtables: pointers to unions in uapi struct xtables: move include to headers include/linux/netfilter.h | 6 +++++ include/uapi/linux/netfilter.h | 8 ++++++ include/uapi/linux/netfilter/x_tables.h | 18 +++++++++++-- include/uapi/linux/netfilter/xt_CT.h | 10 +++++-- include/uapi/linux/netfilter/xt_IDLETIMER.h | 12 +++++++-- include/uapi/linux/netfilter/xt_RATEEST.h | 6 ++++- include/uapi/linux/netfilter/xt_TEE.h | 6 ++++- include/uapi/linux/netfilter/xt_bpf.h | 13 +++++++-- include/uapi/linux/netfilter/xt_connlimit.h | 6 ++++- include/uapi/linux/netfilter/xt_hashlimit.h | 24 ++++++++++++++--- include/uapi/linux/netfilter/xt_limit.h | 6 ++++- include/uapi/linux/netfilter/xt_nfacct.h | 12 +++++++-- include/uapi/linux/netfilter/xt_quota.h | 6 ++++- include/uapi/linux/netfilter/xt_rateest.h | 9 +++++-- include/uapi/linux/netfilter/xt_statistic.h | 7 ++++- .../uapi/linux/netfilter_bridge/ebtables.h | 27 +++++++++++++++---- net/netfilter/xt_bpf.c | 1 - net/netfilter/xt_statistic.c | 1 - 18 files changed, 149 insertions(+), 29 deletions(-) -- 2.34.1

3 months, 2 weeks

Re: [RESEND PATCH] fs/pipe: Introduce a check to skip sleeping processes during pipe read/write

by Manfred Spraul

Hi, (I had to remove many cc, my mail server rejected to send) On 12/26/24 9:11 PM, Oleg Nesterov wrote: > I mostly agree, see my reply to this patch, but... > > On 12/26, Linus Torvalds wrote: >> If the optimization is correct, there is no point to having a config option. >> >> If the optimization is incorrect, there is no point to having the code. >> >> Either way, there's no way we'd ever have a config option for this. > Agreed, > >>> + if (was_full && pipe_check_wq_has_sleeper(&pipe->wr_wait)) >>> wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM); >> End result: you need to explain why the race cannot exist. > Agreed, > >> And I think your patch is simply buggy > Agreed again, but probably my reasoning was wrong, > >> But now waiters use "wait_event_interruptible_exclusive()" explicitly >> outside the pipe mutex, so the waiters and wakers aren't actually >> serialized. > This is what I don't understand... Could you spell ? > I _think_ that > > wait_event_whatever(WQ, CONDITION); > vs > > CONDITION = 1; > if (wq_has_sleeper(WQ)) > wake_up_xxx(WQ, ...); > > is fine. This pattern is documented in wait.h: https://elixir.bootlin.com/linux/v6.12.6/source/include/linux/wait.h#L96 Thus if there an issue, then the documentation should be updated. > Both wq_has_sleeper() and wait_event_whatever()->prepare_to_wait_event() > have the necessary barriers to serialize the waiters and wakers. > > Damn I am sure I missed something ;) Actually: Does this work universally? I.e. can we add the optimization into __wake_up()? But I do not understand this comment (from 2.6.0) https://git.kernel.org/pub/scm/linux/kernel/git/tglx/history.git/commit/ker… > /* * Note: we use "set_current_state()" _after_ the wait-queue add, * > because we need a memory barrier there on SMP, so that any * > wake-function that tests for the wait-queue being active * will be > guaranteed to see waitqueue addition _or_ subsequent * tests in this > thread will see the wakeup having taken place. * * The spin_unlock() > itself is semi-permeable and only protects * one way (it only protects > stuff inside the critical region and * stops them from bleeding out - > it would still allow subsequent * loads to move into the the critical > region). */ voidprepare_to_wait > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/prepare_to_wait>(wait_queue_head_t > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/wait_queue_head_t>*q,wait_queue_t > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/wait_queue_t>*wait > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/wait>,intstate) { > unsignedlongflags; wait > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/wait>->flags&=~WQ_FLAG_EXCLUSIVE > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/WQ_FLAG_EXCLUSIVE>; > spin_lock_irqsave > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/spin_lock_irqsave>(&q->lock,flags); > if(list_empty > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/list_empty>(&wait > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/wait>->task_list > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/task_list>)) > __add_wait_queue > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/__add_wait_queue>(q,wait > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/wait>); > set_current_state > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/set_current_state>(state); > spin_unlock_irqrestore > <https://elixir.bootlin.com/linux/v2.6.0/C/ident/spin_unlock_irqrestore>(&q->lock,flags); > } set_current_state() now uses smp_store_mb(), which is a memory barrier _after_ the store. Thus I do not see what enforces that the store happens before the store for the __add_wait_queue(). -- Manfred

5 months, 1 week

[RESEND PATCH] fs/pipe: Introduce a check to skip sleeping processes during pipe read/write

by WangYuli

When a user calls the read/write system call and passes a pipe descriptor, the pipe_read/pipe_write functions are invoked: 1. pipe_read(): 1). Checks if the pipe is valid and if there is any data in the pipe buffer. 2). Waits for data: *If there is no data in the pipe and the write end is still open, the current process enters a sleep state (wait_event()) until data is written. *If the write end is closed, return 0. 3). Reads data: *Wakes up the process and copies data from the pipe's memory buffer to user space. *When the buffer is full, the writing process will go to sleep, waiting for the pipe state to change to be awakened (using the wake_up_interruptible_sync_poll() mechanism). Once data is read from the buffer, the writing process can continue writing, and the reading process can continue reading new data. 4). Returns the number of bytes read upon successful read. 2. pipe_write(): 1). Checks if the pipe is valid and if there is any available space in the pipe buffer. 2). Waits for buffer space: *If the pipe buffer is full and the reading process has not read any data, pipe_write() may put the current process to sleep until there is space in the buffer. *If the read end of the pipe is closed (no process is waiting to read), an error code -EPIPE is returned, and a SIGPIPE signal may be sent to the process. 3). Writes data: *If there is enough space in the pipe buffer, pipe_write() copies data from the user space buffer to the kernel buffer of the pipe (using copy_from_user()). *If the amount of data the user requests to write is larger than the available space in the buffer, multiple writes may be required, or the process may wait for new space to be freed. 4). Wakes up waiting reading processes: *After the data is successfully written, pipe_write() wakes up any processes that may be waiting to read data (using the wake_up_interruptible_sync_poll() mechanism). 5). Returns the number of bytes successfully written. Check if there are any waiting processes in the process wait queue by introducing wq_has_sleeper() when waking up processes for pipe read/write operations. If no processes are waiting, there's no need to execute wake_up_interruptible_sync_poll(), thus avoiding unnecessary wake-ups. Unnecessary wake-ups can lead to context switches, where a process is woken up to handle I/O events even when there is no immediate need. Only wake up processes when there are actually waiting processes to reduce context switches and system overhead by checking with wq_has_sleeper(). Additionally, by reducing unnecessary synchronization and wake-up operations, wq_has_sleeper() can decrease system resource waste and lock contention, improving overall system performance. For pipe read/write operations, this eliminates ineffective scheduling and enhances concurrency. It's important to note that enabling this option means invoking wq_has_sleeper() to check for sleeping processes in the wait queue for every read or write operation. While this is a lightweight operation, it still incurs some overhead. In low-load or single-task scenarios, this overhead may not yield significant benefits and could even introduce minor performance degradation. UnixBench Pipe benchmark results on Zhaoxin KX-U6780A processor: With the option disabled: Single-core: 841.8, Multi-core (8): 4621.6 With the option enabled: Single-core: 877.8, Multi-core (8): 4854.7 Single-core performance improved by 4.1%, multi-core performance improved by 4.8%. Co-developed-by: Shengjin Yu <yushengjin(a)uniontech.com> Signed-off-by: Shengjin Yu <yushengjin(a)uniontech.com> Co-developed-by: Dandan Zhang <zhangdandan(a)uniontech.com> Signed-off-by: Dandan Zhang <zhangdandan(a)uniontech.com> Tested-by: Dandan Zhang <zhangdandan(a)uniontech.com> Signed-off-by: WangYuli <wangyuli(a)uniontech.com> --- fs/Kconfig | 13 +++++++++++++ fs/pipe.c | 21 +++++++++++++++------ 2 files changed, 28 insertions(+), 6 deletions(-) diff --git a/fs/Kconfig b/fs/Kconfig index 64d420e3c475..0dacc46a73fe 100644 --- a/fs/Kconfig +++ b/fs/Kconfig @@ -429,4 +429,17 @@ source "fs/unicode/Kconfig" config IO_WQ bool +config PIPE_SKIP_SLEEPER + bool "Skip sleeping processes during pipe read/write" + default n + help + This option introduces a check whether the sleep queue will + be awakened during pipe read/write. + + It often leads to a performance improvement. However, in + low-load or single-task scenarios, it may introduce minor + performance overhead. + + If unsure, say N. + endmenu diff --git a/fs/pipe.c b/fs/pipe.c index 12b22c2723b7..c085333ae72c 100644 --- a/fs/pipe.c +++ b/fs/pipe.c @@ -247,6 +247,15 @@ static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe, return tail; } +static inline bool +pipe_check_wq_has_sleeper(struct wait_queue_head *wq_head) +{ + if (IS_ENABLED(CONFIG_PIPE_SKIP_SLEEPER)) + return wq_has_sleeper(wq_head); + else + return true; +} + static ssize_t pipe_read(struct kiocb *iocb, struct iov_iter *to) { @@ -377,7 +386,7 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to) * _very_ unlikely case that the pipe was full, but we got * no data. */ - if (unlikely(was_full)) + if (unlikely(was_full) && pipe_check_wq_has_sleeper(&pipe->wr_wait)) wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM); kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT); @@ -398,9 +407,9 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to) wake_next_reader = false; mutex_unlock(&pipe->mutex); - if (was_full) + if (was_full && pipe_check_wq_has_sleeper(&pipe->wr_wait)) wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM); - if (wake_next_reader) + if (wake_next_reader && pipe_check_wq_has_sleeper(&pipe->rd_wait)) wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM); kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT); if (ret > 0) @@ -573,7 +582,7 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from) * become empty while we dropped the lock. */ mutex_unlock(&pipe->mutex); - if (was_empty) + if (was_empty && pipe_check_wq_has_sleeper(&pipe->rd_wait)) wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM); kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN); wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe)); @@ -598,10 +607,10 @@ pipe_write(struct kiocb *iocb, struct iov_iter *from) * Epoll nonsensically wants a wakeup whether the pipe * was already empty or not. */ - if (was_empty || pipe->poll_usage) + if ((was_empty || pipe->poll_usage) && pipe_check_wq_has_sleeper(&pipe->rd_wait)) wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM); kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN); - if (wake_next_writer) + if (wake_next_writer && pipe_check_wq_has_sleeper(&pipe->wr_wait)) wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM); if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) { int err = file_update_time(filp); -- 2.45.2

5 months, 1 week

2025

2024

2023

2022

linux-morello December 2024