----- On Feb 28, 2019, at 3:42 AM, Michael Kerrisk mtk.manpages@xxxxxxxxx wrote: > On 12/6/18 3:42 PM, Mathieu Desnoyers wrote: >> [ Michael, rseq(2) was merged into 4.18. Can you have a look at this >> patch which adds rseq documentation to the man-pages project ? ] > Hi Matthieu > > Sorry for the long delay. I've merged this page into a private > branch and have done quite a lot of editing. I have many > questions :-). No worries, thanks for looking into it! > > In the first instance, I think it is probably best to have > a free-form text discussion rather than firing patches > back and forward. Could you take a look at the questions below > and respond? Sure, > > Thanks, > > Michael > > > RSEQ(2) Linux Programmer's Manual RSEQ(2) > > NAME > rseq - Restartable sequences and CPU number cache > > SYNOPSIS > #include <linux/rseq.h> > > int rseq(struct rseq *rseq, uint32_t rseq_len, int flags, uint32_t sig); > > DESCRIPTION > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │Imagine you are someone who is pretty new to this │ > │idea... What is notably lacking from this page is │ > │an overview explaining: │ > │ │ > │ * What a restartable sequence actually is. │ > │ │ > │ * An outline of the steps to perform when using │ > │ restartable sequences / rseq(2). │ > │ │ > │I.e., something along the lines of Jon Corbet's │ > │https://lwn.net/Articles/697979/. Can you come up │ > │with something? (Part of it might be at the start of │ > │this page, and the rest in NOTES; it need not be all │ > │in one place.) │ > └─────────────────────────────────────────────────────┘ We recently published a blog post about rseq, which might contain just the right level of information we are looking for here: https://www.efficios.com/blog/2019/02/08/linux-restartable-sequences/ Could something along the following lines work ? "A restartable sequence is a sequence of instructions guaranteed to be executed atomically with respect to other threads and signal handlers on the current CPU. If its execution does not complete atomically, the kernel changes the execution flow by jumping to an abort handler defined by user-space for that restartable sequence. Using restartable sequences requires to register a __rseq_abi thread-local storage data structure (struct rseq) through the rseq(2) system call. Only one __rseq_abi can be registered per thread, so user-space libraries and applications must follow a user-space ABI defining how to share this resource. The ABI defining how to share this resource between applications and libraries is defined by the C library. The __rseq_abi contains a rseq_cs field which points to the currently executing critical section. For each thread, a single rseq critical section can run at any given point. Each critical section need to be implemented in assembly." > The rseq() ABI accelerates user-space operations on per-CPU data by > defining a shared data structure ABI between each user-space thread and > the kernel. > > It allows user-space to perform update operations on per-CPU data with‐ > out requiring heavy-weight atomic operations. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │In the following para: "a hardware execution con‐ │ > │text"? What is the contrast being drawn here? It │ > │would be good to state it more explicitly. │ > └─────────────────────────────────────────────────────┘ Here I'm trying to clarify what we mean by "CPU" in this document. We define a CPU as having its own number returned by sched_getcpu(), which I think is sometimes referred to as "logical cpu". This is the current hyperthread on the current core, on the current "physical CPU", in the current socket. > The term CPU used in this documentation refers to a hardware execution > context. > > Restartable sequences are atomic with respect to preemption (making it > atomic with respect to other threads running on the same CPU), as well > as signal delivery (user-space execution contexts nested over the same > thread). They either complete atomically with respect to preemption on > the current CPU and signal delivery, or they are aborted. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │In the preceding sentence, we need a definition of │ > │"current CPU". │ > └─────────────────────────────────────────────────────┘ Not sure how to word it. If a thread or signal handler execution context can possibly run and issue, for instance, "sched_getcpu()" between the beginning and the end of the critical section and get the same logical CPU number as the current thread, then we are guaranteed to abort. Of course, sched_getcpu() is just one way to get the CPU number, considering that we can also read it from the __rseq_abi cpu_id and cpu_id_start fields. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │In the following, does "It is" means "Restartable │ > │sequences are"? │ > └─────────────────────────────────────────────────────┘ > It is suited for update operations on per-CPU data. Yes. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │In the following, does "It is" means "Restartable │ > │sequences are"? │ > └─────────────────────────────────────────────────────┘ "Restartable sequences can be..." > It can be used on data structures shared between threads within a > process, and on data structures shared between threads across different > processes. > > Some examples of operations that can be accelerated or improved by this > ABI: > > · Memory allocator per-CPU free-lists > > · Querying the current CPU number > > · Incrementing per-CPU counters > > · Modifying data protected by per-CPU spinlocks > > · Inserting/removing elements in per-CPU linked-lists > > · Writing/reading per-CPU ring buffers content > > · Accurately reading performance monitoring unit counters with respect > to thread migration > > Restartable sequences must not perform system calls. Doing so may > result in termination of the process by a segmentation fault. > > The rseq argument is a pointer to the thread-local rseq structure to be > shared between kernel and user-space. The layout of this structure is > shown below. > > The rseq_len argument is the size of the struct rseq to register. > > The flags argument is 0 for registration, or RSEQ_FLAG_UNREGISTER for > unregistration. > > The sig argument is the 32-bit signature to be expected before the > abort handler code. > > The rseq structure > The struct rseq is aligned on a 32-byte boundary. This structure is > extensible. Its size is passed as parameter to the rseq() system call. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │Below, I added the structure definition (in abbrevi‐ │ > │ated form). Is there any reason not to do this? │ > └─────────────────────────────────────────────────────┘ It seems appropriate. > > struct rseq { > __u32 cpu_id_start; > __u32 cpu_id; > union { > __u64 ptr64; > #ifdef __LP64__ > __u64 ptr; > #else > .... > #endif > } rseq_cs; > __u32 flags; > } __attribute__((aligned(4 * sizeof(__u64)))); > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │In the text below, I think it would be helpful to │ > │explicitly note which of these fields are set by the │ > │kernel (on return from the reseq() call) and which │ > │are set by the caller (before calling rseq()). Is │ > │the following correct: │ > │ │ > │ cpu_id_start - initialized by caller to possible │ > │ CPU number (e.g., 0), updated by kernel │ > │ on return │ "initialized by caller to possible CPU number (e.g., 0), updated by the kernel on return, and updated by the kernel on return after thread migration to a different CPU" > │ │ > │ cpu_id - initialized to -1 by caller, │ > │ updated by kernel on return │ "initialized to -1 by caller, updated by the kernel on return, and updated by the kernel on return after thread migration to a different CPU" > │ │ > │ rseq_cs - initialized by caller, either to NULL │ > │ or a pointer to an 'rseq_cs' structure │ > │ that is initialized by the caller │ "initialized by caller to NULL, then, after returning from successful registration, updated to a pointer to an "rseq_cs" structure by user-space. Set to NULL by the kernel when it restarts a rseq critical section, when it preempts or deliver a signal outside of the range targeted by the rseq_cs. Set to NULL by user-space before reclaiming memory that contains the targeted struct rseq_cs." > │ │ > │ flags - initialized by caller, used by kernel │ > └─────────────────────────────────────────────────────┘ > > The structure fields are as follows: > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │In the following paragraph, and in later places, I │ > │changed "current thread" to "calling thread". Okay? │ > └─────────────────────────────────────────────────────┘ Yes. > > cpu_id_start > Optimistic cache of the CPU number on which the calling thread > is running. The value in this field is guaranteed to always be > a possible CPU number, even when rseq is not initialized. The > value it contains should always be confirmed by reading the > cpu_id field. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │What does the last sentence mean? │ > └─────────────────────────────────────────────────────┘ It means the caller thread can always use __rseq_abi.cpu_id_start to index an array of per-cpu data and this won't cause an out-of-bound access on load, but it does not mean it really contains the current CPU number. For instance, if rseq registration failed, it will contain "0". Therefore, it's fine to use cpu_is_start to fetch per-cpu data, but the cpu_id field should be used to compare the cpu_is_start value, so the case where rseq is not registered is caught. In that case, cpu_id_start=0, but cpu_id=-1 or -2, which differ, and therefore the critical section needs to jump to the abort handler. > > This field is an optimistic cache in the sense that it is always > guaranteed to hold a valid CPU number in the range [0..(nr_pos‐ > sible_cpus - 1)]. It can therefore be loaded by user-space and > used as an offset in per-CPU data structures without having to > check whether its value is within the valid bounds compared to > the number of possible CPUs in the system. > > For user-space applications executed on a kernel without rseq > support, the cpu_id_start field stays initialized at 0, which is > indeed a valid CPU number. It is therefore valid to use it as > an offset in per-CPU data structures, and only validate whether > it's actually the current CPU number by comparing it with the > cpu_id field within the rseq critical section. > > If the kernel does not provide rseq support, that cpu_id field > stays initialized at -1, so the comparison always fails, as > intended. It is then up to user-space to use a fall-back mecha‐ > nism, considering that rseq is not available. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │The last sentence is rather difficult to grok. Can │ > │we say some more here? │ > └─────────────────────────────────────────────────────┘ Perhaps we could use the explanation I've written above in my reply ? > > cpu_id Cache of the CPU number on which the calling thread is running. > -1 if uninitialized. > > rseq_cs > The rseq_cs field is a pointer to a struct rseq_cs (described > below). It is NULL when no rseq assembly block critical section > is active for the calling thread. Setting it to point to a > critical section descriptor (struct rseq_cs) marks the beginning > of the critical section. > > flags Flags indicating the restart behavior for the calling thread. > This is mainly used for debugging purposes. Can be either: > > RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT Inhibit instruction sequence block restart on preemption for this thread. > > RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL Inhibit instruction sequence block restart on signal delivery for this thread. Restart on signal can only be inhibited when restart on preemption and restart on migration are inhibited too, else it will terminate the offending process with a segmentation fault. > > RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE Inhibit instruction sequence block restart on migration for this thread. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │Each of the above values needs an explanation. │ > │ │ > │Is it correct that only one of the values may be │ > │specified in 'flags'? I ask because in the 'rseq_cs' │ > │structure below, the 'flags' field is a bit mask │ > │where any combination of these flags may be ORed │ > │together. │ > │ │ > └─────────────────────────────────────────────────────┘ Those are also masks and can be ORed. > > The rseq_cs structure > The struct rseq_cs is aligned on a 32-byte boundary and has a fixed > size of 32 bytes. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │Below, I added the structure definition (in abbrevi‐ │ > │ated form). Is there any reason not to do this? │ > └─────────────────────────────────────────────────────┘ It's fine. > > struct rseq_cs { > __u32 version; > __u32 flags; > __u64 start_ip; > __u64 post_commit_offset; > __u64 abort_ip; > } __attribute__((aligned(4 * sizeof(__u64)))); > > The structure fields are as follows: > > version > Version of this structure. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │What does 'version' need to be initialized to? │ > └─────────────────────────────────────────────────────┘ Currently version needs to be 0. Eventually, if we implement support for new flags to rseq(), we could add feature flags which register support for newer versions of struct rseq_cs. > > flags Flags indicating the restart behavior of this structure. Can be > a combination of: > > RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT Inhibit instruction sequence block restart on preemption for this thread. > > RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL Inhibit instruction sequence block restart on signal delivery for this thread. Restart on signal can only be inhibited when restart on preemption and restart on migration are inhibited too, else it will terminate the offending process with a segmentation fault. > > RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE Inhibit instruction sequence block restart on migration for this thread. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │Each of the above values needs an explanation. │ > └─────────────────────────────────────────────────────┘ > > start_ip > Instruction pointer address of the first instruction of the > sequence of consecutive assembly instructions. > > post_commit_offset > Offset (from start_ip address) of the address after the last > instruction of the sequence of consecutive assembly instruc‐ > tions. > > abort_ip > Instruction pointer address where to move the execution flow in > case of abort of the sequence of consecutive assembly instruc‐ > tions. > > NOTES > A single library per process should keep the rseq structure in a > thread-local storage variable. The cpu_id field should be initialized > to -1, and the cpu_id_start field should be initialized to a possible > CPU value (typically 0). The part above is not quite right. All applications/libraries wishing to register rseq must follow the ABI specified by the C library. It can be defined within more that a single application/library, but in the end only one symbol will be chosen for the process's global symbol table. > > Each thread is responsible for registering and unregistering its rseq > structure. No more than one rseq structure address can be registered > per thread at a given time. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │In the following paragraph, what is the difference │ > │between "freed" and "reclaim"? I'm supposing they │ > │mean the same thing, but it's not clear. And if they │ > │do mean the same thing, then the first two sentences │ > │appear to contain contradictory information. │ > └─────────────────────────────────────────────────────┘ The mean the same thing, and they are subtly not contradictory. The first states that memory of a _registered_ rseq object must not be freed before the thread exits. The second states that memory of a rseq object must not be freed before it is unregistered or the thread exits. Do you have an alternative wording in mind to make this clearer ? > > Memory of a registered rseq object must not be freed before the thread > exits. Reclaim of rseq object's memory must only be done after either > an explicit rseq unregistration is performed or after the thread exits. > Keep in mind that the implementation of the Thread-Local Storage (C > language __thread) lifetime does not guarantee existence of the TLS > area up until the thread exits. > > In a typical usage scenario, the thread registering the rseq structure > will be performing loads and stores from/to that structure. It is how‐ > ever also allowed to read that structure from other threads. The rseq > field updates performed by the kernel provide relaxed atomicity seman‐ > tics, which guarantee that other threads performing relaxed atomic > reads of the CPU number cache will always observe a consistent value. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │In the preceding paragraph, can we reasonably add │ > │some words to explain "relaxed atomicity semantics" │ > │and "relaxed atomic reads"? │ > └─────────────────────────────────────────────────────┘ Not sure how to word this exactly, but here it means the stores and loads need to be done atomically, but don't require nor provide any ordering guarantees with respect to other loads/stores (no memory barriers). > > RETURN VALUE > A return value of 0 indicates success. On error, -1 is returned, and > errno is set appropriately. > > ERRORS > EBUSY Restartable sequence is already registered for this thread. > > EFAULT rseq is an invalid address. > > EINVAL Either flags contains an invalid value, or rseq contains an > address which is not appropriately aligned, or rseq_len contains > a size that does not match the size received on registration. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │The last case "rseq_len contains a size that does │ > │not match the size received on registration" can │ > │occur only on RSEQ_FLAG_UNREGISTER, tight? │ > └─────────────────────────────────────────────────────┘ > > ENOSYS The rseq() system call is not implemented by this kernel. > > EPERM The sig argument on unregistration does not match the signature > received on registration. > > VERSIONS > The rseq() system call was added in Linux 4.18. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │What is the current state of library support? │ > └─────────────────────────────────────────────────────┘ After going through a few RFC rounds, it's been posted as non-rfc a few weeks ago. It is pending review from glibc maintainers. I currently aim for inclusion of the rseq TLS registration by glibc for glibc 2.30: https://sourceware.org/ml/libc-alpha/2019-02/msg00317.html https://sourceware.org/ml/libc-alpha/2019-02/msg00320.html https://sourceware.org/ml/libc-alpha/2019-02/msg00319.html https://sourceware.org/ml/libc-alpha/2019-02/msg00318.html https://sourceware.org/ml/libc-alpha/2019-02/msg00321.html Note that the C library will define a user-space ABI which states how applications/libraries wishing to register the rseq TLS need to behave so they are compatible with the C library when it gets updated to a new version providing rseq registration support. It seems like an important point to document, perhaps even here in the rseq(2) man page. > > CONFORMING TO > rseq() is Linux-specific. > > ┌─────────────────────────────────────────────────────┐ > │FIXME │ > ├─────────────────────────────────────────────────────┤ > │Is there any example code that can reasonably be │ > │included in this manual page? Or some example code │ > │that can be referred to? │ > └─────────────────────────────────────────────────────┘ > The per-cpu counter example we have here seems compact enough: https://www.efficios.com/blog/2019/02/08/linux-restartable-sequences/ Thanks, Mathieu > SEE ALSO > sched_getcpu(3), membarrier(2) > > -- > Michael Kerrisk > Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/ > Linux/UNIX System Programming Training: http://man7.org/training/ -- Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
