kthreads.txt Barret Rhoden What Are They, And Why? ------------------------------- Eventually a thread of execution in the kernel will want to block. This means that the thread is unable to make forward progress and something else ought to run - the common case for this is when we wait on an IO operation. This gets trickier when a function does not know if a function it calls will block or not. Sometimes they do, sometimes they don't. The critical feature is not that we want to save the registers, but that we want to preserve the stack and be able to use it independently of whatever else we do on that core in the interim time. If we knew we would be done with and return from whatever_else() before we needed to continue the current thread of execution, we could simply call the function. Instead, we want to be able to run the old context independently of what else is running (which may be a process). We call this suspended context and the associated information a kthread, managed by a struct kthread. It's the bare minimum needed for the kernel to stop and restart a thread of execution. It holds the registers, stack pointer, PC, struct proc* if applicable, stacktop, and little else. There is no silly_state / floating point state, or anything else. Its address space is determined from which process context (possibly none) that was running. We also get a few other benefits, such as the ability to pick and choose which kthreads to run where and when. Users of kthreads should not assume that the core_id() stayed the same across blocking calls. We can also use this infrastructure in other cases where we might want to start on a new stack. One example is when we deal with low memory. We may have to do a lot of work, but only need to do a little to allow the original thread (that might have failed on a page_alloc) to keep running, while we want the memory freer to keep running too (or later) from where it left off. In essence, we want to fork, work, and yield or run on another core. The kthread is just a means of suspending a call stack and a context for a little while. Side Note: ----------- Right now, blocking a kthread is an explicit action. Some function realizes it can't make progress (like waiting on a block device), so it sleeps on something (for now a semaphore), and gets woken up when it receives its signal. This differs from processes, which can be stopped and suspended at any moment (pagefault is the classic example). In the future, we could make kthreads be preemptable (timer interrupt goes off, and we choose to suspend a kthread), but even then kthreads still have the ability to turn off interrupts for tricky situations (like suspending the kthread). The analog in the process code is disabling notifications, which dramatically complicates its functions (compare the save and pop functions for _ros_tf and _kernel_tf). Furthermore, when a process disables notifications, it still doesn't mean it is running without interruptions (it looks like that to the vcore). When the kernel disables interrupts, it really is running. What About Events? ------------------------------- Why not just be event driven for all IO? Why do we need these kernel threads? In short, IO isn't as simple as "I just want a block and when it's done, run a function." While that is what the block device driver will do, the subsystems actually needing the IO are much simpler if they are threaded. Consider the potentially numerous blocking IO calls involved in opening a file. Having a continuation for each one of those points in the call graph seems like a real pain to code. Perhaps I'm not seeing it, but if you're looking for a simple, light mechanism for keeping track of what work you need to do, just use a stack. Programming is much simpler, and it costs a page plus a small data structure. Note that this doesn't mean that all IO needs to use kthreads, just that some will really benefit from it. I plan to make the "last part" of some IO calls more event driven. Basically, it's all just a toolbox, and you should use what you need. Freeing Stacks and Structs ------------------------------- When we restart a kthread, we have to be careful about freeing the old stack and the struct kthread. We need to delay the freeing of both of these until after we longjmp to the new kthread. We can't free the kthread before popping it, and we are on the stack we need to free (until we pop to the new stack). To deal with this, we have a "spare" kthread per core, which gets assigned as the spare when we restart a previous kthread. When making/suspending a kthread, we'll use this spare. When restarting one, we'll free the old spare if it exists and put ours there. One drawback is that we potentially waste a chunk of memory (1 page + a bit per core, worst case), but it is a nice, simple solution. Also, it will cut down on contention for free pages and the kthread_kcache, though this won't help with serious contention issues (which we'll deal with eventually). What To Run Next? ------------------------------- When a kthread suspends, what do we run next? And how do we know what to run next? For now, we call smp_idle() - it is what you do when you have nothing else to do, or don't know what to do. We could consider having sleep_on() take a function pointer, but when we start hopping stacks, passing that info gets tricky. And we need to make a decision about which function to call quickly (in the code. I don't trust the compiler much). We can store the function pointer at the bottom of the future stack and extract it from there. Or we could put it in per_cpu_info. Or we can send ourselves a routine kernel message. Regardless of where we put it, we ought to call smp_idle() (or something similar) before calling it, since we need to make sure that whatever we call right after jumping stacks never returns. It's more flexible to allow a function that returns for the func *, so we'll use smp_idle() as a level of indirection. Semaphore Stuff ------------------------------- We use the semaphore (defined in kthread.h) for kthreads to sleep on and wait for a signal. It is possible that the signal wins the race and beats the call to sleep_on(). The semaphore handles this by "returning false." You'll notice that we don't actually call __down_sem(), but instead "build it in" to sleep_on(). I didn't want to deal with returning a bool (even if it was an inline), because I want to minimize the amount of stuff we do with potential stack variables (I don't trust the register variable). As soon as we unlock, the kthread could be restarted (in theory), and it could start to clobber the stack in later function calls. So it is possible that we lose the semaphore race and shouldn't sleep. We unwind the sleep prep work. An alternative was to only do the prep work if we won the race, but that would mean we have to do a lot of work in that delicate period of "I'm on the queue but it is unlocked" - work that requires touching the stack. Or we could just hold the lock for a longer period of time, which I don't care to do. What we do now is try and down the semaphore early (the early bailout), and if it fails then try to sleep (unlocked). If it then loses the race (unlikely), it can manually unwind. Note that a lot of this is probably needless worry - we have interrupts disabled for most of sleep_on(), though arguably we can be a little more careful with pcpui->spare and move the disable_irq() down to right before setjmp(). What's the Deal with Stacks/Stacktops? ------------------------------- When the kernel traps from userspace, it needs to know what to set the kernel stack pointer to. In x86, it looks in the TSS. In riscv, we have a data structure tracking that info (core_stacktops). One thing I considered was migrating the kernel from its boot stacks (x86, just core0, riscv, all the cores have one). Instead, we just make sure the tables/TSS are up to date right away (before interrupts or traps can come in for x86, and right away for riscv). These boot stacks aren't particularly special, just note they are in the program data/bss sections and were never originally added to a free list. But they can be freed later on. This might be an issue in some places, but those places ought to be fixed. There is also some implications about PGSIZE stacks (specifically in the asserts, how we alloc only one page, etc). The bootstacks are bigger than a page (for now), but in general we don't want to have giant stacks (and shouldn't need them - note linux runs with 4KB stacks). In the future (long range, when we're 64 bit), I'd like to put all kernel stacks high in the address space, with guard pages after them. This would require a certain "quiet migration" to the new locations for the bootstacks (though not a new page - just a different virtual address for the stacks (not their page-alloced KVA). A bunch of minor things would need to change for that, so don't hold your breath. So what about stacktop? It's just the top of the stack, but sometimes it is the stack we were on (when suspending the kthread), other times kthread->stacktop is just a scrap page's top. What's important when suspending is that the current stack is not used in future traps - that it doesn't get clobbered. That's why we need to find a new stack and set it as the current stacktop. We also need to 'save' the stack page of the old kthread - we don't want it to be freed, since we need it later. When starting a kthread, I don't particularly care about which stack is now the default stack. The sleep_on() assumes it was the kthread's, so unless we always have a default one that is only used very briefly and never blocked on, (which requires a stack jump), we ought to just have a kthread run with its stack as the default stacktop. When restarting a kthread, we eventually will use its stack, instead of the current one, but we can't free the current stack until after we actually longjmp() to it. This is the same problem as with the struct kthread dealloc. So we can have the kthread (which we want to free later) hold on to the page we wanted to dealloc. Likewise, when we would need a fresh kthread, we also need a page to use as the default stacktop. So if we had a cached kthread, we then use the page that kthread was pointing to. NOTE: the spare kthread struct is not holding the stack it was originally saved with. Instead, it is saving the page of the stack that was running when that kthread was reactivated. It's spare storage for both the struct and the page, but they aren't linked in any meaningful way (like it is the stack of the page). That linkage is only true when a kthread is being used (like in a semaphore queue). Current and Process Contexts ------------------------------- When a kthread is suspended, should the core stay in process context (if it was before)? Short answer: yes. For vcore local calls (process context, trapped on the calling core), we're giving the core back, so we can avoid TLB shootdowns. Though we do have to incref (which writes a cache line in the proc struct), since we are storing a reference to the proc (and will try to load its cr3 later). While this sucks, keep in mind this is for a blocking IO call (where we couldn't find the page in any cache, etc). It might be a scalability bottleneck, but it also might not matter in any real case. For async calls, it is less clear. We might want to keep processing that process's syscalls, so it'd be easier to keep its cr3 loaded. Though it's not as clear how we get from smp_idle() to a workable function and if it is useful to be in process context until we start processing those functions again. Keep in mind that normally, smp_idle() shouldn't be in any process's context. I'll probably write something later that abandons any context before halting to make sure processes die appropriately. But there are still some unresolved issues that depend on what exactly we want to do. While it is tempting to say that we stay in process context if it was local, but not if it is async, there is an added complication. The function calling sleep_on() doesn't care about whether it is on a process-allocated core or not. This is solvable by using per_cpu_info(), and will probably work its way into a future patch, regardless of whether or not we stay in process context for async calls. As a final case, what will we do for processes that were interrupted by something that wants to block, but wasn't servicing a syscall? We probably shouldn't have these (I don't have a good example of when we'd want it, and a bunch of reasons why we wouldn't), but if we do, then it might be okay anyway - the kthread is just holding that proc alive for a bit. Page faults are a bit different - they are something the process wants at least. I was thinking more about unrelated async events. Still, shouldn't be a big deal. Kmsgs and Kthreads ------------------------------- Is there a way to mix kernel messages and kthreads? What's the difference, and can one do the other? A kthread is a suspended call-stack and context (thread), stopped in the middle of its work. Kernel messages are about starting fresh - "hey core X, run this function." A kmsg can very easily be a tool used to restart a kthread (either locally or on another core). We do this in the test code, if you're curious how it could work. Note we use the semaphore to deal with races. In test_kthreads(), we're actually using the kmsg to up the semaphore. You just as easily could up the semaphore in one core (possibly in response to a kmsg, though more likely due to an interrupt), and then send the kthread to another core to restart via a kmsg. There's no reason you can't separate the __up_sem() and the running of the kthread - the semaphore just protects you from missing the signal. Perhaps you'll want to rerun the kthread on the physical core it was suspended on! (cache locality, and it might be a legit option to allow processes to say it's okay to take their vcore). Note this may require more bookkeeping in the struct kthread. There is another complication: the way we've been talking about kmsgs (starting fresh), we are talking about *routine* messages. One requirement for routine messages that do not return is that they handle process state. The current kmsgs, such as __death and __preempt are built so they can handle acting on whichever process is currently running. Likewise, __launch_kthread() needs to handle the cases that arise when it runs on a core that was about to run a process (as can often happen with proc_restartcore(), which calls process_routine_kmsg()). Basically, if it was a _S, it just yields the process, similar to what happens in Linux (call schedule() on the way out, from what I recall). If it was a _M, things are a bit more complicated, since this should only happen if the kthread is for that process (and probably a bunch of other things - like they said it was okay to interrupt their vcore to finish the syscall). Note - this might not be accurate anymore (see discussions on current_ctx). To a certain extent, routine kmsgs don't seem like a nice fit, when we really want to be calling schedule(). Though if you think of it as the enactment of a previous scheduling decision (like other kmsgs (__death())), then it makes more sense. The scheduling decision (as of now) was made in the interrupt handler when it decided to send the kernel msg. In the future, we could split this into having the handler make the kthread active, and have the scheduler called to decide where and when to run the kthread. Current_ctx, Returning Twice, and Blocking -------------------------------- One of the reasons for decoupling kthreads from a vcore or the notion of a kernel thread per user processs/task is so that when the kernel blocks (on a syscall or wherever), it can return to the process. This is the essence of the asynchronous kernel/syscall interface (though it's not limited to syscalls (pagefaults!!)). Here is what we want it to be able to handle: - When a process traps (syscall, trap, or actual interrupt), the process regains control when the kernel is done or when it blocks. - Any kernel path can block at any time. - Kernel control paths need to not "return twice", but we don't want to have to go through acrobatics in the code to prevent this. There are a couple of approaches I considered, and it involves the nature of "current_ctx", and a brutal bug. Current_ctx (formerly current_ctx) is a pointer to the trapframe of the process that was interrupted/trapped, and is what user context ought to be running on this core if we return. Current_ctx is 'made' when the kernel saves the context at the top of the interrupt stack (aka 'stacktop'). Then the kernel's call path proceeds down the same stack. This call path may get blocked in a kthread. When we block, we want to restart the current_ctx. There is a coupling between the kthread's stack and the storage of current_ctx (contents, not the pointer (which is in pcpui)). This coupling presents a problem when we are in userspace and get interrupted, and that interrupt wants to restart a kthread. In this case, current_ctx points to the interrupt stack, but then we want to switch to the kthread's stack. This is okay. When that kthread wants to block again, it needs to switch back to another stack. Up until this commit, it was jumping to the top of the old stack it was on, clobbering current_ctx (took about 8-10 hours to figure this out). While we could just make sure to save space for current_ctx, it doesn't solve the problem: namely that the current_ctx concept is not bound to a specific kernel stack (kthread or otherwise). We could have cases where more than one kthread starts up on a core and we end up freeing the page that holds current_ctx (since it is a stack we no longer need). We don't want to bother keeping stacks around just to hold the current_ctx. Part of the nature of this weird coupling is that a given kthread might or might not have the current_ctx at the top of its stack. What a pain in the ass... The right answer is to decouple current_ctx from kthread stacks. There are two ways to do this. In both ways, current_ctx retains its role of the context the kernel restarts (or saves) when it goes back to a process, and is independent of blocking kthreads. SPOILER: solution 1 is not the one I picked 1) All traps/interrupts come in on one stack per core. That stack never changes (regardless of blocking), and current_ctx is stored at the top. Kthreads sort of 'dispatch' / turn into threads from this event-like handling code. This actually sounds really cool! 2) The contents of current_ctx get stored in per-cpu-info (pcpui), thereby clearly decoupling it from any execution context. Any thread of execution can block without any special treatment (though interrupt handlers shouldn't do this). We handle the "returning twice" problem at the point of return. One nice thing about 1) is that it might make stack management easier (we wouldn't need to keep a spare page, since it's the default core stack). 2) is also tricky since we need to change some entry point code to write the TF to pcpui (or at least copy-out for now). The main problem with 1) is that you need to know and have code to handle when you "become" a kthread and are allowed to block. It also prevents us making changes such that all executing contexts are kthreads (which sort of is what is going on, even if they don't have a struct yet). While considering 1), here's something I wanted to say: "every thread of execution, including a KMSG, needs to always return (and thus not block), or never return (and be allowed to block)." To "become" a kthread, we'd need to have code that jumps stacks, and once it jumps it can never return. It would have to go back to some place such as smp_idle(). The jumping stacks isn't a problem, and whatever we jump to would just have to have smp_idle() at the end. The problem is that this is a pain in the ass to work with in reality. But wait! Don't we do that with batched syscalls right now? Yes (though we should be using kmsgs instead of the hacked together workqueue spread across smp_idle() and syscall.c), and it is a pain in the ass. It is doable with syscalls because we have that clearly defined point (submitting vs processing). But what about other handlers, such as the page fault handler? It could block, and lots of other handlers could block too. All of those would need to have a jump point (in trap.c). We aren't even handling events anymore, we are immediately jumping to other stacks, using our "event handler" to hold current_ctx and handle how we return to current_ctx. Don't forget about other code (like the boot code) that wants to block. Simply put, option 1 creates a layer that is a pain to work with, cuts down on the flexibility of the kernel to block when it wants, and doesn't handle the issue at its source. The issue about having a defined point in the code that you can't return back across (which is where 1 would jump stacks) is about "returning twice." Imagine a syscall that doesn't block. It traps into the kernel, does its work, then returns. Now imagine a syscall that blocks. Most of these calls are going to block on occasion, but not always (imagine the read was filled from the page cache). These calls really need to handle both situations. So in one instance, the call blocks. Since we're async, we return to userspace early (pop the current_ctx). Now, when that kthread unblocks, its code is going to want to finish and unroll its stack, then pop back to userspace. This is the 'returning twice' problem. Note that a *kthread* never returns twice. This is what makes the idea of magic jumping points we can't return back across (and tying that to how we block in the kernel) painful. The way I initially dealt with this was by always calling smp_idle(), and having smp_idle decide what to do. I also used it as a place to dispatch batched syscalls, which is what made smp_idle() more attractive. However, after a bit, I realized the real nature of returning twice: current_ctx. If we forget about the batching for a second, all we really need to do is not return twice. The best place to do that is at the place where we consider returning to userspace: proc_restartcore(). Instead of calling smp_idle() all the time (which was in essence a "you can now block" point), and checking for current_ctx to return, just check in restartcore to see if there is a tf to restart. If there isn't, then we smp_idle(). And don't forget to handle the cases where we want to start and scp_ctx (which we ought to point current_ctx to in proc_run()). As a side note, we ought to use kmsgs for batched syscalls - it will help with preemption latencies. At least for all but the first syscall (which can be called directly). Instead of sending a retval via current_ctx about how many started, just put that info in the syscall struct's flags (which might help the remote syscall case - no need for a response message, though there are still a few differences (no failure model other than death)). Note that page faults will still be tricky, but at least now we don't have to worry about points of no return. We just check if there is a current_ctx to restart. The tricky part is communicating that the PF was sorted when there wasn't an explicit syscall made. Aborting Syscalls (2013-11-22) ------------------------------- On occasion, userspace would like to abort a syscall, specifically ones that are listening on sockets/conversations where no one will ever answer. We have limited support for aborting syscalls. Kthreads that are in rendez_sleep() (common for anything in the 9ns chunk of the kernel, which includes any conversation listens) can be aborted. They'll return with an error string to userspace. The easier part is the rules for kernel code to be abortable: - Restore your invariants with waserror() before calling rendez_sleep(). - That's really it. So if you're holding a qlock, put your qunlock() code and any other unwinding (such as a kfree()) in a waserror() catch. As it happens, it looks like plan9 already did that (at least for the rendez in listen). And, as always, you can't hold a spinlock when blocking, regardless of aborting calls or anything. I don't want arbitrary sleeps to be abortable. For instance, if a kthread is waiting on an arbitrary semaphore/qlock, we won't allow an abort. The reasoning is that the kthread will eventually acquire the qlock - we're not waiting on external sources to wake up. That's not 100% true - a kthread could be blocked on a qlock, and the qlock holder could be abortable. In the future, we could build some sort of "abort inheritance", usable by root or something (danger of aborting another process's kthread). Alternatively, we could make qlocks abortable too, though that would require all qlocking code to be unwindable. The harder part to syscall aborting is safely waking a kthread. There are several layers to go through from uthread or syscall down to the condition variable a kthread is sleeping on. Given a uthread, find its syscall. Given a syscall, find its kthread. Given the kthread, find the CV. And during all of these, syscalls complete concurrently, kthreads get repurposed for other syscalls, CVs could be freed (though that doesn't happen). Syscalls are often on stacks, so when they complete, the memory is both gibberish and potentially in use. Ultimately, I decided on a system of "safe abort attempts", where it is harmless to be wrong with an attempt. Instead of dealing with the races associated with memory freeing and syscalls completing, the aborts will only work if it is safe to work (using a lookup via pointer, and only dereferencing if the lookup succeeds). As it stands now, all abortable kthreads/sleepers/syscalls are on a per-proc list, and we can lookup by struct syscall*. They are only on the list when they are abortable (the CV can be poked), and the invariant is that when they are on the list, they are in a state that can be safely aborted: the kthread is working on the syscall, it hasn't unwound, it is still in rendez_sleep(), the CV is safe, etc. The details of this protection are sorted out with __reg_abortable_cv() and dereg_abortable_cv() (since it's really the condition variable that we're trying to find). So from the kernel side, nothing bad can happen if you ask to abort an arbitrary struct syscall*. The actual abort takes the "write/signal, then wake" method. The aborter tracks down the kthread via the lookup, the success of which guarantees the sleeper is in rendez_sleep() (or similar sleep paths), marks "SC_ABORT", (barriers), and attempts to wake the kthread (cv_broadcast, since we need to be sure we woke the right kthread). On the user side, we set an alarm to run an event handler that will cancel our syscall. The alarm stuff is fairly standard (runs in vcore context). Userspace no longer has the concern of syscalls completing while they abort, since the kernel will only abort syscalls that are abortable. However, it may have issues (in theory) with aborting future syscalls. If the alarm goes off when the uthread is in another later syscall (which may happen to use the same struct syscall*), then we could accidentally abort the wrong call. There's an aspect of time associated with the first abort alarm handler. This is relatively easy to handle: just turn off the alarm before reusing that syscall struct for a syscall. This relies on a property of the alarms: that when deregistering completes, the alarm handler will not be running concurrently. Incidentally, there is *another* minor trick here: the uthread when adjusting the alarm will issue a syscall, possibly reusing its old sysc*, but that will be *after* deregistering its original alarm: the point at which we could have potentially accidentally cancelled an arbitrary syscall. Also note that the call to change the kernel alarm wouldn't actually block and become abortable, but regardless, we're safe. There are a couple downsides to the "safe abort attempts" approach. We can only abort syscalls when they are at a certain point - if they aren't currently sleeping, the call will fail. Technically, the abort could take effect later on in the life of a syscall (the aborter flags the kthread to abort concurrent with the kthread waking up naturally, and then the call aborts on the next rendez_sleep that needs to block). Related to this limitation, userspace must keep attempting to cancel a syscall until it succeeds. It may also be told an abort succeeded, even if the call actually completes (the aborter flags the kthread, the rendez wakes naturally, and the kthread never blocks again). Ultimately, we can't "fire and forget" our abort attempt. It's not a huge problem though, and is less of a problem than my older approaches that didn't have this problem. For instance, the original idea I had was for userspace to flag the syscall (flags |= SC_ABORT). It could do this at any time. Whenever the kthread was going to block in an abortable location (e.g. rendez_sleep()), it would see the flag and abort. It might already be asleep, so we would also provide a syscall that would 'kick' the kthread responsible for some sysc*, to wake it up to see the flag and abort. The first problem was writing to the sysc flags. Unless we know the memory is actually the syscall we want, this could result in randomly writing to memory (such as a uthread's stack). I ran into similar issues in the kernel: you can't touch a kthread struct unless you know it is the kthread you want. Once I started dealing with the syscall -> kthread mapping, it became clear I'd need a per-proc lookup service in the kernel, which acts as a way to lock a reference to the kthread. I could solve the 'kthread memory safety' problem by looking up by reference, similar to how pid2proc works. Analogously, by changing the interface for sys_abort_syscall() to be a "lookup" approach, I solve the struct syscall * memory problem. As a smaller note, I considered registering every kthread with the process right away (specifically, when we link the syscall to the kthread->sysc) for the sysc->kthread lookup service. But this would get expensive, since every syscall pays the lookup tax (and we'd need to worry about scaling). We want syscalls to be fast, but the infrequent aborts can be expensive. The obvious change was to only save the abortable kthreads. The tradeoff is that we can't flag syscalls for aborting unless they are in an abortable state. This requires multiple pokes by userspace. In theory, they would have to deal with that scenario anyways (in case they attempt to abort before we even register in the first place). As another side note, if userspace ever has a struct syscall allocator, for use in async (non-uthread stack-based) syscalls, we'll need to not reuse a syscall struct until after the cancel alarm has been disarmed. Right now we do this by not having the uthread issue another syscall til after the disarm, since uthread stack-based syscalls are inherently bound to the uthread. A simple solution would be to have a per-uthread syscall struct, which that uthread uses preferentially, and the sysc is only freed when the uthread is freed. Not only would this scale better than accessing the sysc allocator for every syscall, but also there is no worry of reuse til the uthread disarms and exits. It is a userspace bug for a uthread to set the alarm and not unset it before either making a syscall or exiting. The root issue of that potential bug is that someone (alarm handler) holds a pointer to a uthread, with the intent of cancelling its syscall, and we need to somehow take back that pointer (cancel the alarm) before reusing the syscall or freeing the uthread. I considered not making the alarm guarantee that when the cancel returns, the handler isn't running concurrently. We could handle the races in the alarm handler and in the cancel code, but it's an added hassle that isn't clearly needed. This does mean we have to run the alarm handlers serially, while holding the alarm lock. I'm fine with this, for now. Perhaps if users want more concurrency, their handlers can spawn or wake up a uthread. It is also worth noting that many rendez_sleep() calls actually return right away. This is common if some data is already in the queue (or whatever the condition is that we want to conditionally sleep on). Since registration is a little bit heavier than just locking the CV, I use the classic "check, signal, check again" style, where we check cond, then register, and then check cond for real. The initial check is the optimization, while the "signal, then check" is the true synchronization. I use this style all over the place (check out the event delivery with concurrent vcore yields code). Because of this optimization, we have a slightly odd interface: __reg is called with the CV lock held, and dereg_ is not. There are some lock ordering issues. Without the optimization, we could simply make the order {list lock, CV lock}, so that the aborter can use the list lock to keep a kthread/cv alive (one of the struct cv_lookup_elm in the code, to be precise) while it cv_broadcasts. However, the "check first" optimization would need to lock and unlock the CV a couple times, which seems excessive. So we switch the lock order to {CV, list lock}, and the aborter doesn't hold the list lock while signalling the CV. Instead, it keeps the cle alive with a flag that dereg_ spins on. This spinwait is why dereg can't hold the CV lock: it would create a circular dependency.