How does epoll know socket is ready in kernel? - linux-kernel

I didn't find any hints in epool source code about how epoll knows socket is ready for read/write.
Does epoll register a callback in the kernel?
Does epool register a signal in the kernel for read/write?
Or something else?
Many thanks.

Short answer
Not only for epoll but in general for "blocking I/O" (the same mechanism is used by read() syscall, for example), kernel uses waitqueues (don't confuse them with workqueues which is totally different mechanism). If you check ep_poll() implementation, it's even documented in comments.
Some not-so-interesting details
In order to put current thread to sleep on waitqueue, one would normally use wait_event_interruptible() call. epoll_wait does not do that, however. Instead it kind off re-implements what this call would do by adding itself to the waitqueue with __add_wait_queue_exclusive(), putting itself to sleep with set_current_state(TASK_INTERRUPTIBLE) and checking what was the cause of being woken up in a loop. The end result is the same - the current thread will be put to interruptible sleep which may be terminated either by sending signal (in which case epoll_wait will return EINTR) or when woken up by ep_poll_callback through waitqueues mechanism.

Related

Interrupt a kernel module when a user process terminates/receives a signal?

I am working on a kernel module where I need to be "aware" that a given process has crashed.
Right now my approach is to set up a periodic timer interrupt in the kernel module; on every timer interrupt, I check the task_struct.state and task_struct.exitstate values for that process.
I am wondering if there's a way to set up an interrupt in the kernel module that would go off when the process terminates, or, when the process receives a given signal (e.g., SIGINT or SIGHUP).
Thanks!
EDIT: A catch here is that I can't modify the user application. Or at least, it would be a much tougher sell to the customer if I place additional requirements/constraints on s/w from another vendor...
You could have your module create a character device node and then open that node from your userspace process. It's only about a dozen lines of boilerplate to register a simple cdev in your module. Your cdev's open method will get called when the process opens the device node and the release method will be called when the device node is closed. If a process exits, either intentionally or because of a signal, all open file descriptors are closed by the kernel. So you can be certain that release will be called. This avoids any need to poll the process status and you can avoid modifying any kernel code outside of your module.
You could also setup a watchdog style system, where your process must write one byte to the device every so often. Have the write method of the cdev reset a timer. If too much time passes without a write and the timer expires, it is assumed the process has somehow failed, even if it hasn't crashed and terminated. For instance a programming bug that allowed for a mutex deadlock or placed the process into an infinite loop.
There is a point in the kernel code where signals are delivered to user processes. You could patch that, check the process name, and signal a condition variable if it matches. This would just catch signals, not intentional process exits. IMHO, this is much uglier and you'll need to deal with maintaining a kernel patch. But it's not that hard, there's a single point, I don't recall what function, sorry, where one can insert the necessary code and it will catch all signals.

implementing blocking syscalls in Linux

I would like to understand how implementing blocking I/O syscalls is different from non-blocking? Googling it didn't help much, any links or references would be greatly appreciated.
Thanks.
http://faculty.salina.k-state.edu/tim/ossg/Device/blocking.html
Blocking syscall will put the task (calling thread) to sleep (block it from running on CPU), and syscall will return only after event (or timeout). Non-blocking syscall will not block thread, it just checks in-kernel states and immediately returns.
More detailed description: http://www.makelinux.net/ldd3/chp-6-sect-2
one important issue: how does a driver respond if it cannot immediately satisfy the request? A call to read may come when no data is available, but more is expected in the future. Or a process could attempt to write, but your device is not ready to accept the data, because your output buffer is full. The calling process usually does not care about such issues; the programmer simply expects to call read or write and have the call return after the necessary work has been done. So, in such cases, your driver should (by default) block the process, putting it to sleep until the request can proceed. ....
There are several forms of wait_event kernel functions to block the caller thread, check include/linux/wait.h; thread can be waked up by different ways, for example with wake_up/wake_up_interruptible.

Why is there a call to mdelay(1) when resetting interrupt affinities?

I'm trying to change the code that brings down a cpu, and got into something I don't completely understand:
One of the things that happen after a core is removed from cpu_online_mask, is the resetting of the interrupt affinities.
This is being done in the fixup_irqs() function, found in /arch/x86/kernel/irq.c.
The function resets interrupt affinities, then calls to mdelay(1) (which simply waits for 1 millisecond), and finally turns to handle possibly "lost" interrupts.
My question is: why is the call to mdelay(1) necessary? what can happen without it?
My guess is that it takes time for the rerouting in the APIC to take effect... but I'm sure that there is a more convincing explanation for this.
Thanks!
In a nut shell, there is a race condition in fixup_irq() - the function starts by going over all the IRQs routed to the CPU that is being offlined and tells the HW to route them to somewhere else.
The thing is, the process of changing this interrupt routing is not atomic or instantaneous. The transaction that changes the routing on the PIC chip might race with a transaction that sends an interrupt - and that might take some cycles to arrive, so you might end up with:
Tell the APIC to send interrupts to some other CPU, not me.
Interrupt!
So what the code does is basically:
Tell the APIC to send interrupts to some other CPU not me.
Wait a bit.Enough so that the interrupt re-route would be guaranteed to finalize. ( How to know how much time is enough to wait? maybe its documented in the APIC spec, maybe its internal knowledge some Intel VLSI engineer revealed to their Linux people - I don't know :-)
Check if an interrupt occurred by reading a register on the APIC that latches when an interrupt was sent and if you find any, send it to the proper target as an IPI.
Now we know no interrupt will really get to us.

BUG: Scheduling while atomic .... using sysfs_notify()

I have a kernel module that uses hrtimers to notify userspace when the timer has fired. I understand I can just use userspace timers, but it is emulating a driver that will actually talk to hardware in the future. Every once in a while I get a BUG: Scheduling while atomic. After doing some research I am assuming that the hrtimer.function that I register as a callback, is being called from an interrupt routine by the kernel internals (making my callback function in an "Atomic Context"). Then when I call sysfs_notify() within the callback, I get the kernel bug, because sysfs_notify() acquires a mutex.
1) Is this a correct assumption?
If this is correct, I have seen that there is a function called sys_notify_dirent() that I can use to notify userspace from an atomic context. But according to this source:
http://linux.derkeiler.com/Mailing-Lists/Kernel/2009-10/msg07510.html
It can only be called from a "process" context, and not an interrupt context (due to the spinlock).
2) Could someone explain the difference between process, interrupt, and atomic context?
3) If this cannot be used in an interrupt context, what is an alternative to notifying userspace in this context?
Correct, sysfs_notify() cannot be called from atomic context. And yes, sysfs_notify_dirent() appears to be safe to call from atomic context. The source you cite is a bug report that notices in an old kernel version that statement wasn't actually true, along with a patch to fix it. It now appears to be safe to call.
Follow the source code in gpiolib_sysfs.c, and you'll notice that sysfs_notify_dirent() eventually calls schedule_work(), which defers the actual call to sysfs_notify(), which is exactly what the comments to your question are advising you to do. It's just wrapped inside the convenience function.

Why kernel code/thread executing in interrupt context cannot sleep?

I am reading following article by Robert Love
http://www.linuxjournal.com/article/6916
that says
"...Let's discuss the fact that work queues run in process context. This is in contrast to the other bottom-half mechanisms, which all run in interrupt context. Code running in interrupt context is unable to sleep, or block, because interrupt context does not have a backing process with which to reschedule. Therefore, because interrupt handlers are not associated with a process, there is nothing for the scheduler to put to sleep and, more importantly, nothing for the scheduler to wake up..."
I don't get it. AFAIK, scheduler in the kernel is O(1), that is implemented through the bitmap. So what stops the scehduler from putting interrupt context to sleep and taking next schedulable process and passing it the control?
So what stops the scehduler from putting interrupt context to sleep and taking next schedulable process and passing it the control?
The problem is that the interrupt context is not a process, and therefore cannot be put to sleep.
When an interrupt occurs, the processor saves the registers onto the stack and jumps to the start of the interrupt service routine. This means that when the interrupt handler is running, it is running in the context of the process that was executing when the interrupt occurred. The interrupt is executing on that process's stack, and when the interrupt handler completes, that process will resume executing.
If you tried to sleep or block inside an interrupt handler, you would wind up not only stopping the interrupt handler, but also the process it interrupted. This could be dangerous, as the interrupt handler has no way of knowing what the interrupted process was doing, or even if it is safe for that process to be suspended.
A simple scenario where things could go wrong would be a deadlock between the interrupt handler and the process it interrupts.
Process1 enters kernel mode.
Process1 acquires LockA.
Interrupt occurs.
ISR starts executing using Process1's stack.
ISR tries to acquire LockA.
ISR calls sleep to wait for LockA to be released.
At this point, you have a deadlock. Process1 can't resume execution until the ISR is done with its stack. But the ISR is blocked waiting for Process1 to release LockA.
I think it's a design idea.
Sure, you can design a system that you can sleep in interrupt, but except to make to the system hard to comprehend and complicated(many many situation you have to take into account), that's does not help anything. So from a design view, declare interrupt handler as can not sleep is very clear and easy to implement.
From Robert Love (a kernel hacker):
http://permalink.gmane.org/gmane.linux.kernel.kernelnewbies/1791
You cannot sleep in an interrupt handler because interrupts do not have
a backing process context, and thus there is nothing to reschedule back
into. In other words, interrupt handlers are not associated with a task,
so there is nothing to "put to sleep" and (more importantly) "nothing to
wake up". They must run atomically.
This is not unlike other operating systems. In most operating systems,
interrupts are not threaded. Bottom halves often are, however.
The reason the page fault handler can sleep is that it is invoked only
by code that is running in process context. Because the kernel's own
memory is not pagable, only user-space memory accesses can result in a
page fault. Thus, only a few certain places (such as calls to
copy_{to,from}_user()) can cause a page fault within the kernel. Those
places must all be made by code that can sleep (i.e., process context,
no locks, et cetera).
Because the thread switching infrastructure is unusable at that point. When servicing an interrupt, only stuff of higher priority can execute - See the Intel Software Developer's Manual on interrupt, task and processor priority. If you did allow another thread to execute (which you imply in your question that it would be easy to do), you wouldn't be able to let it do anything - if it caused a page fault, you'd have to use services in the kernel that are unusable while the interrupt is being serviced (see below for why).
Typically, your only goal in an interrupt routine is to get the device to stop interrupting and queue something at a lower interrupt level (in unix this is typically a non-interrupt level, but for Windows, it's dispatch, apc or passive level) to do the heavy lifting where you have access to more features of the kernel/os. See - Implementing a handler.
It's a property of how O/S's have to work, not something inherent in Linux. An interrupt routine can execute at any point so the state of what you interrupted is inconsistent. If you interrupted the thread scheduling code, its state is inconsistent so you can't be sure you can "sleep" and switch threads. Even if you protect the thread switching code from being interrupted, thread switching is a very high level feature of the O/S and if you protected everything it relies on, an interrupt becomes more of a suggestion than the imperative implied by its name.
So what stops the scehduler from putting interrupt context to sleep and taking next schedulable process and passing it the control?
Scheduling happens on timer interrupts. The basic rule is that only one interrupt can be open at a time, so if you go to sleep in the "got data from device X" interrupt, the timer interrupt cannot run to schedule it out.
Interrupts also happen many times and overlap. If you put the "got data" interrupt to sleep, and then get more data, what happens? It's confusing (and fragile) enough that the catch-all rule is: no sleeping in interrupts. You will do it wrong.
Disallowing an interrupt handler to block is a design choice. When some data is on the device, the interrupt handler intercepts the current process, prepares the transfer of the data and enables the interrupt; before the handler enables the current interrupt, the device has to hang. We want keep our I/O busy and our system responsive, then we had better not block the interrupt handler.
I don't think the "unstable states" are an essential reason. Processes, no matter they are in user-mode or kernel-mode, should be aware that they may be interrupted by interrupts. If some kernel-mode data structure will be accessed by both interrupt handler and the current process, and race condition exists, then the current process should disable local interrupts, and moreover for multi-processor architectures, spinlocks should be used to during the critical sections.
I also don't think if the interrupt handler were blocked, it cannot be waken up. When we say "block", basically it means that the blocked process is waiting for some event/resource, so it links itself into some wait-queue for that event/resource. Whenever the resource is released, the releasing process is responsible for waking up the waiting process(es).
However, the really annoying thing is that the blocked process can do nothing during the blocking time; it did nothing wrong for this punishment, which is unfair. And nobody could surely predict the blocking time, so the innocent process has to wait for unclear reason and for unlimited time.
Even if you could put an ISR to sleep, you wouldn't want to do it. You want your ISRs to be as fast as possible to reduce the risk of missing subsequent interrupts.
The linux kernel has two ways to allocate interrupt stack. One is on the kernel stack of the interrupted process, the other is a dedicated interrupt stack per CPU. If the interrupt context is saved on the dedicated interrupt stack per CPU, then indeed the interrupt context is completely not associated with any process. The "current" macro will produce an invalid pointer to current running process, since the "current" macro with some architecture are computed with the stack pointer. The stack pointer in the interrupt context may point to the dedicated interrupt stack, not the kernel stack of some process.
By nature, the question is whether in interrupt handler you can get a valid "current" (address to the current process task_structure), if yes, it's possible to modify the content there accordingly to make it into "sleep" state, which can be back by scheduler later if the state get changed somehow. The answer may be hardware-dependent.
But in ARM, it's impossible since 'current' is irrelevant to process under interrupt mode. See the code below:
#linux/arch/arm/include/asm/thread_info.h
94 static inline struct thread_info *current_thread_info(void)
95 {
96 register unsigned long sp asm ("sp");
97 return (struct thread_info *)(sp & ~(THREAD_SIZE - 1));
98 }
sp in USER mode and SVC mode are the "same" ("same" here not mean they're equal, instead, user mode's sp point to user space stack, while svc mode's sp r13_svc point to the kernel stack, where the user process's task_structure was updated at previous task switch, When a system call occurs, the process enter kernel space again, when the sp (sp_svc) is still not changed, these 2 sp are associated with each other, in this sense, they're 'same'), So under SVC mode, kernel code can get the valid 'current'. But under other privileged modes, say interrupt mode, sp is 'different', point to dedicated address defined in cpu_init(). The 'current' calculated under these mode will be irrelevant to the interrupted process, accessing it will result in unexpected behaviors. That's why it's always said that system call can sleep but interrupt handler can't, system call works on process context but interrupt not.
High-level interrupt handlers mask the operations of all lower-priority interrupts, including those of the system timer interrupt. Consequently, the interrupt handler must avoid involving itself in an activity that might cause it to sleep. If the handler sleeps, then the system may hang because the timer is masked and incapable of scheduling the sleeping thread.
Does this make sense?
If a higher-level interrupt routine gets to the point where the next thing it must do has to happen after a period of time, then it needs to put a request into the timer queue, asking that another interrupt routine be run (at lower priority level) some time later.
When that interrupt routine runs, it would then raise priority level back to the level of the original interrupt routine, and continue execution. This has the same effect as a sleep.
It is just a design/implementation choices in Linux OS. The advantage of this design is simple, but it may not be good for real time OS requirements.
Other OSes have other designs/implementations.
For example, in Solaris, the interrupts could have different priorities, that allows most of devices interrupts are invoked in interrupt threads. The interrupt threads allows sleep because each of interrupt threads has separate stack in the context of the thread.
The interrupt threads design is good for real time threads which should have higher priorities than interrupts.

Resources