Messages in this thread | | | Date | Tue, 19 Aug 2003 08:17:19 -0400 (EDT) | From | "Richard B. Johnson" <> | Subject | Re: File access |
| |
On Tue, 19 Aug 2003, Simon Haynes wrote:
> I actually had a character interface which I use for on the fly > configuration. I have now implemented some code which uses a user process to > pass the configuration to the driver. I have however run into problems trying > to write files from the driver. I have tried implementing a user process > which performs a blocking read. The user process is blocked with > interruptible_sleep_on and is woken by the main part of the driver when it > needs to write. The problem is I then need to stall the main part of the > driver while the data gets written out. My problem is that this write needs > to happen from an interrupt handler or a timer process. I cannot seem to > block these with interruptible_sleep_on, the kernel crashes. I guess you > cannot use this in these cases ? I have also tried semaphores without much > success. I have looked for the howto but failed there also. > > Could you please tell me where I could find this FAQ. > > Many Thanks > > Simon. >
You may want to post this somewhere. I don't have a web-page. We are treated like prisoners here ;;;)) Just kidding, Thor (network spy).
Cheers, Dick Johnson
Penguin : Linux version 2.4.20 on an i686 machine (797.90 BogoMips). Note 96.31% of all statistics are fiction.
File I/O from inside the kernel, FAQ............
As cited many times, the kernel is not a process. It is designed to perform functions on behalf of a calling process.
Because it is not a process, it does not have a context. A file descriptor without a context with which to associate the file descriptor is worthless. For instance, all processes are created with file descriptors, STDIN_FILENO, STDOUT_FILENO, and STDERR_FILENO (0, 1, and 2), which are associated with some I/O device. If these file-descriptor numbers were not associated with a process context, every task would do I/O from the same terminal.
At any instant, the context of a process is represented by the object called "current". This, currently, is a pointer to a structure which contains the elements necessary to resume execution of the process when the kernel call returns. It also contains the member values which uniquely define the process and allow the kernel to perform tasks on behalf of the calling process.
To perform file I/O within the kernel requires a process context. You can either steal one or you can create one. Stealing a process context has the great potential of corrupting the task from which you stole the context. This is because I/O would be performed into its resources and you will very well overwrite buffered data that that process is using.
Creating a process context involves creating a kernel thread. This thread will function properly and can perform file I/O. However, you can never use any C runtime library functions in the kernel so you have to perform all primitive file I/O yourself. For instance, you call sys_open(), you can never use open() or fopen(), etc.
Here is some module-code that starts up, then runs down a kernel- thread. This is snipped from a module that I wrote so it is incomplete, meant only as a template.
#ifndef __KERNEL__ #define __KERNEL__ #endif
#ifndef MODULE #define MODULE #endif
#include <linux/module.h> #include <linux/slab.h> #include <linux/ioport.h> #include <linux/poll.h> #include <linux/sched.h> #include <linux/init.h> #include <asm/atomic.h> #include <asm/delay.h> #include <asm/io.h> /*-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=*/ /* * This kernel thread runs forever as long as the module is installed. */ static int gpib_thread(void *unused) { int doit; unsigned long flags; exit_files(current); daemonize(); spin_lock_irq(¤t->sigmask_lock); sigemptyset(¤t->blocked); recalc_sigpending(current); spin_unlock_irq(¤t->sigmask_lock); memcpy(current->comm, task_name, sizeof(task_name)); DEB(printk("gpib_thread\n")); for(;;) { // Kernel thread code goes here.
if(!!signal_pending(current)) #ifdef NEW_THREAD_EXIT complete_and_exit(&info->quit, 0); #else up_and_exit(&info->quit, 0); #endif } }
/*-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=*/ /* * Initialize and register the module */ int __init init_module() { #ifdef NEW_THREAD_EXIT init_completion(&info->quit); #else init_MUTEX_LOCKED(&info->quit); #endif info->pid = kernel_thread(gpib_thread, NULL, CLONE_FS | CLONE_FILES); } /*-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=*/ /* * Release the module. */ void cleanup_module() { if(!!(result = kill_proc(info->pid, SIGTERM, 1))) { printk(A_sig, info->dev); return; } #ifdef NEW_THREAD_EXIT wait_for_completion(&info->quit); #else down(&info->quit); #endif }
Now, the only reason you would ever perform file I/O within the kernel is as an "intellectual exercise". There is absolutely no reason whatsoever to perform user-mode tasks within the kernel.
The way Unix/Linux is designed will always favor performing I/O from user space. That's how it's supposed to be done with such an Operating System. If you are actually designing something that requires file I/O in the kernel, the design is __defective__.
There is only one Operating System that I know of in which there was some performance to be gained by performing I/O from within the kernel. That was VAX/VMS. Any such I/O was performed within the context of the swapper process. The VAX/VMS kernel maintains its own context. There are disadvantages, for instance every interrupt generated a context-switch. There are advantages, the interrupted task did not necessarily get the CPU back immediately after an interrupt, etc.
The Unix model shares kernel code between all tasks, but the tasks themselves get most of the CPU and are forced to give up the CPU only when waiting for I/O or, in the case of a CPU-Hog, when a context switch is forced by a timer.
If you need to read parameters from a file, when installing a kernel module, you simply do this in user-space at the time your user-space process installs the module. That's what ioctl() functions are for. Also, even if you continually need to do file I/O, you still do it from user-mode. Your kernel module just does the things that can't be done from user-mode code.
If you need a user-mode response to an interrupt, it's easy. Your user-mode code sleeps in select() or poll(). The driver/module code does whatever is necessary to handle the immediate needs of the hardware, then executes wake_up_interruptible(). This wakes up your task and it processes the data received in the ISR.
There is no advantage running user-mode tasks within the kernel because you have to do the same thing anyway. There is no way for an interrupt to attach() or callback() a task. No real-world Operating System uses interrupt "thunks", where some user gets the CPU directly from an interrupt. This cannot happen because, with no context-switch in an interrupt, there is no way to return to the interrupted task. It cannot switch to another task. If it did (by making some kernel changes), you would never be able to get back to the original interrupted task.
Over and over again I have users state; "I MUST do file I/O in the kernel...". This has always been a result of an incomplete understanding of what the kernel is and what the kernel does. Some look at driver code and say; "It's just 'C' code. I can do this..." Then they code a module and triple-fault the kernel.
The kernel is not a "high-performance" API where you can execute some code with a performance advantage. The kernel is common-code that can be executed by all tasks. That code is executed within the context of the calling process. This makes it fast. The kernel does file I/O through code that can't be modified by user tasks so it follows the rules necessary to maintain file-systems. The kernel, therefore, is a mechanism for maintaining the sanity of a system. The actual work is done by the process itself.
| |