Thanks! ! As I understand , a TLB flush is an expenive operation.
Have anybody any idea if there is a TLB flush (on x86) when there is a context switch between two kernel threads ? My guess is that there is no need for such a TLB flush (on x86) since the kernel thread does not need to access virtual addresses of the user space. (But I am not sure about it, because a kernel thread DOES have access to 1GB virtual address space of the kernel ). (To be more precise,and If I remember well from Love's kernel book, the task struct has 2 mm_struct members : mm and mm_acive; and in case of kernel_thread mm is NULL ,but active_mm gets the value of previous thread , but in fact it needs ONLY the kernel address space part of it because anyhow a kernel thread is NOT allowed to use the user space address space.) And to verify one more thing, if I may : when there is a context switch between 2 ***user space*** threads,is there ALWAYS a TLB flush? Rafi G. On 11/22/05, Muli Ben-Yehuda <[EMAIL PROTECTED]> wrote: > On Tue, Nov 22, 2005 at 12:02:12PM +0200, Rafi Gordon wrote: > > Is the cost , in terms of CPU , when there occurs a context switch > > between two kernel threads is the same as when > > there is a context switch between two user space threads ? > > userspace threads are mapped to kernel threads in Linux (more > accurately, userspace threads are represented by tasks in a 1-1 > mapping). > > > Is this cost the same when there is a context switch between a kernel > > thread and a user space thread ? > > No. A switch between userspace threads in Linux requires switching to > kernel space first. There's no support by default for userspace > threading. > > > If you look at the implementation of kernel_thread() method (in i386 it's in > > arch/i386/kernel/process.c), you will see that it ultimately > > calls do_fork() ; in fact also when you create a user space process/thread > > by > > fork() or pthread_create() you also ultimately call do_fork(). > > > > There are of course some differences between kernel threads > > and user space threads. > > for example : a kernel thread does not have any need for access to memory > > of user space. > > > > But do these differences cause a conetxt switch between 2 kernel > > threads to be quicker ? > > Yes. Since there's no need to switch address spaces or priviledge > levels to ctx switch between two kernel threads, it's just th overhead > of the scheduler and ctx_switch itself. > > Cheers, > Muli > -- > Muli Ben-Yehuda > http://www.mulix.org | http://mulix.livejournal.com/ > > ================================================================To unsubscribe, send mail to [EMAIL PROTECTED] with the word "unsubscribe" in the message body, e.g., run the command echo unsubscribe | mail [EMAIL PROTECTED]
