The kthread_create() API is supposed to allow you to create threads

with more than the normal amount of stack pages, however the stack
pointer always wound up being initialized using KSTACK_PAGES. It
should be using td->td_kstack_pages instead. This means that although
the vm subsystem would give you all the stack pages you asked for,
%esp would always be initialized as if you had just 2 pages, and
the rest would go to waste.

I wanted to use the 'give me more stack pages' feature of kthread_create()
because the Intel 2200BG NDIS driver does an alloca() of about 5000 bytes,
which wrecks the stack with the default 2 page size, and I was baffled
that no matter how much code I shoved into thread contexts with
allegedly larger stacks, the thing would still crash unless I changed
KSTACK_PAGES.

Note: this bug is present in _ALL_ arches at this point. Peter has
promised to merge this fix into all of them.

1 files changed, 4 insertions, 3 deletions

diff --git a/sys/i386/i386/vm_machdep.c b/sys/i386/i386/vm_machdep.c
index 9828918..4c3fa5d 100644
--- a/sys/i386/i386/vm_machdep.c
+++ b/sys/i386/i386/vm_machdep.c
@@ -172,7 +172,8 @@ cpu_fork(td1, p2, td2, flags)
 #endif
 
 	/* Point the pcb to the top of the stack */
-	pcb2 = (struct pcb *)(td2->td_kstack + KSTACK_PAGES * PAGE_SIZE) - 1;
+	pcb2 = (struct pcb *)(td2->td_kstack +
+	    td2->td_kstack_pages * PAGE_SIZE) - 1;
 	td2->td_pcb = pcb2;
 
 	/* Copy p1's pcb */
@@ -346,8 +347,8 @@ void
 cpu_thread_setup(struct thread *td)
 {
 
-	td->td_pcb =
-	     (struct pcb *)(td->td_kstack + KSTACK_PAGES * PAGE_SIZE) - 1;
+	td->td_pcb = (struct pcb *)(td->td_kstack +
+	    td->td_kstack_pages * PAGE_SIZE) - 1;
 	td->td_frame = (struct trapframe *)((caddr_t)td->td_pcb - 16) - 1;
 	td->td_pcb->pcb_ext = NULL; 
 }