Don't free the cursor boundary tag during vmem_destroy().

[jhb]

Summary

The cursor boundary tag is statically allocated in the vmem instead
of from the vmem_bt_zone. Avoid placing it on the freetags list
when removing segments from the vmem in vmem_destroy so that isn't
freed to the vmem_bt_zone.

Test Plan

• kldunload of if_cxgbe no longer panics under a kernel with INVARIANTS
• I wrote a simple kernel module that did vmem_create on load and vmem_destroy on unload. It was able to provoke the panic (uma_zfree complained about a misaligned pointer) and no longer panics on unload.

[jhb]

sys/kern/subr_vmem.c

848

There are some other ways to handle this such as doing TAILQ_REMOVE instead of bt_remseg and keeping the while loop instead of TAILQ_FOREACH_SAFE. I'm open to fixing it however you prefer.

[jhb]

Test kernel module:

#include <sys/param.h>
#include <sys/kernel.h>
#include <sys/malloc.h>
#include <sys/module.h>
#include <sys/vmem.h>

static vmem_t *test_vmem;

static int
vmem_load(void)
{

	test_vmem = vmem_create("T4TLS key map", 0x258aaa80, 67108864,
		    32, 0, M_FIRSTFIT | M_WAITOK);
	if (test_vmem == NULL)
		return (ENOMEM);
	return (0);
}

static int
vmem_unload(void)
{

	if (test_vmem)
		vmem_destroy(test_vmem);
	return (0);
}

static int
mod_event(struct module *mod, int what, void *arg)
{

	switch (what) {
	case MOD_LOAD:
		return (vmem_load());
	case MOD_UNLOAD:
		return (vmem_unload());
	default:
		return (EOPNOTSUPP);
	}
}

static moduledata_t vmem_mod = {
	"vmem_test",
	mod_event,
	NULL
};

DECLARE_MODULE(vmem_test, vmem_mod, SI_SUB_LAST, SI_ORDER_ANY);

[markj]

sys/kern/subr_vmem.c

848

I think I would remove the cursor before the loop, if only to avoid an extra local var, but I'm happy with the change either way.

[jhb]

sys/kern/subr_vmem.c

848

Yeah that thought occurred to me this morning as well. It would be a smaller diff by just adding a single line before the loop. I'll test it.

[jhb]

• Simplify the vmem fix.