d/freebsd-nq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Remove the support for lazy cr3 switching from i386.

Browse Source 
amd64 has already this micro-optimization removed.

Submitted by:	kib
This commit is contained in:
Attilio Rao 2011-04-30 23:02:17 +00:00
parent 5d991209fd
commit 9734077245
7 changed files with 2 additions and 219 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
sys/i386/i386/apic_vector.s
View File

@ -357,20 +357,4 @@ IDTVEC(rendezvous)
POP_FRAME
iret
/*
* Clean up when we lose out on the lazy context switch optimization.
* ie: when we are about to release a PTD but a cpu is still borrowing it.
*/
SUPERALIGN_TEXT
IDTVEC(lazypmap)
PUSH_FRAME
SET_KERNEL_SREGS
cld
call pmap_lazyfix_action
movl lapic, %eax
movl $0, LA_EOI(%eax) /* End Of Interrupt to APIC */
POP_FRAME
iret
#endif /* SMP */

3
sys/i386/i386/db_trace.c
View File

@ -312,8 +312,7 @@ db_nextframe(struct i386_frame **fp, db_addr_t *ip, struct thread *td)
frame_type = TRAP_TIMERINT;
else if (strcmp(name, "Xcpustop") == 0 ||
strcmp(name, "Xrendezvous") == 0 ||
strcmp(name, "Xipi_intr_bitmap_handler") == 0 ||
strcmp(name, "Xlazypmap") == 0)
strcmp(name, "Xipi_intr_bitmap_handler") == 0)
frame_type = TRAP_INTERRUPT;
}

7
sys/i386/i386/mp_machdep.c
View File

@ -165,7 +165,6 @@ u_long *ipi_invlrng_counts[MAXCPU];
u_long *ipi_invlpg_counts[MAXCPU];
u_long *ipi_invlcache_counts[MAXCPU];
u_long *ipi_rendezvous_counts[MAXCPU];
u_long *ipi_lazypmap_counts[MAXCPU];
static u_long *ipi_hardclock_counts[MAXCPU];
#endif
@ -552,10 +551,6 @@ cpu_mp_start(void)
setidt(IPI_INVLCACHE, IDTVEC(invlcache),
SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
/* Install an inter-CPU IPI for lazy pmap release */
setidt(IPI_LAZYPMAP, IDTVEC(lazypmap),
SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
/* Install an inter-CPU IPI for all-CPU rendezvous */
setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous),
SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
@ -1682,8 +1677,6 @@ mp_ipi_intrcnt(void *dummy)
intrcnt_add(buf, &ipi_ast_counts[i]);
snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
intrcnt_add(buf, &ipi_rendezvous_counts[i]);
snprintf(buf, sizeof(buf), "cpu%d:lazypmap", i);
intrcnt_add(buf, &ipi_lazypmap_counts[i]);
snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
intrcnt_add(buf, &ipi_hardclock_counts[i]);
}

93
sys/i386/i386/pmap.c
View File

@ -1881,98 +1881,6 @@ pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags)
* Pmap allocation/deallocation routines.
***************************************************/
#ifdef SMP
/*
* Deal with a SMP shootdown of other users of the pmap that we are
* trying to dispose of. This can be a bit hairy.
*/
static cpumask_t *lazymask;
static u_int lazyptd;
static volatile u_int lazywait;
void pmap_lazyfix_action(void);
void
pmap_lazyfix_action(void)
{
cpumask_t mymask = PCPU_GET(cpumask);
#ifdef COUNT_IPIS
(*ipi_lazypmap_counts[PCPU_GET(cpuid)])++;
#endif
if (rcr3() == lazyptd)
load_cr3(PCPU_GET(curpcb)->pcb_cr3);
atomic_clear_int(lazymask, mymask);
atomic_store_rel_int(&lazywait, 1);
}
static void
pmap_lazyfix_self(cpumask_t mymask)
{
if (rcr3() == lazyptd)
load_cr3(PCPU_GET(curpcb)->pcb_cr3);
atomic_clear_int(lazymask, mymask);
}
static void
pmap_lazyfix(pmap_t pmap)
{
cpumask_t mymask, mask;
u_int spins;
while ((mask = pmap->pm_active) != 0) {
spins = 50000000;
mask = mask & -mask; /* Find least significant set bit */
mtx_lock_spin(&smp_ipi_mtx);
#ifdef PAE
lazyptd = vtophys(pmap->pm_pdpt);
#else
lazyptd = vtophys(pmap->pm_pdir);
#endif
mymask = PCPU_GET(cpumask);
if (mask == mymask) {
lazymask = &pmap->pm_active;
pmap_lazyfix_self(mymask);
} else {
atomic_store_rel_int((u_int *)&lazymask,
(u_int)&pmap->pm_active);
atomic_store_rel_int(&lazywait, 0);
ipi_selected(mask, IPI_LAZYPMAP);
while (lazywait == 0) {
ia32_pause();
if (--spins == 0)
break;
}
}
mtx_unlock_spin(&smp_ipi_mtx);
if (spins == 0)
printf("pmap_lazyfix: spun for 50000000\n");
}
}
#else /* SMP */
/*
* Cleaning up on uniprocessor is easy. For various reasons, we're
* unlikely to have to even execute this code, including the fact
* that the cleanup is deferred until the parent does a wait(2), which
* means that another userland process has run.
*/
static void
pmap_lazyfix(pmap_t pmap)
{
u_int cr3;
cr3 = vtophys(pmap->pm_pdir);
if (cr3 == rcr3()) {
load_cr3(PCPU_GET(curpcb)->pcb_cr3);
pmap->pm_active &= ~(PCPU_GET(cpumask));
}
}
#endif /* SMP */
/*
* Release any resources held by the given physical map.
* Called when a pmap initialized by pmap_pinit is being released.
@ -1990,7 +1898,6 @@ pmap_release(pmap_t pmap)
KASSERT(pmap->pm_root == NULL,
("pmap_release: pmap has reserved page table page(s)"));
pmap_lazyfix(pmap);
mtx_lock_spin(&allpmaps_lock);
LIST_REMOVE(pmap, pm_list);
mtx_unlock_spin(&allpmaps_lock);

6
sys/i386/i386/swtch.s
View File

@ -174,12 +174,6 @@ ENTRY(cpu_switch)
/* switch address space */
movl PCB_CR3(%edx),%eax
#ifdef PAE
cmpl %eax,IdlePDPT /* Kernel address space? */
#else
cmpl %eax,IdlePTD /* Kernel address space? */
#endif
je sw0
READ_CR3(%ebx) /* The same address space? */
cmpl %ebx,%eax
je sw0

4
sys/i386/include/smp.h
View File

@ -42,7 +42,6 @@ extern u_long *ipi_invlrng_counts[MAXCPU];
extern u_long *ipi_invlpg_counts[MAXCPU];
extern u_long *ipi_invlcache_counts[MAXCPU];
extern u_long *ipi_rendezvous_counts[MAXCPU];
extern u_long *ipi_lazypmap_counts[MAXCPU];
#endif
/* IPI handlers */
@ -53,8 +52,7 @@ inthand_t
IDTVEC(invlcache), /* Write back and invalidate cache */
IDTVEC(ipi_intr_bitmap_handler), /* Bitmap based IPIs */
IDTVEC(cpustop), /* CPU stops & waits to be restarted */
IDTVEC(rendezvous), /* handle CPU rendezvous */
IDTVEC(lazypmap); /* handle lazy pmap release */
IDTVEC(rendezvous); /* handle CPU rendezvous */
/* functions in mp_machdep.c */
void cpu_add(u_int apic_id, char boot_cpu);

92
sys/i386/xen/pmap.c
View File

@ -1681,98 +1681,6 @@ pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags)
* Pmap allocation/deallocation routines.
***************************************************/
#ifdef SMP
/*
* Deal with a SMP shootdown of other users of the pmap that we are
* trying to dispose of. This can be a bit hairy.
*/
static cpumask_t *lazymask;
static u_int lazyptd;
static volatile u_int lazywait;
void pmap_lazyfix_action(void);
void
pmap_lazyfix_action(void)
{
cpumask_t mymask = PCPU_GET(cpumask);
#ifdef COUNT_IPIS
(*ipi_lazypmap_counts[PCPU_GET(cpuid)])++;
#endif
if (rcr3() == lazyptd)
load_cr3(PCPU_GET(curpcb)->pcb_cr3);
atomic_clear_int(lazymask, mymask);
atomic_store_rel_int(&lazywait, 1);
}
static void
pmap_lazyfix_self(cpumask_t mymask)
{
if (rcr3() == lazyptd)
load_cr3(PCPU_GET(curpcb)->pcb_cr3);
atomic_clear_int(lazymask, mymask);
}
static void
pmap_lazyfix(pmap_t pmap)
{
cpumask_t mymask, mask;
u_int spins;
while ((mask = pmap->pm_active) != 0) {
spins = 50000000;
mask = mask & -mask; /* Find least significant set bit */
mtx_lock_spin(&smp_ipi_mtx);
#ifdef PAE
lazyptd = vtophys(pmap->pm_pdpt);
#else
lazyptd = vtophys(pmap->pm_pdir);
#endif
mymask = PCPU_GET(cpumask);
if (mask == mymask) {
lazymask = &pmap->pm_active;
pmap_lazyfix_self(mymask);
} else {
atomic_store_rel_int((u_int *)&lazymask,
(u_int)&pmap->pm_active);
atomic_store_rel_int(&lazywait, 0);
ipi_selected(mask, IPI_LAZYPMAP);
while (lazywait == 0) {
ia32_pause();
if (--spins == 0)
break;
}
}
mtx_unlock_spin(&smp_ipi_mtx);
if (spins == 0)
printf("pmap_lazyfix: spun for 50000000\n");
}
}
#else /* SMP */
/*
* Cleaning up on uniprocessor is easy. For various reasons, we're
* unlikely to have to even execute this code, including the fact
* that the cleanup is deferred until the parent does a wait(2), which
* means that another userland process has run.
*/
static void
pmap_lazyfix(pmap_t pmap)
{
u_int cr3;
cr3 = vtophys(pmap->pm_pdir);
if (cr3 == rcr3()) {
load_cr3(PCPU_GET(curpcb)->pcb_cr3);
pmap->pm_active &= ~(PCPU_GET(cpumask));
}
}
#endif /* SMP */
/*
* Release any resources held by the given physical map.
* Called when a pmap initialized by pmap_pinit is being released.