Part IV of ad's performance improvements for aarch64

- Implement pmap_growkernel(), and update kernel pmap's stats with atomics.

- Then, pmap_kenter_pa() and pmap_kremove() no longer need to allocate
  memory nor take pm_lock, because they only modify L3 PTEs.

- Then, pm_lock and pp_lock can be adaptive mutexes at IPL_NONE which are
  cheaper than spin mutexes.

- Take the pmap's lock in pmap_extract() if not the kernel's pmap, otherwise
  pmap_extract() might see inconsistent state.
This commit is contained in:
skrll 2020-08-12 13:36:36 +00:00
parent 739b4cc834
commit 0a3d29f4ed
2 changed files with 195 additions and 143 deletions

View File

@ -1,4 +1,4 @@
/* $NetBSD: pmap.c,v 1.85 2020/08/09 09:11:41 skrll Exp $ */
/* $NetBSD: pmap.c,v 1.86 2020/08/12 13:36:36 skrll Exp $ */
/*
* Copyright (c) 2017 Ryo Shimizu <ryo@nerv.org>
@ -27,7 +27,7 @@
*/
#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.85 2020/08/09 09:11:41 skrll Exp $");
__KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.86 2020/08/12 13:36:36 skrll Exp $");
#include "opt_arm_debug.h"
#include "opt_ddb.h"
@ -195,6 +195,8 @@ static pt_entry_t _pmap_pte_adjust_cacheflags(pt_entry_t, u_int);
static void _pmap_remove(struct pmap *, vaddr_t, vaddr_t, bool,
struct pv_entry **);
static int _pmap_enter(struct pmap *, vaddr_t, paddr_t, vm_prot_t, u_int, bool);
static int _pmap_get_pdp(struct pmap *, vaddr_t, bool, int, paddr_t *,
struct vm_page **, bool *);
static struct pmap kernel_pmap __cacheline_aligned;
@ -215,27 +217,27 @@ static inline void
pmap_pv_lock(struct pmap_page *pp)
{
mutex_spin_enter(&pp->pp_pvlock);
mutex_enter(&pp->pp_pvlock);
}
static inline void
pmap_pv_unlock(struct pmap_page *pp)
{
mutex_spin_exit(&pp->pp_pvlock);
mutex_exit(&pp->pp_pvlock);
}
static inline void
pm_lock(struct pmap *pm)
{
mutex_spin_enter(&pm->pm_lock);
mutex_enter(&pm->pm_lock);
}
static inline void
pm_unlock(struct pmap *pm)
{
mutex_spin_exit(&pm->pm_lock);
mutex_exit(&pm->pm_lock);
}
static bool
@ -249,13 +251,13 @@ pm_reverse_lock(struct pmap *pm, struct pmap_page *pp)
if (pm != pmap_kernel())
pmap_reference(pm);
mutex_spin_exit(&pp->pp_pvlock);
mutex_spin_enter(&pm->pm_lock);
mutex_exit(&pp->pp_pvlock);
mutex_enter(&pm->pm_lock);
/* nothing, just wait for lock */
mutex_spin_exit(&pm->pm_lock);
mutex_exit(&pm->pm_lock);
if (pm != pmap_kernel())
pmap_destroy(pm);
mutex_spin_enter(&pp->pp_pvlock);
mutex_enter(&pp->pp_pvlock);
return false;
}
@ -471,26 +473,34 @@ pmap_bootstrap(vaddr_t vstart, vaddr_t vend)
kpm->pm_l0table_pa = l0pa;
kpm->pm_activated = true;
LIST_INIT(&kpm->pm_vmlist);
mutex_init(&kpm->pm_lock, MUTEX_DEFAULT, IPL_VM);
mutex_init(&kpm->pm_lock, MUTEX_DEFAULT, IPL_NONE);
CTASSERT(sizeof(kpm->pm_stats.wired_count) == sizeof(long));
CTASSERT(sizeof(kpm->pm_stats.resident_count) == sizeof(long));
#define PMSTAT_INC_WIRED_COUNT(pm) do { \
KASSERT(mutex_owned(&(pm)->pm_lock)); \
(pm)->pm_stats.wired_count++; \
} while (/* CONSTCOND */ 0);
#define PMSTAT_DEC_WIRED_COUNT(pm) do{ \
KASSERT(mutex_owned(&(pm)->pm_lock)); \
(pm)->pm_stats.wired_count--; \
} while (/* CONSTCOND */ 0);
#define PMSTAT_INC_RESIDENT_COUNT(pm) do { \
KASSERT(mutex_owned(&(pm)->pm_lock)); \
(pm)->pm_stats.resident_count++; \
} while (/* CONSTCOND */ 0);
#define PMSTAT_DEC_RESIDENT_COUNT(pm) do { \
KASSERT(mutex_owned(&(pm)->pm_lock)); \
(pm)->pm_stats.resident_count--; \
} while (/* CONSTCOND */ 0);
}
static inline void
_pmap_adj_wired_count(struct pmap *pm, int adj)
{
if (pm == pmap_kernel()) {
atomic_add_long(&pm->pm_stats.wired_count, adj);
} else {
KASSERT(mutex_owned(&pm->pm_lock));
pm->pm_stats.wired_count += adj;
}
}
static inline void
_pmap_adj_resident_count(struct pmap *pm, int adj)
{
if (pm == pmap_kernel()) {
atomic_add_long(&pm->pm_stats.resident_count, adj);
} else {
KASSERT(mutex_owned(&pm->pm_lock));
pm->pm_stats.resident_count += adj;
}
}
inline static int
@ -709,18 +719,34 @@ _pmap_free_pdp_all(struct pmap *pm)
vaddr_t
pmap_growkernel(vaddr_t maxkvaddr)
{
struct pmap *pm = pmap_kernel();
struct vm_page *pg;
bool l3only = true;
int error;
vaddr_t va;
paddr_t pa;
UVMHIST_FUNC(__func__);
UVMHIST_CALLED(pmaphist);
UVMHIST_LOG(pmaphist, "maxkvaddr=%llx, pmap_maxkvaddr=%llx",
maxkvaddr, pmap_maxkvaddr, 0, 0);
mutex_enter(&pm->pm_lock);
for (va = pmap_maxkvaddr & L2_FRAME; va <= maxkvaddr; va += L2_SIZE) {
error = _pmap_get_pdp(pm, va, false, 0, &pa, &pg, &l3only);
if (error != 0) {
panic("%s: cannot allocate L3 table error=%d",
__func__, error);
}
}
aarch64_tlbi_by_asid(pm->pm_asid);
kasan_shadow_map((void *)pmap_maxkvaddr,
(size_t)(maxkvaddr - pmap_maxkvaddr));
(size_t)(va - pmap_maxkvaddr));
pmap_maxkvaddr = va;
mutex_exit(&pm->pm_lock);
pmap_maxkvaddr = maxkvaddr;
return maxkvaddr;
return va;
}
bool
@ -738,7 +764,7 @@ pmap_extract_coherency(struct pmap *pm, vaddr_t va, paddr_t *pap,
paddr_t pa;
vsize_t blocksize = 0;
int space;
bool coherency;
bool coherency, valid;
extern char __kernel_text[];
extern char _end[];
@ -781,12 +807,17 @@ pmap_extract_coherency(struct pmap *pm, vaddr_t va, paddr_t *pap,
* because the page may be in an access fault state due to
* reference bit emulation.
*/
if (pm != pmap_kernel())
mutex_enter(&pm->pm_lock);
ptep = _pmap_pte_lookup_bs(pm, va, &blocksize);
if (ptep == NULL)
return false;
pte = *ptep;
if (!lxpde_valid(pte))
valid = (ptep != NULL && lxpde_valid(pte = *ptep));
if (pm != pmap_kernel())
mutex_exit(&pm->pm_lock);
if (!valid) {
return false;
}
pa = lxpde_pa(pte) + (va & (blocksize - 1));
switch (pte & LX_BLKPAG_ATTR_MASK) {
@ -834,6 +865,8 @@ _pmap_pte_lookup_bs(struct pmap *pm, vaddr_t va, vsize_t *bs)
vsize_t blocksize;
unsigned int idx;
KASSERT(pm == pmap_kernel() || mutex_owned(&pm->pm_lock));
/*
* traverse L0 -> L1 -> L2 -> L3
*/
@ -1220,9 +1253,7 @@ pmap_kremove(vaddr_t va, vsize_t size)
KDASSERT(!IN_KSEG_ADDR(va));
KDASSERT(IN_RANGE(va, VM_MIN_KERNEL_ADDRESS, VM_MAX_KERNEL_ADDRESS));
pm_lock(kpm);
_pmap_remove(kpm, va, va + size, true, NULL);
pm_unlock(kpm);
}
static void
@ -1455,7 +1486,7 @@ pmap_create(void)
pm->pm_idlepdp = 0;
pm->pm_asid = -1;
LIST_INIT(&pm->pm_vmlist);
mutex_init(&pm->pm_lock, MUTEX_DEFAULT, IPL_VM);
mutex_init(&pm->pm_lock, MUTEX_DEFAULT, IPL_NONE);
pm->pm_l0table_pa = pmap_alloc_pdp(pm, NULL, 0, true);
KASSERT(pm->pm_l0table_pa != POOL_PADDR_INVALID);
@ -1620,17 +1651,94 @@ _pmap_pdp_delref(struct pmap *pm, paddr_t pdppa, bool do_free_pdp)
return removed;
}
/*
* traverse L0 -> L1 -> L2 -> L3 table with growing pdp if needed.
*/
static int
_pmap_get_pdp(struct pmap *pm, vaddr_t va, bool kenter, int flags,
paddr_t *pap, struct vm_page **pgp, bool *l3only)
{
pd_entry_t *l0, *l1, *l2;
struct vm_page *pdppg, *pdppg0;
paddr_t pdppa, pdppa0;
unsigned int idx;
pd_entry_t pde;
KASSERT(kenter || mutex_owned(&pm->pm_lock));
l0 = pm->pm_l0table;
idx = l0pde_index(va);
pde = l0[idx];
if (!l0pde_valid(pde)) {
KASSERT(!kenter);
/* no need to increment L0 occupancy. L0 page never freed */
pdppa = pmap_alloc_pdp(pm, &pdppg, flags, false); /* L1 pdp */
if (pdppa == POOL_PADDR_INVALID) {
return ENOMEM;
}
atomic_swap_64(&l0[idx], pdppa | L0_TABLE);
_pmap_pdp_setparent(pm, pdppg, &l0[idx]);
*l3only = false;
} else {
pdppa = l0pde_pa(pde);
pdppg = NULL;
}
l1 = (void *)AARCH64_PA_TO_KVA(pdppa);
idx = l1pde_index(va);
pde = l1[idx];
if (!l1pde_valid(pde)) {
KASSERT(!kenter);
pdppa0 = pdppa;
pdppg0 = pdppg;
pdppa = pmap_alloc_pdp(pm, &pdppg, flags, false); /* L2 pdp */
if (pdppa == POOL_PADDR_INVALID) {
return ENOMEM;
}
atomic_swap_64(&l1[idx], pdppa | L1_TABLE);
_pmap_pdp_addref(pm, pdppa0, pdppg0); /* L1 occupancy++ */
_pmap_pdp_setparent(pm, pdppg, &l1[idx]);
*l3only = false;
} else {
pdppa = l1pde_pa(pde);
pdppg = NULL;
}
l2 = (void *)AARCH64_PA_TO_KVA(pdppa);
idx = l2pde_index(va);
pde = l2[idx];
if (!l2pde_valid(pde)) {
KASSERT(!kenter);
pdppa0 = pdppa;
pdppg0 = pdppg;
pdppa = pmap_alloc_pdp(pm, &pdppg, flags, false); /* L3 pdp */
if (pdppa == POOL_PADDR_INVALID) {
return ENOMEM;
}
atomic_swap_64(&l2[idx], pdppa | L2_TABLE);
_pmap_pdp_addref(pm, pdppa0, pdppg0); /* L2 occupancy++ */
_pmap_pdp_setparent(pm, pdppg, &l2[idx]);
*l3only = false;
} else {
pdppa = l2pde_pa(pde);
pdppg = NULL;
}
*pap = pdppa;
*pgp = pdppg;
return 0;
}
static int
_pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
u_int flags, bool kenter)
{
struct vm_page *pdppg, *pdppg0;
struct vm_page *pdppg;
struct pmap_page *pp, *opp, *pps[2];
struct pv_entry *spv, *opv = NULL;
pd_entry_t pde;
pt_entry_t attr, pte, opte, *ptep;
pd_entry_t *l0, *l1, *l2, *l3;
paddr_t pdppa, pdppa0;
pd_entry_t *l3;
paddr_t pdppa;
uint32_t mdattr;
unsigned int idx;
int error = 0;
@ -1673,6 +1781,8 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
if (kenter) {
pp = NULL;
spv = NULL;
need_enter_pv = false;
} else {
struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
if (pg != NULL) {
@ -1692,12 +1802,10 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
PMAP_COUNT(unmanaged_mappings);
#endif /* __HAVE_PMAP_PV_TRACK */
}
}
if (pp != NULL) {
/*
* allocate pv in advance of pm_lock() to avoid locking myself.
* pool_cache_get() may call pmap_kenter() internally.
* allocate pv in advance of pm_lock().
*/
spv = pool_cache_get(&_pmap_pv_pool, PR_NOWAIT);
need_enter_pv = true;
@ -1707,9 +1815,8 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
}
pm_lock(pm);
if (pm->pm_idlepdp >= PDPSWEEP_TRIGGER) {
if (_pmap_sweep_pdp(pm) != 0) {
if (pm->pm_idlepdp >= PDPSWEEP_TRIGGER &&
_pmap_sweep_pdp(pm) != 0) {
/* several L1-L3 page table pages have been freed */
aarch64_tlbi_by_asid(pm->pm_asid);
}
@ -1718,76 +1825,15 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
/*
* traverse L0 -> L1 -> L2 -> L3 table with growing pdp if needed.
*/
l0 = pm->pm_l0table;
idx = l0pde_index(va);
pde = l0[idx];
if (!l0pde_valid(pde)) {
/* no need to increment L0 occupancy. L0 page never freed */
pdppa = pmap_alloc_pdp(pm, &pdppg, flags, false); /* L1 pdp */
if (pdppa == POOL_PADDR_INVALID) {
error = _pmap_get_pdp(pm, va, kenter, flags, &pdppa, &pdppg, &l3only);
if (error != 0) {
if (flags & PMAP_CANFAIL) {
error = ENOMEM;
goto fail0;
}
pm_unlock(pm);
panic("%s: cannot allocate L1 table", __func__);
panic("%s: cannot allocate L3 table error=%d", __func__,
error);
}
atomic_swap_64(&l0[idx], pdppa | L0_TABLE);
_pmap_pdp_setparent(pm, pdppg, &l0[idx]);
l3only = false;
} else {
pdppa = l0pde_pa(pde);
pdppg = NULL;
}
l1 = (void *)AARCH64_PA_TO_KVA(pdppa);
idx = l1pde_index(va);
pde = l1[idx];
if (!l1pde_valid(pde)) {
pdppa0 = pdppa;
pdppg0 = pdppg;
pdppa = pmap_alloc_pdp(pm, &pdppg, flags, false); /* L2 pdp */
if (pdppa == POOL_PADDR_INVALID) {
if (flags & PMAP_CANFAIL) {
error = ENOMEM;
goto fail0;
}
pm_unlock(pm);
panic("%s: cannot allocate L2 table", __func__);
}
atomic_swap_64(&l1[idx], pdppa | L1_TABLE);
_pmap_pdp_addref(pm, pdppa0, pdppg0); /* L1 occupancy++ */
_pmap_pdp_setparent(pm, pdppg, &l1[idx]);
l3only = false;
} else {
pdppa = l1pde_pa(pde);
pdppg = NULL;
}
l2 = (void *)AARCH64_PA_TO_KVA(pdppa);
idx = l2pde_index(va);
pde = l2[idx];
if (!l2pde_valid(pde)) {
pdppa0 = pdppa;
pdppg0 = pdppg;
pdppa = pmap_alloc_pdp(pm, &pdppg, flags, false); /* L3 pdp */
if (pdppa == POOL_PADDR_INVALID) {
if (flags & PMAP_CANFAIL) {
error = ENOMEM;
goto fail0;
}
pm_unlock(pm);
panic("%s: cannot allocate L3 table", __func__);
}
atomic_swap_64(&l2[idx], pdppa | L2_TABLE);
_pmap_pdp_addref(pm, pdppa0, pdppg0); /* L2 occupancy++ */
_pmap_pdp_setparent(pm, pdppg, &l2[idx]);
l3only = false;
} else {
pdppa = l2pde_pa(pde);
pdppg = NULL;
}
l3 = (void *)AARCH64_PA_TO_KVA(pdppa);
idx = l3pte_index(va);
@ -1806,9 +1852,9 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
KASSERT(!kenter); /* pmap_kenter_pa() cannot override */
if (opte & LX_BLKPAG_OS_WIRED) {
PMSTAT_DEC_WIRED_COUNT(pm);
_pmap_adj_wired_count(pm, -1);
}
PMSTAT_DEC_RESIDENT_COUNT(pm);
_pmap_adj_resident_count(pm, -1);
#ifdef PMAPCOUNTERS
PMAP_COUNT(remappings);
if (user) {
@ -1873,6 +1919,7 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
mdattr = VM_PROT_READ | VM_PROT_WRITE;
if (need_enter_pv) {
KASSERT(!kenter);
error = _pmap_enter_pv(pp, pm, &spv, va, ptep, pa, flags);
if (error != 0) {
/*
@ -1895,6 +1942,7 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
if (pp != NULL) {
/* update referenced/modified flags */
KASSERT(!kenter);
pp->pp_pv.pv_va |= (flags & (VM_PROT_READ | VM_PROT_WRITE));
mdattr &= (uint32_t)pp->pp_pv.pv_va;
}
@ -1940,9 +1988,9 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
}
if (pte & LX_BLKPAG_OS_WIRED) {
PMSTAT_INC_WIRED_COUNT(pm);
_pmap_adj_wired_count(pm, 1);
}
PMSTAT_INC_RESIDENT_COUNT(pm);
_pmap_adj_resident_count(pm, 1);
fail1:
if (pps[1] != NULL)
@ -1950,6 +1998,7 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
if (pps[0] != NULL)
pmap_pv_unlock(pps[0]);
fail0:
if (!kenter) {
pm_unlock(pm);
/* spare pv was not used. discard */
@ -1958,6 +2007,7 @@ _pmap_enter(struct pmap *pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
if (opv != NULL)
pool_cache_put(&_pmap_pv_pool, opv);
}
return error;
}
@ -1993,6 +2043,8 @@ _pmap_remove(struct pmap *pm, vaddr_t sva, vaddr_t eva, bool kremove,
UVMHIST_LOG(pmaphist, "pm=%p, sva=%016lx, eva=%016lx, kremovemode=%d",
pm, sva, eva, kremove);
KASSERT(kremove || mutex_owned(&pm->pm_lock));
for (va = sva; (va < eva) && (pm->pm_stats.resident_count != 0);
va = (va + blocksize) & ~(blocksize - 1)) {
@ -2038,9 +2090,9 @@ _pmap_remove(struct pmap *pm, vaddr_t sva, vaddr_t eva, bool kremove,
}
if ((pte & LX_BLKPAG_OS_WIRED) != 0) {
PMSTAT_DEC_WIRED_COUNT(pm);
_pmap_adj_wired_count(pm, -1);
}
PMSTAT_DEC_RESIDENT_COUNT(pm);
_pmap_adj_resident_count(pm, -1);
}
}
@ -2093,9 +2145,9 @@ pmap_page_remove(struct pmap_page *pp, vm_prot_t prot)
trunc_page(pv->pv_va), true);
if ((opte & LX_BLKPAG_OS_WIRED) != 0) {
PMSTAT_DEC_WIRED_COUNT(pv->pv_pmap);
_pmap_adj_wired_count(pv->pv_pmap, -1);
}
PMSTAT_DEC_RESIDENT_COUNT(pv->pv_pmap);
_pmap_adj_resident_count(pv->pv_pmap, -1);
}
pvtmp = _pmap_remove_pv(pp, pm, trunc_page(pv->pv_va), opte);
if (pvtmp == NULL) {
@ -2216,7 +2268,7 @@ pmap_unwire(struct pmap *pm, vaddr_t va)
pte &= ~LX_BLKPAG_OS_WIRED;
atomic_swap_64(ptep, pte);
PMSTAT_DEC_WIRED_COUNT(pm);
_pmap_adj_wired_count(pm, -1);
}
pm_unlock(pm);
}

View File

@ -1,4 +1,4 @@
/* $NetBSD: pmap.h,v 1.41 2020/07/16 11:36:35 skrll Exp $ */
/* $NetBSD: pmap.h,v 1.42 2020/08/12 13:36:36 skrll Exp $ */
/*-
* Copyright (c) 2014 The NetBSD Foundation, Inc.
@ -111,7 +111,7 @@ struct vm_page_md {
#define PMAP_PAGE_INIT(pp) \
do { \
mutex_init(&(pp)->pp_pvlock, MUTEX_NODEBUG, IPL_VM); \
mutex_init(&(pp)->pp_pvlock, MUTEX_NODEBUG, IPL_NONE); \
(pp)->pp_pv.pv_next = NULL; \
(pp)->pp_pv.pv_pmap = NULL; \
(pp)->pp_pv.pv_va = 0; \