Index: sparc64/sparc64/pmap.c =================================================================== --- sparc64/sparc64/pmap.c (.../head/sys) (revision 227753) +++ sparc64/sparc64/pmap.c (.../user/attilio/vmcontention/sys) (revision 227756) @@ -71,6 +71,7 @@ #include #include #include +#include #include #include #include @@ -175,6 +176,11 @@ extern int tl1_immu_miss_patch_tsb_mask_2[]; /* + * MD page read/write lock. + */ +struct rwlock md_page_rwlock; + +/* * If user pmap is processed with pmap_remove and with pmap_remove and the * resident count drops to 0, there are no more pages to remove, so we * need not continue. @@ -666,6 +672,11 @@ CPU_FILL(&pm->pm_active); /* + * Initialize the MD page lock. + */ + rw_init(&md_page_rwlock, "MD page"); + + /* * Flush all non-locked TLB entries possibly left over by the * firmware. */ @@ -875,7 +886,7 @@ struct tte *tp; int color; - mtx_assert(&vm_page_queue_mtx, MA_OWNED); + MDPAGE_ASSERT_WLOCKED(); KASSERT((m->flags & PG_FICTITIOUS) == 0, ("pmap_cache_enter: fake page")); PMAP_STATS_INC(pmap_ncache_enter); @@ -950,7 +961,7 @@ struct tte *tp; int color; - mtx_assert(&vm_page_queue_mtx, MA_OWNED); + MDPAGE_ASSERT_WLOCKED(); CTR3(KTR_PMAP, "pmap_cache_remove: m=%p va=%#lx c=%d", m, va, m->md.colors[DCACHE_COLOR(va)]); KASSERT((m->flags & PG_FICTITIOUS) == 0, @@ -1025,7 +1036,7 @@ vm_page_t om; u_long data; - mtx_assert(&vm_page_queue_mtx, MA_OWNED); + MDPAGE_ASSERT_WLOCKED(); PMAP_STATS_INC(pmap_nkenter); tp = tsb_kvtotte(va); CTR4(KTR_PMAP, "pmap_kenter: va=%#lx pa=%#lx tp=%p data=%#lx", @@ -1087,7 +1098,7 @@ struct tte *tp; vm_page_t m; - mtx_assert(&vm_page_queue_mtx, MA_OWNED); + MDPAGE_ASSERT_WLOCKED(); PMAP_STATS_INC(pmap_nkremove); tp = tsb_kvtotte(va); CTR3(KTR_PMAP, "pmap_kremove: va=%#lx tp=%p data=%#lx", va, tp, @@ -1138,19 +1149,16 @@ pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) { vm_offset_t va; - int locked; PMAP_STATS_INC(pmap_nqenter); va = sva; - if (!(locked = mtx_owned(&vm_page_queue_mtx))) - vm_page_lock_queues(); + MDPAGE_WLOCK(); while (count-- > 0) { pmap_kenter(va, *m); va += PAGE_SIZE; m++; } - if (!locked) - vm_page_unlock_queues(); + MDPAGE_WUNLOCK(); tlb_range_demap(kernel_pmap, sva, va); } @@ -1162,18 +1170,15 @@ pmap_qremove(vm_offset_t sva, int count) { vm_offset_t va; - int locked; PMAP_STATS_INC(pmap_nqremove); va = sva; - if (!(locked = mtx_owned(&vm_page_queue_mtx))) - vm_page_lock_queues(); + MDPAGE_WLOCK(); while (count-- > 0) { pmap_kremove(va); va += PAGE_SIZE; } - if (!locked) - vm_page_unlock_queues(); + MDPAGE_WUNLOCK(); tlb_range_demap(kernel_pmap, sva, va); } @@ -1321,7 +1326,7 @@ vm_page_t m; u_long data; - mtx_assert(&vm_page_queue_mtx, MA_OWNED); + MDPAGE_ASSERT_WLOCKED(); data = atomic_readandclear_long(&tp->tte_data); if ((data & TD_FAKE) == 0) { m = PHYS_TO_VM_PAGE(TD_PA(data)); @@ -1358,7 +1363,7 @@ pm->pm_context[curcpu], start, end); if (PMAP_REMOVE_DONE(pm)) return; - vm_page_lock_queues(); + MDPAGE_WLOCK(); PMAP_LOCK(pm); if (end - start > PMAP_TSB_THRESH) { tsb_foreach(pm, NULL, start, end, pmap_remove_tte); @@ -1371,7 +1376,7 @@ tlb_range_demap(pm, start, end - 1); } PMAP_UNLOCK(pm); - vm_page_unlock_queues(); + MDPAGE_WUNLOCK(); } void @@ -1384,7 +1389,7 @@ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); - vm_page_lock_queues(); + MDPAGE_WLOCK(); for (tp = TAILQ_FIRST(&m->md.tte_list); tp != NULL; tp = tpn) { tpn = TAILQ_NEXT(tp, tte_link); if ((tp->tte_data & TD_PV) == 0) @@ -1406,8 +1411,8 @@ TTE_ZERO(tp); PMAP_UNLOCK(pm); } + MDPAGE_WUNLOCK(); vm_page_aflag_clear(m, PGA_WRITEABLE); - vm_page_unlock_queues(); } static int @@ -1469,11 +1474,11 @@ vm_prot_t prot, boolean_t wired) { - vm_page_lock_queues(); + MDPAGE_WLOCK(); PMAP_LOCK(pm); pmap_enter_locked(pm, va, m, prot, wired); - vm_page_unlock_queues(); PMAP_UNLOCK(pm); + MDPAGE_WUNLOCK(); } /* @@ -1492,7 +1497,7 @@ vm_page_t real; u_long data; - mtx_assert(&vm_page_queue_mtx, MA_OWNED); + MDPAGE_ASSERT_WLOCKED(); PMAP_LOCK_ASSERT(pm, MA_OWNED); KASSERT((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) != 0 || VM_OBJECT_LOCKED(m->object), @@ -1635,27 +1640,27 @@ psize = atop(end - start); m = m_start; - vm_page_lock_queues(); + MDPAGE_WLOCK(); PMAP_LOCK(pm); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { pmap_enter_locked(pm, start + ptoa(diff), m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), FALSE); m = TAILQ_NEXT(m, listq); } - vm_page_unlock_queues(); PMAP_UNLOCK(pm); + MDPAGE_WUNLOCK(); } void pmap_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m, vm_prot_t prot) { - vm_page_lock_queues(); + MDPAGE_WLOCK(); PMAP_LOCK(pm); pmap_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), FALSE); - vm_page_unlock_queues(); PMAP_UNLOCK(pm); + MDPAGE_WUNLOCK(); } void @@ -1700,6 +1705,8 @@ vm_page_t m; u_long data; + MDPAGE_ASSERT_WLOCKED(); + if ((tp->tte_data & TD_FAKE) != 0) return (1); if (tsb_tte_lookup(dst_pmap, va) == NULL) { @@ -1720,7 +1727,7 @@ if (dst_addr != src_addr) return; - vm_page_lock_queues(); + MDPAGE_WLOCK(); if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); @@ -1738,9 +1745,9 @@ pmap_copy_tte(src_pmap, dst_pmap, tp, va); tlb_range_demap(dst_pmap, src_addr, src_addr + len - 1); } - vm_page_unlock_queues(); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); + MDPAGE_WUNLOCK(); } void @@ -1937,7 +1944,7 @@ ("pmap_page_exists_quick: page %p is not managed", m)); loops = 0; rv = FALSE; - vm_page_lock_queues(); + MDPAGE_RLOCK(); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; @@ -1948,7 +1955,7 @@ if (++loops >= 16) break; } - vm_page_unlock_queues(); + MDPAGE_RUNLOCK(); return (rv); } @@ -1965,11 +1972,11 @@ count = 0; if ((m->oflags & VPO_UNMANAGED) != 0) return (count); - vm_page_lock_queues(); + MDPAGE_RLOCK(); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) if ((tp->tte_data & (TD_PV | TD_WIRED)) == (TD_PV | TD_WIRED)) count++; - vm_page_unlock_queues(); + MDPAGE_RUNLOCK(); return (count); } @@ -1996,13 +2003,13 @@ rv = FALSE; if ((m->oflags & VPO_UNMANAGED) != 0) return (rv); - vm_page_lock_queues(); + MDPAGE_RLOCK(); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) if ((tp->tte_data & TD_PV) != 0) { rv = TRUE; break; } - vm_page_unlock_queues(); + MDPAGE_RUNLOCK(); return (rv); } @@ -2028,7 +2035,7 @@ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); count = 0; - vm_page_lock_queues(); + MDPAGE_WLOCK(); if ((tp = TAILQ_FIRST(&m->md.tte_list)) != NULL) { tpf = tp; do { @@ -2042,7 +2049,7 @@ break; } while ((tp = tpn) != NULL && tp != tpf); } - vm_page_unlock_queues(); + MDPAGE_WUNLOCK(); return (count); } @@ -2050,6 +2057,7 @@ pmap_is_modified(vm_page_t m) { struct tte *tp; + pmap_t pmap; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, @@ -2065,16 +2073,19 @@ if ((m->oflags & VPO_BUSY) == 0 && (m->aflags & PGA_WRITEABLE) == 0) return (rv); - vm_page_lock_queues(); + MDPAGE_RLOCK(); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; - if ((tp->tte_data & TD_W) != 0) { + pmap = TTE_GET_PMAP(tp); + PMAP_LOCK(pmap); + if ((tp->tte_data & TD_W) != 0) rv = TRUE; + PMAP_UNLOCK(pmap); + if (rv) break; - } } - vm_page_unlock_queues(); + MDPAGE_RUNLOCK(); return (rv); } @@ -2108,7 +2119,7 @@ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); rv = FALSE; - vm_page_lock_queues(); + MDPAGE_RLOCK(); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; @@ -2117,7 +2128,7 @@ break; } } - vm_page_unlock_queues(); + MDPAGE_RUNLOCK(); return (rv); } @@ -2140,7 +2151,7 @@ */ if ((m->aflags & PGA_WRITEABLE) == 0) return; - vm_page_lock_queues(); + MDPAGE_RLOCK(); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; @@ -2148,7 +2159,7 @@ if ((data & TD_W) != 0) tlb_page_demap(TTE_GET_PMAP(tp), TTE_GET_VA(tp)); } - vm_page_unlock_queues(); + MDPAGE_RUNLOCK(); } void @@ -2159,7 +2170,7 @@ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_reference: page %p is not managed", m)); - vm_page_lock_queues(); + MDPAGE_RLOCK(); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; @@ -2167,13 +2178,14 @@ if ((data & TD_REF) != 0) tlb_page_demap(TTE_GET_PMAP(tp), TTE_GET_VA(tp)); } - vm_page_unlock_queues(); + MDPAGE_RUNLOCK(); } void pmap_remove_write(vm_page_t m) { struct tte *tp; + pmap_t pmap; u_long data; KASSERT((m->oflags & VPO_UNMANAGED) == 0, @@ -2188,18 +2200,21 @@ if ((m->oflags & VPO_BUSY) == 0 && (m->aflags & PGA_WRITEABLE) == 0) return; - vm_page_lock_queues(); + MDPAGE_RLOCK(); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; + pmap = TTE_GET_PMAP(tp); + PMAP_LOCK(pmap); data = atomic_clear_long(&tp->tte_data, TD_SW | TD_W); if ((data & TD_W) != 0) { vm_page_dirty(m); - tlb_page_demap(TTE_GET_PMAP(tp), TTE_GET_VA(tp)); + tlb_page_demap(pmap, TTE_GET_VA(tp)); } + PMAP_UNLOCK(pmap); } + MDPAGE_RUNLOCK(); vm_page_aflag_clear(m, PGA_WRITEABLE); - vm_page_unlock_queues(); } int Index: sparc64/sparc64/tsb.c =================================================================== --- sparc64/sparc64/tsb.c (.../head/sys) (revision 227753) +++ sparc64/sparc64/tsb.c (.../user/attilio/vmcontention/sys) (revision 227756) @@ -40,6 +40,7 @@ #include #include #include +#include #include #include #include @@ -131,7 +132,7 @@ PMAP_STATS_INC(tsb_nenter_u_oc); } - mtx_assert(&vm_page_queue_mtx, MA_OWNED); + MDPAGE_ASSERT_WLOCKED(); PMAP_LOCK_ASSERT(pm, MA_OWNED); if (pm == kernel_pmap) { PMAP_STATS_INC(tsb_nenter_k); Index: sparc64/include/pmap.h =================================================================== --- sparc64/include/pmap.h (.../head/sys) (revision 227753) +++ sparc64/include/pmap.h (.../user/attilio/vmcontention/sys) (revision 227756) @@ -43,6 +43,7 @@ #include #include #include +#include #include #include @@ -78,6 +79,12 @@ #define PMAP_TRYLOCK(pmap) mtx_trylock(&(pmap)->pm_mtx) #define PMAP_UNLOCK(pmap) mtx_unlock(&(pmap)->pm_mtx) +#define MDPAGE_ASSERT_WLOCKED() rw_assert(&md_page_rwlock, RA_WLOCKED) +#define MDPAGE_RLOCK() rw_rlock(&md_page_rwlock) +#define MDPAGE_WLOCK() rw_wlock(&md_page_rwlock) +#define MDPAGE_RUNLOCK() rw_runlock(&md_page_rwlock) +#define MDPAGE_WUNLOCK() rw_wunlock(&md_page_rwlock) + #define pmap_page_get_memattr(m) VM_MEMATTR_DEFAULT #define pmap_page_set_memattr(m, ma) (void)0 @@ -101,6 +108,7 @@ extern struct pmap kernel_pmap_store; #define kernel_pmap (&kernel_pmap_store) +extern struct rwlock md_page_rwlock; extern vm_paddr_t phys_avail[]; extern vm_offset_t virtual_avail; extern vm_offset_t virtual_end;