Store the physical address of each pcb in struct mdproc, and use this in

order to speed up window spills, rather than doing an inline pmap_extract
(well, pseg_get).

ok claudio@ kettenis@

diff --git a/sys/arch/sparc64/include/cpu.h b/sys/arch/sparc64/include/cpu.h
index 6586476..de9c7cc 100644 (file)
--- a/sys/arch/sparc64/include/cpu.h
+++ b/sys/arch/sparc64/include/cpu.h
@@ -1,4 +1,4 @@
-/*     $OpenBSD: cpu.h,v 1.106 2024/03/29 21:09:04 miod Exp $  */
+/*     $OpenBSD: cpu.h,v 1.107 2024/03/29 21:27:53 miod Exp $  */
 /*     $NetBSD: cpu.h,v 1.28 2001/06/14 22:56:58 thorpej Exp $ */
 
 /*
@@ -113,6 +113,7 @@ struct cpu_info {
        /* Most important fields first */
        struct proc             *ci_curproc;
        struct pcb              *ci_cpcb;       /* also initial stack */
+       paddr_t                 ci_cpcbpaddr;
        struct cpu_info         *ci_next;
 
        struct proc             *ci_fpproc;

diff --git a/sys/arch/sparc64/include/proc.h b/sys/arch/sparc64/include/proc.h
index 721d856..31739fa 100644 (file)
--- a/sys/arch/sparc64/include/proc.h
+++ b/sys/arch/sparc64/include/proc.h
@@ -47,4 +47,5 @@ struct mdproc {
        struct  trapframe *md_tf;       /* trap/syscall registers */
        struct  fpstate *md_fpstate;    /* fpu state, if any; always resident */
        volatile int md_astpending;
+       paddr_t md_pcbpaddr;
 };

diff --git a/sys/arch/sparc64/sparc64/db_interface.c b/sys/arch/sparc64/sparc64/db_interface.c
index 570285b..0351dfc 100644 (file)
--- a/sys/arch/sparc64/sparc64/db_interface.c
+++ b/sys/arch/sparc64/sparc64/db_interface.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: db_interface.c,v 1.62 2024/02/23 18:19:03 cheloha Exp $       */
+/*     $OpenBSD: db_interface.c,v 1.63 2024/03/29 21:27:53 miod Exp $  */
 /*     $NetBSD: db_interface.c,v 1.61 2001/07/31 06:55:47 eeh Exp $ */
 
 /*
@@ -1009,6 +1009,7 @@ db_setpcb(db_expr_t addr, int have_addr, db_expr_t count, char *modif)
                if (p->p_stat && p->p_tid == addr) {
                        curproc = p;
                        curpcb = (struct pcb*)p->p_addr;
+                       curcpu()->ci_cpcbpaddr = p->p_md.md_pcbpaddr;
                        if (p->p_vmspace->vm_map.pmap->pm_ctx) {
                                switchtoctx(p->p_vmspace->vm_map.pmap->pm_ctx);
                                return;

diff --git a/sys/arch/sparc64/sparc64/genassym.cf b/sys/arch/sparc64/sparc64/genassym.cf
index 9ac3644..ec32585 100644 (file)
--- a/sys/arch/sparc64/sparc64/genassym.cf
+++ b/sys/arch/sparc64/sparc64/genassym.cf
@@ -1,4 +1,4 @@
-#      $OpenBSD: genassym.cf,v 1.43 2024/03/29 21:14:56 miod Exp $
+#      $OpenBSD: genassym.cf,v 1.44 2024/03/29 21:27:53 miod Exp $
 #      $NetBSD: genassym.cf,v 1.23 2001/08/08 00:09:30 eeh Exp $
 
 #
@@ -91,6 +91,7 @@ member        p_vmspace
 member p_cpu
 member P_FPSTATE       p_md.md_fpstate
 member P_MD_ASTPENDING p_md.md_astpending
+member P_MD_PCBPADDR   p_md.md_pcbpaddr
 
 export SONPROC
 
@@ -116,6 +117,7 @@ struct      cpu_info
 member ci_self
 member ci_curproc
 member ci_cpcb
+member ci_cpcbpaddr
 member ci_fpproc
 member ci_handled_intr_level
 member ci_intrpending

diff --git a/sys/arch/sparc64/sparc64/locore.s b/sys/arch/sparc64/sparc64/locore.s
index e7d4fe3..d34a856 100644 (file)
--- a/sys/arch/sparc64/sparc64/locore.s
+++ b/sys/arch/sparc64/sparc64/locore.s
@@ -1,4 +1,4 @@
-/*     $OpenBSD: locore.s,v 1.213 2024/03/29 21:25:55 miod Exp $       */
+/*     $OpenBSD: locore.s,v 1.214 2024/03/29 21:27:53 miod Exp $       */
 /*     $NetBSD: locore.s,v 1.137 2001/08/13 06:10:10 jdolecek Exp $    */
 
 /*
@@ -157,7 +157,7 @@ sun4u_mtp_patch_end:
 #ifdef SUN4V
 
 #define GET_CPUINFO_PA(ci) \
-       mov     0x10, ci                                ;\
+       mov     0x10, ci                        ;\
        ldxa    [ci] ASI_SCRATCHPAD, ci
 
 #define GET_MMFSA(mmfsa) \
@@ -1927,61 +1927,9 @@ winfixspill:
         wrpr   %g5, 0, %cwp                            !  return here and write out all dirty windows.
 #endif /* 1 */
        wrpr    %g2, 0, %tl                             ! Restore trap level for now XXXX
-       GET_CPCB(%g6)
-#ifdef DEBUG
-       set     0x12, %g5                               ! debug
-       sethi   %hi(DATA_START), %g7                    ! debug
-       stb     %g5, [%g7 + 0x20]                       ! debug
-       CHKPT %g5,%g7,0x11
-#endif /* DEBUG */
-
-       /*
-        * Traverse kernel map to find paddr of cpcb and only us ASI_PHYS_CACHED to
-        * prevent any faults while saving the windows.  BTW if it isn't mapped, we
-        * will trap and hopefully panic.
-        */
-
-!      ba      0f                                      ! DEBUG -- don't use phys addresses
-        wr     %g0, ASI_NUCLEUS, %asi                  ! In case of problems finding PA
-       sethi   %hi(ctxbusy), %g1
-       ldx     [%g1 + %lo(ctxbusy)], %g1               ! Load start of ctxbusy
-#ifdef DEBUG
-       srax    %g6, HOLESHIFT, %g7                     ! Check for valid address
-       brz,pt  %g7, 1f                                 ! Should be zero or -1
-        addcc  %g7, 1, %g7                                     ! Make -1 -> 0
-       tnz     %xcc, 1                                 ! Invalid address??? How did this happen?
-1:
-#endif /* DEBUG */
-       srlx    %g6, STSHIFT, %g7
-       ldx     [%g1], %g1                              ! Load pointer to kernel_pmap
-       and     %g7, STMASK, %g7
-       sll     %g7, 3, %g7
-       add     %g7, %g1, %g1
-       ldxa    [%g1] ASI_PHYS_CACHED, %g1              ! Load pointer to directory
-
-       srlx    %g6, PDSHIFT, %g7                       ! Do page directory
-       and     %g7, PDMASK, %g7
-       sll     %g7, 3, %g7
-       brz,pn  %g1, 0f
-        add    %g7, %g1, %g1
-       ldxa    [%g1] ASI_PHYS_CACHED, %g1
-
-       srlx    %g6, PTSHIFT, %g7                       ! Convert to ptab offset
-       and     %g7, PTMASK, %g7
-       brz     %g1, 0f
-        sll    %g7, 3, %g7
-       add     %g1, %g7, %g7
-       ldxa    [%g7] ASI_PHYS_CACHED, %g7              ! This one is not
-       brgez   %g7, 0f
-        srlx   %g7, PGSHIFT, %g7                       ! Isolate PA part
-       sll     %g6, 32-PGSHIFT, %g6                    ! And offset
-       sllx    %g7, PGSHIFT+17, %g7                    ! There are 17 bits to the left of the PA in the TTE
-       srl     %g6, 32-PGSHIFT, %g6
-       srax    %g7, 17, %g7
-       or      %g7, %g6, %g6                           ! Then combine them to form PA
-
+       GET_CPUINFO_VA(%g6)
+       ldx     [%g6 + CI_CPCBPADDR], %g6
        wr      %g0, ASI_PHYS_CACHED, %asi              ! Use ASI_PHYS_CACHED to prevent possible page faults
-0:
        /*
         * Now save all user windows to cpcb.
         */
@@ -3145,41 +3093,8 @@ pcbspill_others:
 
 pcbspill:
        GET_CPUINFO_PA(%g6)
-       wr      %g0, ASI_PHYS_CACHED, %asi
-       ldxa    [%g6 + CI_CPCB] %asi, %g6
-
-       sethi   %hi(ctxbusy), %g1
-       ldx     [%g1 + %lo(ctxbusy)], %g1
-       ldx     [%g1], %g1
-
-       srlx    %g6, STSHIFT, %g7
-       and     %g7, STMASK, %g7
-       sll     %g7, 3, %g7
-       add     %g7, %g1, %g1
-       ldxa    [%g1] ASI_PHYS_CACHED, %g1              ! Load pointer to directory
-
-       srlx    %g6, PDSHIFT, %g7                       ! Do page directory
-       and     %g7, PDMASK, %g7
-       sll     %g7, 3, %g7
-       brz,pn  %g1, pcbspill_fail
-        add    %g7, %g1, %g1
-       ldxa    [%g1] ASI_PHYS_CACHED, %g1
-
-       srlx    %g6, PTSHIFT, %g7                       ! Convert to ptab offset
-       and     %g7, PTMASK, %g7
-       brz     %g1, pcbspill_fail
-        sll    %g7, 3, %g7
-       add     %g1, %g7, %g7
-       ldxa    [%g7] ASI_PHYS_CACHED, %g7              ! This one is not
-       brgez   %g7, pcbspill_fail
-        srlx   %g7, PGSHIFT, %g7                       ! Isolate PA part
-       sll     %g6, 32-PGSHIFT, %g6                    ! And offset
-       sllx    %g7, PGSHIFT+8, %g7                     ! There are 8 bits to the left of the PA in the TTE
-       srl     %g6, 32-PGSHIFT, %g6
-       srax    %g7, 8, %g7
-       or      %g7, %g6, %g6                           ! Then combine them to form PA
-
-!      wr      %g0, ASI_PHYS_CACHED, %asi              ! Use ASI_PHYS_CACHED to prevent possible page faults
+       wr      %g0, ASI_PHYS_CACHED, %asi              ! Use ASI_PHYS_CACHED to prevent possible page faults
+       ldxa    [%g6 + CI_CPCBPADDR] %asi, %g6
        
        lduba   [%g6 + PCB_NSAVED] %asi, %g7
        sllx    %g7, 7, %g5
@@ -3197,10 +3112,6 @@ pcbspill:
        retry
        NOTREACHED
 
-pcbspill_fail:
-       db_enter()
-       NOTREACHED
-
 
 sun4v_datatrap:
        GET_MMFSA(%g3)
@@ -5885,9 +5796,11 @@ ENTRY(cpu_switchto)
         * We also must load up the `in' and `local' registers.
         */
 Lsw_load:
-       /* set new cpcb */
-       stx     %i1, [%g7 + CI_CURPROC] ! curproc = newproc;
-       stx     %l1, [%g7 + CI_CPCB]    ! cpcb = newpcb;
+       /* set new cpcb and cpcbpaddr */
+       stx     %i1, [%g7 + CI_CURPROC]         ! curproc = newproc;
+       ldx     [%i1 + P_MD_PCBPADDR], %o2
+       stx     %l1, [%g7 + CI_CPCB]            ! cpcb = newpcb;
+       stx     %o2, [%g7 + CI_CPCBPADDR]
 
        ldx     [%l1 + PCB_SP], %i6
        ldx     [%l1 + PCB_PC], %i7

diff --git a/sys/arch/sparc64/sparc64/machdep.c b/sys/arch/sparc64/sparc64/machdep.c
index ddd1c5c..68cadc2 100644 (file)
--- a/sys/arch/sparc64/sparc64/machdep.c
+++ b/sys/arch/sparc64/sparc64/machdep.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: machdep.c,v 1.214 2024/03/29 21:26:38 miod Exp $      */
+/*     $OpenBSD: machdep.c,v 1.215 2024/03/29 21:27:53 miod Exp $      */
 /*     $NetBSD: machdep.c,v 1.108 2001/07/24 19:30:14 eeh Exp $ */
 
 /*-
@@ -205,6 +205,8 @@ cpu_startup(void)
 #endif
 
        proc0.p_addr = proc0paddr;
+       (void)pmap_extract(pmap_kernel(), (vaddr_t)proc0paddr,
+           &proc0.p_md.md_pcbpaddr);
 
        /*
         * Good {morning,afternoon,evening,night}.

diff --git a/sys/arch/sparc64/sparc64/pmap.c b/sys/arch/sparc64/sparc64/pmap.c
index 0b9082b..c770206 100644 (file)
--- a/sys/arch/sparc64/sparc64/pmap.c
+++ b/sys/arch/sparc64/sparc64/pmap.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: pmap.c,v 1.115 2024/03/29 21:26:38 miod Exp $ */
+/*     $OpenBSD: pmap.c,v 1.116 2024/03/29 21:27:53 miod Exp $ */
 /*     $NetBSD: pmap.c,v 1.107 2001/08/31 16:47:41 eeh Exp $   */
 /*
  * 
@@ -1174,6 +1174,8 @@ remap_data:
                cpus->ci_next = NULL; /* Redundant, I know. */
                cpus->ci_curproc = &proc0;
                cpus->ci_cpcb = (struct pcb *)u0[0]; /* Need better source */
+               cpus->ci_cpcbpaddr = pseg_get(pmap_kernel(), u0[0]) &
+                   TLB_PA_MASK;
                cpus->ci_upaid = cpu_myid();
                cpus->ci_cpuid = 0;
                cpus->ci_flags = CPUF_RUNNING;

diff --git a/sys/arch/sparc64/sparc64/vm_machdep.c b/sys/arch/sparc64/sparc64/vm_machdep.c
index ea8c23b..73bf75e 100644 (file)
--- a/sys/arch/sparc64/sparc64/vm_machdep.c
+++ b/sys/arch/sparc64/sparc64/vm_machdep.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: vm_machdep.c,v 1.44 2024/03/29 21:12:58 miod Exp $    */
+/*     $OpenBSD: vm_machdep.c,v 1.45 2024/03/29 21:27:53 miod Exp $    */
 /*     $NetBSD: vm_machdep.c,v 1.38 2001/06/30 00:02:20 eeh Exp $ */
 
 /*
@@ -103,6 +103,13 @@ cpu_fork(struct proc *p1, struct proc *p2, void *stack, void *tcb,
        size_t pcbsz;
        extern struct proc proc0;
 
+       /*
+        * Cache the physical address of the pcb, to speed up window
+        * spills in locore.
+        */
+       (void)pmap_extract(pmap_kernel(), (vaddr_t)npcb,
+           &p2->p_md.md_pcbpaddr);
+
        /*
         * Save all user registers to p1's stack or, in the case of
         * user registers and invalid stack pointers, to opcb.