This adds a shadow buffer for the SLBs and regsiters it with PHYP.
Only the bolted SLB entries (top 3) are shadowed.
The SLB shadow buffer tells the hypervisor what the kernel needs to
have in the SLB for the kernel to be able to function.  The hypervisor
can use this information to speed up partition context switches.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
        DEFINE(PACA_STARTPURR, offsetof(struct paca_struct, startpurr));
        DEFINE(PACA_USER_TIME, offsetof(struct paca_struct, user_time));
        DEFINE(PACA_SYSTEM_TIME, offsetof(struct paca_struct, system_time));
+       DEFINE(PACA_SLBSHADOWPTR, offsetof(struct paca_struct, slb_shadow_ptr));
 
        DEFINE(LPPACASRR0, offsetof(struct lppaca, saved_srr0));
        DEFINE(LPPACASRR1, offsetof(struct lppaca, saved_srr1));
        DEFINE(LPPACAANYINT, offsetof(struct lppaca, int_dword.any_int));
        DEFINE(LPPACADECRINT, offsetof(struct lppaca, int_dword.fields.decr_int));
+       DEFINE(SLBSHADOW_SAVEAREA, offsetof(struct slb_shadow, save_area));
 #endif /* CONFIG_PPC64 */
 
        /* RTAS */
 
  * The code which creates the new task context is in 'copy_thread'
  * in arch/powerpc/kernel/process.c 
  */
+#define SHADOW_SLB_BOLTED_STACK_ESID \
+               (SLBSHADOW_SAVEAREA + 0x10*(SLB_NUM_BOLTED-1))
+#define SHADOW_SLB_BOLTED_STACK_VSID \
+               (SLBSHADOW_SAVEAREA + 0x10*(SLB_NUM_BOLTED-1) + 8)
+
        .align  7
 _GLOBAL(_switch)
        mflr    r0
        ld      r7,KSP_VSID(r4) /* Get new stack's VSID */
        oris    r0,r6,(SLB_ESID_V)@h
        ori     r0,r0,(SLB_NUM_BOLTED-1)@l
+
+       /* Update the last bolted SLB */
+       ld      r9,PACA_SLBSHADOWPTR(r13)
+       li      r12,0
+       std     r12,SHADOW_SLB_BOLTED_STACK_ESID(r9) /* Clear ESID */
+       std     r7,SHADOW_SLB_BOLTED_STACK_VSID(r9)  /* Save VSID */
+       std     r0,SHADOW_SLB_BOLTED_STACK_ESID(r9)  /* Save ESID */
+
        slbie   r6
        slbie   r6              /* Workaround POWER5 < DD2.1 issue */
        slbmte  r7,r0
 
 #include <asm/lppaca.h>
 #include <asm/iseries/it_lp_reg_save.h>
 #include <asm/paca.h>
+#include <asm/mmu.h>
 
 
 /* This symbol is provided by the linker - let it fill in the paca
        },
 };
 
+/*
+ * 3 persistent SLBs are registered here.  The buffer will be zero
+ * initially, hence will all be invaild until we actually write them.
+ */
+struct slb_shadow slb_shadow[] __cacheline_aligned = {
+       [0 ... (NR_CPUS-1)] = {
+               .persistent = SLB_NUM_BOLTED,
+               .buffer_length = sizeof(struct slb_shadow),
+       },
+};
+
 /* The Paca is an array with one entry per processor.  Each contains an
  * lppaca, which contains the information shared between the
  * hypervisor and Linux.
        .lock_token = 0x8000,                                               \
        .paca_index = (number),         /* Paca Index */                    \
        .kernel_toc = (unsigned long)(&__toc_start) + 0x8000UL,             \
-       .hw_cpu_id = 0xffff,
+       .hw_cpu_id = 0xffff,                                                \
+       .slb_shadow_ptr = &slb_shadow[number],
 
 #ifdef CONFIG_PPC_ISERIES
 #define PACA_INIT_ISERIES(number)                                          \
 
 #include <asm/paca.h>
 #include <asm/cputable.h>
 #include <asm/cacheflush.h>
+#include <asm/smp.h>
+#include <linux/compiler.h>
 
 #ifdef DEBUG
 #define DBG(fmt...) udbg_printf(fmt)
        return (get_kernel_vsid(ea) << SLB_VSID_SHIFT) | flags;
 }
 
-static inline void create_slbe(unsigned long ea, unsigned long flags,
-                              unsigned long entry)
+static inline void slb_shadow_update(unsigned long esid, unsigned long vsid,
+                                    unsigned long entry)
 {
+       /*
+        * Clear the ESID first so the entry is not valid while we are
+        * updating it.
+        */
+       get_slb_shadow()->save_area[entry].esid = 0;
+       barrier();
+       get_slb_shadow()->save_area[entry].vsid = vsid;
+       barrier();
+       get_slb_shadow()->save_area[entry].esid = esid;
+
+}
+
+static inline void create_shadowed_slbe(unsigned long ea, unsigned long flags,
+                                       unsigned long entry)
+{
+       /*
+        * Updating the shadow buffer before writing the SLB ensures
+        * we don't get a stale entry here if we get preempted by PHYP
+        * between these two statements.
+        */
+       slb_shadow_update(mk_esid_data(ea, entry), mk_vsid_data(ea, flags),
+                         entry);
+
        asm volatile("slbmte  %0,%1" :
                     : "r" (mk_vsid_data(ea, flags)),
                       "r" (mk_esid_data(ea, entry))
        if ((ksp_esid_data & ESID_MASK) == PAGE_OFFSET)
                ksp_esid_data &= ~SLB_ESID_V;
 
+       /* Only third entry (stack) may change here so only resave that */
+       slb_shadow_update(ksp_esid_data,
+                         mk_vsid_data(ksp_esid_data, lflags), 2);
+
        /* We need to do this all in asm, so we're sure we don't touch
         * the stack between the slbia and rebolting it. */
        asm volatile("isync\n"
        asm volatile("isync":::"memory");
        asm volatile("slbmte  %0,%0"::"r" (0) : "memory");
        asm volatile("isync; slbia; isync":::"memory");
-       create_slbe(PAGE_OFFSET, lflags, 0);
+       create_shadowed_slbe(PAGE_OFFSET, lflags, 0);
 
-       create_slbe(VMALLOC_START, vflags, 1);
+       create_shadowed_slbe(VMALLOC_START, vflags, 1);
 
        /* We don't bolt the stack for the time being - we're in boot,
         * so the stack is in the bolted segment.  By the time it goes
 
 void vpa_init(int cpu)
 {
        int hwcpu = get_hard_smp_processor_id(cpu);
-       unsigned long vpa = __pa(&lppaca[cpu]);
+       unsigned long addr;
        long ret;
 
        if (cpu_has_feature(CPU_FTR_ALTIVEC))
                lppaca[cpu].vmxregs_in_use = 1;
 
-       ret = register_vpa(hwcpu, vpa);
+       addr = __pa(&lppaca[cpu]);
+       ret = register_vpa(hwcpu, addr);
 
-       if (ret)
+       if (ret) {
                printk(KERN_ERR "WARNING: vpa_init: VPA registration for "
                                "cpu %d (hw %d) of area %lx returns %ld\n",
-                               cpu, hwcpu, vpa, ret);
+                               cpu, hwcpu, addr, ret);
+               return;
+       }
+       /*
+        * PAPR says this feature is SLB-Buffer but firmware never
+        * reports that.  All SPLPAR support SLB shadow buffer.
+        */
+       addr = __pa(&slb_shadow[cpu]);
+       if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
+               ret = register_slb_shadow(hwcpu, addr);
+               if (ret)
+                       printk(KERN_ERR
+                              "WARNING: vpa_init: SLB shadow buffer "
+                              "registration for cpu %d (hw %d) of area %lx "
+                              "returns %ld\n", cpu, hwcpu, addr, ret);
+       }
 }
 
 long pSeries_lpar_hpte_insert(unsigned long hpte_group,
 
        return vpa_call(0x1, cpu, vpa);
 }
 
+static inline long unregister_slb_shadow(unsigned long cpu, unsigned long vpa)
+{
+       return vpa_call(0x7, cpu, vpa);
+}
+
+static inline long register_slb_shadow(unsigned long cpu, unsigned long vpa)
+{
+       return vpa_call(0x3, cpu, vpa);
+}
+
 extern void vpa_init(int cpu);
 
 static inline long plpar_pte_enter(unsigned long flags,
 
 {
        /* Don't risk a hypervisor call if we're crashing */
        if (firmware_has_feature(FW_FEATURE_SPLPAR) && !crash_shutdown) {
-               unsigned long vpa = __pa(get_lppaca());
+               unsigned long addr;
 
-               if (unregister_vpa(hard_smp_processor_id(), vpa)) {
+               addr = __pa(get_slb_shadow());
+               if (unregister_slb_shadow(hard_smp_processor_id(), addr))
+                       printk("SLB shadow buffer deregistration of "
+                              "cpu %u (hw_cpu_id %d) failed\n",
+                              smp_processor_id(),
+                              hard_smp_processor_id());
+
+               addr = __pa(get_lppaca());
+               if (unregister_vpa(hard_smp_processor_id(), addr)) {
                        printk("VPA deregistration of cpu %u (hw_cpu_id %d) "
                                        "failed\n", smp_processor_id(),
                                        hard_smp_processor_id());
 
 //
 //
 //----------------------------------------------------------------------------
+#include <linux/cache.h>
 #include <asm/types.h>
+#include <asm/mmu.h>
 
 /* The Hypervisor barfs if the lppaca crosses a page boundary.  A 1k
  * alignment is sufficient to prevent this */
 
 extern struct lppaca lppaca[];
 
+/*
+ * SLB shadow buffer structure as defined in the PAPR.  The save_area
+ * contains adjacent ESID and VSID pairs for each shadowed SLB.  The
+ * ESID is stored in the lower 64bits, then the VSID.
+ */
+struct slb_shadow {
+       u32     persistent;             // Number of persistent SLBs    x00-x03
+       u32     buffer_length;          // Total shadow buffer length   x04-x07
+       u64     reserved;               // Alignment                    x08-x0f
+       struct  {
+               u64     esid;
+               u64     vsid;
+       } save_area[SLB_NUM_BOLTED];    //                              x10-x40
+} ____cacheline_aligned;
+
+extern struct slb_shadow slb_shadow[];
+
 #endif /* __KERNEL__ */
 #endif /* _ASM_POWERPC_LPPACA_H */
 
 register struct paca_struct *local_paca asm("r13");
 #define get_paca()     local_paca
 #define get_lppaca()   (get_paca()->lppaca_ptr)
+#define get_slb_shadow()       (get_paca()->slb_shadow_ptr)
 
 struct task_struct;
 
        u64 user_time;                  /* accumulated usermode TB ticks */
        u64 system_time;                /* accumulated system TB ticks */
        u64 startpurr;                  /* PURR/TB value snapshot */
+
+       struct slb_shadow *slb_shadow_ptr;
 };
 
 extern struct paca_struct paca[];