[PATCH] kernel-side context switch code for spufs
authorMark Nutter <mnutter@us.ibm.com>
Tue, 15 Nov 2005 20:53:50 +0000 (15:53 -0500)
committerPaul Mackerras <paulus@samba.org>
Mon, 9 Jan 2006 03:49:21 +0000 (14:49 +1100)
This adds the code needed to perform a context switch from
spufs, following the recommended 76-step sequence.

Signed-off-by: Arnd Bergmann <arndb@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
arch/powerpc/platforms/cell/spufs/switch.c
include/asm-powerpc/spu_csa.h

index 6804342..70345b0 100644 (file)
 #include "spu_save_dump.h"
 #include "spu_restore_dump.h"
 
+#if 0
+#define POLL_WHILE_TRUE(_c) {                          \
+    do {                                               \
+    } while (_c);                                      \
+  }
+#else
+#define RELAX_SPIN_COUNT                               1000
+#define POLL_WHILE_TRUE(_c) {                          \
+    do {                                               \
+       int _i;                                         \
+       for (_i=0; _i<RELAX_SPIN_COUNT && (_c); _i++) { \
+           cpu_relax();                                \
+       }                                               \
+       if (unlikely(_c)) yield();                      \
+       else break;                                     \
+    } while (_c);                                      \
+  }
+#endif                         /* debug */
+
+#define POLL_WHILE_FALSE(_c)   POLL_WHILE_TRUE(!(_c))
+
+static inline void acquire_spu_lock(struct spu *spu)
+{
+       /* Save, Step 1:
+        * Restore, Step 1:
+        *    Acquire SPU-specific mutual exclusion lock.
+        *    TBD.
+        */
+}
+
+static inline void release_spu_lock(struct spu *spu)
+{
+       /* Restore, Step 76:
+        *    Release SPU-specific mutual exclusion lock.
+        *    TBD.
+        */
+}
+
+static inline int check_spu_isolate(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       u32 isolate_state;
+
+       /* Save, Step 2:
+        * Save, Step 6:
+        *     If SPU_Status[E,L,IS] any field is '1', this
+        *     SPU is in isolate state and cannot be context
+        *     saved at this time.
+        */
+       isolate_state = SPU_STATUS_ISOLATED_STATE |
+           SPU_STATUS_ISOLATED_LOAD_STAUTUS | SPU_STATUS_ISOLATED_EXIT_STAUTUS;
+       return (in_be32(&prob->spu_status_R) & isolate_state) ? 1 : 0;
+}
+
+static inline void disable_interrupts(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Save, Step 3:
+        * Restore, Step 2:
+        *     Save INT_Mask_class0 in CSA.
+        *     Write INT_MASK_class0 with value of 0.
+        *     Save INT_Mask_class1 in CSA.
+        *     Write INT_MASK_class1 with value of 0.
+        *     Save INT_Mask_class2 in CSA.
+        *     Write INT_MASK_class2 with value of 0.
+        */
+       spin_lock_irq(&spu->register_lock);
+       if (csa) {
+               csa->priv1.int_mask_class0_RW =
+                   in_be64(&priv1->int_mask_class0_RW);
+               csa->priv1.int_mask_class1_RW =
+                   in_be64(&priv1->int_mask_class1_RW);
+               csa->priv1.int_mask_class2_RW =
+                   in_be64(&priv1->int_mask_class2_RW);
+       }
+       out_be64(&priv1->int_mask_class0_RW, 0UL);
+       out_be64(&priv1->int_mask_class1_RW, 0UL);
+       out_be64(&priv1->int_mask_class2_RW, 0UL);
+       eieio();
+       spin_unlock_irq(&spu->register_lock);
+}
+
+static inline void set_watchdog_timer(struct spu_state *csa, struct spu *spu)
+{
+       /* Save, Step 4:
+        * Restore, Step 25.
+        *    Set a software watchdog timer, which specifies the
+        *    maximum allowable time for a context save sequence.
+        *
+        *    For present, this implementation will not set a global
+        *    watchdog timer, as virtualization & variable system load
+        *    may cause unpredictable execution times.
+        */
+}
+
+static inline void inhibit_user_access(struct spu_state *csa, struct spu *spu)
+{
+       /* Save, Step 5:
+        * Restore, Step 3:
+        *     Inhibit user-space access (if provided) to this
+        *     SPU by unmapping the virtual pages assigned to
+        *     the SPU memory-mapped I/O (MMIO) for problem
+        *     state. TBD.
+        */
+}
+
+static inline void set_switch_pending(struct spu_state *csa, struct spu *spu)
+{
+       /* Save, Step 7:
+        * Restore, Step 5:
+        *     Set a software context switch pending flag.
+        */
+       set_bit(SPU_CONTEXT_SWITCH_PENDING_nr, &spu->flags);
+       mb();
+}
+
+static inline void save_mfc_cntl(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 8:
+        *     Read and save MFC_CNTL[Ss].
+        */
+       if (csa) {
+               csa->priv2.mfc_control_RW = in_be64(&priv2->mfc_control_RW) &
+                   MFC_CNTL_SUSPEND_DMA_STATUS_MASK;
+       }
+}
+
+static inline void save_spu_runcntl(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 9:
+        *     Save SPU_Runcntl in the CSA.  This value contains
+        *     the "Application Desired State".
+        */
+       csa->prob.spu_runcntl_RW = in_be32(&prob->spu_runcntl_RW);
+}
+
+static inline void save_mfc_sr1(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Save, Step 10:
+        *     Save MFC_SR1 in the CSA.
+        */
+       csa->priv1.mfc_sr1_RW = in_be64(&priv1->mfc_sr1_RW);
+}
+
+static inline void save_spu_status(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 11:
+        *     Read SPU_Status[R], and save to CSA.
+        */
+       if ((in_be32(&prob->spu_status_R) & SPU_STATUS_RUNNING) == 0) {
+               csa->prob.spu_status_R = in_be32(&prob->spu_status_R);
+       } else {
+               u32 stopped;
+
+               out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_STOP);
+               eieio();
+               POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                               SPU_STATUS_RUNNING);
+               stopped =
+                   SPU_STATUS_INVALID_INSTR | SPU_STATUS_SINGLE_STEP |
+                   SPU_STATUS_STOPPED_BY_HALT | SPU_STATUS_STOPPED_BY_STOP;
+               if ((in_be32(&prob->spu_status_R) & stopped) == 0)
+                       csa->prob.spu_status_R = SPU_STATUS_RUNNING;
+               else
+                       csa->prob.spu_status_R = in_be32(&prob->spu_status_R);
+       }
+}
+
+static inline void save_mfc_decr(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 12:
+        *     Read MFC_CNTL[Ds].  Update saved copy of
+        *     CSA.MFC_CNTL[Ds].
+        */
+       if (in_be64(&priv2->mfc_control_RW) & MFC_CNTL_DECREMENTER_RUNNING) {
+               csa->priv2.mfc_control_RW |= MFC_CNTL_DECREMENTER_RUNNING;
+               csa->suspend_time = get_cycles();
+               out_be64(&priv2->spu_chnlcntptr_RW, 7ULL);
+               eieio();
+               csa->spu_chnldata_RW[7] = in_be64(&priv2->spu_chnldata_RW);
+               eieio();
+       }
+}
+
+static inline void halt_mfc_decr(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 13:
+        *     Write MFC_CNTL[Dh] set to a '1' to halt
+        *     the decrementer.
+        */
+       out_be64(&priv2->mfc_control_RW, MFC_CNTL_DECREMENTER_HALTED);
+       eieio();
+}
+
+static inline void save_timebase(struct spu_state *csa, struct spu *spu)
+{
+       /* Save, Step 14:
+        *    Read PPE Timebase High and Timebase low registers
+        *    and save in CSA.  TBD.
+        */
+       csa->suspend_time = get_cycles();
+}
+
+static inline void remove_other_spu_access(struct spu_state *csa,
+                                          struct spu *spu)
+{
+       /* Save, Step 15:
+        *     Remove other SPU access to this SPU by unmapping
+        *     this SPU's pages from their address space.  TBD.
+        */
+}
+
+static inline void do_mfc_mssync(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 16:
+        * Restore, Step 11.
+        *     Write SPU_MSSync register. Poll SPU_MSSync[P]
+        *     for a value of 0.
+        */
+       out_be64(&prob->spc_mssync_RW, 1UL);
+       POLL_WHILE_TRUE(in_be64(&prob->spc_mssync_RW) & MS_SYNC_PENDING);
+}
+
+static inline void issue_mfc_tlbie(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Save, Step 17:
+        * Restore, Step 12.
+        * Restore, Step 48.
+        *     Write TLB_Invalidate_Entry[IS,VPN,L,Lp]=0 register.
+        *     Then issue a PPE sync instruction.
+        */
+       out_be64(&priv1->tlb_invalidate_entry_W, 0UL);
+       mb();
+}
+
+static inline void handle_pending_interrupts(struct spu_state *csa,
+                                            struct spu *spu)
+{
+       /* Save, Step 18:
+        *     Handle any pending interrupts from this SPU
+        *     here.  This is OS or hypervisor specific.  One
+        *     option is to re-enable interrupts to handle any
+        *     pending interrupts, with the interrupt handlers
+        *     recognizing the software Context Switch Pending
+        *     flag, to ensure the SPU execution or MFC command
+        *     queue is not restarted.  TBD.
+        */
+}
+
+static inline void save_mfc_queues(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       int i;
+
+       /* Save, Step 19:
+        *     If MFC_Cntl[Se]=0 then save
+        *     MFC command queues.
+        */
+       if ((in_be64(&priv2->mfc_control_RW) & MFC_CNTL_DMA_QUEUES_EMPTY) == 0) {
+               for (i = 0; i < 8; i++) {
+                       csa->priv2.puq[i].mfc_cq_data0_RW =
+                           in_be64(&priv2->puq[i].mfc_cq_data0_RW);
+                       csa->priv2.puq[i].mfc_cq_data1_RW =
+                           in_be64(&priv2->puq[i].mfc_cq_data1_RW);
+                       csa->priv2.puq[i].mfc_cq_data2_RW =
+                           in_be64(&priv2->puq[i].mfc_cq_data2_RW);
+                       csa->priv2.puq[i].mfc_cq_data3_RW =
+                           in_be64(&priv2->puq[i].mfc_cq_data3_RW);
+               }
+               for (i = 0; i < 16; i++) {
+                       csa->priv2.spuq[i].mfc_cq_data0_RW =
+                           in_be64(&priv2->spuq[i].mfc_cq_data0_RW);
+                       csa->priv2.spuq[i].mfc_cq_data1_RW =
+                           in_be64(&priv2->spuq[i].mfc_cq_data1_RW);
+                       csa->priv2.spuq[i].mfc_cq_data2_RW =
+                           in_be64(&priv2->spuq[i].mfc_cq_data2_RW);
+                       csa->priv2.spuq[i].mfc_cq_data3_RW =
+                           in_be64(&priv2->spuq[i].mfc_cq_data3_RW);
+               }
+       }
+}
+
+static inline void save_ppu_querymask(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 20:
+        *     Save the PPU_QueryMask register
+        *     in the CSA.
+        */
+       csa->prob.dma_querymask_RW = in_be32(&prob->dma_querymask_RW);
+}
+
+static inline void save_ppu_querytype(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 21:
+        *     Save the PPU_QueryType register
+        *     in the CSA.
+        */
+       csa->prob.dma_querytype_RW = in_be32(&prob->dma_querytype_RW);
+}
+
+static inline void save_mfc_csr_tsq(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 22:
+        *     Save the MFC_CSR_TSQ register
+        *     in the LSCSA.
+        */
+       csa->priv2.spu_tag_status_query_RW =
+           in_be64(&priv2->spu_tag_status_query_RW);
+}
+
+static inline void save_mfc_csr_cmd(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 23:
+        *     Save the MFC_CSR_CMD1 and MFC_CSR_CMD2
+        *     registers in the CSA.
+        */
+       csa->priv2.spu_cmd_buf1_RW = in_be64(&priv2->spu_cmd_buf1_RW);
+       csa->priv2.spu_cmd_buf2_RW = in_be64(&priv2->spu_cmd_buf2_RW);
+}
+
+static inline void save_mfc_csr_ato(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 24:
+        *     Save the MFC_CSR_ATO register in
+        *     the CSA.
+        */
+       csa->priv2.spu_atomic_status_RW = in_be64(&priv2->spu_atomic_status_RW);
+}
+
+static inline void save_mfc_tclass_id(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Save, Step 25:
+        *     Save the MFC_TCLASS_ID register in
+        *     the CSA.
+        */
+       csa->priv1.mfc_tclass_id_RW = in_be64(&priv1->mfc_tclass_id_RW);
+}
+
+static inline void set_mfc_tclass_id(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Save, Step 26:
+        * Restore, Step 23.
+        *     Write the MFC_TCLASS_ID register with
+        *     the value 0x10000000.
+        */
+       out_be64(&priv1->mfc_tclass_id_RW, 0x10000000);
+       eieio();
+}
+
+static inline void purge_mfc_queue(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 27:
+        * Restore, Step 14.
+        *     Write MFC_CNTL[Pc]=1 (purge queue).
+        */
+       out_be64(&priv2->mfc_control_RW, MFC_CNTL_PURGE_DMA_REQUEST);
+       eieio();
+}
+
+static inline void wait_purge_complete(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 28:
+        *     Poll MFC_CNTL[Ps] until value '11' is read
+        *     (purge complete).
+        */
+       POLL_WHILE_FALSE(in_be64(&priv2->mfc_control_RW) &
+                        MFC_CNTL_PURGE_DMA_COMPLETE);
+}
+
+static inline void save_mfc_slbs(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       int i;
+
+       /* Save, Step 29:
+        *     If MFC_SR1[R]='1', save SLBs in CSA.
+        */
+       if (in_be64(&priv1->mfc_sr1_RW) & MFC_STATE1_RELOCATE_MASK) {
+               csa->priv2.slb_index_W = in_be64(&priv2->slb_index_W);
+               for (i = 0; i < 8; i++) {
+                       out_be64(&priv2->slb_index_W, i);
+                       eieio();
+                       csa->slb_esid_RW[i] = in_be64(&priv2->slb_esid_RW);
+                       csa->slb_vsid_RW[i] = in_be64(&priv2->slb_vsid_RW);
+                       eieio();
+               }
+       }
+}
+
+static inline void setup_mfc_sr1(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Save, Step 30:
+        * Restore, Step 18:
+        *     Write MFC_SR1 with MFC_SR1[D=0,S=1] and
+        *     MFC_SR1[TL,R,Pr,T] set correctly for the
+        *     OS specific environment.
+        *
+        *     Implementation note: The SPU-side code
+        *     for save/restore is privileged, so the
+        *     MFC_SR1[Pr] bit is not set.
+        *
+        */
+       out_be64(&priv1->mfc_sr1_RW, (MFC_STATE1_MASTER_RUN_CONTROL_MASK |
+                                     MFC_STATE1_RELOCATE_MASK |
+                                     MFC_STATE1_BUS_TLBIE_MASK));
+}
+
+static inline void save_spu_npc(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 31:
+        *     Save SPU_NPC in the CSA.
+        */
+       csa->prob.spu_npc_RW = in_be32(&prob->spu_npc_RW);
+}
+
+static inline void save_spu_privcntl(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 32:
+        *     Save SPU_PrivCntl in the CSA.
+        */
+       csa->priv2.spu_privcntl_RW = in_be64(&priv2->spu_privcntl_RW);
+}
+
+static inline void reset_spu_privcntl(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 33:
+        * Restore, Step 16:
+        *     Write SPU_PrivCntl[S,Le,A] fields reset to 0.
+        */
+       out_be64(&priv2->spu_privcntl_RW, 0UL);
+       eieio();
+}
+
+static inline void save_spu_lslr(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 34:
+        *     Save SPU_LSLR in the CSA.
+        */
+       csa->priv2.spu_lslr_RW = in_be64(&priv2->spu_lslr_RW);
+}
+
+static inline void reset_spu_lslr(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 35:
+        * Restore, Step 17.
+        *     Reset SPU_LSLR.
+        */
+       out_be64(&priv2->spu_lslr_RW, LS_ADDR_MASK);
+       eieio();
+}
+
+static inline void save_spu_cfg(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 36:
+        *     Save SPU_Cfg in the CSA.
+        */
+       csa->priv2.spu_cfg_RW = in_be64(&priv2->spu_cfg_RW);
+}
+
+static inline void save_pm_trace(struct spu_state *csa, struct spu *spu)
+{
+       /* Save, Step 37:
+        *     Save PM_Trace_Tag_Wait_Mask in the CSA.
+        *     Not performed by this implementation.
+        */
+}
+
+static inline void save_mfc_rag(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Save, Step 38:
+        *     Save RA_GROUP_ID register and the
+        *     RA_ENABLE reigster in the CSA.
+        */
+       csa->priv1.resource_allocation_groupID_RW =
+           in_be64(&priv1->resource_allocation_groupID_RW);
+       csa->priv1.resource_allocation_enable_RW =
+           in_be64(&priv1->resource_allocation_enable_RW);
+}
+
+static inline void save_ppu_mb_stat(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 39:
+        *     Save MB_Stat register in the CSA.
+        */
+       csa->prob.mb_stat_R = in_be32(&prob->mb_stat_R);
+}
+
+static inline void save_ppu_mb(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 40:
+        *     Save the PPU_MB register in the CSA.
+        */
+       csa->prob.pu_mb_R = in_be32(&prob->pu_mb_R);
+}
+
+static inline void save_ppuint_mb(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 41:
+        *     Save the PPUINT_MB register in the CSA.
+        */
+       csa->priv2.puint_mb_R = in_be64(&priv2->puint_mb_R);
+}
+
+static inline void save_ch_part1(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       u64 idx, ch_indices[7] = { 0UL, 1UL, 3UL, 4UL, 24UL, 25UL, 27UL };
+       int i;
+
+       /* Save, Step 42:
+        *     Save the following CH: [0,1,3,4,24,25,27]
+        */
+       for (i = 0; i < 7; i++) {
+               idx = ch_indices[i];
+               out_be64(&priv2->spu_chnlcntptr_RW, idx);
+               eieio();
+               csa->spu_chnldata_RW[idx] = in_be64(&priv2->spu_chnldata_RW);
+               csa->spu_chnlcnt_RW[idx] = in_be64(&priv2->spu_chnlcnt_RW);
+               out_be64(&priv2->spu_chnldata_RW, 0UL);
+               out_be64(&priv2->spu_chnlcnt_RW, 0UL);
+               eieio();
+       }
+}
+
+static inline void save_spu_mb(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       int i;
+
+       /* Save, Step 43:
+        *     Save SPU Read Mailbox Channel.
+        */
+       out_be64(&priv2->spu_chnlcntptr_RW, 29UL);
+       eieio();
+       csa->spu_chnlcnt_RW[29] = in_be64(&priv2->spu_chnlcnt_RW);
+       for (i = 0; i < 4; i++) {
+               csa->pu_mailbox_data[i] = in_be64(&priv2->spu_chnldata_RW);
+       }
+       out_be64(&priv2->spu_chnlcnt_RW, 0UL);
+       eieio();
+}
+
+static inline void save_mfc_cmd(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 44:
+        *     Save MFC_CMD Channel.
+        */
+       out_be64(&priv2->spu_chnlcntptr_RW, 21UL);
+       eieio();
+       csa->spu_chnlcnt_RW[21] = in_be64(&priv2->spu_chnlcnt_RW);
+       eieio();
+}
+
+static inline void reset_ch(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       u64 ch_indices[4] = { 21UL, 23UL, 28UL, 30UL };
+       u64 ch_counts[4] = { 16UL, 1UL, 1UL, 1UL };
+       u64 idx;
+       int i;
+
+       /* Save, Step 45:
+        *     Reset the following CH: [21, 23, 28, 30]
+        */
+       for (i = 0; i < 4; i++) {
+               idx = ch_indices[i];
+               out_be64(&priv2->spu_chnlcntptr_RW, idx);
+               eieio();
+               out_be64(&priv2->spu_chnlcnt_RW, ch_counts[i]);
+               eieio();
+       }
+}
+
+static inline void resume_mfc_queue(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 46:
+        * Restore, Step 25.
+        *     Write MFC_CNTL[Sc]=0 (resume queue processing).
+        */
+       out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESUME_DMA_QUEUE);
+}
+
+static inline void invalidate_slbs(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Save, Step 45:
+        * Restore, Step 19:
+        *     If MFC_SR1[R]=1, write 0 to SLB_Invalidate_All.
+        */
+       if (in_be64(&priv1->mfc_sr1_RW) & MFC_STATE1_RELOCATE_MASK) {
+               out_be64(&priv2->slb_invalidate_all_W, 0UL);
+               eieio();
+       }
+}
+
+static inline void get_kernel_slb(u64 ea, u64 slb[2])
+{
+       slb[0] = (get_kernel_vsid(ea) << SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
+       slb[1] = (ea & ESID_MASK) | SLB_ESID_V;
+
+       /* Large pages are used for kernel text/data, but not vmalloc.  */
+       if (cpu_has_feature(CPU_FTR_16M_PAGE)
+           && REGION_ID(ea) == KERNEL_REGION_ID)
+               slb[0] |= SLB_VSID_L;
+}
+
+static inline void load_mfc_slb(struct spu *spu, u64 slb[2], int slbe)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       out_be64(&priv2->slb_index_W, slbe);
+       eieio();
+       out_be64(&priv2->slb_vsid_RW, slb[0]);
+       out_be64(&priv2->slb_esid_RW, slb[1]);
+       eieio();
+}
+
+static inline void setup_mfc_slbs(struct spu_state *csa, struct spu *spu)
+{
+       u64 code_slb[2];
+       u64 lscsa_slb[2];
+
+       /* Save, Step 47:
+        * Restore, Step 30.
+        *     If MFC_SR1[R]=1, write 0 to SLB_Invalidate_All
+        *     register, then initialize SLB_VSID and SLB_ESID
+        *     to provide access to SPU context save code and
+        *     LSCSA.
+        *
+        *     This implementation places both the context
+        *     switch code and LSCSA in kernel address space.
+        *
+        *     Further this implementation assumes that the
+        *     MFC_SR1[R]=1 (in other words, assume that
+        *     translation is desired by OS environment).
+        */
+       invalidate_slbs(csa, spu);
+       get_kernel_slb((unsigned long)&spu_save_code[0], code_slb);
+       get_kernel_slb((unsigned long)csa->lscsa, lscsa_slb);
+       load_mfc_slb(spu, code_slb, 0);
+       if ((lscsa_slb[0] != code_slb[0]) || (lscsa_slb[1] != code_slb[1]))
+               load_mfc_slb(spu, lscsa_slb, 1);
+}
+
+static inline void set_switch_active(struct spu_state *csa, struct spu *spu)
+{
+       /* Save, Step 48:
+        * Restore, Step 23.
+        *     Change the software context switch pending flag
+        *     to context switch active.
+        */
+       set_bit(SPU_CONTEXT_SWITCH_ACTIVE_nr, &spu->flags);
+       clear_bit(SPU_CONTEXT_SWITCH_PENDING_nr, &spu->flags);
+       mb();
+}
+
+static inline void enable_interrupts(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+       unsigned long class1_mask = CLASS1_ENABLE_SEGMENT_FAULT_INTR |
+           CLASS1_ENABLE_STORAGE_FAULT_INTR;
+
+       /* Save, Step 49:
+        * Restore, Step 22:
+        *     Reset and then enable interrupts, as
+        *     needed by OS.
+        *
+        *     This implementation enables only class1
+        *     (translation) interrupts.
+        */
+       spin_lock_irq(&spu->register_lock);
+       out_be64(&priv1->int_stat_class0_RW, ~(0UL));
+       out_be64(&priv1->int_stat_class1_RW, ~(0UL));
+       out_be64(&priv1->int_stat_class2_RW, ~(0UL));
+       out_be64(&priv1->int_mask_class0_RW, 0UL);
+       out_be64(&priv1->int_mask_class1_RW, class1_mask);
+       out_be64(&priv1->int_mask_class2_RW, 0UL);
+       spin_unlock_irq(&spu->register_lock);
+}
+
+static inline int send_mfc_dma(struct spu *spu, unsigned long ea,
+                              unsigned int ls_offset, unsigned int size,
+                              unsigned int tag, unsigned int rclass,
+                              unsigned int cmd)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       union mfc_tag_size_class_cmd command;
+       unsigned int transfer_size;
+       volatile unsigned int status = 0x0;
+
+       while (size > 0) {
+               transfer_size =
+                   (size > MFC_MAX_DMA_SIZE) ? MFC_MAX_DMA_SIZE : size;
+               command.u.mfc_size = transfer_size;
+               command.u.mfc_tag = tag;
+               command.u.mfc_rclassid = rclass;
+               command.u.mfc_cmd = cmd;
+               do {
+                       out_be32(&prob->mfc_lsa_W, ls_offset);
+                       out_be64(&prob->mfc_ea_W, ea);
+                       out_be64(&prob->mfc_union_W.all64, command.all64);
+                       status =
+                           in_be32(&prob->mfc_union_W.by32.mfc_class_cmd32);
+                       if (unlikely(status & 0x2)) {
+                               cpu_relax();
+                       }
+               } while (status & 0x3);
+               size -= transfer_size;
+               ea += transfer_size;
+               ls_offset += transfer_size;
+       }
+       return 0;
+}
+
+static inline void save_ls_16kb(struct spu_state *csa, struct spu *spu)
+{
+       unsigned long addr = (unsigned long)&csa->lscsa->ls[0];
+       unsigned int ls_offset = 0x0;
+       unsigned int size = 16384;
+       unsigned int tag = 0;
+       unsigned int rclass = 0;
+       unsigned int cmd = MFC_PUT_CMD;
+
+       /* Save, Step 50:
+        *     Issue a DMA command to copy the first 16K bytes
+        *     of local storage to the CSA.
+        */
+       send_mfc_dma(spu, addr, ls_offset, size, tag, rclass, cmd);
+}
+
+static inline void set_spu_npc(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 51:
+        * Restore, Step 31.
+        *     Write SPU_NPC[IE]=0 and SPU_NPC[LSA] to entry
+        *     point address of context save code in local
+        *     storage.
+        *
+        *     This implementation uses SPU-side save/restore
+        *     programs with entry points at LSA of 0.
+        */
+       out_be32(&prob->spu_npc_RW, 0);
+       eieio();
+}
+
+static inline void set_signot1(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       union {
+               u64 ull;
+               u32 ui[2];
+       } addr64;
+
+       /* Save, Step 52:
+        * Restore, Step 32:
+        *    Write SPU_Sig_Notify_1 register with upper 32-bits
+        *    of the CSA.LSCSA effective address.
+        */
+       addr64.ull = (u64) csa->lscsa;
+       out_be32(&prob->signal_notify1, addr64.ui[0]);
+       eieio();
+}
+
+static inline void set_signot2(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       union {
+               u64 ull;
+               u32 ui[2];
+       } addr64;
+
+       /* Save, Step 53:
+        * Restore, Step 33:
+        *    Write SPU_Sig_Notify_2 register with lower 32-bits
+        *    of the CSA.LSCSA effective address.
+        */
+       addr64.ull = (u64) csa->lscsa;
+       out_be32(&prob->signal_notify2, addr64.ui[1]);
+       eieio();
+}
+
+static inline void send_save_code(struct spu_state *csa, struct spu *spu)
+{
+       unsigned long addr = (unsigned long)&spu_save_code[0];
+       unsigned int ls_offset = 0x0;
+       unsigned int size = sizeof(spu_save_code);
+       unsigned int tag = 0;
+       unsigned int rclass = 0;
+       unsigned int cmd = MFC_GETFS_CMD;
+
+       /* Save, Step 54:
+        *     Issue a DMA command to copy context save code
+        *     to local storage and start SPU.
+        */
+       send_mfc_dma(spu, addr, ls_offset, size, tag, rclass, cmd);
+}
+
+static inline void set_ppu_querymask(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Save, Step 55:
+        * Restore, Step 38.
+        *     Write PPU_QueryMask=1 (enable Tag Group 0)
+        *     and issue eieio instruction.
+        */
+       out_be32(&prob->dma_querymask_RW, MFC_TAGID_TO_TAGMASK(0));
+       eieio();
+}
+
+static inline void wait_tag_complete(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+       struct spu_problem __iomem *prob = spu->problem;
+       u32 mask = MFC_TAGID_TO_TAGMASK(0);
+       unsigned long flags;
+
+       /* Save, Step 56:
+        * Restore, Step 39.
+        * Restore, Step 39.
+        * Restore, Step 46.
+        *     Poll PPU_TagStatus[gn] until 01 (Tag group 0 complete)
+        *     or write PPU_QueryType[TS]=01 and wait for Tag Group
+        *     Complete Interrupt.  Write INT_Stat_Class0 or
+        *     INT_Stat_Class2 with value of 'handled'.
+        */
+       POLL_WHILE_FALSE(in_be32(&prob->dma_tagstatus_R) & mask);
+
+       local_irq_save(flags);
+       out_be64(&priv1->int_stat_class0_RW, ~(0UL));
+       out_be64(&priv1->int_stat_class2_RW, ~(0UL));
+       local_irq_restore(flags);
+}
+
+static inline void wait_spu_stopped(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+       struct spu_problem __iomem *prob = spu->problem;
+       unsigned long flags;
+
+       /* Save, Step 57:
+        * Restore, Step 40.
+        *     Poll until SPU_Status[R]=0 or wait for SPU Class 0
+        *     or SPU Class 2 interrupt.  Write INT_Stat_class0
+        *     or INT_Stat_class2 with value of handled.
+        */
+       POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) & SPU_STATUS_RUNNING);
+
+       local_irq_save(flags);
+       out_be64(&priv1->int_stat_class0_RW, ~(0UL));
+       out_be64(&priv1->int_stat_class2_RW, ~(0UL));
+       local_irq_restore(flags);
+}
+
+static inline int check_save_status(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       u32 complete;
+
+       /* Save, Step 54:
+        *     If SPU_Status[P]=1 and SPU_Status[SC] = "success",
+        *     context save succeeded, otherwise context save
+        *     failed.
+        */
+       complete = ((SPU_SAVE_COMPLETE << SPU_STOP_STATUS_SHIFT) |
+                   SPU_STATUS_STOPPED_BY_STOP);
+       return (in_be32(&prob->spu_status_R) != complete) ? 1 : 0;
+}
+
+static inline void terminate_spu_app(struct spu_state *csa, struct spu *spu)
+{
+       /* Restore, Step 4:
+        *    If required, notify the "using application" that
+        *    the SPU task has been terminated.  TBD.
+        */
+}
+
+static inline void suspend_mfc(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 7:
+        * Restore, Step 47.
+        *     Write MFC_Cntl[Dh,Sc]='1','1' to suspend
+        *     the queue and halt the decrementer.
+        */
+       out_be64(&priv2->mfc_control_RW, MFC_CNTL_SUSPEND_DMA_QUEUE |
+                MFC_CNTL_DECREMENTER_HALTED);
+       eieio();
+}
+
+static inline void wait_suspend_mfc_complete(struct spu_state *csa,
+                                            struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 8:
+        * Restore, Step 47.
+        *     Poll MFC_CNTL[Ss] until 11 is returned.
+        */
+       POLL_WHILE_FALSE(in_be64(&priv2->mfc_control_RW) &
+                        MFC_CNTL_SUSPEND_COMPLETE);
+}
+
+static inline int suspend_spe(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Restore, Step 9:
+        *    If SPU_Status[R]=1, stop SPU execution
+        *    and wait for stop to complete.
+        *
+        *    Returns       1 if SPU_Status[R]=1 on entry.
+        *                  0 otherwise
+        */
+       if (in_be32(&prob->spu_status_R) & SPU_STATUS_RUNNING) {
+               if (in_be32(&prob->spu_status_R) &
+                   SPU_STATUS_ISOLATED_EXIT_STAUTUS) {
+                       POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                                       SPU_STATUS_RUNNING);
+               }
+               if ((in_be32(&prob->spu_status_R) &
+                    SPU_STATUS_ISOLATED_LOAD_STAUTUS)
+                   || (in_be32(&prob->spu_status_R) &
+                       SPU_STATUS_ISOLATED_STATE)) {
+                       out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_STOP);
+                       eieio();
+                       POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                                       SPU_STATUS_RUNNING);
+                       out_be32(&prob->spu_runcntl_RW, 0x2);
+                       eieio();
+                       POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                                       SPU_STATUS_RUNNING);
+               }
+               if (in_be32(&prob->spu_status_R) &
+                   SPU_STATUS_WAITING_FOR_CHANNEL) {
+                       out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_STOP);
+                       eieio();
+                       POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                                       SPU_STATUS_RUNNING);
+               }
+               return 1;
+       }
+       return 0;
+}
+
+static inline void clear_spu_status(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Restore, Step 10:
+        *    If SPU_Status[R]=0 and SPU_Status[E,L,IS]=1,
+        *    release SPU from isolate state.
+        */
+       if (!(in_be32(&prob->spu_status_R) & SPU_STATUS_RUNNING)) {
+               if (in_be32(&prob->spu_status_R) &
+                   SPU_STATUS_ISOLATED_EXIT_STAUTUS) {
+                       out_be64(&priv1->mfc_sr1_RW,
+                                MFC_STATE1_MASTER_RUN_CONTROL_MASK);
+                       eieio();
+                       out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_RUNNABLE);
+                       eieio();
+                       POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                                       SPU_STATUS_RUNNING);
+               }
+               if ((in_be32(&prob->spu_status_R) &
+                    SPU_STATUS_ISOLATED_LOAD_STAUTUS)
+                   || (in_be32(&prob->spu_status_R) &
+                       SPU_STATUS_ISOLATED_STATE)) {
+                       out_be64(&priv1->mfc_sr1_RW,
+                                MFC_STATE1_MASTER_RUN_CONTROL_MASK);
+                       eieio();
+                       out_be32(&prob->spu_runcntl_RW, 0x2);
+                       eieio();
+                       POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                                       SPU_STATUS_RUNNING);
+               }
+       }
+}
+
+static inline void reset_ch_part1(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       u64 ch_indices[7] = { 0UL, 1UL, 3UL, 4UL, 24UL, 25UL, 27UL };
+       u64 idx;
+       int i;
+
+       /* Restore, Step 20:
+        *     Reset the following CH: [0,1,3,4,24,25,27]
+        */
+       for (i = 0; i < 7; i++) {
+               idx = ch_indices[i];
+               out_be64(&priv2->spu_chnlcntptr_RW, idx);
+               eieio();
+               out_be64(&priv2->spu_chnldata_RW, 0UL);
+               out_be64(&priv2->spu_chnlcnt_RW, 0UL);
+               eieio();
+       }
+}
+
+static inline void reset_ch_part2(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       u64 ch_indices[5] = { 21UL, 23UL, 28UL, 29UL, 30UL };
+       u64 ch_counts[5] = { 16UL, 1UL, 1UL, 0UL, 1UL };
+       u64 idx;
+       int i;
+
+       /* Restore, Step 21:
+        *     Reset the following CH: [21, 23, 28, 29, 30]
+        */
+       for (i = 0; i < 5; i++) {
+               idx = ch_indices[i];
+               out_be64(&priv2->spu_chnlcntptr_RW, idx);
+               eieio();
+               out_be64(&priv2->spu_chnlcnt_RW, ch_counts[i]);
+               eieio();
+       }
+}
+
+static inline void setup_spu_status_part1(struct spu_state *csa,
+                                         struct spu *spu)
+{
+       u32 status_P = SPU_STATUS_STOPPED_BY_STOP;
+       u32 status_I = SPU_STATUS_INVALID_INSTR;
+       u32 status_H = SPU_STATUS_STOPPED_BY_HALT;
+       u32 status_S = SPU_STATUS_SINGLE_STEP;
+       u32 status_S_I = SPU_STATUS_SINGLE_STEP | SPU_STATUS_INVALID_INSTR;
+       u32 status_S_P = SPU_STATUS_SINGLE_STEP | SPU_STATUS_STOPPED_BY_STOP;
+       u32 status_P_H = SPU_STATUS_STOPPED_BY_HALT |SPU_STATUS_STOPPED_BY_STOP;
+       u32 status_P_I = SPU_STATUS_STOPPED_BY_STOP |SPU_STATUS_INVALID_INSTR;
+       u32 status_code;
+
+       /* Restore, Step 27:
+        *     If the CSA.SPU_Status[I,S,H,P]=1 then add the correct
+        *     instruction sequence to the end of the SPU based restore
+        *     code (after the "context restored" stop and signal) to
+        *     restore the correct SPU status.
+        *
+        *     NOTE: Rather than modifying the SPU executable, we
+        *     instead add a new 'stopped_status' field to the
+        *     LSCSA.  The SPU-side restore reads this field and
+        *     takes the appropriate action when exiting.
+        */
+
+       status_code =
+           (csa->prob.spu_status_R >> SPU_STOP_STATUS_SHIFT) & 0xFFFF;
+       if ((csa->prob.spu_status_R & status_P_I) == status_P_I) {
+
+               /* SPU_Status[P,I]=1 - Illegal Instruction followed
+                * by Stop and Signal instruction, followed by 'br -4'.
+                *
+                */
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_P_I;
+               csa->lscsa->stopped_status.slot[1] = status_code;
+
+       } else if ((csa->prob.spu_status_R & status_P_H) == status_P_H) {
+
+               /* SPU_Status[P,H]=1 - Halt Conditional, followed
+                * by Stop and Signal instruction, followed by
+                * 'br -4'.
+                */
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_P_H;
+               csa->lscsa->stopped_status.slot[1] = status_code;
+
+       } else if ((csa->prob.spu_status_R & status_S_P) == status_S_P) {
+
+               /* SPU_Status[S,P]=1 - Stop and Signal instruction
+                * followed by 'br -4'.
+                */
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_S_P;
+               csa->lscsa->stopped_status.slot[1] = status_code;
+
+       } else if ((csa->prob.spu_status_R & status_S_I) == status_S_I) {
+
+               /* SPU_Status[S,I]=1 - Illegal instruction followed
+                * by 'br -4'.
+                */
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_S_I;
+               csa->lscsa->stopped_status.slot[1] = status_code;
+
+       } else if ((csa->prob.spu_status_R & status_P) == status_P) {
+
+               /* SPU_Status[P]=1 - Stop and Signal instruction
+                * followed by 'br -4'.
+                */
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_P;
+               csa->lscsa->stopped_status.slot[1] = status_code;
+
+       } else if ((csa->prob.spu_status_R & status_H) == status_H) {
+
+               /* SPU_Status[H]=1 - Halt Conditional, followed
+                * by 'br -4'.
+                */
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_H;
+
+       } else if ((csa->prob.spu_status_R & status_S) == status_S) {
+
+               /* SPU_Status[S]=1 - Two nop instructions.
+                */
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_S;
+
+       } else if ((csa->prob.spu_status_R & status_I) == status_I) {
+
+               /* SPU_Status[I]=1 - Illegal instruction followed
+                * by 'br -4'.
+                */
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_I;
+
+       }
+}
+
+static inline void setup_spu_status_part2(struct spu_state *csa,
+                                         struct spu *spu)
+{
+       u32 mask;
+
+       /* Restore, Step 28:
+        *     If the CSA.SPU_Status[I,S,H,P,R]=0 then
+        *     add a 'br *' instruction to the end of
+        *     the SPU based restore code.
+        *
+        *     NOTE: Rather than modifying the SPU executable, we
+        *     instead add a new 'stopped_status' field to the
+        *     LSCSA.  The SPU-side restore reads this field and
+        *     takes the appropriate action when exiting.
+        */
+       mask = SPU_STATUS_INVALID_INSTR |
+           SPU_STATUS_SINGLE_STEP |
+           SPU_STATUS_STOPPED_BY_HALT |
+           SPU_STATUS_STOPPED_BY_STOP | SPU_STATUS_RUNNING;
+       if (!(csa->prob.spu_status_R & mask)) {
+               csa->lscsa->stopped_status.slot[0] = SPU_STOPPED_STATUS_R;
+       }
+}
+
+static inline void restore_mfc_rag(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Restore, Step 29:
+        *     Restore RA_GROUP_ID register and the
+        *     RA_ENABLE reigster from the CSA.
+        */
+       out_be64(&priv1->resource_allocation_groupID_RW,
+                csa->priv1.resource_allocation_groupID_RW);
+       out_be64(&priv1->resource_allocation_enable_RW,
+                csa->priv1.resource_allocation_enable_RW);
+}
+
+static inline void send_restore_code(struct spu_state *csa, struct spu *spu)
+{
+       unsigned long addr = (unsigned long)&spu_restore_code[0];
+       unsigned int ls_offset = 0x0;
+       unsigned int size = sizeof(spu_restore_code);
+       unsigned int tag = 0;
+       unsigned int rclass = 0;
+       unsigned int cmd = MFC_GETFS_CMD;
+
+       /* Restore, Step 37:
+        *     Issue MFC DMA command to copy context
+        *     restore code to local storage.
+        */
+       send_mfc_dma(spu, addr, ls_offset, size, tag, rclass, cmd);
+}
+
+static inline void setup_decr(struct spu_state *csa, struct spu *spu)
+{
+       /* Restore, Step 34:
+        *     If CSA.MFC_CNTL[Ds]=1 (decrementer was
+        *     running) then adjust decrementer, set
+        *     decrementer running status in LSCSA,
+        *     and set decrementer "wrapped" status
+        *     in LSCSA.
+        */
+       if (csa->priv2.mfc_control_RW & MFC_CNTL_DECREMENTER_RUNNING) {
+               cycles_t resume_time = get_cycles();
+               cycles_t delta_time = resume_time - csa->suspend_time;
+
+               csa->lscsa->decr.slot[0] = delta_time;
+       }
+}
+
+static inline void setup_ppu_mb(struct spu_state *csa, struct spu *spu)
+{
+       /* Restore, Step 35:
+        *     Copy the CSA.PU_MB data into the LSCSA.
+        */
+       csa->lscsa->ppu_mb.slot[0] = csa->prob.pu_mb_R;
+}
+
+static inline void setup_ppuint_mb(struct spu_state *csa, struct spu *spu)
+{
+       /* Restore, Step 36:
+        *     Copy the CSA.PUINT_MB data into the LSCSA.
+        */
+       csa->lscsa->ppuint_mb.slot[0] = csa->priv2.puint_mb_R;
+}
+
+static inline int check_restore_status(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       u32 complete;
+
+       /* Restore, Step 40:
+        *     If SPU_Status[P]=1 and SPU_Status[SC] = "success",
+        *     context restore succeeded, otherwise context restore
+        *     failed.
+        */
+       complete = ((SPU_RESTORE_COMPLETE << SPU_STOP_STATUS_SHIFT) |
+                   SPU_STATUS_STOPPED_BY_STOP);
+       return (in_be32(&prob->spu_status_R) != complete) ? 1 : 0;
+}
+
+static inline void restore_spu_privcntl(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 41:
+        *     Restore SPU_PrivCntl from the CSA.
+        */
+       out_be64(&priv2->spu_privcntl_RW, csa->priv2.spu_privcntl_RW);
+       eieio();
+}
+
+static inline void restore_status_part1(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       u32 mask;
+
+       /* Restore, Step 42:
+        *     If any CSA.SPU_Status[I,S,H,P]=1, then
+        *     restore the error or single step state.
+        */
+       mask = SPU_STATUS_INVALID_INSTR |
+           SPU_STATUS_SINGLE_STEP |
+           SPU_STATUS_STOPPED_BY_HALT | SPU_STATUS_STOPPED_BY_STOP;
+       if (csa->prob.spu_status_R & mask) {
+               out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_RUNNABLE);
+               eieio();
+               POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                               SPU_STATUS_RUNNING);
+       }
+}
+
+static inline void restore_status_part2(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       u32 mask;
+
+       /* Restore, Step 43:
+        *     If all CSA.SPU_Status[I,S,H,P,R]=0 then write
+        *     SPU_RunCntl[R0R1]='01', wait for SPU_Status[R]=1,
+        *     then write '00' to SPU_RunCntl[R0R1] and wait
+        *     for SPU_Status[R]=0.
+        */
+       mask = SPU_STATUS_INVALID_INSTR |
+           SPU_STATUS_SINGLE_STEP |
+           SPU_STATUS_STOPPED_BY_HALT |
+           SPU_STATUS_STOPPED_BY_STOP | SPU_STATUS_RUNNING;
+       if (!(csa->prob.spu_status_R & mask)) {
+               out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_RUNNABLE);
+               eieio();
+               POLL_WHILE_FALSE(in_be32(&prob->spu_status_R) &
+                                SPU_STATUS_RUNNING);
+               out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_STOP);
+               eieio();
+               POLL_WHILE_TRUE(in_be32(&prob->spu_status_R) &
+                               SPU_STATUS_RUNNING);
+       }
+}
+
+static inline void restore_ls_16kb(struct spu_state *csa, struct spu *spu)
+{
+       unsigned long addr = (unsigned long)&csa->lscsa->ls[0];
+       unsigned int ls_offset = 0x0;
+       unsigned int size = 16384;
+       unsigned int tag = 0;
+       unsigned int rclass = 0;
+       unsigned int cmd = MFC_GET_CMD;
+
+       /* Restore, Step 44:
+        *     Issue a DMA command to restore the first
+        *     16kb of local storage from CSA.
+        */
+       send_mfc_dma(spu, addr, ls_offset, size, tag, rclass, cmd);
+}
+
+static inline void clear_interrupts(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Restore, Step 49:
+        *     Write INT_MASK_class0 with value of 0.
+        *     Write INT_MASK_class1 with value of 0.
+        *     Write INT_MASK_class2 with value of 0.
+        *     Write INT_STAT_class0 with value of -1.
+        *     Write INT_STAT_class1 with value of -1.
+        *     Write INT_STAT_class2 with value of -1.
+        */
+       spin_lock_irq(&spu->register_lock);
+       out_be64(&priv1->int_mask_class0_RW, 0UL);
+       out_be64(&priv1->int_mask_class1_RW, 0UL);
+       out_be64(&priv1->int_mask_class2_RW, 0UL);
+       out_be64(&priv1->int_stat_class0_RW, ~(0UL));
+       out_be64(&priv1->int_stat_class1_RW, ~(0UL));
+       out_be64(&priv1->int_stat_class2_RW, ~(0UL));
+       spin_unlock_irq(&spu->register_lock);
+}
+
+static inline void restore_mfc_queues(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       int i;
+
+       /* Restore, Step 50:
+        *     If MFC_Cntl[Se]!=0 then restore
+        *     MFC command queues.
+        */
+       if ((csa->priv2.mfc_control_RW & MFC_CNTL_DMA_QUEUES_EMPTY_MASK) == 0) {
+               for (i = 0; i < 8; i++) {
+                       out_be64(&priv2->puq[i].mfc_cq_data0_RW,
+                                csa->priv2.puq[i].mfc_cq_data0_RW);
+                       out_be64(&priv2->puq[i].mfc_cq_data1_RW,
+                                csa->priv2.puq[i].mfc_cq_data1_RW);
+                       out_be64(&priv2->puq[i].mfc_cq_data2_RW,
+                                csa->priv2.puq[i].mfc_cq_data2_RW);
+                       out_be64(&priv2->puq[i].mfc_cq_data3_RW,
+                                csa->priv2.puq[i].mfc_cq_data3_RW);
+               }
+               for (i = 0; i < 16; i++) {
+                       out_be64(&priv2->spuq[i].mfc_cq_data0_RW,
+                                csa->priv2.spuq[i].mfc_cq_data0_RW);
+                       out_be64(&priv2->spuq[i].mfc_cq_data1_RW,
+                                csa->priv2.spuq[i].mfc_cq_data1_RW);
+                       out_be64(&priv2->spuq[i].mfc_cq_data2_RW,
+                                csa->priv2.spuq[i].mfc_cq_data2_RW);
+                       out_be64(&priv2->spuq[i].mfc_cq_data3_RW,
+                                csa->priv2.spuq[i].mfc_cq_data3_RW);
+               }
+       }
+       eieio();
+}
+
+static inline void restore_ppu_querymask(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Restore, Step 51:
+        *     Restore the PPU_QueryMask register from CSA.
+        */
+       out_be32(&prob->dma_querymask_RW, csa->prob.dma_querymask_RW);
+       eieio();
+}
+
+static inline void restore_ppu_querytype(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Restore, Step 52:
+        *     Restore the PPU_QueryType register from CSA.
+        */
+       out_be32(&prob->dma_querytype_RW, csa->prob.dma_querytype_RW);
+       eieio();
+}
+
+static inline void restore_mfc_csr_tsq(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 53:
+        *     Restore the MFC_CSR_TSQ register from CSA.
+        */
+       out_be64(&priv2->spu_tag_status_query_RW,
+                csa->priv2.spu_tag_status_query_RW);
+       eieio();
+}
+
+static inline void restore_mfc_csr_cmd(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 54:
+        *     Restore the MFC_CSR_CMD1 and MFC_CSR_CMD2
+        *     registers from CSA.
+        */
+       out_be64(&priv2->spu_cmd_buf1_RW, csa->priv2.spu_cmd_buf1_RW);
+       out_be64(&priv2->spu_cmd_buf2_RW, csa->priv2.spu_cmd_buf2_RW);
+       eieio();
+}
+
+static inline void restore_mfc_csr_ato(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 55:
+        *     Restore the MFC_CSR_ATO register from CSA.
+        */
+       out_be64(&priv2->spu_atomic_status_RW, csa->priv2.spu_atomic_status_RW);
+}
+
+static inline void restore_mfc_tclass_id(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Restore, Step 56:
+        *     Restore the MFC_TCLASS_ID register from CSA.
+        */
+       out_be64(&priv1->mfc_tclass_id_RW, csa->priv1.mfc_tclass_id_RW);
+       eieio();
+}
+
+static inline void set_llr_event(struct spu_state *csa, struct spu *spu)
+{
+       u64 ch0_cnt, ch0_data;
+       u64 ch1_data;
+
+       /* Restore, Step 57:
+        *    Set the Lock Line Reservation Lost Event by:
+        *      1. OR CSA.SPU_Event_Status with bit 21 (Lr) set to 1.
+        *      2. If CSA.SPU_Channel_0_Count=0 and
+        *         CSA.SPU_Wr_Event_Mask[Lr]=1 and
+        *         CSA.SPU_Event_Status[Lr]=0 then set
+        *         CSA.SPU_Event_Status_Count=1.
+        */
+       ch0_cnt = csa->spu_chnlcnt_RW[0];
+       ch0_data = csa->spu_chnldata_RW[0];
+       ch1_data = csa->spu_chnldata_RW[1];
+       csa->spu_chnldata_RW[0] |= MFC_LLR_LOST_EVENT;
+       if ((ch0_cnt == 0) && !(ch0_data & MFC_LLR_LOST_EVENT) &&
+           (ch1_data & MFC_LLR_LOST_EVENT)) {
+               csa->spu_chnlcnt_RW[0] = 1;
+       }
+}
+
+static inline void restore_decr_wrapped(struct spu_state *csa, struct spu *spu)
+{
+       /* Restore, Step 58:
+        *     If the status of the CSA software decrementer
+        *     "wrapped" flag is set, OR in a '1' to
+        *     CSA.SPU_Event_Status[Tm].
+        */
+       if (csa->lscsa->decr_status.slot[0] == 1) {
+               csa->spu_chnldata_RW[0] |= 0x20;
+       }
+       if ((csa->lscsa->decr_status.slot[0] == 1) &&
+           (csa->spu_chnlcnt_RW[0] == 0 &&
+            ((csa->spu_chnldata_RW[2] & 0x20) == 0x0) &&
+            ((csa->spu_chnldata_RW[0] & 0x20) != 0x1))) {
+               csa->spu_chnlcnt_RW[0] = 1;
+       }
+}
+
+static inline void restore_ch_part1(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       u64 idx, ch_indices[7] = { 0UL, 1UL, 3UL, 4UL, 24UL, 25UL, 27UL };
+       int i;
+
+       /* Restore, Step 59:
+        *     Restore the following CH: [0,1,3,4,24,25,27]
+        */
+       for (i = 0; i < 7; i++) {
+               idx = ch_indices[i];
+               out_be64(&priv2->spu_chnlcntptr_RW, idx);
+               eieio();
+               out_be64(&priv2->spu_chnldata_RW, csa->spu_chnldata_RW[idx]);
+               out_be64(&priv2->spu_chnlcnt_RW, csa->spu_chnlcnt_RW[idx]);
+               eieio();
+       }
+}
+
+static inline void restore_ch_part2(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       u64 ch_indices[3] = { 9UL, 21UL, 23UL };
+       u64 ch_counts[3] = { 1UL, 16UL, 1UL };
+       u64 idx;
+       int i;
+
+       /* Restore, Step 60:
+        *     Restore the following CH: [9,21,23].
+        */
+       ch_counts[0] = 1UL;
+       ch_counts[1] = csa->spu_chnlcnt_RW[21];
+       ch_counts[2] = 1UL;
+       for (i = 0; i < 3; i++) {
+               idx = ch_indices[i];
+               out_be64(&priv2->spu_chnlcntptr_RW, idx);
+               eieio();
+               out_be64(&priv2->spu_chnlcnt_RW, ch_counts[i]);
+               eieio();
+       }
+}
+
+static inline void restore_spu_lslr(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 61:
+        *     Restore the SPU_LSLR register from CSA.
+        */
+       out_be64(&priv2->spu_lslr_RW, csa->priv2.spu_lslr_RW);
+       eieio();
+}
+
+static inline void restore_spu_cfg(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 62:
+        *     Restore the SPU_Cfg register from CSA.
+        */
+       out_be64(&priv2->spu_cfg_RW, csa->priv2.spu_cfg_RW);
+       eieio();
+}
+
+static inline void restore_pm_trace(struct spu_state *csa, struct spu *spu)
+{
+       /* Restore, Step 63:
+        *     Restore PM_Trace_Tag_Wait_Mask from CSA.
+        *     Not performed by this implementation.
+        */
+}
+
+static inline void restore_spu_npc(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Restore, Step 64:
+        *     Restore SPU_NPC from CSA.
+        */
+       out_be32(&prob->spu_npc_RW, csa->prob.spu_npc_RW);
+       eieio();
+}
+
+static inline void restore_spu_mb(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       int i;
+
+       /* Restore, Step 65:
+        *     Restore MFC_RdSPU_MB from CSA.
+        */
+       out_be64(&priv2->spu_chnlcntptr_RW, 29UL);
+       eieio();
+       out_be64(&priv2->spu_chnlcnt_RW, csa->spu_chnlcnt_RW[29]);
+       for (i = 0; i < 4; i++) {
+               out_be64(&priv2->spu_chnldata_RW, csa->pu_mailbox_data[i]);
+       }
+       eieio();
+}
+
+static inline void check_ppu_mb_stat(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+       u32 dummy = 0;
+
+       /* Restore, Step 66:
+        *     If CSA.MB_Stat[P]=0 (mailbox empty) then
+        *     read from the PPU_MB register.
+        */
+       if ((csa->prob.mb_stat_R & 0xFF) == 0) {
+               dummy = in_be32(&prob->pu_mb_R);
+               eieio();
+       }
+}
+
+static inline void check_ppuint_mb_stat(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       u64 dummy = 0UL;
+
+       /* Restore, Step 66:
+        *     If CSA.MB_Stat[I]=0 (mailbox empty) then
+        *     read from the PPUINT_MB register.
+        */
+       if ((csa->prob.mb_stat_R & 0xFF0000) == 0) {
+               dummy = in_be64(&priv2->puint_mb_R);
+               eieio();
+               out_be64(&priv1->int_stat_class2_RW,
+                        CLASS2_ENABLE_MAILBOX_INTR);
+               eieio();
+       }
+}
+
+static inline void restore_mfc_slbs(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+       int i;
+
+       /* Restore, Step 68:
+        *     If MFC_SR1[R]='1', restore SLBs from CSA.
+        */
+       if (csa->priv1.mfc_sr1_RW & MFC_STATE1_RELOCATE_MASK) {
+               for (i = 0; i < 8; i++) {
+                       out_be64(&priv2->slb_index_W, i);
+                       eieio();
+                       out_be64(&priv2->slb_esid_RW, csa->slb_esid_RW[i]);
+                       out_be64(&priv2->slb_vsid_RW, csa->slb_vsid_RW[i]);
+                       eieio();
+               }
+               out_be64(&priv2->slb_index_W, csa->priv2.slb_index_W);
+               eieio();
+       }
+}
+
+static inline void restore_mfc_sr1(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Restore, Step 69:
+        *     Restore the MFC_SR1 register from CSA.
+        */
+       out_be64(&priv1->mfc_sr1_RW, csa->priv1.mfc_sr1_RW);
+       eieio();
+}
+
+static inline void restore_other_spu_access(struct spu_state *csa,
+                                           struct spu *spu)
+{
+       /* Restore, Step 70:
+        *     Restore other SPU mappings to this SPU. TBD.
+        */
+}
+
+static inline void restore_spu_runcntl(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_problem __iomem *prob = spu->problem;
+
+       /* Restore, Step 71:
+        *     If CSA.SPU_Status[R]=1 then write
+        *     SPU_RunCntl[R0R1]='01'.
+        */
+       if (csa->prob.spu_status_R & SPU_STATUS_RUNNING) {
+               out_be32(&prob->spu_runcntl_RW, SPU_RUNCNTL_RUNNABLE);
+               eieio();
+       }
+}
+
+static inline void restore_mfc_cntl(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+       /* Restore, Step 72:
+        *    Restore the MFC_CNTL register for the CSA.
+        */
+       out_be64(&priv2->mfc_control_RW, csa->priv2.mfc_control_RW);
+       eieio();
+}
+
+static inline void enable_user_access(struct spu_state *csa, struct spu *spu)
+{
+       /* Restore, Step 73:
+        *     Enable user-space access (if provided) to this
+        *     SPU by mapping the virtual pages assigned to
+        *     the SPU memory-mapped I/O (MMIO) for problem
+        *     state. TBD.
+        */
+}
+
+static inline void reset_switch_active(struct spu_state *csa, struct spu *spu)
+{
+       /* Restore, Step 74:
+        *     Reset the "context switch active" flag.
+        */
+       clear_bit(SPU_CONTEXT_SWITCH_ACTIVE_nr, &spu->flags);
+       mb();
+}
+
+static inline void reenable_interrupts(struct spu_state *csa, struct spu *spu)
+{
+       struct spu_priv1 __iomem *priv1 = spu->priv1;
+
+       /* Restore, Step 75:
+        *     Re-enable SPU interrupts.
+        */
+       spin_lock_irq(&spu->register_lock);
+       out_be64(&priv1->int_mask_class0_RW, csa->priv1.int_mask_class0_RW);
+       out_be64(&priv1->int_mask_class1_RW, csa->priv1.int_mask_class1_RW);
+       out_be64(&priv1->int_mask_class2_RW, csa->priv1.int_mask_class2_RW);
+       spin_unlock_irq(&spu->register_lock);
+}
+
+static int quiece_spu(struct spu_state *prev, struct spu *spu)
+{
+       /*
+        * Combined steps 2-18 of SPU context save sequence, which
+        * quiesce the SPU state (disable SPU execution, MFC command
+        * queues, decrementer, SPU interrupts, etc.).
+        *
+        * Returns      0 on success.
+        *              2 if failed step 2.
+        *              6 if failed step 6.
+        */
+
+       if (check_spu_isolate(prev, spu)) {     /* Step 2. */
+               return 2;
+       }
+       disable_interrupts(prev, spu);          /* Step 3. */
+       set_watchdog_timer(prev, spu);          /* Step 4. */
+       inhibit_user_access(prev, spu);         /* Step 5. */
+       if (check_spu_isolate(prev, spu)) {     /* Step 6. */
+               return 6;
+       }
+       set_switch_pending(prev, spu);          /* Step 7. */
+       save_mfc_cntl(prev, spu);               /* Step 8. */
+       save_spu_runcntl(prev, spu);            /* Step 9. */
+       save_mfc_sr1(prev, spu);                /* Step 10. */
+       save_spu_status(prev, spu);             /* Step 11. */
+       save_mfc_decr(prev, spu);               /* Step 12. */
+       halt_mfc_decr(prev, spu);               /* Step 13. */
+       save_timebase(prev, spu);               /* Step 14. */
+       remove_other_spu_access(prev, spu);     /* Step 15. */
+       do_mfc_mssync(prev, spu);               /* Step 16. */
+       issue_mfc_tlbie(prev, spu);             /* Step 17. */
+       handle_pending_interrupts(prev, spu);   /* Step 18. */
+
+       return 0;
+}
+
+static void save_csa(struct spu_state *prev, struct spu *spu)
+{
+       /*
+        * Combine steps 19-44 of SPU context save sequence, which
+        * save regions of the privileged & problem state areas.
+        */
+
+       save_mfc_queues(prev, spu);     /* Step 19. */
+       save_ppu_querymask(prev, spu);  /* Step 20. */
+       save_ppu_querytype(prev, spu);  /* Step 21. */
+       save_mfc_csr_tsq(prev, spu);    /* Step 22. */
+       save_mfc_csr_cmd(prev, spu);    /* Step 23. */
+       save_mfc_csr_ato(prev, spu);    /* Step 24. */
+       save_mfc_tclass_id(prev, spu);  /* Step 25. */
+       set_mfc_tclass_id(prev, spu);   /* Step 26. */
+       purge_mfc_queue(prev, spu);     /* Step 27. */
+       wait_purge_complete(prev, spu); /* Step 28. */
+       save_mfc_slbs(prev, spu);       /* Step 29. */
+       setup_mfc_sr1(prev, spu);       /* Step 30. */
+       save_spu_npc(prev, spu);        /* Step 31. */
+       save_spu_privcntl(prev, spu);   /* Step 32. */
+       reset_spu_privcntl(prev, spu);  /* Step 33. */
+       save_spu_lslr(prev, spu);       /* Step 34. */
+       reset_spu_lslr(prev, spu);      /* Step 35. */
+       save_spu_cfg(prev, spu);        /* Step 36. */
+       save_pm_trace(prev, spu);       /* Step 37. */
+       save_mfc_rag(prev, spu);        /* Step 38. */
+       save_ppu_mb_stat(prev, spu);    /* Step 39. */
+       save_ppu_mb(prev, spu);         /* Step 40. */
+       save_ppuint_mb(prev, spu);      /* Step 41. */
+       save_ch_part1(prev, spu);       /* Step 42. */
+       save_spu_mb(prev, spu);         /* Step 43. */
+       save_mfc_cmd(prev, spu);        /* Step 44. */
+       reset_ch(prev, spu);            /* Step 45. */
+}
+
+static void save_lscsa(struct spu_state *prev, struct spu *spu)
+{
+       /*
+        * Perform steps 46-57 of SPU context save sequence,
+        * which save regions of the local store and register
+        * file.
+        */
+
+       resume_mfc_queue(prev, spu);    /* Step 46. */
+       setup_mfc_slbs(prev, spu);      /* Step 47. */
+       set_switch_active(prev, spu);   /* Step 48. */
+       enable_interrupts(prev, spu);   /* Step 49. */
+       save_ls_16kb(prev, spu);        /* Step 50. */
+       set_spu_npc(prev, spu);         /* Step 51. */
+       set_signot1(prev, spu);         /* Step 52. */
+       set_signot2(prev, spu);         /* Step 53. */
+       send_save_code(prev, spu);      /* Step 54. */
+       set_ppu_querymask(prev, spu);   /* Step 55. */
+       wait_tag_complete(prev, spu);   /* Step 56. */
+       wait_spu_stopped(prev, spu);    /* Step 57. */
+}
+
+static void harvest(struct spu_state *prev, struct spu *spu)
+{
+       /*
+        * Perform steps 2-25 of SPU context restore sequence,
+        * which resets an SPU either after a failed save, or
+        * when using SPU for first time.
+        */
+
+       disable_interrupts(prev, spu);          /* Step 2.  */
+       inhibit_user_access(prev, spu);         /* Step 3.  */
+       terminate_spu_app(prev, spu);           /* Step 4.  */
+       set_switch_pending(prev, spu);          /* Step 5.  */
+       remove_other_spu_access(prev, spu);     /* Step 6.  */
+       suspend_mfc(prev, spu);                 /* Step 7.  */
+       wait_suspend_mfc_complete(prev, spu);   /* Step 8.  */
+       if (!suspend_spe(prev, spu))            /* Step 9.  */
+               clear_spu_status(prev, spu);    /* Step 10. */
+       do_mfc_mssync(prev, spu);               /* Step 11. */
+       issue_mfc_tlbie(prev, spu);             /* Step 12. */
+       handle_pending_interrupts(prev, spu);   /* Step 13. */
+       purge_mfc_queue(prev, spu);             /* Step 14. */
+       wait_purge_complete(prev, spu);         /* Step 15. */
+       reset_spu_privcntl(prev, spu);          /* Step 16. */
+       reset_spu_lslr(prev, spu);              /* Step 17. */
+       setup_mfc_sr1(prev, spu);               /* Step 18. */
+       invalidate_slbs(prev, spu);             /* Step 19. */
+       reset_ch_part1(prev, spu);              /* Step 20. */
+       reset_ch_part2(prev, spu);              /* Step 21. */
+       enable_interrupts(prev, spu);           /* Step 22. */
+       set_switch_active(prev, spu);           /* Step 23. */
+       set_mfc_tclass_id(prev, spu);           /* Step 24. */
+       resume_mfc_queue(prev, spu);            /* Step 25. */
+}
+
+static void restore_lscsa(struct spu_state *next, struct spu *spu)
+{
+       /*
+        * Perform steps 26-40 of SPU context restore sequence,
+        * which restores regions of the local store and register
+        * file.
+        */
+
+       set_watchdog_timer(next, spu);          /* Step 26. */
+       setup_spu_status_part1(next, spu);      /* Step 27. */
+       setup_spu_status_part2(next, spu);      /* Step 28. */
+       restore_mfc_rag(next, spu);             /* Step 29. */
+       setup_mfc_slbs(next, spu);              /* Step 30. */
+       set_spu_npc(next, spu);                 /* Step 31. */
+       set_signot1(next, spu);                 /* Step 32. */
+       set_signot2(next, spu);                 /* Step 33. */
+       setup_decr(next, spu);                  /* Step 34. */
+       setup_ppu_mb(next, spu);                /* Step 35. */
+       setup_ppuint_mb(next, spu);             /* Step 36. */
+       send_restore_code(next, spu);           /* Step 37. */
+       set_ppu_querymask(next, spu);           /* Step 38. */
+       wait_tag_complete(next, spu);           /* Step 39. */
+       wait_spu_stopped(next, spu);            /* Step 40. */
+}
+
+static void restore_csa(struct spu_state *next, struct spu *spu)
+{
+       /*
+        * Combine steps 41-76 of SPU context restore sequence, which
+        * restore regions of the privileged & problem state areas.
+        */
+
+       restore_spu_privcntl(next, spu);        /* Step 41. */
+       restore_status_part1(next, spu);        /* Step 42. */
+       restore_status_part2(next, spu);        /* Step 43. */
+       restore_ls_16kb(next, spu);             /* Step 44. */
+       wait_tag_complete(next, spu);           /* Step 45. */
+       suspend_mfc(next, spu);                 /* Step 46. */
+       wait_suspend_mfc_complete(next, spu);   /* Step 47. */
+       issue_mfc_tlbie(next, spu);             /* Step 48. */
+       clear_interrupts(next, spu);            /* Step 49. */
+       restore_mfc_queues(next, spu);          /* Step 50. */
+       restore_ppu_querymask(next, spu);       /* Step 51. */
+       restore_ppu_querytype(next, spu);       /* Step 52. */
+       restore_mfc_csr_tsq(next, spu);         /* Step 53. */
+       restore_mfc_csr_cmd(next, spu);         /* Step 54. */
+       restore_mfc_csr_ato(next, spu);         /* Step 55. */
+       restore_mfc_tclass_id(next, spu);       /* Step 56. */
+       set_llr_event(next, spu);               /* Step 57. */
+       restore_decr_wrapped(next, spu);        /* Step 58. */
+       restore_ch_part1(next, spu);            /* Step 59. */
+       restore_ch_part2(next, spu);            /* Step 60. */
+       restore_spu_lslr(next, spu);            /* Step 61. */
+       restore_spu_cfg(next, spu);             /* Step 62. */
+       restore_pm_trace(next, spu);            /* Step 63. */
+       restore_spu_npc(next, spu);             /* Step 64. */
+       restore_spu_mb(next, spu);              /* Step 65. */
+       check_ppu_mb_stat(next, spu);           /* Step 66. */
+       check_ppuint_mb_stat(next, spu);        /* Step 67. */
+       restore_mfc_slbs(next, spu);            /* Step 68. */
+       restore_mfc_sr1(next, spu);             /* Step 69. */
+       restore_other_spu_access(next, spu);    /* Step 70. */
+       restore_spu_runcntl(next, spu);         /* Step 71. */
+       restore_mfc_cntl(next, spu);            /* Step 72. */
+       enable_user_access(next, spu);          /* Step 73. */
+       reset_switch_active(next, spu);         /* Step 74. */
+       reenable_interrupts(next, spu);         /* Step 75. */
+}
+
+static int __do_spu_save(struct spu_state *prev, struct spu *spu)
+{
+       int rc;
+
+       /*
+        * SPU context save can be broken into three phases:
+        *
+        *     (a) quiesce [steps 2-16].
+        *     (b) save of CSA, performed by PPE [steps 17-42]
+        *     (c) save of LSCSA, mostly performed by SPU [steps 43-52].
+        *
+        * Returns      0 on success.
+        *              2,6 if failed to quiece SPU
+        *              53 if SPU-side of save failed.
+        */
+
+       rc = quiece_spu(prev, spu);             /* Steps 2-16. */
+       switch (rc) {
+       default:
+       case 2:
+       case 6:
+               harvest(prev, spu);
+               return rc;
+               break;
+       case 0:
+               break;
+       }
+       save_csa(prev, spu);                    /* Steps 17-43. */
+       save_lscsa(prev, spu);                  /* Steps 44-53. */
+       return check_save_status(prev, spu);    /* Step 54.     */
+}
+
+static int __do_spu_restore(struct spu_state *next, struct spu *spu)
+{
+       int rc;
+
+       /*
+        * SPU context restore can be broken into three phases:
+        *
+        *    (a) harvest (or reset) SPU [steps 2-24].
+        *    (b) restore LSCSA [steps 25-40], mostly performed by SPU.
+        *    (c) restore CSA [steps 41-76], performed by PPE.
+        *
+        * The 'harvest' step is not performed here, but rather
+        * as needed below.
+        */
+
+       restore_lscsa(next, spu);               /* Steps 24-39. */
+       rc = check_restore_status(next, spu);   /* Step 40.     */
+       switch (rc) {
+       default:
+               /* Failed. Return now. */
+               return rc;
+               break;
+       case 0:
+               /* Fall through to next step. */
+               break;
+       }
+       restore_csa(next, spu);
+
+       return 0;
+}
+
 /**
  * spu_save - SPU context save, with locking.
  * @prev: pointer to SPU context save area, to be saved.
  */
 int spu_save(struct spu_state *prev, struct spu *spu)
 {
-       /* XXX missing */
+       int rc;
 
-       return 0;
+       acquire_spu_lock(spu);          /* Step 1.     */
+       rc = __do_spu_save(prev, spu);  /* Steps 2-53. */
+       release_spu_lock(spu);
+
+       return rc;
 }
 
 /**
@@ -77,9 +2094,14 @@ int spu_save(struct spu_state *prev, struct spu *spu)
  */
 int spu_restore(struct spu_state *new, struct spu *spu)
 {
-       /* XXX missing */
+       int rc;
 
-       return 0;
+       acquire_spu_lock(spu);
+       harvest(NULL, spu);
+       rc = __do_spu_restore(new, spu);
+       release_spu_lock(spu);
+
+       return rc;
 }
 
 /**
@@ -93,9 +2115,17 @@ int spu_restore(struct spu_state *new, struct spu *spu)
  */
 int spu_switch(struct spu_state *prev, struct spu_state *new, struct spu *spu)
 {
-       /* XXX missing */
+       int rc;
 
-       return 0;
+       acquire_spu_lock(spu);          /* Save, Step 1.     */
+       rc = __do_spu_save(prev, spu);  /* Save, Steps 2-53. */
+       if (rc != 0) {
+               harvest(prev, spu);
+       }
+       rc = __do_spu_restore(new, spu);
+       release_spu_lock(spu);
+
+       return rc;
 }
 
 static void init_prob(struct spu_state *csa)
index d1d537d..989a068 100644 (file)
@@ -200,7 +200,6 @@ struct spu_priv2_collapsed {
        u64 spu_chnlcnt_RW;
        u64 spu_chnldata_RW;
        u64 spu_cfg_RW;
-       u64 spu_pm_trace_tag_status_RW;
        u64 spu_tag_status_query_RW;
        u64 spu_cmd_buf1_RW;
        u64 spu_cmd_buf2_RW;