105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter/*
205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * spu_save.c
305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * (C) Copyright IBM Corp. 2005
505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * SPU-side context save sequence outlined in
705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * Synergistic Processor Element Book IV
805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * Author: Mark Nutter <mnutter@us.ibm.com>
1005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
1105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * This program is free software; you can redistribute it and/or modify
1205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * it under the terms of the GNU General Public License as published by
1305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * the Free Software Foundation; either version 2, or (at your option)
1405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * any later version.
1505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
1605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * This program is distributed in the hope that it will be useful,
1705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * but WITHOUT ANY WARRANTY; without even the implied warranty of
1805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
1905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * GNU General Public License for more details.
2005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
2105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * You should have received a copy of the GNU General Public License
2205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * along with this program; if not, write to the Free Software
2305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
2405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
2505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter */
2605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
2705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
2805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter#ifndef LS_SIZE
2905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter#define LS_SIZE                 0x40000	/* 256K (in bytes) */
3005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter#endif
3105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
3205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nuttertypedef unsigned int u32;
3305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nuttertypedef unsigned long long u64;
3405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
3505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter#include <spu_intrinsics.h>
3605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter#include <asm/spu_csa.h>
3705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter#include "spu_utils.h"
3805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
3905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void save_event_mask(void)
4005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
4105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int offset;
4205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
4305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 2:
4405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Read the SPU_RdEventMsk channel and save to the LSCSA.
4505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
4605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	offset = LSCSA_QW_OFFSET(event_mask);
477a896dc5f4a369193256653535aa7e2b521c611dSebastian Siewior	regs_spill[offset].slot[0] = spu_readch(SPU_RdEventMask);
4805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
4905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
5005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void save_tag_mask(void)
5105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
5205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int offset;
5305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
5405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 3:
5505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Read the SPU_RdTagMsk channel and save to the LSCSA.
5605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
5705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	offset = LSCSA_QW_OFFSET(tag_mask);
5805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	regs_spill[offset].slot[0] = spu_readch(MFC_RdTagMask);
5905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
6005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
6105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void save_upper_240kb(addr64 lscsa_ea)
6205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
6305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int ls = 16384;
6405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int list = (unsigned int)&dma_list[0];
6505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int size = sizeof(dma_list);
6605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int tag_id = 0;
6705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int cmd = 0x24;	/* PUTL */
6805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
6905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 7:
7005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Enqueue the PUTL command (tag 0) to the MFC SPU command
7105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    queue to transfer the remaining 240 kb of LS to CSA.
7205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
7305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_LSA, ls);
7405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_EAH, lscsa_ea.ui[0]);
7505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_EAL, list);
7605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_Size, size);
7705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_TagID, tag_id);
7805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_Cmd, cmd);
7905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
8005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
8105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void save_fpcr(void)
8205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
8305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	// vector unsigned int fpcr;
8405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int offset;
8505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
8605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 9:
8705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Issue the floating-point status and control register
8805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    read instruction, and save to the LSCSA.
8905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
9005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	offset = LSCSA_QW_OFFSET(fpcr);
9105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	regs_spill[offset].v = spu_mffpscr();
9205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
9305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
9405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void save_decr(void)
9505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
9605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int offset;
9705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
9805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 10:
9905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Read and save the SPU_RdDec channel data to
10005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    the LSCSA.
10105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
10205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	offset = LSCSA_QW_OFFSET(decr);
10305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	regs_spill[offset].slot[0] = spu_readch(SPU_RdDec);
10405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
10505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
10605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void save_srr0(void)
10705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
10805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int offset;
10905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
11005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 11:
11105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Read and save the SPU_WSRR0 channel data to
11205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    the LSCSA.
11305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
11405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	offset = LSCSA_QW_OFFSET(srr0);
11505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	regs_spill[offset].slot[0] = spu_readch(SPU_RdSRR0);
11605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
11705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
11805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void spill_regs_to_mem(addr64 lscsa_ea)
11905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
12005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int ls = (unsigned int)&regs_spill[0];
12105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int size = sizeof(regs_spill);
12205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int tag_id = 0;
12305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int cmd = 0x20;	/* PUT */
12405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
12505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 13:
12605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Enqueue a PUT command (tag 0) to send the LSCSA
12705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    to the CSA.
12805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
12905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_LSA, ls);
13005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_EAH, lscsa_ea.ui[0]);
13105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_EAL, lscsa_ea.ui[1]);
13205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_Size, size);
13305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_TagID, tag_id);
13405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_Cmd, cmd);
13505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
13605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
13705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void enqueue_sync(addr64 lscsa_ea)
13805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
13905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int tag_id = 0;
14005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	unsigned int cmd = 0xCC;
14105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
14205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 14:
14305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Enqueue an MFC_SYNC command (tag 0).
14405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
14505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_TagID, tag_id);
14605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_writech(MFC_Cmd, cmd);
14705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
14805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
14905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterstatic inline void save_complete(void)
15005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
15105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Save, Step 18:
15205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    Issue a stop-and-signal instruction indicating
15305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    "save complete".  Note: This function will not
15405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 *    return!!
15505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	 */
15605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spu_stop(SPU_SAVE_COMPLETE);
15705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
15805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
15905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter/**
16005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * main - entry point for SPU-side context save.
16105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
16205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter * This code deviates from the documented sequence as follows:
16305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *
16405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *      1. The EA for LSCSA is passed from PPE in the
16505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *         signal notification channels.
16605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter *      2. All 128 registers are saved by crt0.o.
16705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter */
16805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutterint main()
16905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter{
17005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	addr64 lscsa_ea;
17105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
17205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	lscsa_ea.ui[0] = spu_readch(SPU_RdSigNotify1);
17305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	lscsa_ea.ui[1] = spu_readch(SPU_RdSigNotify2);
17405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
17505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Step 1: done by exit(). */
17605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	save_event_mask();	/* Step 2.  */
17705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	save_tag_mask();	/* Step 3.  */
17805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	set_event_mask();	/* Step 4.  */
17905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	set_tag_mask();		/* Step 5.  */
18005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	build_dma_list(lscsa_ea);	/* Step 6.  */
18105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	save_upper_240kb(lscsa_ea);	/* Step 7.  */
18205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	/* Step 8: done by exit(). */
18305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	save_fpcr();		/* Step 9.  */
18405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	save_decr();		/* Step 10. */
18505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	save_srr0();		/* Step 11. */
18605b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	enqueue_putllc(lscsa_ea);	/* Step 12. */
18705b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	spill_regs_to_mem(lscsa_ea);	/* Step 13. */
18805b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	enqueue_sync(lscsa_ea);	/* Step 14. */
18905b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	set_tag_update();	/* Step 15. */
19005b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	read_tag_status();	/* Step 16. */
19105b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	read_llar_status();	/* Step 17. */
19205b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	save_complete();	/* Step 18. */
19305b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter
19405b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter	return 0;
19505b841174c289ca62a6b42d883b8791d9ac3a4bdMark Nutter}
196