1/*
2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
3 *
4 *   This program is free software; you can redistribute it and/or
5 *   modify it under the terms of the GNU General Public License
6 *   as published by the Free Software Foundation, version 2.
7 *
8 *   This program is distributed in the hope that it will be useful, but
9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11 *   NON INFRINGEMENT.  See the GNU General Public License for
12 *   more details.
13 */
14
15#include <asm/page.h>
16#include <asm/cacheflush.h>
17#include <arch/icache.h>
18#include <arch/spr_def.h>
19
20
21void __flush_icache_range(unsigned long start, unsigned long end)
22{
23	invalidate_icache((const void *)start, end - start, PAGE_SIZE);
24}
25
26
27/* Force a load instruction to issue. */
28static inline void force_load(char *p)
29{
30	*(volatile char *)p;
31}
32
33/*
34 * Flush and invalidate a VA range that is homed remotely on a single
35 * core (if "!hfh") or homed via hash-for-home (if "hfh"), waiting
36 * until the memory controller holds the flushed values.
37 */
38void finv_buffer_remote(void *buffer, size_t size, int hfh)
39{
40	char *p, *base;
41	size_t step_size, load_count;
42
43	/*
44	 * On TILEPro the striping granularity is a fixed 8KB; on
45	 * TILE-Gx it is configurable, and we rely on the fact that
46	 * the hypervisor always configures maximum striping, so that
47	 * bits 9 and 10 of the PA are part of the stripe function, so
48	 * every 512 bytes we hit a striping boundary.
49	 *
50	 */
51#ifdef __tilegx__
52	const unsigned long STRIPE_WIDTH = 512;
53#else
54	const unsigned long STRIPE_WIDTH = 8192;
55#endif
56
57#ifdef __tilegx__
58	/*
59	 * On TILE-Gx, we must disable the dstream prefetcher before doing
60	 * a cache flush; otherwise, we could end up with data in the cache
61	 * that we don't want there.  Note that normally we'd do an mf
62	 * after the SPR write to disabling the prefetcher, but we do one
63	 * below, before any further loads, so there's no need to do it
64	 * here.
65	 */
66	uint_reg_t old_dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
67	__insn_mtspr(SPR_DSTREAM_PF, 0);
68#endif
69
70	/*
71	 * Flush and invalidate the buffer out of the local L1/L2
72	 * and request the home cache to flush and invalidate as well.
73	 */
74	__finv_buffer(buffer, size);
75
76	/*
77	 * Wait for the home cache to acknowledge that it has processed
78	 * all the flush-and-invalidate requests.  This does not mean
79	 * that the flushed data has reached the memory controller yet,
80	 * but it does mean the home cache is processing the flushes.
81	 */
82	__insn_mf();
83
84	/*
85	 * Issue a load to the last cache line, which can't complete
86	 * until all the previously-issued flushes to the same memory
87	 * controller have also completed.  If we weren't striping
88	 * memory, that one load would be sufficient, but since we may
89	 * be, we also need to back up to the last load issued to
90	 * another memory controller, which would be the point where
91	 * we crossed a "striping" boundary (the granularity of striping
92	 * across memory controllers).  Keep backing up and doing this
93	 * until we are before the beginning of the buffer, or have
94	 * hit all the controllers.
95	 *
96	 * If we are flushing a hash-for-home buffer, it's even worse.
97	 * Each line may be homed on a different tile, and each tile
98	 * may have up to four lines that are on different
99	 * controllers.  So as we walk backwards, we have to touch
100	 * enough cache lines to satisfy these constraints.  In
101	 * practice this ends up being close enough to "load from
102	 * every cache line on a full memory stripe on each
103	 * controller" that we simply do that, to simplify the logic.
104	 *
105	 * On TILE-Gx the hash-for-home function is much more complex,
106	 * with the upshot being we can't readily guarantee we have
107	 * hit both entries in the 128-entry AMT that were hit by any
108	 * load in the entire range, so we just re-load them all.
109	 * With larger buffers, we may want to consider using a hypervisor
110	 * trap to issue loads directly to each hash-for-home tile for
111	 * each controller (doing it from Linux would trash the TLB).
112	 */
113	if (hfh) {
114		step_size = L2_CACHE_BYTES;
115#ifdef __tilegx__
116		load_count = (size + L2_CACHE_BYTES - 1) / L2_CACHE_BYTES;
117#else
118		load_count = (STRIPE_WIDTH / L2_CACHE_BYTES) *
119			      (1 << CHIP_LOG_NUM_MSHIMS());
120#endif
121	} else {
122		step_size = STRIPE_WIDTH;
123		load_count = (1 << CHIP_LOG_NUM_MSHIMS());
124	}
125
126	/* Load the last byte of the buffer. */
127	p = (char *)buffer + size - 1;
128	force_load(p);
129
130	/* Bump down to the end of the previous stripe or cache line. */
131	p -= step_size;
132	p = (char *)((unsigned long)p | (step_size - 1));
133
134	/* Figure out how far back we need to go. */
135	base = p - (step_size * (load_count - 2));
136	if ((unsigned long)base < (unsigned long)buffer)
137		base = buffer;
138
139	/*
140	 * Fire all the loads we need.  The MAF only has eight entries
141	 * so we can have at most eight outstanding loads, so we
142	 * unroll by that amount.
143	 */
144#pragma unroll 8
145	for (; p >= base; p -= step_size)
146		force_load(p);
147
148	/*
149	 * Repeat, but with inv's instead of loads, to get rid of the
150	 * data we just loaded into our own cache and the old home L3.
151	 * No need to unroll since inv's don't target a register.
152	 */
153	p = (char *)buffer + size - 1;
154	__insn_inv(p);
155	p -= step_size;
156	p = (char *)((unsigned long)p | (step_size - 1));
157	for (; p >= base; p -= step_size)
158		__insn_inv(p);
159
160	/* Wait for the load+inv's (and thus finvs) to have completed. */
161	__insn_mf();
162
163#ifdef __tilegx__
164	/* Reenable the prefetcher. */
165	__insn_mtspr(SPR_DSTREAM_PF, old_dstream_pf);
166#endif
167}
168