2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
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.
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
15 #include <linux/export.h>
17 #include <asm/cacheflush.h>
18 #include <arch/icache.h>
19 #include <arch/spr_def.h>
22 void __flush_icache_range(unsigned long start, unsigned long end)
24 invalidate_icache((const void *)start, end - start, PAGE_SIZE);
28 /* Force a load instruction to issue. */
29 static inline void force_load(char *p)
35 * Flush and invalidate a VA range that is homed remotely on a single
36 * core (if "!hfh") or homed via hash-for-home (if "hfh"), waiting
37 * until the memory controller holds the flushed values.
39 void __attribute__((optimize("omit-frame-pointer")))
40 finv_buffer_remote(void *buffer, size_t size, int hfh)
43 size_t step_size, load_count;
46 * On TILEPro the striping granularity is a fixed 8KB; on
47 * TILE-Gx it is configurable, and we rely on the fact that
48 * the hypervisor always configures maximum striping, so that
49 * bits 9 and 10 of the PA are part of the stripe function, so
50 * every 512 bytes we hit a striping boundary.
54 const unsigned long STRIPE_WIDTH = 512;
56 const unsigned long STRIPE_WIDTH = 8192;
61 * On TILE-Gx, we must disable the dstream prefetcher before doing
62 * a cache flush; otherwise, we could end up with data in the cache
63 * that we don't want there. Note that normally we'd do an mf
64 * after the SPR write to disabling the prefetcher, but we do one
65 * below, before any further loads, so there's no need to do it
68 uint_reg_t old_dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
69 __insn_mtspr(SPR_DSTREAM_PF, 0);
73 * Flush and invalidate the buffer out of the local L1/L2
74 * and request the home cache to flush and invalidate as well.
76 __finv_buffer(buffer, size);
79 * Wait for the home cache to acknowledge that it has processed
80 * all the flush-and-invalidate requests. This does not mean
81 * that the flushed data has reached the memory controller yet,
82 * but it does mean the home cache is processing the flushes.
87 * Issue a load to the last cache line, which can't complete
88 * until all the previously-issued flushes to the same memory
89 * controller have also completed. If we weren't striping
90 * memory, that one load would be sufficient, but since we may
91 * be, we also need to back up to the last load issued to
92 * another memory controller, which would be the point where
93 * we crossed a "striping" boundary (the granularity of striping
94 * across memory controllers). Keep backing up and doing this
95 * until we are before the beginning of the buffer, or have
96 * hit all the controllers.
98 * If we are flushing a hash-for-home buffer, it's even worse.
99 * Each line may be homed on a different tile, and each tile
100 * may have up to four lines that are on different
101 * controllers. So as we walk backwards, we have to touch
102 * enough cache lines to satisfy these constraints. In
103 * practice this ends up being close enough to "load from
104 * every cache line on a full memory stripe on each
105 * controller" that we simply do that, to simplify the logic.
107 * On TILE-Gx the hash-for-home function is much more complex,
108 * with the upshot being we can't readily guarantee we have
109 * hit both entries in the 128-entry AMT that were hit by any
110 * load in the entire range, so we just re-load them all.
111 * With larger buffers, we may want to consider using a hypervisor
112 * trap to issue loads directly to each hash-for-home tile for
113 * each controller (doing it from Linux would trash the TLB).
116 step_size = L2_CACHE_BYTES;
118 load_count = (size + L2_CACHE_BYTES - 1) / L2_CACHE_BYTES;
120 load_count = (STRIPE_WIDTH / L2_CACHE_BYTES) *
121 (1 << CHIP_LOG_NUM_MSHIMS());
124 step_size = STRIPE_WIDTH;
125 load_count = (1 << CHIP_LOG_NUM_MSHIMS());
128 /* Load the last byte of the buffer. */
129 p = (char *)buffer + size - 1;
132 /* Bump down to the end of the previous stripe or cache line. */
134 p = (char *)((unsigned long)p | (step_size - 1));
136 /* Figure out how far back we need to go. */
137 base = p - (step_size * (load_count - 2));
138 if ((unsigned long)base < (unsigned long)buffer)
142 * Fire all the loads we need. The MAF only has eight entries
143 * so we can have at most eight outstanding loads, so we
144 * unroll by that amount.
147 for (; p >= base; p -= step_size)
151 * Repeat, but with finv's instead of loads, to get rid of the
152 * data we just loaded into our own cache and the old home L3.
153 * No need to unroll since finv's don't target a register.
154 * The finv's are guaranteed not to actually flush the data in
155 * the buffer back to their home, since we just read it, so the
156 * lines are clean in cache; we will only invalidate those lines.
158 p = (char *)buffer + size - 1;
161 p = (char *)((unsigned long)p | (step_size - 1));
162 for (; p >= base; p -= step_size)
165 /* Wait for these finv's (and thus the first finvs) to be done. */
169 /* Reenable the prefetcher. */
170 __insn_mtspr(SPR_DSTREAM_PF, old_dstream_pf);
173 EXPORT_SYMBOL_GPL(finv_buffer_remote);