/*
* ARC700 VIPT Cache Management
*
* Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
* published by the Free Software Foundation.
*
* vineetg: May 2011: for Non-aliasing VIPT D-cache following can be NOPs
* -flush_cache_dup_mm (fork)
* -likewise for flush_cache_mm (exit/execve)
* -likewise for flush_cache_range,flush_cache_page (munmap, exit, COW-break)
*
* vineetg: Apr 2011
* -Now that MMU can support larger pg sz (16K), the determiniation of
* aliasing shd not be based on assumption of 8k pg
*
* vineetg: Mar 2011
* -optimised version of flush_icache_range( ) for making I/D coherent
* when vaddr is available (agnostic of num of aliases)
*
* vineetg: Mar 2011
* -Added documentation about I-cache aliasing on ARC700 and the way it
* was handled up until MMU V2.
* -Spotted a three year old bug when killing the 4 aliases, which needs
* bottom 2 bits, so we need to do paddr | {0x00, 0x01, 0x02, 0x03}
* instead of paddr | {0x00, 0x01, 0x10, 0x11}
* (Rajesh you owe me one now)
*
* vineetg: Dec 2010
* -Off-by-one error when computing num_of_lines to flush
* This broke signal handling with bionic which uses synthetic sigret stub
*
* vineetg: Mar 2010
* -GCC can't generate ZOL for core cache flush loops.
* Conv them into iterations based as opposed to while (start < end) types
*
* Vineetg: July 2009
* -In I-cache flush routine we used to chk for aliasing for every line INV.
* Instead now we setup routines per cache geometry and invoke them
* via function pointers.
*
* Vineetg: Jan 2009
* -Cache Line flush routines used to flush an extra line beyond end addr
* because check was while (end >= start) instead of (end > start)
* =Some call sites had to work around by doing -1, -4 etc to end param
* =Some callers didnt care. This was spec bad in case of INV routines
* which would discard valid data (cause of the horrible ext2 bug
* in ARC IDE driver)
*
* vineetg: June 11th 2008: Fixed flush_icache_range( )
* -Since ARC700 caches are not coherent (I$ doesnt snoop D$) both need
* to be flushed, which it was not doing.
* -load_module( ) passes vmalloc addr (Kernel Virtual Addr) to the API,
* however ARC cache maintenance OPs require PHY addr. Thus need to do
* vmalloc_to_phy.
* -Also added optimisation there, that for range > PAGE SIZE we flush the
* entire cache in one shot rather than line by line. For e.g. a module
* with Code sz 600k, old code flushed 600k worth of cache (line-by-line),
* while cache is only 16 or 32k.
*/
#include <linux/module.h>
#include <linux/mm.h>
#include <linux/sched.h>
#include <linux/cache.h>
#include <linux/mmu_context.h>
#include <linux/syscalls.h>
#include <linux/uaccess.h>
#include <linux/pagemap.h>
#include <asm/cacheflush.h>
#include <asm/cachectl.h>
#include <asm/setup.h>
char *arc_cache_mumbojumbo(int c, char *buf, int len)
{
int n = 0;
#define PR_CACHE(p, cfg, str) \
if (!(p)->ver) \
n += scnprintf(buf + n, len - n, str"\t\t: N/A\n"); \
else \
n += scnprintf(buf + n, len - n, \
str"\t\t: %uK, %dway/set, %uB Line, %s%s%s\n", \
(p)->sz_k, (p)->assoc, (p)->line_len, \
(p)->vipt ? "VIPT" : "PIPT", \
(p)->alias ? " aliasing" : "", \
IS_ENABLED(cfg) ? "" : " (not used)");
PR_CACHE(&cpuinfo_arc700[c].icache, CONFIG_ARC_HAS_ICACHE, "I-Cache");
PR_CACHE(&cpuinfo_arc700[c].dcache, CONFIG_ARC_HAS_DCACHE, "D-Cache");
return buf;
}
/*
* Read the Cache Build Confuration Registers, Decode them and save into
* the cpuinfo structure for later use.
* No Validation done here, simply read/convert the BCRs
*/
void read_decode_cache_bcr(void)
{
struct cpuinfo_arc_cache *p_ic, *p_dc;
unsigned int cpu = smp_processor_id();
struct bcr_cache {
#ifdef CONFIG_CPU_BIG_ENDIAN
unsigned int pad:12, line_len:4, sz:4, config:4, ver:8;
#else
unsigned int ver:8, config:4, sz:4, line_len:4, pad:12;
#endif
} ibcr, dbcr;
p_ic = &cpuinfo_arc700[cpu].icache;
READ_BCR(ARC_REG_IC_BCR, ibcr);
if (!ibcr.ver)
goto dc_chk;
BUG_ON(ibcr.config != 3);
p_ic->assoc = 2; /* Fixed to 2w set assoc */
p_ic->line_len = 8 << ibcr.line_len;
p_ic->sz_k = 1 << (ibcr.sz - 1);
p_ic->ver = ibcr.ver;
p_ic->vipt = 1;
p_ic->alias = p_ic->sz_k/p_ic->assoc/TO_KB(PAGE_SIZE) > 1;
dc_chk:
p_dc = &cpuinfo_arc700[cpu].dcache;
READ_BCR(ARC_REG_DC_BCR, dbcr);
if (!dbcr.ver)
return;
BUG_ON(dbcr.config != 2);
p_dc->assoc = 4; /* Fixed to 4w set assoc */
p_dc->line_len = 16 << dbcr.line_len;
p_dc->sz_k = 1 << (dbcr.sz - 1);
p_dc->ver = dbcr.ver;
p_dc->vipt = 1;
p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
}
/*
* 1. Validate the Cache Geomtery (compile time config matches hardware)
* 2. If I-cache suffers from aliasing, setup work arounds (difft flush rtn)
* (aliasing D-cache configurations are not supported YET)
* 3. Enable the Caches, setup default flush mode for D-Cache
* 3. Calculate the SHMLBA used by user space
*/
void arc_cache_init(void)
{
unsigned int __maybe_unused cpu = smp_processor_id();
char str[256];
printk(arc_cache_mumbojumbo(0, str, sizeof(str)));
if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
if (!ic->ver)
panic("cache support enabled but non-existent cache\n");
if (ic->line_len != L1_CACHE_BYTES)
panic("ICache line [%d] != kernel Config [%d]",
ic->line_len, L1_CACHE_BYTES);
if (ic->ver != CONFIG_ARC_MMU_VER)
panic("Cache ver [%d] doesn't match MMU ver [%d]\n",
ic->ver, CONFIG_ARC_MMU_VER);
}
if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
int handled;
if (!dc->ver)
panic("cache support enabled but non-existent cache\n");
if (dc->line_len != L1_CACHE_BYTES)
panic("DCache line [%d] != kernel Config [%d]",
dc->line_len, L1_CACHE_BYTES);
/* check for D-Cache aliasing */
handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
if (dc->alias && !handled)
panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
else if (!dc->alias && handled)
panic("Don't need CONFIG_ARC_CACHE_VIPT_ALIASING\n");
}
}
#define OP_INV 0x1
#define OP_FLUSH 0x2
#define OP_FLUSH_N_INV 0x3
#define OP_INV_IC 0x4
/*
* Common Helper for Line Operations on {I,D}-Cache
*/
static inline void __cache_line_loop(unsigned long paddr, unsigned long vaddr,
unsigned long sz, const int cacheop)
{
unsigned int aux_cmd, aux_tag;
int num_lines;
const int full_page_op = __builtin_constant_p(sz) && sz == PAGE_SIZE;
if (cacheop == OP_INV_IC) {
aux_cmd = ARC_REG_IC_IVIL;
#if (CONFIG_ARC_MMU_VER > 2)
aux_tag = ARC_REG_IC_PTAG;
#endif
}
else {
/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
aux_cmd = cacheop & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
#if (CONFIG_ARC_MMU_VER > 2)
aux_tag = ARC_REG_DC_PTAG;
#endif
}
/* Ensure we properly floor/ceil the non-line aligned/sized requests
* and have @paddr - aligned to cache line and integral @num_lines.
* This however can be avoided for page sized since:
* -@paddr will be cache-line aligned already (being page aligned)
* -@sz will be integral multiple of line size (being page sized).
*/
if (!full_page_op) {
sz += paddr & ~CACHE_LINE_MASK;
paddr &= CACHE_LINE_MASK;
vaddr &= CACHE_LINE_MASK;
}
num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
#if (CONFIG_ARC_MMU_VER <= 2)
/* MMUv2 and before: paddr contains stuffed vaddrs bits */
paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
#else
/* if V-P const for loop, PTAG can be written once outside loop */
if (full_page_op)
write_aux_reg(aux_tag, paddr);
#endif
while (num_lines-- > 0) {
#if (CONFIG_ARC_MMU_VER > 2)
/* MMUv3, cache ops require paddr seperately */
if