linux/arch/arc/mm/cache_arc700.c

749 lines
22 KiB
C
Raw Normal View History

/*
* 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 cpu_id, char *buf, int len)
{
int n = 0;
unsigned int c = smp_processor_id();
#define PR_CACHE(p, enb, 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) VIPT, %dway set-asc, %ub Line %s\n", \
TO_KB((p)->sz), (p)->assoc, (p)->line_len, \
enb ? "" : "DISABLED (kernel-build)"); \
}
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 __cpuinit read_decode_cache_bcr(void)
{
struct bcr_cache ibcr, dbcr;
struct cpuinfo_arc_cache *p_ic, *p_dc;
unsigned int cpu = smp_processor_id();
p_ic = &cpuinfo_arc700[cpu].icache;
READ_BCR(ARC_REG_IC_BCR, ibcr);
if (ibcr.config == 0x3)
p_ic->assoc = 2;
p_ic->line_len = 8 << ibcr.line_len;
p_ic->sz = 0x200 << ibcr.sz;
p_ic->ver = ibcr.ver;
p_dc = &cpuinfo_arc700[cpu].dcache;
READ_BCR(ARC_REG_DC_BCR, dbcr);
if (dbcr.config == 0x2)
p_dc->assoc = 4;
p_dc->line_len = 16 << dbcr.line_len;
p_dc->sz = 0x200 << dbcr.sz;
p_dc->ver = dbcr.ver;
}
/*
* 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 __cpuinit arc_cache_init(void)
{
unsigned int temp;
unsigned int cpu = smp_processor_id();
struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
int way_pg_ratio = way_pg_ratio;
int dcache_does_alias;
char str[256];
printk(arc_cache_mumbojumbo(0, str, sizeof(str)));
if (!ic->ver)
goto chk_dc;
#ifdef CONFIG_ARC_HAS_ICACHE
/* 1. Confirm some of I-cache params which Linux assumes */
if ((ic->assoc != ARC_ICACHE_WAYS) ||
(ic->line_len != ARC_ICACHE_LINE_LEN)) {
panic("Cache H/W doesn't match kernel Config");
}
#if (CONFIG_ARC_MMU_VER > 2)
if (ic->ver != 3) {
if (running_on_hw)
panic("Cache ver doesn't match MMU ver\n");
/* For ISS - suggest the toggles to use */
pr_err("Use -prop=icache_version=3,-prop=dcache_version=3\n");
}
#endif
#endif
/* Enable/disable I-Cache */
temp = read_aux_reg(ARC_REG_IC_CTRL);
#ifdef CONFIG_ARC_HAS_ICACHE
temp &= ~IC_CTRL_CACHE_DISABLE;
#else
temp |= IC_CTRL_CACHE_DISABLE;
#endif
write_aux_reg(ARC_REG_IC_CTRL, temp);
chk_dc:
if (!dc->ver)
return;
#ifdef CONFIG_ARC_HAS_DCACHE
if ((dc->assoc != ARC_DCACHE_WAYS) ||
(dc->line_len != ARC_DCACHE_LINE_LEN)) {
panic("Cache H/W doesn't match kernel Config");
}
dcache_does_alias = (dc->sz / ARC_DCACHE_WAYS) > PAGE_SIZE;
/* check for D-Cache aliasing */
if (dcache_does_alias && !cache_is_vipt_aliasing())
panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
else if (!dcache_does_alias && cache_is_vipt_aliasing())
panic("Don't need CONFIG_ARC_CACHE_VIPT_ALIASING\n");
#endif
/* Set the default Invalidate Mode to "simpy discard dirty lines"
* as this is more frequent then flush before invalidate
* Ofcourse we toggle this default behviour when desired
*/
temp = read_aux_reg(ARC_REG_DC_CTRL);
temp &= ~DC_CTRL_INV_MODE_FLUSH;
#ifdef CONFIG_ARC_HAS_DCACHE
/* Enable D-Cache: Clear Bit 0 */
write_aux_reg(ARC_REG_DC_CTRL, temp & ~IC_CTRL_CACHE_DISABLE);
#else
/* Flush D cache */
write_aux_reg(ARC_REG_DC_FLSH, 0x1);
/* Disable D cache */
write_aux_reg(ARC_REG_DC_CTRL, temp | IC_CTRL_CACHE_DISABLE);
#endif
return;
}
#define OP_INV 0x1
#define OP_FLUSH 0x2
#define OP_FLUSH_N_INV 0x3
#ifdef CONFIG_ARC_HAS_DCACHE
/***************************************************************
* Machine specific helpers for Entire D-Cache or Per Line ops
*/
static inline void wait_for_flush(void)
{
while (read_aux_reg(ARC_REG_DC_CTRL) & DC_CTRL_FLUSH_STATUS)
;
}
/*
* Operation on Entire D-Cache
* @cacheop = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
* Note that constant propagation ensures all the checks are gone
* in generated code
*/
static inline void __dc_entire_op(const int cacheop)
{
unsigned long flags, tmp = tmp;
int aux;
local_irq_save(flags);
if (cacheop == OP_FLUSH_N_INV) {
/* Dcache provides 2 cmd: FLUSH or INV
* INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
* flush-n-inv is achieved by INV cmd but with IM=1
* Default INV sub-mode is DISCARD, which needs to be toggled
*/
tmp = read_aux_reg(ARC_REG_DC_CTRL);
write_aux_reg(ARC_REG_DC_CTRL, tmp | DC_CTRL_INV_MODE_FLUSH);
}
if (cacheop & OP_INV) /* Inv or flush-n-inv use same cmd reg */
aux = ARC_REG_DC_IVDC;
else
aux = ARC_REG_DC_FLSH;
write_aux_reg(aux, 0x1);
if (cacheop & OP_FLUSH) /* flush / flush-n-inv both wait */
wait_for_flush();
/* Switch back the DISCARD ONLY Invalidate mode */
if (cacheop == OP_FLUSH_N_INV)
write_aux_reg(ARC_REG_DC_CTRL, tmp & ~DC_CTRL_INV_MODE_FLUSH);
local_irq_restore(flags);
}
/*
* Per Line Operation on D-Cache
* Doesn't deal with type-of-op/IRQ-disabling/waiting-for-flush-to-complete
* It's sole purpose is to help gcc generate ZOL
* (aliasing VIPT dcache flushing needs both vaddr and paddr)
*/
static inline void __dc_line_loop(unsigned long paddr, unsigned long vaddr,
unsigned long sz, const int aux_reg)
{
int num_lines;
/* 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 (!(__builtin_constant_p(sz) && sz == PAGE_SIZE)) {
sz += paddr & ~DCACHE_LINE_MASK;
paddr &= DCACHE_LINE_MASK;
vaddr &= DCACHE_LINE_MASK;
}
num_lines = DIV_ROUND_UP(sz, ARC_DCACHE_LINE_LEN);
#if (CONFIG_ARC_MMU_VER <= 2)
paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
#endif
while (num_lines-- > 0) {
#if (CONFIG_ARC_MMU_VER > 2)
/*
* Just as for I$, in MMU v3, D$ ops also require
* "tag" bits in DC_PTAG, "index" bits in FLDL,IVDL ops
*/
write_aux_reg(ARC_REG_DC_PTAG, paddr);
write_aux_reg(aux_reg, vaddr);
vaddr += ARC_DCACHE_LINE_LEN;
#else
/* paddr contains stuffed vaddrs bits */
write_aux_reg(aux_reg, paddr);
#endif
paddr += ARC_DCACHE_LINE_LEN;
}
}
/* For kernel mappings cache operation: index is same as paddr */
#define __dc_line_op_k(p, sz, op) __dc_line_op(p, p, sz, op)
/*
* D-Cache : Per Line INV (discard or wback+discard) or FLUSH (wback)
*/
static inline void __dc_line_op(unsigned long paddr, unsigned long vaddr,
unsigned long sz, const int cacheop)
{
unsigned long flags, tmp = tmp;
int aux;
local_irq_save(flags);
if (cacheop == OP_FLUSH_N_INV) {
/*
* Dcache provides 2 cmd: FLUSH or INV
* INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
* flush-n-inv is achieved by INV cmd but with IM=1
* Default INV sub-mode is DISCARD, which needs to be toggled
*/
tmp = read_aux_reg(ARC_REG_DC_CTRL);
write_aux_reg(ARC_REG_DC_CTRL, tmp | DC_CTRL_INV_MODE_FLUSH);
}
if (cacheop & OP_INV) /* Inv / flush-n-inv use same cmd reg */
aux = ARC_REG_DC_IVDL;
else
aux = ARC_REG_DC_FLDL;
__dc_line_loop(paddr, vaddr, sz, aux);
if (cacheop & OP_FLUSH) /* flush / flush-n-inv both wait */
wait_for_flush();
/* Switch back the DISCARD ONLY Invalidate mode */
if (cacheop == OP_FLUSH_N_INV)
write_aux_reg(ARC_REG_DC_CTRL, tmp & ~DC_CTRL_INV_MODE_FLUSH);
local_irq_restore(flags);
}
#else
#define __dc_entire_op(cacheop)
#define __dc_line_op(paddr, vaddr, sz, cacheop)
#define __dc_line_op_k(paddr, sz, cacheop)
#endif /* CONFIG_ARC_HAS_DCACHE */
#ifdef CONFIG_ARC_HAS_ICACHE
/*
* I-Cache Aliasing in ARC700 VIPT caches
*
* ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
* The orig Cache Management Module "CDU" only required paddr to invalidate a
* certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
* Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
* the exact same line.
*
* However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
* paddr alone could not be used to correctly index the cache.
*
* ------------------
* MMU v1/v2 (Fixed Page Size 8k)
* ------------------
* The solution was to provide CDU with these additonal vaddr bits. These
* would be bits [x:13], x would depend on cache-geometry, 13 comes from
* standard page size of 8k.
* H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
* of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
* orig 5 bits of paddr were anyways ignored by CDU line ops, as they
* represent the offset within cache-line. The adv of using this "clumsy"
* interface for additional info was no new reg was needed in CDU programming
* model.
*
* 17:13 represented the max num of bits passable, actual bits needed were
* fewer, based on the num-of-aliases possible.
* -for 2 alias possibility, only bit 13 needed (32K cache)
* -for 4 alias possibility, bits 14:13 needed (64K cache)
*
* ------------------
* MMU v3
* ------------------
* This ver of MMU supports variable page sizes (1k-16k): although Linux will
* only support 8k (default), 16k and 4k.
* However from hardware perspective, smaller page sizes aggrevate aliasing
* meaning more vaddr bits needed to disambiguate the cache-line-op ;
* the existing scheme of piggybacking won't work for certain configurations.
* Two new registers IC_PTAG and DC_PTAG inttoduced.
* "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
*/
/***********************************************************
* Machine specific helper for per line I-Cache invalidate.
*/
static void __ic_line_inv_vaddr(unsigned long paddr, unsigned long vaddr,
unsigned long sz)
{
unsigned long flags;
int num_lines;
/*
* Ensure we properly floor/ceil the non-line aligned/sized requests:
* However page sized flushes can be compile time optimised.
* -@paddr will be cache-line aligned already (being page aligned)
* -@sz will be integral multiple of line size (being page sized).
*/
if (!(__builtin_constant_p(sz) && sz == PAGE_SIZE)) {
sz += paddr & ~ICACHE_LINE_MASK;
paddr &= ICACHE_LINE_MASK;
vaddr &= ICACHE_LINE_MASK;
}
num_lines = DIV_ROUND_UP(sz, ARC_ICACHE_LINE_LEN);
#if (CONFIG_ARC_MMU_VER <= 2)
/* bits 17:13 of vaddr go as bits 4:0 of paddr */
paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
#endif
local_irq_save(flags);
while (num_lines-- > 0) {
#if (CONFIG_ARC_MMU_VER > 2)
/* tag comes from phy addr */
write_aux_reg(ARC_REG_IC_PTAG, paddr);
/* index bits come from vaddr */
write_aux_reg(ARC_REG_IC_IVIL, vaddr);
vaddr += ARC_ICACHE_LINE_LEN;
#else
/* paddr contains stuffed vaddrs bits */
write_aux_reg(ARC_REG_IC_IVIL, paddr);
#endif
paddr += ARC_ICACHE_LINE_LEN;
}
local_irq_restore(flags);
}
#else
#define __ic_line_inv_vaddr(pstart, vstart, sz)
#endif /* CONFIG_ARC_HAS_ICACHE */
/***********************************************************
* Exported APIs
*/
/*
* Handle cache congruency of kernel and userspace mappings of page when kernel
* writes-to/reads-from
*
* The idea is to defer flushing of kernel mapping after a WRITE, possible if:
* -dcache is NOT aliasing, hence any U/K-mappings of page are congruent
* -U-mapping doesn't exist yet for page (finalised in update_mmu_cache)
* -In SMP, if hardware caches are coherent
*
* There's a corollary case, where kernel READs from a userspace mapped page.
* If the U-mapping is not congruent to to K-mapping, former needs flushing.
*/
void flush_dcache_page(struct page *page)
{
struct address_space *mapping;
if (!cache_is_vipt_aliasing()) {
set_bit(PG_arch_1, &page->flags);
return;
}
/* don't handle anon pages here */
mapping = page_mapping(page);
if (!mapping)
return;
/*
* pagecache page, file not yet mapped to userspace
* Make a note that K-mapping is dirty
*/
if (!mapping_mapped(mapping)) {
set_bit(PG_arch_1, &page->flags);
} else if (page_mapped(page)) {
/* kernel reading from page with U-mapping */
void *paddr = page_address(page);
unsigned long vaddr = page->index << PAGE_CACHE_SHIFT;
if (addr_not_cache_congruent(paddr, vaddr))
__flush_dcache_page(paddr, vaddr);
}
}
EXPORT_SYMBOL(flush_dcache_page);
void dma_cache_wback_inv(unsigned long start, unsigned long sz)
{
__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
}
EXPORT_SYMBOL(dma_cache_wback_inv);
void dma_cache_inv(unsigned long start, unsigned long sz)
{
__dc_line_op_k(start, sz, OP_INV);
}
EXPORT_SYMBOL(dma_cache_inv);
void dma_cache_wback(unsigned long start, unsigned long sz)
{
__dc_line_op_k(start, sz, OP_FLUSH);
}
EXPORT_SYMBOL(dma_cache_wback);
/*
* This is API for making I/D Caches consistent when modifying
* kernel code (loadable modules, kprobes, kgdb...)
* This is called on insmod, with kernel virtual address for CODE of
* the module. ARC cache maintenance ops require PHY address thus we
* need to convert vmalloc addr to PHY addr
*/
void flush_icache_range(unsigned long kstart, unsigned long kend)
{
unsigned int tot_sz, off, sz;
unsigned long phy, pfn;
/* printk("Kernel Cache Cohenercy: %lx to %lx\n",kstart, kend); */
/* This is not the right API for user virtual address */
if (kstart < TASK_SIZE) {
BUG_ON("Flush icache range for user virtual addr space");
return;
}
/* Shortcut for bigger flush ranges.
* Here we don't care if this was kernel virtual or phy addr
*/
tot_sz = kend - kstart;
if (tot_sz > PAGE_SIZE) {
flush_cache_all();
return;
}
/* Case: Kernel Phy addr (0x8000_0000 onwards) */
if (likely(kstart > PAGE_OFFSET)) {
/*
* The 2nd arg despite being paddr will be used to index icache
* This is OK since no alternate virtual mappings will exist
* given the callers for this case: kprobe/kgdb in built-in
* kernel code only.
*/
__sync_icache_dcache(kstart, kstart, kend - kstart);
return;
}
/*
* Case: Kernel Vaddr (0x7000_0000 to 0x7fff_ffff)
* (1) ARC Cache Maintenance ops only take Phy addr, hence special
* handling of kernel vaddr.
*
* (2) Despite @tot_sz being < PAGE_SIZE (bigger cases handled already),
* it still needs to handle a 2 page scenario, where the range
* straddles across 2 virtual pages and hence need for loop
*/
while (tot_sz > 0) {
off = kstart % PAGE_SIZE;
pfn = vmalloc_to_pfn((void *)kstart);
phy = (pfn << PAGE_SHIFT) + off;
sz = min_t(unsigned int, tot_sz, PAGE_SIZE - off);
__sync_icache_dcache(phy, kstart, sz);
kstart += sz;
tot_sz -= sz;
}
}
/*
* General purpose helper to make I and D cache lines consistent.
* @paddr is phy addr of region
* @vaddr is typically user or kernel vaddr (vmalloc)
* Howver in one instance, flush_icache_range() by kprobe (for a breakpt in
* builtin kernel code) @vaddr will be paddr only, meaning CDU operation will
* use a paddr to index the cache (despite VIPT). This is fine since since a
* built-in kernel page will not have any virtual mappings (not even kernel)
* kprobe on loadable module is different as it will have kvaddr.
*/
void __sync_icache_dcache(unsigned long paddr, unsigned long vaddr, int len)
{
unsigned long flags;
local_irq_save(flags);
__ic_line_inv_vaddr(paddr, vaddr, len);
__dc_line_op(paddr, vaddr, len, OP_FLUSH);
local_irq_restore(flags);
}
/* wrapper to compile time eliminate alignment checks in flush loop */
void __inv_icache_page(unsigned long paddr, unsigned long vaddr)
{
__ic_line_inv_vaddr(paddr, vaddr, PAGE_SIZE);
}
/*
* wrapper to clearout kernel or userspace mappings of a page
* For kernel mappings @vaddr == @paddr
*/
void ___flush_dcache_page(unsigned long paddr, unsigned long vaddr)
ARC: [mm] Lazy D-cache flush (non aliasing VIPT) flush_dcache_page( ) is MM hook to ensure that a page has consistent views between kernel and userspace. Thus it is called when * kernel writes to a page which at some later point could get mapped to userspace (so kernel mapping needs to be flushed-n-inv) * kernel is about to read from a page with possible userspace mappings (so userspace mappings needs to be made coherent with kernel ones) However for Non aliasing VIPT dcache, any userspace mapping will always be congruent to kernel mapping. Thus d-cache need need not be flushed at all (or delayed indefinitely). The only reason it does need to be flushed is when mapping code pages. Since icache doesn't snoop dcache, those dirty dcache lines need to be written back to memory and icache line invalidated so that icache lines fetch will get the right data. Decent gains on LMBench fork/exec/sh and File I/O micro-benchmarks. (1) FPGA @ 80 MHZ Processor, Processes - times in microseconds - smaller is better ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- 3.9-rc6-a Linux 3.9.0-r 80 4.79 8.72 66.7 116. 239. 8.39 30.4 4798 14.K 34.K 3.9-rc6-b Linux 3.9.0-r 80 4.79 8.62 65.4 111. 239. 8.35 29.0 3995 12.K 30.K 3.9-rc7-c Linux 3.9.0-r 80 4.79 9.00 66.1 106. 239. 8.61 30.4 2858 10.K 24.K ^^^^ ^^^^ ^^^ File & VM system latencies in microseconds - smaller is better ------------------------------------------------------------------------------- Host OS 0K File 10K File Mmap Prot Page 100fd Create Delete Create Delete Latency Fault Fault selct --------- ------------- ------ ------ ------ ------ ------- ----- ------- ----- 3.9-rc6-a Linux 3.9.0-r 317.8 204.2 1122.3 375.1 3522.0 4.288 20.7 126.8 3.9-rc6-b Linux 3.9.0-r 298.7 223.0 1141.6 367.8 3531.0 4.866 20.9 126.4 3.9-rc7-c Linux 3.9.0-r 278.4 179.2 862.1 339.3 3705.0 3.223 20.3 126.6 ^^^^^ ^^^^^ ^^^^^ ^^^^ (2) Customer Silicon @ 500 MHz (166 MHz mem) ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- abilis-ba Linux 3.9.0-r 497 0.71 1.38 4.58 12.0 35.5 1.40 3.89 2070 5525 13.K abilis-ca Linux 3.9.0-r 497 0.71 1.40 4.61 11.8 35.6 1.37 3.92 1411 4317 10.K ^^^^ ^^^^ ^^^ Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
2013-04-16 16:40:48 +08:00
{
__dc_line_op(paddr, vaddr & PAGE_MASK, PAGE_SIZE, OP_FLUSH_N_INV);
ARC: [mm] Lazy D-cache flush (non aliasing VIPT) flush_dcache_page( ) is MM hook to ensure that a page has consistent views between kernel and userspace. Thus it is called when * kernel writes to a page which at some later point could get mapped to userspace (so kernel mapping needs to be flushed-n-inv) * kernel is about to read from a page with possible userspace mappings (so userspace mappings needs to be made coherent with kernel ones) However for Non aliasing VIPT dcache, any userspace mapping will always be congruent to kernel mapping. Thus d-cache need need not be flushed at all (or delayed indefinitely). The only reason it does need to be flushed is when mapping code pages. Since icache doesn't snoop dcache, those dirty dcache lines need to be written back to memory and icache line invalidated so that icache lines fetch will get the right data. Decent gains on LMBench fork/exec/sh and File I/O micro-benchmarks. (1) FPGA @ 80 MHZ Processor, Processes - times in microseconds - smaller is better ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- 3.9-rc6-a Linux 3.9.0-r 80 4.79 8.72 66.7 116. 239. 8.39 30.4 4798 14.K 34.K 3.9-rc6-b Linux 3.9.0-r 80 4.79 8.62 65.4 111. 239. 8.35 29.0 3995 12.K 30.K 3.9-rc7-c Linux 3.9.0-r 80 4.79 9.00 66.1 106. 239. 8.61 30.4 2858 10.K 24.K ^^^^ ^^^^ ^^^ File & VM system latencies in microseconds - smaller is better ------------------------------------------------------------------------------- Host OS 0K File 10K File Mmap Prot Page 100fd Create Delete Create Delete Latency Fault Fault selct --------- ------------- ------ ------ ------ ------ ------- ----- ------- ----- 3.9-rc6-a Linux 3.9.0-r 317.8 204.2 1122.3 375.1 3522.0 4.288 20.7 126.8 3.9-rc6-b Linux 3.9.0-r 298.7 223.0 1141.6 367.8 3531.0 4.866 20.9 126.4 3.9-rc7-c Linux 3.9.0-r 278.4 179.2 862.1 339.3 3705.0 3.223 20.3 126.6 ^^^^^ ^^^^^ ^^^^^ ^^^^ (2) Customer Silicon @ 500 MHz (166 MHz mem) ------------------------------------------------------------------------------ Host OS Mhz null null open slct sig sig fork exec sh call I/O stat clos TCP inst hndl proc proc proc --------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- abilis-ba Linux 3.9.0-r 497 0.71 1.38 4.58 12.0 35.5 1.40 3.89 2070 5525 13.K abilis-ca Linux 3.9.0-r 497 0.71 1.40 4.61 11.8 35.6 1.37 3.92 1411 4317 10.K ^^^^ ^^^^ ^^^ Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
2013-04-16 16:40:48 +08:00
}
void flush_icache_all(void)
{
unsigned long flags;
local_irq_save(flags);
write_aux_reg(ARC_REG_IC_IVIC, 1);
/* lr will not complete till the icache inv operation is not over */
read_aux_reg(ARC_REG_IC_CTRL);
local_irq_restore(flags);
}
noinline void flush_cache_all(void)
{
unsigned long flags;
local_irq_save(flags);
flush_icache_all();
__dc_entire_op(OP_FLUSH_N_INV);
local_irq_restore(flags);
}
#ifdef CONFIG_ARC_CACHE_VIPT_ALIASING
void flush_cache_mm(struct mm_struct *mm)
{
flush_cache_all();
}
void flush_cache_page(struct vm_area_struct *vma, unsigned long u_vaddr,
unsigned long pfn)
{
unsigned int paddr = pfn << PAGE_SHIFT;
__sync_icache_dcache(paddr, u_vaddr, PAGE_SIZE);
}
void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
unsigned long end)
{
flush_cache_all();
}
void copy_user_highpage(struct page *to, struct page *from,
unsigned long u_vaddr, struct vm_area_struct *vma)
{
void *kfrom = page_address(from);
void *kto = page_address(to);
int clean_src_k_mappings = 0;
/*
* If SRC page was already mapped in userspace AND it's U-mapping is
* not congruent with K-mapping, sync former to physical page so that
* K-mapping in memcpy below, sees the right data
*
* Note that while @u_vaddr refers to DST page's userspace vaddr, it is
* equally valid for SRC page as well
*/
if (page_mapped(from) && addr_not_cache_congruent(kfrom, u_vaddr)) {
__flush_dcache_page(kfrom, u_vaddr);
clean_src_k_mappings = 1;
}
copy_page(kto, kfrom);
/*
* Mark DST page K-mapping as dirty for a later finalization by
* update_mmu_cache(). Although the finalization could have been done
* here as well (given that both vaddr/paddr are available).
* But update_mmu_cache() already has code to do that for other
* non copied user pages (e.g. read faults which wire in pagecache page
* directly).
*/
set_bit(PG_arch_1, &to->flags);
/*
* if SRC was already usermapped and non-congruent to kernel mapping
* sync the kernel mapping back to physical page
*/
if (clean_src_k_mappings) {
__flush_dcache_page(kfrom, kfrom);
} else {
set_bit(PG_arch_1, &from->flags);
}
}
void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
{
clear_page(to);
set_bit(PG_arch_1, &page->flags);
}
void flush_anon_page(struct vm_area_struct *vma, struct page *page,
unsigned long u_vaddr)
{
/* TBD: do we really need to clear the kernel mapping */
__flush_dcache_page(page_address(page), u_vaddr);
__flush_dcache_page(page_address(page), page_address(page));
}
#endif
/**********************************************************************
* Explicit Cache flush request from user space via syscall
* Needed for JITs which generate code on the fly
*/
SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
{
/* TBD: optimize this */
flush_cache_all();
return 0;
}