staticint read_decode_cache_bcr_arcv2(int c, char *buf, int len)
{ struct cpuinfo_arc_cache *p_slc = &slc_info; struct bcr_identity ident; struct bcr_generic sbcr; struct bcr_clust_cfg cbcr; struct bcr_volatile vol; int n = 0;
READ_BCR(ARC_REG_SLC_BCR, sbcr); if (sbcr.ver) { struct bcr_slc_cfg slc_cfg;
READ_BCR(ARC_REG_SLC_CFG, slc_cfg);
p_slc->sz_k = 128 << slc_cfg.sz;
l2_line_sz = p_slc->line_len = (slc_cfg.lsz == 0) ? 128 : 64;
n += scnprintf(buf + n, len - n, "SLC\t\t: %uK, %uB Line%s\n",
p_slc->sz_k, p_slc->line_len, IS_USED_RUN(slc_enable));
}
READ_BCR(ARC_REG_CLUSTER_BCR, cbcr); if (cbcr.c) {
ioc_exists = 1;
/* * As for today we don't support both IOC and ZONE_HIGHMEM enabled * simultaneously. This happens because as of today IOC aperture covers * only ZONE_NORMAL (low mem) and any dma transactions outside this * region won't be HW coherent. * If we want to use both IOC and ZONE_HIGHMEM we can use * bounce_buffer to handle dma transactions to HIGHMEM. * Also it is possible to modify dma_direct cache ops or increase IOC * aperture size if we are planning to use HIGHMEM without PAE.
*/ if (IS_ENABLED(CONFIG_HIGHMEM) || is_pae40_enabled())
ioc_enable = 0;
} else {
ioc_enable = 0;
}
READ_BCR(AUX_IDENTITY, ident);
/* HS 2.0 didn't have AUX_VOL */ if (ident.family > 0x51) {
READ_BCR(AUX_VOL, vol);
perip_base = vol.start << 28; /* HS 3.0 has limit and strict-ordering fields */ if (ident.family > 0x52)
perip_end = (vol.limit << 28) - 1;
}
n += scnprintf(buf + n, len - n, "Peripherals\t: %#lx%s%s\n",
perip_base,
IS_AVAIL3(ioc_exists, ioc_enable, ", IO-Coherency (per-device) "));
return n;
}
int arc_cache_mumbojumbo(int c, char *buf, int len)
{ struct cpuinfo_arc_cache *p_ic = &ic_info, *p_dc = &dc_info; struct bcr_cache ibcr, dbcr; int vipt, assoc; int n = 0;
READ_BCR(ARC_REG_IC_BCR, ibcr); if (!ibcr.ver) goto dc_chk;
/* * Cache Flush programming model * * ARC700 MMUv3 I$ and D$ are both VIPT and can potentially alias. * Programming model requires both paddr and vaddr irrespecive of aliasing * considerations: * - vaddr in {I,D}C_IV?L * - paddr in {I,D}C_PTAG * * In HS38x (MMUv4), D$ is PIPT, I$ is VIPT and can still alias. * Programming model is different for aliasing vs. non-aliasing I$ * - D$ / Non-aliasing I$: only paddr in {I,D}C_IV?L * - Aliasing I$: same as ARC700 above (so MMUv3 routine used for MMUv4 I$) * * - If PAE40 is enabled, independent of aliasing considerations, the higher * bits needs to be written into PTAG_HI
*/
/* Ensure we properly floor/ceil the non-line aligned/sized requests * and have @paddr - aligned to cache line and integral @num_lines. * This however can be avoided for page sized since: * -@paddr will be cache-line aligned already (being page aligned) * -@sz will be integral multiple of line size (being page sized).
*/ if (!full_page) {
sz += paddr & ~CACHE_LINE_MASK;
paddr &= CACHE_LINE_MASK;
vaddr &= CACHE_LINE_MASK;
}
num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
/* * MMUv3, cache ops require paddr in PTAG reg * if V-P const for loop, PTAG can be written once outside loop
*/ if (full_page)
write_aux_reg(aux_tag, paddr);
/* * This is technically for MMU v4, using the MMU v3 programming model * Special work for HS38 aliasing I-cache configuration with PAE40 * - upper 8 bits of paddr need to be written into PTAG_HI * - (and needs to be written before the lower 32 bits) * Note that PTAG_HI is hoisted outside the line loop
*/ if (is_pae40_enabled() && op == OP_INV_IC)
write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
while (num_lines-- > 0) { if (!full_page) {
write_aux_reg(aux_tag, paddr);
paddr += L1_CACHE_BYTES;
}
if (op == OP_INV_IC) {
aux_cmd = ARC_REG_IC_IVIL;
} else { /* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
}
/* 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) {
sz += paddr & ~CACHE_LINE_MASK;
paddr &= CACHE_LINE_MASK;
}
num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
/* * For HS38 PAE40 configuration * - upper 8 bits of paddr need to be written into PTAG_HI * - (and needs to be written before the lower 32 bits)
*/ if (is_pae40_enabled()) { if (op == OP_INV_IC) /* * Non aliasing I-cache in HS38, * aliasing I-cache handled in __cache_line_loop_v3()
*/
write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32); else
write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
}
/* * optimized flush operation which takes a region as opposed to iterating per line
*/ staticinline void __cache_line_loop_v4(phys_addr_t paddr, unsignedlong vaddr, unsignedlong sz, constint op, constint full_page)
{ unsignedint s, e;
/* Only for Non aliasing I-cache in HS38 */ if (op == OP_INV_IC) {
s = ARC_REG_IC_IVIR;
e = ARC_REG_IC_ENDR;
} else {
s = ARC_REG_DC_STARTR;
e = ARC_REG_DC_ENDR;
}
if (!full_page) { /* for any leading gap between @paddr and start of cache line */
sz += paddr & ~CACHE_LINE_MASK;
paddr &= CACHE_LINE_MASK;
/* * account for any trailing gap to end of cache line * this is equivalent to DIV_ROUND_UP() in line ops above
*/
sz += L1_CACHE_BYTES - 1;
}
if (is_pae40_enabled()) { /* TBD: check if crossing 4TB boundary */ if (op == OP_INV_IC)
write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32); else
write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
}
/* ENDR needs to be set ahead of START */
write_aux_reg(e, paddr + sz); /* ENDR is exclusive */
write_aux_reg(s, paddr);
/*************************************************************** * Machine specific helpers for Entire D-Cache or Per Line ops
*/
#ifndef USE_RGN_FLSH /* * this version avoids extra read/write of DC_CTRL for flush or invalid ops * in the non region flush regime (such as for ARCompact)
*/ staticinlinevoid __before_dc_op(constint op)
{ if (op == OP_FLUSH_N_INV) { /* Dcache provides 2 cmd: FLUSH or INV * INV in turn has sub-modes: DISCARD or FLUSH-BEFORE * flush-n-inv is achieved by INV cmd but with IM=1 * So toggle INV sub-mode depending on op request and default
*/ constunsignedint ctl = ARC_REG_DC_CTRL;
write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
}
}
/* flush / flush-n-inv both wait */ while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
;
/* Switch back to default Invalidate mode */ if (op == OP_FLUSH_N_INV)
write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
}
}
/* * Operation on Entire D-Cache * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV} * Note that constant propagation ensures all the checks are gone * in generated code
*/ staticinlinevoid __dc_entire_op(constint op)
{ int aux;
__before_dc_op(op);
if (op & 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);
__after_dc_op(op);
}
staticinlinevoid __dc_disable(void)
{ constint r = ARC_REG_DC_CTRL;
static noinline void slc_op_rgn(phys_addr_t paddr, unsignedlong sz, constint op)
{ #ifdef CONFIG_ISA_ARCV2 /* * SLC is shared between all cores and concurrent aux operations from * multiple cores need to be serialized using a spinlock * A concurrent operation can be silently ignored and/or the old/new * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop * below)
*/ static DEFINE_SPINLOCK(lock); unsignedlong flags; unsignedint ctrl;
phys_addr_t end;
spin_lock_irqsave(&lock, flags);
/* * The Region Flush operation is specified by CTRL.RGN_OP[11..9] * - b'000 (default) is Flush, * - b'001 is Invalidate if CTRL.IM == 0 * - b'001 is Flush-n-Invalidate if CTRL.IM == 1
*/
ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
/* Don't rely on default value of IM bit */ if (!(op & OP_FLUSH)) /* i.e. OP_INV */
ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */ else
ctrl |= SLC_CTRL_IM;
if (op & OP_INV)
ctrl |= SLC_CTRL_RGN_OP_INV; /* Inv or flush-n-inv */ else
ctrl &= ~SLC_CTRL_RGN_OP_INV;
write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
/* * Lower bits are ignored, no need to clip * END needs to be setup before START (latter triggers the operation) * END can't be same as START, so add (l2_line_sz - 1) to sz
*/
end = paddr + sz + l2_line_sz - 1; if (is_pae40_enabled())
write_aux_reg(ARC_REG_SLC_RGN_END1, upper_32_bits(end));
/* Make sure "busy" bit reports correct stataus, see STAR 9001165532 */
read_aux_reg(ARC_REG_SLC_CTRL);
while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
spin_unlock_irqrestore(&lock, flags); #endif
}
static __maybe_unused noinline void slc_op_line(phys_addr_t paddr, unsignedlong sz, constint op)
{ #ifdef CONFIG_ISA_ARCV2 /* * SLC is shared between all cores and concurrent aux operations from * multiple cores need to be serialized using a spinlock * A concurrent operation can be silently ignored and/or the old/new * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop * below)
*/ static DEFINE_SPINLOCK(lock);
/* Don't rely on default value of IM bit */ if (!(op & OP_FLUSH)) /* i.e. OP_INV */
ctrl &= ~SLC_CTRL_IM; /* clear IM: Disable flush before Inv */ else
ctrl |= SLC_CTRL_IM;
write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
cmd = op & OP_INV ? ARC_AUX_SLC_IVDL : ARC_AUX_SLC_FLDL;
/* * DMA ops for systems with L1 cache only * Make memory coherent with L1 cache by flushing/invalidating L1 lines
*/ staticvoid __dma_cache_wback_inv_l1(phys_addr_t start, unsignedlong sz)
{
__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
}
/* * DMA ops for systems with both L1 and L2 caches, but without IOC * Both L1 and L2 lines need to be explicitly flushed/invalidated
*/ staticvoid __dma_cache_wback_inv_slc(phys_addr_t start, unsignedlong sz)
{
__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
slc_op(start, sz, OP_FLUSH_N_INV);
}
/* * 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(unsignedlong kstart, unsignedlong kend)
{ unsignedint tot_sz;
WARN(kstart < TASK_SIZE, "%s() can't handle user vaddr", __func__);
/* 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) { unsignedint off, sz; unsignedlong phy, pfn;
/* * General purpose helper to make I and D cache lines consistent. * @paddr is phy addr of region * @vaddr is typically user vaddr (breakpoint) or kernel vaddr (vmalloc) * However in one instance, when called 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 a * builtin kernel page will not have any virtual mappings. * kprobe on loadable module will be kernel vaddr.
*/ void __sync_icache_dcache(phys_addr_t paddr, unsignedlong vaddr, int len)
{
__dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
__ic_line_inv_vaddr(paddr, vaddr, len);
}
/* wrapper to compile time eliminate alignment checks in flush loop */ void __inv_icache_pages(phys_addr_t paddr, unsignedlong vaddr, unsigned nr)
{
__ic_line_inv_vaddr(paddr, vaddr, nr * PAGE_SIZE);
}
/* * wrapper to clearout kernel or userspace mappings of a page * For kernel mappings @vaddr == @paddr
*/ void __flush_dcache_pages(phys_addr_t paddr, unsignedlong vaddr, unsigned nr)
{
__dc_line_op(paddr, vaddr & PAGE_MASK, nr * PAGE_SIZE, OP_FLUSH_N_INV);
}
/********************************************************************** * 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;
}
/* * IO-Coherency (IOC) setup rules: * * 1. Needs to be at system level, so only once by Master core * Non-Masters need not be accessing caches at that time * - They are either HALT_ON_RESET and kick started much later or * - if run on reset, need to ensure that arc_platform_smp_wait_to_boot() * doesn't perturb caches or coherency unit * * 2. caches (L1 and SLC) need to be purged (flush+inv) before setting up IOC, * otherwise any straggler data might behave strangely post IOC enabling * * 3. All Caches need to be disabled when setting up IOC to elide any in-flight * Coherency transactions
*/ static noinline void __init arc_ioc_setup(void)
{ unsignedint ioc_base, mem_sz;
/* * If IOC was already enabled (due to bootloader) it technically needs to * be reconfigured with aperture base,size corresponding to Linux memory map * which will certainly be different than uboot's. But disabling and * reenabling IOC when DMA might be potentially active is tricky business. * To avoid random memory issues later, just panic here and ask user to * upgrade bootloader to one which doesn't enable IOC
*/ if (read_aux_reg(ARC_REG_IO_COH_ENABLE) & ARC_IO_COH_ENABLE_BIT)
panic("IOC already enabled, please upgrade bootloader!\n");
/* * Cache related boot time checks/setups only needed on master CPU: * - Geometry checks (kernel build and hardware agree: e.g. L1_CACHE_BYTES) * Assume SMP only, so all cores will have same cache config. A check on * one core suffices for all * - IOC setup / dma callbacks only need to be done once
*/ static noinline void __init arc_cache_init_master(void)
{ if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) { struct cpuinfo_arc_cache *ic = &ic_info;
if (!ic->line_len)
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);
/* * In MMU v4 (HS38x) the aliasing icache config uses IVIL/PTAG * pair to provide vaddr/paddr respectively, just as in MMU v3
*/ if (is_isa_arcv2() && ic->colors > 1)
_cache_line_loop_ic_fn = __cache_line_loop_v3; else
_cache_line_loop_ic_fn = __cache_line_loop;
}
if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) { struct cpuinfo_arc_cache *dc = &dc_info;
if (!dc->line_len)
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 on ARCompact: ARCv2 has PIPT */ if (is_isa_arcompact() && dc->colors > 1) {
panic("Aliasing VIPT cache not supported\n");
}
}
/* * Check that SMP_CACHE_BYTES (and hence ARCH_DMA_MINALIGN) is larger * or equal to any cache line length.
*/
BUILD_BUG_ON_MSG(L1_CACHE_BYTES > SMP_CACHE_BYTES, "SMP_CACHE_BYTES must be >= any cache line length"); if (is_isa_arcv2() && (l2_line_sz > SMP_CACHE_BYTES))
panic("L2 Cache line [%d] > kernel Config [%d]\n",
l2_line_sz, SMP_CACHE_BYTES);
/* Note that SLC disable not formally supported till HS 3.0 */ if (is_isa_arcv2() && l2_line_sz && !slc_enable)
arc_slc_disable();
if (is_isa_arcv2() && ioc_exists)
arc_ioc_setup();
if (is_isa_arcv2() && l2_line_sz && slc_enable) {
__dma_cache_wback_inv = __dma_cache_wback_inv_slc;
__dma_cache_inv = __dma_cache_inv_slc;
__dma_cache_wback = __dma_cache_wback_slc;
} else {
__dma_cache_wback_inv = __dma_cache_wback_inv_l1;
__dma_cache_inv = __dma_cache_inv_l1;
__dma_cache_wback = __dma_cache_wback_l1;
} /* * In case of IOC (say IOC+SLC case), pointers above could still be set * but end up not being relevant as the first function in chain is not * called at all for devices using coherent DMA. * arch_sync_dma_for_cpu() -> dma_cache_*() -> __dma_cache_*()
*/
}
void __ref arc_cache_init(void)
{ unsignedint __maybe_unused cpu = smp_processor_id();
if (!cpu)
arc_cache_init_master();
/* * In PAE regime, TLB and cache maintenance ops take wider addresses * And even if PAE is not enabled in kernel, the upper 32-bits still need * to be zeroed to keep the ops sane. * As an optimization for more common !PAE enabled case, zero them out * once at init, rather than checking/setting to 0 for every runtime op
*/ if (is_isa_arcv2() && pae40_exist_but_not_enab()) {
if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE))
write_aux_reg(ARC_REG_IC_PTAG_HI, 0);
if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE))
write_aux_reg(ARC_REG_DC_PTAG_HI, 0);
if (l2_line_sz) {
write_aux_reg(ARC_REG_SLC_RGN_END1, 0);
write_aux_reg(ARC_REG_SLC_RGN_START1, 0);
}
}
}
Messung V0.5
¤ Dauer der Verarbeitung: 0.15 Sekunden
(vorverarbeitet)
¤
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.