static void __init cps_gen_set_top_bit(u32 **pp, struct uasm_label **pl,
struct uasm_reloc **pr,
unsigned r_addr, int lbl)
{
uasm_i_lui(pp, t0, uasm_rel_hi(0x80000000));
uasm_build_label(pl, *pp, lbl);
uasm_i_ll(pp, t1, 0, r_addr);
uasm_i_or(pp, t1, t1, t0);
uasm_i_sc(pp, t1, 0, r_addr);
uasm_il_beqz(pp, pr, t1, lbl);
uasm_i_nop(pp);
}
static void __init cps_gen_cache_routine(u32 **pp, struct uasm_label **pl,
struct uasm_reloc **pr,
const struct cache_desc *cache,
unsigned op, int lbl)
{
unsigned cache_size = cache->ways << cache->waybit;
unsigned i;
const unsigned unroll_lines = 32;
/* If the cache isn't present this function has it easy */
if (cache->flags & MIPS_CACHE_NOT_PRESENT)
return;
/* Load base address */
UASM_i_LA(pp, t0, (long)CKSEG0);
/* Calculate end address */
if (cache_size < 0x8000)
uasm_i_addiu(pp, t1, t0, cache_size);
else
UASM_i_LA(pp, t1, (long)(CKSEG0 + cache_size));
/* Start of cache op loop */
uasm_build_label(pl, *pp, lbl);
/* Generate the cache ops */
for (i = 0; i < unroll_lines; i++) {
if (cpu_has_mips_r6) {
uasm_i_cache(pp, op, 0, t0);
uasm_i_addiu(pp, t0, t0, cache->linesz);
} else {
uasm_i_cache(pp, op, i * cache->linesz, t0);
}
}
if (!cpu_has_mips_r6)
/* Update the base address */
uasm_i_addiu(pp, t0, t0, unroll_lines * cache->linesz);
/* Loop if we haven't reached the end address yet */
uasm_il_bne(pp, pr, t0, t1, lbl);
uasm_i_nop(pp);
}
static void * __init cps_gen_entry_code(unsigned cpu, enum cps_pm_state state)
{
struct uasm_label *l = labels;
struct uasm_reloc *r = relocs;
u32 *buf, *p;
const unsigned r_online = a0;
const unsigned r_nc_count = a1;
const unsigned r_pcohctl = t7;
const unsigned max_instrs = 256;
unsigned cpc_cmd;
enum {
lbl_incready = 1,
lbl_poll_cont,
lbl_secondary_hang,
lbl_disable_coherence,
lbl_flush_fsb,
lbl_invicache,
lbl_flushdcache,
lbl_hang,
lbl_set_cont,
lbl_secondary_cont,
lbl_decready,
};
/* Allocate a buffer to hold the generated code */
p = buf = kcalloc(max_instrs, sizeof(u32), GFP_KERNEL);
if (!buf)
return NULL;
/* Clear labels & relocs ready for (re)use */
memset(labels, 0, sizeof(labels));
memset(relocs, 0, sizeof(relocs));
if (state == CPS_PM_POWER_GATED) {
/* Power gating relies upon CPS SMP */
if (!mips_cps_smp_in_use())
goto out_err;
/*
* Save CPU state. Note the non-standard calling convention
* with the return address placed in v0 to avoid clobbering
* the ra register before it is saved.
*/
UASM_i_LA(&p, t0, (long)mips_cps_pm_save);
uasm_i_jalr(&p, v0, t0);
uasm_i_nop(&p);
}
/*
* Load addresses of required CM & CPC registers. This is done early
* because they're needed in both the enable & disable coherence steps
* but in the coupled case the enable step will only run on one VPE.
*/
UASM_i_LA(&p, r_pcohctl, (long)_gcmp_base + GCMPCLCBOFS(COHCTL));
if (coupled_coherence) {
/* Increment ready_count */
uasm_i_sync(&p, stype_ordering);
uasm_build_label(&l, p, lbl_incready);
uasm_i_ll(&p, t1, 0, r_nc_count);
uasm_i_addiu(&p, t2, t1, 1);
uasm_i_sc(&p, t2, 0, r_nc_count);
uasm_il_beqz(&p, &r, t2, lbl_incready);
uasm_i_addiu(&p, t1, t1, 1);
/* Ordering barrier */
uasm_i_sync(&p, stype_ordering);
/*
* If this is the last VPE to become ready for non-coherence
* then it should branch below.
*/
uasm_il_beq(&p, &r, t1, r_online, lbl_disable_coherence);
uasm_i_nop(&p);
if (state < CPS_PM_POWER_GATED) {
/*
* Otherwise this is not the last VPE to become ready
* for non-coherence. It needs to wait until coherence
* has been disabled before proceeding, which it will do
* by polling for the top bit of ready_count being set.
*/
uasm_i_addiu(&p, t1, zero, -1);
uasm_build_label(&l, p, lbl_poll_cont);
uasm_i_lw(&p, t0, 0, r_nc_count);
uasm_il_bltz(&p, &r, t0, lbl_secondary_cont);
uasm_i_ehb(&p);
uasm_i_yield(&p, zero, t1);
uasm_il_b(&p, &r, lbl_poll_cont);
uasm_i_nop(&p);
} else {
/*
* The core will lose power & this VPE will not continue
* so it can simply halt here.
*/
uasm_i_addiu(&p, t0, zero, TCHALT_H);
uasm_i_mtc0(&p, t0, 2, 4);
uasm_build_label(&l, p, lbl_secondary_hang);
uasm_il_b(&p, &r, lbl_secondary_hang);
uasm_i_nop(&p);
}
}
/*
* This is the point of no return - this VPE will now proceed to
* disable coherence. At this point we *must* be sure that no other
* VPE within the core will interfere with the L1 dcache.
*/
uasm_build_label(&l, p, lbl_disable_coherence);
/* Invalidate the L1 icache */
cps_gen_cache_routine(&p, &l, &r, &cpu_data[cpu].icache,
Index_Invalidate_I, lbl_invicache);
/* Writeback & invalidate the L1 dcache */
cps_gen_cache_routine(&p, &l, &r, &cpu_data[cpu].dcache,
Index_Writeback_Inv_D, lbl_flushdcache);
/* Completion barrier */
uasm_i_sync(&p, stype_memory);
uasm_i_ehb(&p);
/*
* Disable all but self interventions. The load from COHCTL is defined
* by the interAptiv & proAptiv SUMs as ensuring that the operation
* resulting from the preceeding store is complete.
*/
uasm_i_addiu(&p, t0, zero, 1 << cpu_data[cpu].core);
uasm_i_sw(&p, t0, 0, r_pcohctl);
uasm_i_lw(&p, t0, 0, r_pcohctl);
/* Sync to ensure previous interventions are complete */
uasm_i_sync(&p, stype_intervention);
uasm_i_ehb(&p);
/* Disable coherence */
uasm_i_sw(&p, zero, 0, r_pcohctl);
uasm_i_lw(&p, t0, 0, r_pcohctl);
if (state >= CPS_PM_CLOCK_GATED) {
/* TODO: determine whether required based on CPC version */
cps_gen_flush_fsb(&p, &l, &r, &cpu_data[cpu].dcache,
lbl_flush_fsb);
/* Determine the CPC command to issue */
switch (state) {
case CPS_PM_CLOCK_GATED:
cpc_cmd = CPC_Cx_CMD_CLOCKOFF;
break;
case CPS_PM_POWER_GATED:
cpc_cmd = CPC_Cx_CMD_PWRDOWN;
break;
default:
BUG();
goto out_err;
}
/* Issue the CPC command */
UASM_i_LA(&p, t0, (long)addr_cpc_cl_cmd());
uasm_i_addiu(&p, t1, zero, cpc_cmd);
uasm_i_sw(&p, t1, 0, t0);
if (state == CPS_PM_POWER_GATED) {
/* If anything goes wrong just hang */
uasm_build_label(&l, p, lbl_hang);
uasm_il_b(&p, &r, lbl_hang);
uasm_i_nop(&p);
/*
* There's no point generating more code, the core is
* powered down & if powered back up will run from the
* reset vector not from here.
*/
goto gen_done;
}
/* Completion barrier */
uasm_i_sync(&p, stype_memory);
uasm_i_ehb(&p);
}
if (state == CPS_PM_NC_WAIT) {
/*
* At this point it is safe for all VPEs to proceed with
* execution. This VPE will set the top bit of ready_count
* to indicate to the other VPEs that they may continue.
*/
if (coupled_coherence)
cps_gen_set_top_bit(&p, &l, &r, r_nc_count,
lbl_set_cont);
/*
* VPEs which did not disable coherence will continue
* executing, after coherence has been disabled, from this
* point.
*/
uasm_build_label(&l, p, lbl_secondary_cont);
/* Now perform our wait */
uasm_i_wait(&p, 0);
}
/*
* Re-enable coherence. Note that for CPS_PM_NC_WAIT all coupled VPEs
* will run this. The first will actually re-enable coherence & the
* rest will just be performing a rather unusual nop.
*/
uasm_i_addiu(&p, t0, zero, GCMP_CCB_COHCTL_DOMAIN_MSK);
uasm_i_sw(&p, t0, 0, r_pcohctl);
uasm_i_lw(&p, t0, 0, r_pcohctl);
/* Completion barrier */
uasm_i_sync(&p, stype_memory);
uasm_i_ehb(&p);
if (coupled_coherence && (state == CPS_PM_NC_WAIT)) {
/* Decrement ready_count */
uasm_build_label(&l, p, lbl_decready);
uasm_i_sync(&p, stype_ordering);
uasm_i_ll(&p, t1, 0, r_nc_count);
uasm_i_addiu(&p, t2, t1, -1);
uasm_i_sc(&p, t2, 0, r_nc_count);
uasm_il_beqz(&p, &r, t2, lbl_decready);
uasm_i_andi(&p, v0, t1, (1 << fls(smp_num_siblings)) - 1);
/* Ordering barrier */
uasm_i_sync(&p, stype_ordering);
}
if (coupled_coherence && (state == CPS_PM_CLOCK_GATED)) {
/*
* At this point it is safe for all VPEs to proceed with
* execution. This VPE will set the top bit of ready_count
* to indicate to the other VPEs that they may continue.
*/
cps_gen_set_top_bit(&p, &l, &r, r_nc_count, lbl_set_cont);
/*
* This core will be reliant upon another core sending a
* power-up command to the CPC in order to resume operation.
* Thus an arbitrary VPE can't trigger the core leaving the
* idle state and the one that disables coherence might as well
* be the one to re-enable it. The rest will continue from here
* after that has been done.
*/
uasm_build_label(&l, p, lbl_secondary_cont);
/* Ordering barrier */
uasm_i_sync(&p, stype_ordering);
}
/* The core is coherent, time to return to C code */
uasm_i_jr(&p, ra);
uasm_i_nop(&p);
gen_done:
/* Ensure the code didn't exceed the resources allocated for it */
BUG_ON((p - buf) > max_instrs);
BUG_ON((l - labels) > ARRAY_SIZE(labels));
BUG_ON((r - relocs) > ARRAY_SIZE(relocs));
/* Patch branch offsets */
uasm_resolve_relocs(relocs, labels);
/* Flush the icache */
local_flush_icache_range((unsigned long)buf, (unsigned long)p);
return buf;
out_err:
kfree(buf);
return NULL;
}
static void __init cps_gen_flush_fsb(u32 **pp, struct uasm_label **pl,
struct uasm_reloc **pr,
const struct cache_desc *dcache,
int lbl)
{
unsigned i, fsb_size = 8;
unsigned num_loads = (fsb_size * 3) / 2;
unsigned line_stride = 2;
/*
* Ensure that the fill/store buffer (FSB) is not holding the results
* of a prefetch, since if it is then the CPC sequencer may become
* stuck in the D3 (ClrBus) state whilst entering a low power state.
*/
/* TODO: this is interAptiv-specific, generalise it */
/* Preserve perf counter 1 setup */
uasm_i_mfc0(pp, t2, 25, 2); /* PerfCtl1 */
uasm_i_mfc0(pp, t3, 25, 3); /* PerfCnt1 */
/* Setup perf counter 1 to count FSB full pipeline stalls */
uasm_i_addiu(pp, t0, zero, 0x66f);
uasm_i_mtc0(pp, t0, 25, 2); /* PerfCtl1 */
uasm_i_ehb(pp);
uasm_i_mtc0(pp, zero, 25, 3); /* PerfCnt1 */
uasm_i_ehb(pp);
/* Base address for loads */
UASM_i_LA(pp, t0, (long)CKSEG0);
/* Start of clear loop */
uasm_build_label(pl, *pp, lbl);
/* Perform some loads to fill the FSB */
for (i = 0; i < num_loads; i++)
uasm_i_lw(pp, zero, i * dcache->linesz * line_stride, t0);
/*
* Invalidate the new D-cache entries so that the cache will need
* refilling (via the FSB) if the loop is executed again.
*/
for (i = 0; i < num_loads; i++) {
uasm_i_cache(pp, Hit_Invalidate_D,
i * dcache->linesz * line_stride, t0);
uasm_i_cache(pp, Hit_Writeback_Inv_SD,
i * dcache->linesz * line_stride, t0);
}
/* Completion barrier */
uasm_i_sync(pp, stype_memory);
uasm_i_ehb(pp);
/* Check whether the pipeline stalled due to the FSB being full */
uasm_i_mfc0(pp, t1, 25, 3); /* PerfCnt1 */
/* Loop if it didn't */
uasm_il_beqz(pp, pr, t1, lbl);
uasm_i_nop(pp);
/* Restore perf counter 1. The count may well now be wrong... */
uasm_i_mtc0(pp, t2, 25, 2); /* PerfCtl1 */
uasm_i_ehb(pp);
uasm_i_mtc0(pp, t3, 25, 3); /* PerfCnt1 */
uasm_i_ehb(pp);
}
static int __init cps_gen_flush_fsb(u32 **pp, struct uasm_label **pl,
struct uasm_reloc **pr,
const struct cpuinfo_mips *cpu_info,
int lbl)
{
unsigned i, fsb_size = 8;
unsigned num_loads = (fsb_size * 3) / 2;
unsigned line_stride = 2;
unsigned line_size = cpu_info->dcache.linesz;
unsigned perf_counter, perf_event;
unsigned revision = cpu_info->processor_id & PRID_REV_MASK;
/*
* Determine whether this CPU requires an FSB flush, and if so which
* performance counter/event reflect stalls due to a full FSB.
*/
switch (__get_cpu_type(cpu_info->cputype)) {
case CPU_INTERAPTIV:
perf_counter = 1;
perf_event = 51;
break;
case CPU_PROAPTIV:
/* Newer proAptiv cores don't require this workaround */
if (revision >= PRID_REV_ENCODE_332(1, 1, 0))
return 0;
/* On older ones it's unavailable */
return -1;
/* CPUs which do not require the workaround */
case CPU_P5600:
case CPU_I6400:
return 0;
default:
WARN_ONCE(1, "pm-cps: FSB flush unsupported for this CPU\n");
return -1;
}
/*
* Ensure that the fill/store buffer (FSB) is not holding the results
* of a prefetch, since if it is then the CPC sequencer may become
* stuck in the D3 (ClrBus) state whilst entering a low power state.
*/
/* Preserve perf counter setup */
uasm_i_mfc0(pp, t2, 25, (perf_counter * 2) + 0); /* PerfCtlN */
uasm_i_mfc0(pp, t3, 25, (perf_counter * 2) + 1); /* PerfCntN */
/* Setup perf counter to count FSB full pipeline stalls */
uasm_i_addiu(pp, t0, zero, (perf_event << 5) | 0xf);
uasm_i_mtc0(pp, t0, 25, (perf_counter * 2) + 0); /* PerfCtlN */
uasm_i_ehb(pp);
uasm_i_mtc0(pp, zero, 25, (perf_counter * 2) + 1); /* PerfCntN */
uasm_i_ehb(pp);
/* Base address for loads */
UASM_i_LA(pp, t0, (long)CKSEG0);
/* Start of clear loop */
uasm_build_label(pl, *pp, lbl);
/* Perform some loads to fill the FSB */
for (i = 0; i < num_loads; i++)
uasm_i_lw(pp, zero, i * line_size * line_stride, t0);
/*
* Invalidate the new D-cache entries so that the cache will need
* refilling (via the FSB) if the loop is executed again.
*/
for (i = 0; i < num_loads; i++) {
uasm_i_cache(pp, Hit_Invalidate_D,
i * line_size * line_stride, t0);
uasm_i_cache(pp, Hit_Writeback_Inv_SD,
i * line_size * line_stride, t0);
}
/* Completion barrier */
uasm_i_sync(pp, stype_memory);
uasm_i_ehb(pp);
/* Check whether the pipeline stalled due to the FSB being full */
uasm_i_mfc0(pp, t1, 25, (perf_counter * 2) + 1); /* PerfCntN */
/* Loop if it didn't */
uasm_il_beqz(pp, pr, t1, lbl);
uasm_i_nop(pp);
/* Restore perf counter 1. The count may well now be wrong... */
uasm_i_mtc0(pp, t2, 25, (perf_counter * 2) + 0); /* PerfCtlN */
uasm_i_ehb(pp);
uasm_i_mtc0(pp, t3, 25, (perf_counter * 2) + 1); /* PerfCntN */
uasm_i_ehb(pp);
return 0;
}
