/*
* This maps the physical memory to kernel virtual address space, a total
* of max_low_pfn pages, by creating page tables starting from address
* PAGE_OFFSET.
*
* This routine transitions us from using a set of compiled-in large
* pages to using some more precise caching, including removing access
* to code pages mapped at PAGE_OFFSET (executed only at MEM_SV_START)
* marking read-only data as locally cacheable, striping the remaining
* .data and .bss across all the available tiles, and removing access
* to pages above the top of RAM (thus ensuring a page fault from a bad
* virtual address rather than a hypervisor shoot down for accessing
* memory outside the assigned limits).
*/
static void __init kernel_physical_mapping_init(pgd_t *pgd_base)
{
unsigned long long irqmask;
unsigned long address, pfn;
pmd_t *pmd;
pte_t *pte;
int pte_ofs;
const struct cpumask *my_cpu_mask = cpumask_of(smp_processor_id());
struct cpumask kstripe_mask;
int rc, i;
#if CHIP_HAS_CBOX_HOME_MAP()
if (ktext_arg_seen && ktext_hash) {
pr_warning("warning: \"ktext\" boot argument ignored"
" if \"kcache_hash\" sets up text hash-for-home\n");
ktext_small = 0;
}
if (kdata_arg_seen && kdata_hash) {
pr_warning("warning: \"kdata\" boot argument ignored"
" if \"kcache_hash\" sets up data hash-for-home\n");
}
if (kdata_huge && !hash_default) {
pr_warning("warning: disabling \"kdata=huge\"; requires"
" kcache_hash=all or =allbutstack\n");
kdata_huge = 0;
}
#endif
/*
* Set up a mask for cpus to use for kernel striping.
* This is normally all cpus, but minus dataplane cpus if any.
* If the dataplane covers the whole chip, we stripe over
* the whole chip too.
*/
cpumask_copy(&kstripe_mask, cpu_possible_mask);
if (!kdata_arg_seen)
kdata_mask = kstripe_mask;
/* Allocate and fill in L2 page tables */
for (i = 0; i < MAX_NUMNODES; ++i) {
#ifdef CONFIG_HIGHMEM
unsigned long end_pfn = node_lowmem_end_pfn[i];
#else
unsigned long end_pfn = node_end_pfn[i];
#endif
unsigned long end_huge_pfn = 0;
/* Pre-shatter the last huge page to allow per-cpu pages. */
if (kdata_huge)
end_huge_pfn = end_pfn - (HPAGE_SIZE >> PAGE_SHIFT);
pfn = node_start_pfn[i];
/* Allocate enough memory to hold L2 page tables for node. */
init_prealloc_ptes(i, end_pfn - pfn);
address = (unsigned long) pfn_to_kaddr(pfn);
while (pfn < end_pfn) {
BUG_ON(address & (HPAGE_SIZE-1));
pmd = get_pmd(pgtables, address);
pte = get_prealloc_pte(pfn);
if (pfn < end_huge_pfn) {
pgprot_t prot = init_pgprot(address);
*(pte_t *)pmd = pte_mkhuge(pfn_pte(pfn, prot));
for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE;
pfn++, pte_ofs++, address += PAGE_SIZE)
pte[pte_ofs] = pfn_pte(pfn, prot);
} else {
if (kdata_huge)
printk(KERN_DEBUG "pre-shattered huge"
" page at %#lx\n", address);
for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE;
pfn++, pte_ofs++, address += PAGE_SIZE) {
pgprot_t prot = init_pgprot(address);
pte[pte_ofs] = pfn_pte(pfn, prot);
}
assign_pte(pmd, pte);
}
}
}
/*
* Set or check ktext_map now that we have cpu_possible_mask
* and kstripe_mask to work with.
*/
if (ktext_all)
cpumask_copy(&ktext_mask, cpu_possible_mask);
else if (ktext_nondataplane)
ktext_mask = kstripe_mask;
else if (!cpumask_empty(&ktext_mask)) {
/* Sanity-check any mask that was requested */
struct cpumask bad;
cpumask_andnot(&bad, &ktext_mask, cpu_possible_mask);
cpumask_and(&ktext_mask, &ktext_mask, cpu_possible_mask);
if (!cpumask_empty(&bad)) {
char buf[NR_CPUS * 5];
cpulist_scnprintf(buf, sizeof(buf), &bad);
pr_info("ktext: not using unavailable cpus %s\n", buf);
}
if (cpumask_empty(&ktext_mask)) {
pr_warning("ktext: no valid cpus; caching on %d.\n",
smp_processor_id());
cpumask_copy(&ktext_mask,
cpumask_of(smp_processor_id()));
}
}
address = MEM_SV_INTRPT;
pmd = get_pmd(pgtables, address);
pfn = 0; /* code starts at PA 0 */
if (ktext_small) {
/* Allocate an L2 PTE for the kernel text */
int cpu = 0;
pgprot_t prot = construct_pgprot(PAGE_KERNEL_EXEC,
PAGE_HOME_IMMUTABLE);
if (ktext_local) {
if (ktext_nocache)
prot = hv_pte_set_mode(prot,
HV_PTE_MODE_UNCACHED);
else
prot = hv_pte_set_mode(prot,
HV_PTE_MODE_CACHE_NO_L3);
} else {
prot = hv_pte_set_mode(prot,
HV_PTE_MODE_CACHE_TILE_L3);
cpu = cpumask_first(&ktext_mask);
prot = ktext_set_nocache(prot);
}
BUG_ON(address != (unsigned long)_stext);
pte = NULL;
for (; address < (unsigned long)_einittext;
pfn++, address += PAGE_SIZE) {
pte_ofs = pte_index(address);
if (pte_ofs == 0) {
if (pte)
assign_pte(pmd++, pte);
pte = alloc_pte();
}
if (!ktext_local) {
prot = set_remote_cache_cpu(prot, cpu);
cpu = cpumask_next(cpu, &ktext_mask);
if (cpu == NR_CPUS)
cpu = cpumask_first(&ktext_mask);
}
pte[pte_ofs] = pfn_pte(pfn, prot);
}
if (pte)
assign_pte(pmd, pte);
} else {
pte_t pteval = pfn_pte(0, PAGE_KERNEL_EXEC);
pteval = pte_mkhuge(pteval);
#if CHIP_HAS_CBOX_HOME_MAP()
if (ktext_hash) {
pteval = hv_pte_set_mode(pteval,
HV_PTE_MODE_CACHE_HASH_L3);
pteval = ktext_set_nocache(pteval);
} else
#endif /* CHIP_HAS_CBOX_HOME_MAP() */
if (cpumask_weight(&ktext_mask) == 1) {
pteval = set_remote_cache_cpu(pteval,
cpumask_first(&ktext_mask));
pteval = hv_pte_set_mode(pteval,
HV_PTE_MODE_CACHE_TILE_L3);
pteval = ktext_set_nocache(pteval);
} else if (ktext_nocache)
pteval = hv_pte_set_mode(pteval,
HV_PTE_MODE_UNCACHED);
else
pteval = hv_pte_set_mode(pteval,
HV_PTE_MODE_CACHE_NO_L3);
for (; address < (unsigned long)_einittext;
pfn += PFN_DOWN(HPAGE_SIZE), address += HPAGE_SIZE)
*(pte_t *)(pmd++) = pfn_pte(pfn, pteval);
}
/* Set swapper_pgprot here so it is flushed to memory right away. */
swapper_pgprot = init_pgprot((unsigned long)swapper_pg_dir);
/*
* Since we may be changing the caching of the stack and page
* table itself, we invoke an assembly helper to do the
* following steps:
*
* - flush the cache so we start with an empty slate
* - install pgtables[] as the real page table
* - flush the TLB so the new page table takes effect
*/
irqmask = interrupt_mask_save_mask();
interrupt_mask_set_mask(-1ULL);
rc = flush_and_install_context(__pa(pgtables),
init_pgprot((unsigned long)pgtables),
__get_cpu_var(current_asid),
cpumask_bits(my_cpu_mask));
interrupt_mask_restore_mask(irqmask);
BUG_ON(rc != 0);
/* Copy the page table back to the normal swapper_pg_dir. */
memcpy(pgd_base, pgtables, sizeof(pgtables));
__install_page_table(pgd_base, __get_cpu_var(current_asid),
swapper_pgprot);
/*
* We just read swapper_pgprot and thus brought it into the cache,
* with its new home & caching mode. When we start the other CPUs,
* they're going to reference swapper_pgprot via their initial fake
* VA-is-PA mappings, which cache everything locally. At that
* time, if it's in our cache with a conflicting home, the
* simulator's coherence checker will complain. So, flush it out
* of our cache; we're not going to ever use it again anyway.
*/
__insn_finv(&swapper_pgprot);
}
finv_buffer_remote(void *buffer, size_t size, int hfh)
{
char *p, *base;
size_t step_size, load_count;
/*
* On TILEPro the striping granularity is a fixed 8KB; on
* TILE-Gx it is configurable, and we rely on the fact that
* the hypervisor always configures maximum striping, so that
* bits 9 and 10 of the PA are part of the stripe function, so
* every 512 bytes we hit a striping boundary.
*
*/
#ifdef __tilegx__
const unsigned long STRIPE_WIDTH = 512;
#else
const unsigned long STRIPE_WIDTH = 8192;
#endif
#ifdef __tilegx__
/*
* On TILE-Gx, we must disable the dstream prefetcher before doing
* a cache flush; otherwise, we could end up with data in the cache
* that we don't want there. Note that normally we'd do an mf
* after the SPR write to disabling the prefetcher, but we do one
* below, before any further loads, so there's no need to do it
* here.
*/
uint_reg_t old_dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
__insn_mtspr(SPR_DSTREAM_PF, 0);
#endif
/*
* Flush and invalidate the buffer out of the local L1/L2
* and request the home cache to flush and invalidate as well.
*/
__finv_buffer(buffer, size);
/*
* Wait for the home cache to acknowledge that it has processed
* all the flush-and-invalidate requests. This does not mean
* that the flushed data has reached the memory controller yet,
* but it does mean the home cache is processing the flushes.
*/
__insn_mf();
/*
* Issue a load to the last cache line, which can't complete
* until all the previously-issued flushes to the same memory
* controller have also completed. If we weren't striping
* memory, that one load would be sufficient, but since we may
* be, we also need to back up to the last load issued to
* another memory controller, which would be the point where
* we crossed a "striping" boundary (the granularity of striping
* across memory controllers). Keep backing up and doing this
* until we are before the beginning of the buffer, or have
* hit all the controllers.
*
* If we are flushing a hash-for-home buffer, it's even worse.
* Each line may be homed on a different tile, and each tile
* may have up to four lines that are on different
* controllers. So as we walk backwards, we have to touch
* enough cache lines to satisfy these constraints. In
* practice this ends up being close enough to "load from
* every cache line on a full memory stripe on each
* controller" that we simply do that, to simplify the logic.
*
* On TILE-Gx the hash-for-home function is much more complex,
* with the upshot being we can't readily guarantee we have
* hit both entries in the 128-entry AMT that were hit by any
* load in the entire range, so we just re-load them all.
* With larger buffers, we may want to consider using a hypervisor
* trap to issue loads directly to each hash-for-home tile for
* each controller (doing it from Linux would trash the TLB).
*/
if (hfh) {
step_size = L2_CACHE_BYTES;
#ifdef __tilegx__
load_count = (size + L2_CACHE_BYTES - 1) / L2_CACHE_BYTES;
#else
load_count = (STRIPE_WIDTH / L2_CACHE_BYTES) *
(1 << CHIP_LOG_NUM_MSHIMS());
#endif
} else {
step_size = STRIPE_WIDTH;
load_count = (1 << CHIP_LOG_NUM_MSHIMS());
}
/* Load the last byte of the buffer. */
p = (char *)buffer + size - 1;
force_load(p);
/* Bump down to the end of the previous stripe or cache line. */
p -= step_size;
p = (char *)((unsigned long)p | (step_size - 1));
/* Figure out how far back we need to go. */
base = p - (step_size * (load_count - 2));
if ((unsigned long)base < (unsigned long)buffer)
base = buffer;
/* Fire all the loads we need. */
for (; p >= base; p -= step_size)
force_load(p);
/*
* Repeat, but with finv's instead of loads, to get rid of the
* data we just loaded into our own cache and the old home L3.
* The finv's are guaranteed not to actually flush the data in
* the buffer back to their home, since we just read it, so the
* lines are clean in cache; we will only invalidate those lines.
*/
p = (char *)buffer + size - 1;
__insn_finv(p);
p -= step_size;
p = (char *)((unsigned long)p | (step_size - 1));
for (; p >= base; p -= step_size)
__insn_finv(p);
/* Wait for these finv's (and thus the first finvs) to be done. */
__insn_mf();
#ifdef __tilegx__
/* Reenable the prefetcher. */
__insn_mtspr(SPR_DSTREAM_PF, old_dstream_pf);
#endif
}
