[deliverable/linux.git] / arch / tile / mm / homecache.c

/*
 * Copyright 2010 Tilera Corporation. All Rights Reserved.
 *
 *   This program is free software; you can redistribute it and/or
 *   modify it under the terms of the GNU General Public License
 *   as published by the Free Software Foundation, version 2.
 *
 *   This program is distributed in the hope that it will be useful, but
 *   WITHOUT ANY WARRANTY; without even the implied warranty of
 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
 *   NON INFRINGEMENT.  See the GNU General Public License for
 *   more details.
 *
 * This code maintains the "home" for each page in the system.
 */

#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/spinlock.h>
#include <linux/list.h>
#include <linux/bootmem.h>
#include <linux/rmap.h>
#include <linux/pagemap.h>
#include <linux/mutex.h>
#include <linux/interrupt.h>
#include <linux/sysctl.h>
#include <linux/pagevec.h>
#include <linux/ptrace.h>
#include <linux/timex.h>
#include <linux/cache.h>
#include <linux/smp.h>

#include <asm/page.h>
#include <asm/sections.h>
#include <asm/tlbflush.h>
#include <asm/pgalloc.h>
#include <asm/homecache.h>

#include "migrate.h"


#if CHIP_HAS_COHERENT_LOCAL_CACHE()

/*
 * The noallocl2 option suppresses all use of the L2 cache to cache
 * locally from a remote home.  There's no point in using it if we
 * don't have coherent local caching, though.
 */
static int __write_once noallocl2;
static int __init set_noallocl2(char *str)
{
	noallocl2 = 1;
	return 0;
}
early_param("noallocl2", set_noallocl2);

#else

#define noallocl2 0

#endif

/* Provide no-op versions of these routines to keep flush_remote() cleaner. */
#define mark_caches_evicted_start() 0
#define mark_caches_evicted_finish(mask, timestamp) do {} while (0)


/*
 * Update the irq_stat for cpus that we are going to interrupt
 * with TLB or cache flushes.  Also handle removing dataplane cpus
 * from the TLB flush set, and setting dataplane_tlb_state instead.
 */
static void hv_flush_update(const struct cpumask *cache_cpumask,
			    struct cpumask *tlb_cpumask,
			    unsigned long tlb_va, unsigned long tlb_length,
			    HV_Remote_ASID *asids, int asidcount)
{
	struct cpumask mask;
	int i, cpu;

	cpumask_clear(&mask);
	if (cache_cpumask)
		cpumask_or(&mask, &mask, cache_cpumask);
	if (tlb_cpumask && tlb_length) {
		cpumask_or(&mask, &mask, tlb_cpumask);
	}

	for (i = 0; i < asidcount; ++i)
		cpumask_set_cpu(asids[i].y * smp_width + asids[i].x, &mask);

	/*
	 * Don't bother to update atomically; losing a count
	 * here is not that critical.
	 */
	for_each_cpu(cpu, &mask)
		++per_cpu(irq_stat, cpu).irq_hv_flush_count;
}

/*
 * This wrapper function around hv_flush_remote() does several things:
 *
 *  - Provides a return value error-checking panic path, since
 *    there's never any good reason for hv_flush_remote() to fail.
 *  - Accepts a 32-bit PFN rather than a 64-bit PA, which generally
 *    is the type that Linux wants to pass around anyway.
 *  - Centralizes the mark_caches_evicted() handling.
 *  - Canonicalizes that lengths of zero make cpumasks NULL.
 *  - Handles deferring TLB flushes for dataplane tiles.
 *  - Tracks remote interrupts in the per-cpu irq_cpustat_t.
 *
 * Note that we have to wait until the cache flush completes before
 * updating the per-cpu last_cache_flush word, since otherwise another
 * concurrent flush can race, conclude the flush has already
 * completed, and start to use the page while it's still dirty
 * remotely (running concurrently with the actual evict, presumably).
 */
void flush_remote(unsigned long cache_pfn, unsigned long cache_control,
		  const struct cpumask *cache_cpumask_orig,
		  HV_VirtAddr tlb_va, unsigned long tlb_length,
		  unsigned long tlb_pgsize,
		  const struct cpumask *tlb_cpumask_orig,
		  HV_Remote_ASID *asids, int asidcount)
{
	int rc;
	int timestamp = 0;  /* happy compiler */
	struct cpumask cache_cpumask_copy, tlb_cpumask_copy;
	struct cpumask *cache_cpumask, *tlb_cpumask;
	HV_PhysAddr cache_pa;
	char cache_buf[NR_CPUS*5], tlb_buf[NR_CPUS*5];

	mb();   /* provided just to simplify "magic hypervisor" mode */

	/*
	 * Canonicalize and copy the cpumasks.
	 */
	if (cache_cpumask_orig && cache_control) {
		cpumask_copy(&cache_cpumask_copy, cache_cpumask_orig);
		cache_cpumask = &cache_cpumask_copy;
	} else {
		cpumask_clear(&cache_cpumask_copy);
		cache_cpumask = NULL;
	}
	if (cache_cpumask == NULL)
		cache_control = 0;
	if (tlb_cpumask_orig && tlb_length) {
		cpumask_copy(&tlb_cpumask_copy, tlb_cpumask_orig);
		tlb_cpumask = &tlb_cpumask_copy;
	} else {
		cpumask_clear(&tlb_cpumask_copy);
		tlb_cpumask = NULL;
	}

	hv_flush_update(cache_cpumask, tlb_cpumask, tlb_va, tlb_length,
			asids, asidcount);
	cache_pa = (HV_PhysAddr)cache_pfn << PAGE_SHIFT;
	if (cache_control & HV_FLUSH_EVICT_L2)
		timestamp = mark_caches_evicted_start();
	rc = hv_flush_remote(cache_pa, cache_control,
			     cpumask_bits(cache_cpumask),
			     tlb_va, tlb_length, tlb_pgsize,
			     cpumask_bits(tlb_cpumask),
			     asids, asidcount);
	if (cache_control & HV_FLUSH_EVICT_L2)
		mark_caches_evicted_finish(cache_cpumask, timestamp);
	if (rc == 0)
		return;
	cpumask_scnprintf(cache_buf, sizeof(cache_buf), &cache_cpumask_copy);
	cpumask_scnprintf(tlb_buf, sizeof(tlb_buf), &tlb_cpumask_copy);

	pr_err("hv_flush_remote(%#llx, %#lx, %p [%s],"
	       " %#lx, %#lx, %#lx, %p [%s], %p, %d) = %d\n",
	       cache_pa, cache_control, cache_cpumask, cache_buf,
	       (unsigned long)tlb_va, tlb_length, tlb_pgsize,
	       tlb_cpumask, tlb_buf,
	       asids, asidcount, rc);
	panic("Unsafe to continue.");
}

void homecache_evict(const struct cpumask *mask)
{
	flush_remote(0, HV_FLUSH_EVICT_L2, mask, 0, 0, 0, NULL, NULL, 0);
}

/* Return a mask of the cpus whose caches currently own these pages. */
static void homecache_mask(struct page *page, int pages,
			   struct cpumask *home_mask)
{
	int i;
	cpumask_clear(home_mask);
	for (i = 0; i < pages; ++i) {
		int home = page_home(&page[i]);
		if (home == PAGE_HOME_IMMUTABLE ||
		    home == PAGE_HOME_INCOHERENT) {
			cpumask_copy(home_mask, cpu_possible_mask);
			return;
		}
#if CHIP_HAS_CBOX_HOME_MAP()
		if (home == PAGE_HOME_HASH) {
			cpumask_or(home_mask, home_mask, &hash_for_home_map);
			continue;
		}
#endif
		if (home == PAGE_HOME_UNCACHED)
			continue;
		BUG_ON(home < 0 || home >= NR_CPUS);
		cpumask_set_cpu(home, home_mask);
	}
}

/*
 * Return the passed length, or zero if it's long enough that we
 * believe we should evict the whole L2 cache.
 */
static unsigned long cache_flush_length(unsigned long length)
{
	return (length >= CHIP_L2_CACHE_SIZE()) ? HV_FLUSH_EVICT_L2 : length;
}

/* On the simulator, confirm lines have been evicted everywhere. */
static void validate_lines_evicted(unsigned long pfn, size_t length)
{
	sim_syscall(SIM_SYSCALL_VALIDATE_LINES_EVICTED,
		    (HV_PhysAddr)pfn << PAGE_SHIFT, length);
}

/* Flush a page out of whatever cache(s) it is in. */
void homecache_flush_cache(struct page *page, int order)
{
	int pages = 1 << order;
	int length = cache_flush_length(pages * PAGE_SIZE);
	unsigned long pfn = page_to_pfn(page);
	struct cpumask home_mask;

	homecache_mask(page, pages, &home_mask);
	flush_remote(pfn, length, &home_mask, 0, 0, 0, NULL, NULL, 0);
	validate_lines_evicted(pfn, pages * PAGE_SIZE);
}


/* Report the home corresponding to a given PTE. */
static int pte_to_home(pte_t pte)
{
	if (hv_pte_get_nc(pte))
		return PAGE_HOME_IMMUTABLE;
	switch (hv_pte_get_mode(pte)) {
	case HV_PTE_MODE_CACHE_TILE_L3:
		return get_remote_cache_cpu(pte);
	case HV_PTE_MODE_CACHE_NO_L3:
		return PAGE_HOME_INCOHERENT;
	case HV_PTE_MODE_UNCACHED:
		return PAGE_HOME_UNCACHED;
#if CHIP_HAS_CBOX_HOME_MAP()
	case HV_PTE_MODE_CACHE_HASH_L3:
		return PAGE_HOME_HASH;
#endif
	}
	panic("Bad PTE %#llx\n", pte.val);
}

/* Update the home of a PTE if necessary (can also be used for a pgprot_t). */
pte_t pte_set_home(pte_t pte, int home)
{
	/* Check for non-linear file mapping "PTEs" and pass them through. */
	if (pte_file(pte))
		return pte;

#if CHIP_HAS_MMIO()
	/* Check for MMIO mappings and pass them through. */
	if (hv_pte_get_mode(pte) == HV_PTE_MODE_MMIO)
		return pte;
#endif


	/*
	 * Only immutable pages get NC mappings.  If we have a
	 * non-coherent PTE, but the underlying page is not
	 * immutable, it's likely the result of a forced
	 * caching setting running up against ptrace setting
	 * the page to be writable underneath.  In this case,
	 * just keep the PTE coherent.
	 */
	if (hv_pte_get_nc(pte) && home != PAGE_HOME_IMMUTABLE) {
		pte = hv_pte_clear_nc(pte);
		pr_err("non-immutable page incoherently referenced: %#llx\n",
		       pte.val);
	}

	switch (home) {

	case PAGE_HOME_UNCACHED:
		pte = hv_pte_set_mode(pte, HV_PTE_MODE_UNCACHED);
		break;

	case PAGE_HOME_INCOHERENT:
		pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
		break;

	case PAGE_HOME_IMMUTABLE:
		/*
		 * We could home this page anywhere, since it's immutable,
		 * but by default just home it to follow "hash_default".
		 */
		BUG_ON(hv_pte_get_writable(pte));
		if (pte_get_forcecache(pte)) {
			/* Upgrade "force any cpu" to "No L3" for immutable. */
			if (hv_pte_get_mode(pte) == HV_PTE_MODE_CACHE_TILE_L3
			    && pte_get_anyhome(pte)) {
				pte = hv_pte_set_mode(pte,
						      HV_PTE_MODE_CACHE_NO_L3);
			}
		} else
#if CHIP_HAS_CBOX_HOME_MAP()
		if (hash_default)
			pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_HASH_L3);
		else
#endif
			pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
		pte = hv_pte_set_nc(pte);
		break;

#if CHIP_HAS_CBOX_HOME_MAP()
	case PAGE_HOME_HASH:
		pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_HASH_L3);
		break;
#endif

	default:
		BUG_ON(home < 0 || home >= NR_CPUS ||
		       !cpu_is_valid_lotar(home));
		pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
		pte = set_remote_cache_cpu(pte, home);
		break;
	}

#if CHIP_HAS_NC_AND_NOALLOC_BITS()
	if (noallocl2)
		pte = hv_pte_set_no_alloc_l2(pte);

	/* Simplify "no local and no l3" to "uncached" */
	if (hv_pte_get_no_alloc_l2(pte) && hv_pte_get_no_alloc_l1(pte) &&
	    hv_pte_get_mode(pte) == HV_PTE_MODE_CACHE_NO_L3) {
		pte = hv_pte_set_mode(pte, HV_PTE_MODE_UNCACHED);
	}
#endif

	/* Checking this case here gives a better panic than from the hv. */
	BUG_ON(hv_pte_get_mode(pte) == 0);

	return pte;
}

/*
 * The routines in this section are the "static" versions of the normal
 * dynamic homecaching routines; they just set the home cache
 * of a kernel page once, and require a full-chip cache/TLB flush,
 * so they're not suitable for anything but infrequent use.
 */

#if CHIP_HAS_CBOX_HOME_MAP()
static inline int initial_page_home(void) { return PAGE_HOME_HASH; }
#else
static inline int initial_page_home(void) { return 0; }
#endif

int page_home(struct page *page)
{
	if (PageHighMem(page)) {
		return initial_page_home();
	} else {
		unsigned long kva = (unsigned long)page_address(page);
		return pte_to_home(*virt_to_pte(NULL, kva));
	}
}

void homecache_change_page_home(struct page *page, int order, int home)
{
	int i, pages = (1 << order);
	unsigned long kva;

	BUG_ON(PageHighMem(page));
	BUG_ON(page_count(page) > 1);
	BUG_ON(page_mapcount(page) != 0);
	kva = (unsigned long) page_address(page);
	flush_remote(0, HV_FLUSH_EVICT_L2, &cpu_cacheable_map,
		     kva, pages * PAGE_SIZE, PAGE_SIZE, cpu_online_mask,
		     NULL, 0);

	for (i = 0; i < pages; ++i, kva += PAGE_SIZE) {
		pte_t *ptep = virt_to_pte(NULL, kva);
		pte_t pteval = *ptep;
		BUG_ON(!pte_present(pteval) || pte_huge(pteval));
		*ptep = pte_set_home(pteval, home);
	}
}

struct page *homecache_alloc_pages(gfp_t gfp_mask,
				   unsigned int order, int home)
{
	struct page *page;
	BUG_ON(gfp_mask & __GFP_HIGHMEM);   /* must be lowmem */
	page = alloc_pages(gfp_mask, order);
	if (page)
		homecache_change_page_home(page, order, home);
	return page;
}

struct page *homecache_alloc_pages_node(int nid, gfp_t gfp_mask,
					unsigned int order, int home)
{
	struct page *page;
	BUG_ON(gfp_mask & __GFP_HIGHMEM);   /* must be lowmem */
	page = alloc_pages_node(nid, gfp_mask, order);
	if (page)
		homecache_change_page_home(page, order, home);
	return page;
}

void homecache_free_pages(unsigned long addr, unsigned int order)
{
	struct page *page;

	if (addr == 0)
		return;

	VM_BUG_ON(!virt_addr_valid((void *)addr));
	page = virt_to_page((void *)addr);
	if (put_page_testzero(page)) {
		int pages = (1 << order);
		homecache_change_page_home(page, order, initial_page_home());
		while (pages--)
			__free_page(page++);
	}
}
Commit	Line	Data
867e359b CM	1	/*
	2	* Copyright 2010 Tilera Corporation. All Rights Reserved.
	3	*
	4	* This program is free software; you can redistribute it and/or
	5	* modify it under the terms of the GNU General Public License
	6	* as published by the Free Software Foundation, version 2.
	7	*
	8	* This program is distributed in the hope that it will be useful, but
	9	* WITHOUT ANY WARRANTY; without even the implied warranty of
	10	* MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
	11	* NON INFRINGEMENT. See the GNU General Public License for
	12	* more details.
	13	*
	14	* This code maintains the "home" for each page in the system.
	15	*/
	16
	17	#include <linux/kernel.h>
	18	#include <linux/mm.h>
	19	#include <linux/spinlock.h>
	20	#include <linux/list.h>
	21	#include <linux/bootmem.h>
	22	#include <linux/rmap.h>
	23	#include <linux/pagemap.h>
	24	#include <linux/mutex.h>
	25	#include <linux/interrupt.h>
	26	#include <linux/sysctl.h>
	27	#include <linux/pagevec.h>
	28	#include <linux/ptrace.h>
	29	#include <linux/timex.h>
	30	#include <linux/cache.h>
	31	#include <linux/smp.h>
	32
	33	#include <asm/page.h>
	34	#include <asm/sections.h>
	35	#include <asm/tlbflush.h>
	36	#include <asm/pgalloc.h>
	37	#include <asm/homecache.h>
	38
	39	#include "migrate.h"
	40
	41
	42	#if CHIP_HAS_COHERENT_LOCAL_CACHE()
	43
	44	/*
	45	* The noallocl2 option suppresses all use of the L2 cache to cache
	46	* locally from a remote home. There's no point in using it if we
	47	* don't have coherent local caching, though.
	48	*/
0707ad30	49	static int __write_once noallocl2;
867e359b CM	50	static int __init set_noallocl2(char *str)
	51	{
	52	noallocl2 = 1;
	53	return 0;
	54	}
	55	early_param("noallocl2", set_noallocl2);
	56
	57	#else
	58
	59	#define noallocl2 0
	60
	61	#endif
	62
867e359b CM	63	/* Provide no-op versions of these routines to keep flush_remote() cleaner. */
	64	#define mark_caches_evicted_start() 0
	65	#define mark_caches_evicted_finish(mask, timestamp) do {} while (0)
	66
	67
867e359b CM	68	/*
	69	* Update the irq_stat for cpus that we are going to interrupt
	70	* with TLB or cache flushes. Also handle removing dataplane cpus
	71	* from the TLB flush set, and setting dataplane_tlb_state instead.
	72	*/
	73	static void hv_flush_update(const struct cpumask *cache_cpumask,
	74	struct cpumask *tlb_cpumask,
	75	unsigned long tlb_va, unsigned long tlb_length,
	76	HV_Remote_ASID *asids, int asidcount)
	77	{
	78	struct cpumask mask;
	79	int i, cpu;
	80
	81	cpumask_clear(&mask);
	82	if (cache_cpumask)
	83	cpumask_or(&mask, &mask, cache_cpumask);
	84	if (tlb_cpumask && tlb_length) {
	85	cpumask_or(&mask, &mask, tlb_cpumask);
	86	}
	87
	88	for (i = 0; i < asidcount; ++i)
	89	cpumask_set_cpu(asids[i].y * smp_width + asids[i].x, &mask);
	90
	91	/*
	92	* Don't bother to update atomically; losing a count
	93	* here is not that critical.
	94	*/
	95	for_each_cpu(cpu, &mask)
	96	++per_cpu(irq_stat, cpu).irq_hv_flush_count;
	97	}
	98
	99	/*
	100	* This wrapper function around hv_flush_remote() does several things:
	101	*
	102	* - Provides a return value error-checking panic path, since
	103	* there's never any good reason for hv_flush_remote() to fail.
	104	* - Accepts a 32-bit PFN rather than a 64-bit PA, which generally
	105	* is the type that Linux wants to pass around anyway.
	106	* - Centralizes the mark_caches_evicted() handling.
	107	* - Canonicalizes that lengths of zero make cpumasks NULL.
	108	* - Handles deferring TLB flushes for dataplane tiles.
	109	* - Tracks remote interrupts in the per-cpu irq_cpustat_t.
	110	*
	111	* Note that we have to wait until the cache flush completes before
	112	* updating the per-cpu last_cache_flush word, since otherwise another
	113	* concurrent flush can race, conclude the flush has already
	114	* completed, and start to use the page while it's still dirty
	115	* remotely (running concurrently with the actual evict, presumably).
	116	*/
	117	void flush_remote(unsigned long cache_pfn, unsigned long cache_control,
	118	const struct cpumask *cache_cpumask_orig,
	119	HV_VirtAddr tlb_va, unsigned long tlb_length,
	120	unsigned long tlb_pgsize,
	121	const struct cpumask *tlb_cpumask_orig,
	122	HV_Remote_ASID *asids, int asidcount)
	123	{
	124	int rc;
	125	int timestamp = 0; /* happy compiler */
	126	struct cpumask cache_cpumask_copy, tlb_cpumask_copy;
	127	struct cpumask cache_cpumask, tlb_cpumask;
	128	HV_PhysAddr cache_pa;
	129	char cache_buf[NR_CPUS5], tlb_buf[NR_CPUS5];
	130
	131	mb(); /* provided just to simplify "magic hypervisor" mode */
132
133	/*
134	* Canonicalize and copy the cpumasks.
135	*/
136	if (cache_cpumask_orig && cache_control) {
137	cpumask_copy(&cache_cpumask_copy, cache_cpumask_orig);
138	cache_cpumask = &cache_cpumask_copy;
139	} else {
140	cpumask_clear(&cache_cpumask_copy);
141	cache_cpumask = NULL;
142	}
143	if (cache_cpumask == NULL)
144	cache_control = 0;
145	if (tlb_cpumask_orig && tlb_length) {
146	cpumask_copy(&tlb_cpumask_copy, tlb_cpumask_orig);
147	tlb_cpumask = &tlb_cpumask_copy;
148	} else {
149	cpumask_clear(&tlb_cpumask_copy);
150	tlb_cpumask = NULL;
151	}
152
153	hv_flush_update(cache_cpumask, tlb_cpumask, tlb_va, tlb_length,
154	asids, asidcount);
155	cache_pa = (HV_PhysAddr)cache_pfn << PAGE_SHIFT;
156	if (cache_control & HV_FLUSH_EVICT_L2)
157	timestamp = mark_caches_evicted_start();
158	rc = hv_flush_remote(cache_pa, cache_control,
159	cpumask_bits(cache_cpumask),
160	tlb_va, tlb_length, tlb_pgsize,
161	cpumask_bits(tlb_cpumask),
162	asids, asidcount);
163	if (cache_control & HV_FLUSH_EVICT_L2)
164	mark_caches_evicted_finish(cache_cpumask, timestamp);
165	if (rc == 0)
166	return;
167	cpumask_scnprintf(cache_buf, sizeof(cache_buf), &cache_cpumask_copy);
168	cpumask_scnprintf(tlb_buf, sizeof(tlb_buf), &tlb_cpumask_copy);
169
0707ad30	170	pr_err("hv_flush_remote(%#llx, %#lx, %p [%s],"
867e359b CM	171	" %#lx, %#lx, %#lx, %p [%s], %p, %d) = %d\n",
	172	cache_pa, cache_control, cache_cpumask, cache_buf,
	173	(unsigned long)tlb_va, tlb_length, tlb_pgsize,
	174	tlb_cpumask, tlb_buf,
	175	asids, asidcount, rc);
867e359b CM	176	panic("Unsafe to continue.");
	177	}
	178
	179	void homecache_evict(const struct cpumask *mask)
	180	{
	181	flush_remote(0, HV_FLUSH_EVICT_L2, mask, 0, 0, 0, NULL, NULL, 0);
	182	}
	183
	184	/* Return a mask of the cpus whose caches currently own these pages. */
	185	static void homecache_mask(struct page *page, int pages,
	186	struct cpumask *home_mask)
	187	{
	188	int i;
	189	cpumask_clear(home_mask);
	190	for (i = 0; i < pages; ++i) {
	191	int home = page_home(&page[i]);
	192	if (home == PAGE_HOME_IMMUTABLE \|\|
	193	home == PAGE_HOME_INCOHERENT) {
	194	cpumask_copy(home_mask, cpu_possible_mask);
	195	return;
	196	}
	197	#if CHIP_HAS_CBOX_HOME_MAP()
	198	if (home == PAGE_HOME_HASH) {
	199	cpumask_or(home_mask, home_mask, &hash_for_home_map);
	200	continue;
	201	}
	202	#endif
	203	if (home == PAGE_HOME_UNCACHED)
	204	continue;
	205	BUG_ON(home < 0 \|\| home >= NR_CPUS);
	206	cpumask_set_cpu(home, home_mask);
	207	}
	208	}
	209
	210	/*
	211	* Return the passed length, or zero if it's long enough that we
	212	* believe we should evict the whole L2 cache.
	213	*/
	214	static unsigned long cache_flush_length(unsigned long length)
	215	{
	216	return (length >= CHIP_L2_CACHE_SIZE()) ? HV_FLUSH_EVICT_L2 : length;
	217	}
	218
	219	/* On the simulator, confirm lines have been evicted everywhere. */
	220	static void validate_lines_evicted(unsigned long pfn, size_t length)
	221	{
	222	sim_syscall(SIM_SYSCALL_VALIDATE_LINES_EVICTED,
	223	(HV_PhysAddr)pfn << PAGE_SHIFT, length);
	224	}
	225
	226	/* Flush a page out of whatever cache(s) it is in. */
	227	void homecache_flush_cache(struct page *page, int order)
	228	{
	229	int pages = 1 << order;
	230	int length = cache_flush_length(pages * PAGE_SIZE);
	231	unsigned long pfn = page_to_pfn(page);
	232	struct cpumask home_mask;
	233
	234	homecache_mask(page, pages, &home_mask);
	235	flush_remote(pfn, length, &home_mask, 0, 0, 0, NULL, NULL, 0);
	236	validate_lines_evicted(pfn, pages * PAGE_SIZE);
	237	}
	238
	239
240	/* Report the home corresponding to a given PTE. */
241	static int pte_to_home(pte_t pte)
242	{
243	if (hv_pte_get_nc(pte))
244	return PAGE_HOME_IMMUTABLE;
245	switch (hv_pte_get_mode(pte)) {
246	case HV_PTE_MODE_CACHE_TILE_L3:
247	return get_remote_cache_cpu(pte);
248	case HV_PTE_MODE_CACHE_NO_L3:
249	return PAGE_HOME_INCOHERENT;
250	case HV_PTE_MODE_UNCACHED:
251	return PAGE_HOME_UNCACHED;
252	#if CHIP_HAS_CBOX_HOME_MAP()
253	case HV_PTE_MODE_CACHE_HASH_L3:
254	return PAGE_HOME_HASH;
255	#endif
256	}
257	panic("Bad PTE %#llx\n", pte.val);
258	}
259
260	/* Update the home of a PTE if necessary (can also be used for a pgprot_t). */
261	pte_t pte_set_home(pte_t pte, int home)
262	{
263	/* Check for non-linear file mapping "PTEs" and pass them through. */
264	if (pte_file(pte))
265	return pte;
266
267	#if CHIP_HAS_MMIO()
268	/* Check for MMIO mappings and pass them through. */
269	if (hv_pte_get_mode(pte) == HV_PTE_MODE_MMIO)
270	return pte;
271	#endif
272
273
274	/*
275	* Only immutable pages get NC mappings. If we have a
276	* non-coherent PTE, but the underlying page is not
277	* immutable, it's likely the result of a forced
278	* caching setting running up against ptrace setting
279	* the page to be writable underneath. In this case,
280	* just keep the PTE coherent.
281	*/
282	if (hv_pte_get_nc(pte) && home != PAGE_HOME_IMMUTABLE) {
283	pte = hv_pte_clear_nc(pte);
0707ad30	284	pr_err("non-immutable page incoherently referenced: %#llx\n",
867e359b CM	285	pte.val);
	286	}
	287
	288	switch (home) {
	289
	290	case PAGE_HOME_UNCACHED:
	291	pte = hv_pte_set_mode(pte, HV_PTE_MODE_UNCACHED);
	292	break;
	293
	294	case PAGE_HOME_INCOHERENT:
	295	pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
	296	break;
	297
	298	case PAGE_HOME_IMMUTABLE:
	299	/*
	300	* We could home this page anywhere, since it's immutable,
	301	* but by default just home it to follow "hash_default".
	302	*/
	303	BUG_ON(hv_pte_get_writable(pte));
	304	if (pte_get_forcecache(pte)) {
	305	/* Upgrade "force any cpu" to "No L3" for immutable. */
	306	if (hv_pte_get_mode(pte) == HV_PTE_MODE_CACHE_TILE_L3
	307	&& pte_get_anyhome(pte)) {
	308	pte = hv_pte_set_mode(pte,
	309	HV_PTE_MODE_CACHE_NO_L3);
	310	}
	311	} else
	312	#if CHIP_HAS_CBOX_HOME_MAP()
	313	if (hash_default)
	314	pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_HASH_L3);
	315	else
	316	#endif
	317	pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_NO_L3);
	318	pte = hv_pte_set_nc(pte);
	319	break;
	320
	321	#if CHIP_HAS_CBOX_HOME_MAP()
	322	case PAGE_HOME_HASH:
	323	pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_HASH_L3);
	324	break;
	325	#endif
	326
	327	default:
	328	BUG_ON(home < 0 \|\| home >= NR_CPUS \|\|
	329	!cpu_is_valid_lotar(home));
	330	pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
	331	pte = set_remote_cache_cpu(pte, home);
	332	break;
	333	}
	334
	335	#if CHIP_HAS_NC_AND_NOALLOC_BITS()
	336	if (noallocl2)
	337	pte = hv_pte_set_no_alloc_l2(pte);
	338
	339	/* Simplify "no local and no l3" to "uncached" */
	340	if (hv_pte_get_no_alloc_l2(pte) && hv_pte_get_no_alloc_l1(pte) &&
	341	hv_pte_get_mode(pte) == HV_PTE_MODE_CACHE_NO_L3) {
	342	pte = hv_pte_set_mode(pte, HV_PTE_MODE_UNCACHED);
	343	}
	344	#endif
	345
	346	/* Checking this case here gives a better panic than from the hv. */
	347	BUG_ON(hv_pte_get_mode(pte) == 0);
	348
349	return pte;
350	}
351
352	/*
353	* The routines in this section are the "static" versions of the normal
354	* dynamic homecaching routines; they just set the home cache
355	* of a kernel page once, and require a full-chip cache/TLB flush,
356	* so they're not suitable for anything but infrequent use.
357	*/
358
359	#if CHIP_HAS_CBOX_HOME_MAP()
360	static inline int initial_page_home(void) { return PAGE_HOME_HASH; }
361	#else
362	static inline int initial_page_home(void) { return 0; }
363	#endif
364
365	int page_home(struct page *page)
366	{
367	if (PageHighMem(page)) {
368	return initial_page_home();
369	} else {
370	unsigned long kva = (unsigned long)page_address(page);
371	return pte_to_home(*virt_to_pte(NULL, kva));
372	}
373	}
374
375	void homecache_change_page_home(struct page *page, int order, int home)
376	{
377	int i, pages = (1 << order);
378	unsigned long kva;
379
380	BUG_ON(PageHighMem(page));
381	BUG_ON(page_count(page) > 1);
382	BUG_ON(page_mapcount(page) != 0);
383	kva = (unsigned long) page_address(page);
384	flush_remote(0, HV_FLUSH_EVICT_L2, &cpu_cacheable_map,
385	kva, pages * PAGE_SIZE, PAGE_SIZE, cpu_online_mask,
386	NULL, 0);
387
388	for (i = 0; i < pages; ++i, kva += PAGE_SIZE) {
389	pte_t *ptep = virt_to_pte(NULL, kva);
390	pte_t pteval = *ptep;
391	BUG_ON(!pte_present(pteval) \|\| pte_huge(pteval));
392	*ptep = pte_set_home(pteval, home);
393	}
394	}
395
396	struct page *homecache_alloc_pages(gfp_t gfp_mask,
397	unsigned int order, int home)
398	{
399	struct page *page;
400	BUG_ON(gfp_mask & __GFP_HIGHMEM); /* must be lowmem */
401	page = alloc_pages(gfp_mask, order);
402	if (page)
403	homecache_change_page_home(page, order, home);
404	return page;
405	}
406
407	struct page *homecache_alloc_pages_node(int nid, gfp_t gfp_mask,
408	unsigned int order, int home)
409	{
410	struct page *page;
411	BUG_ON(gfp_mask & __GFP_HIGHMEM); /* must be lowmem */
412	page = alloc_pages_node(nid, gfp_mask, order);
413	if (page)
414	homecache_change_page_home(page, order, home);
415	return page;
416	}
417
418	void homecache_free_pages(unsigned long addr, unsigned int order)
419	{
420	struct page *page;
421
422	if (addr == 0)
423	return;
424
425	VM_BUG_ON(!virt_addr_valid((void *)addr));
426	page = virt_to_page((void *)addr);
427	if (put_page_testzero(page)) {
428	int pages = (1 << order);
429	homecache_change_page_home(page, order, initial_page_home());
430	while (pages--)
431	__free_page(page++);
432	}
433	}