linux/arch/tile/mm/homecache.c
Chris Metcalf 76c567fbba arch/tile: support 4KB page size as well as 64KB
The Tilera architecture traditionally supports 64KB page sizes
to improve TLB utilization and improve performance when the
hardware is being used primarily to run a single application.

For more generic server scenarios, it can be beneficial to run
with 4KB page sizes, so this commit allows that to be specified
(by modifying the arch/tile/include/hv/pagesize.h header).

As part of this change, we also re-worked the PTE management
slightly so that PTE writes all go through a __set_pte() function
where we can do some additional validation.  The set_pte_order()
function was eliminated since the "order" argument wasn't being used.

One bug uncovered was in the PCI DMA code, which wasn't properly
flushing the specified range.  This was benign with 64KB pages,
but with 4KB pages we were getting some larger flushes wrong.

The per-cpu memory reservation code also needed updating to
conform with the newer percpu stuff; before it always chose 64KB,
and that was always correct, but with 4KB granularity we now have
to pay closer attention and reserve the amount of memory that will
be requested when the percpu code starts allocating.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>
2011-03-10 13:17:53 -05:00

458 lines
13 KiB
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 <linux/module.h>
#include <asm/page.h>
#include <asm/sections.h>
#include <asm/tlbflush.h>
#include <asm/pgalloc.h>
#include <asm/homecache.h>
#include <arch/sim.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 flush_remote_page(struct page *page, int order)
{
int i, pages = (1 << order);
for (i = 0; i < pages; ++i, ++page) {
void *p = kmap_atomic(page);
int hfh = 0;
int home = page_home(page);
#if CHIP_HAS_CBOX_HOME_MAP()
if (home == PAGE_HOME_HASH)
hfh = 1;
else
#endif
BUG_ON(home < 0 || home >= NR_CPUS);
finv_buffer_remote(p, PAGE_SIZE, hfh);
kunmap_atomic(p);
}
}
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.
* The return value is whether the pages are all coherently cached
* (i.e. none are immutable, incoherent, or uncached).
*/
static int homecache_mask(struct page *page, int pages,
struct cpumask *home_mask)
{
int i;
int cached_coherently = 1;
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 0;
}
#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) {
cached_coherently = 0;
continue;
}
BUG_ON(home < 0 || home >= NR_CPUS);
cpumask_set_cpu(home, home_mask);
}
return cached_coherently;
}
/*
* 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;
}
/* 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);
sim_validate_lines_evicted(PFN_PHYS(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;
}
EXPORT_SYMBOL(pte_set_home);
/*
* 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));
__set_pte(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;
}
EXPORT_SYMBOL(homecache_alloc_pages);
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++);
}
}