sh: Store Queue API rework.

Rewrite the store queue API for a per-cpu interface in the driver
model. The old miscdevice is dropped, due to TASK_SIZE limitations,
and no one was using it anyways.

Carve up and allocate store queue space with a bitmap, back sq
mapping objects with a slab cache, and let userspace worry about
its own prefetching.

Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This commit is contained in:
Paul Mundt 2006-09-27 15:49:57 +09:00
parent 373e68b547
commit d7c30c682a
2 changed files with 261 additions and 324 deletions

View File

@ -1,50 +1,52 @@
/*
* arch/sh/kernel/cpu/sq.c
* arch/sh/kernel/cpu/sh4/sq.c
*
* General management API for SH-4 integrated Store Queues
*
* Copyright (C) 2001, 2002, 2003, 2004 Paul Mundt
* Copyright (C) 2001 - 2006 Paul Mundt
* Copyright (C) 2001, 2002 M. R. Brown
*
* Some of this code has been adopted directly from the old arch/sh/mm/sq.c
* hack that was part of the LinuxDC project. For all intents and purposes,
* this is a completely new interface that really doesn't have much in common
* with the old zone-based approach at all. In fact, it's only listed here for
* general completeness.
*
* This file is subject to the terms and conditions of the GNU General Public
* License. See the file "COPYING" in the main directory of this archive
* for more details.
*/
#include <linux/init.h>
#include <linux/cpu.h>
#include <linux/bitmap.h>
#include <linux/sysdev.h>
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/slab.h>
#include <linux/list.h>
#include <linux/proc_fs.h>
#include <linux/miscdevice.h>
#include <linux/vmalloc.h>
#include <linux/mm.h>
#include <asm/io.h>
#include <asm/page.h>
#include <asm/cacheflush.h>
#include <asm/mmu_context.h>
#include <asm/cpu/sq.h>
static LIST_HEAD(sq_mapping_list);
static DEFINE_SPINLOCK(sq_mapping_lock);
struct sq_mapping;
/**
* sq_flush - Flush (prefetch) the store queue cache
* @addr: the store queue address to flush
*
* Executes a prefetch instruction on the specified store queue cache,
* so that the cached data is written to physical memory.
*/
inline void sq_flush(void *addr)
{
__asm__ __volatile__ ("pref @%0" : : "r" (addr) : "memory");
}
struct sq_mapping {
const char *name;
unsigned long sq_addr;
unsigned long addr;
unsigned int size;
struct sq_mapping *next;
};
static struct sq_mapping *sq_mapping_list;
static DEFINE_SPINLOCK(sq_mapping_lock);
static kmem_cache_t *sq_cache;
static unsigned long *sq_bitmap;
#define store_queue_barrier() \
do { \
(void)ctrl_inl(P4SEG_STORE_QUE); \
ctrl_outl(0, P4SEG_STORE_QUE + 0); \
ctrl_outl(0, P4SEG_STORE_QUE + 8); \
} while (0);
/**
* sq_flush_range - Flush (prefetch) a specific SQ range
@ -57,154 +59,73 @@ inline void sq_flush(void *addr)
void sq_flush_range(unsigned long start, unsigned int len)
{
volatile unsigned long *sq = (unsigned long *)start;
unsigned long dummy;
/* Flush the queues */
for (len >>= 5; len--; sq += 8)
sq_flush((void *)sq);
prefetchw((void *)sq);
/* Wait for completion */
dummy = ctrl_inl(P4SEG_STORE_QUE);
ctrl_outl(0, P4SEG_STORE_QUE + 0);
ctrl_outl(0, P4SEG_STORE_QUE + 8);
store_queue_barrier();
}
static struct sq_mapping *__sq_alloc_mapping(unsigned long virt, unsigned long phys, unsigned long size, const char *name)
static inline void sq_mapping_list_add(struct sq_mapping *map)
{
struct sq_mapping *map;
struct sq_mapping **p, *tmp;
if (virt + size > SQ_ADDRMAX)
return ERR_PTR(-ENOSPC);
spin_lock_irq(&sq_mapping_lock);
map = kmalloc(sizeof(struct sq_mapping), GFP_KERNEL);
if (!map)
return ERR_PTR(-ENOMEM);
p = &sq_mapping_list;
while ((tmp = *p) != NULL)
p = &tmp->next;
INIT_LIST_HEAD(&map->list);
map->next = tmp;
*p = map;
map->sq_addr = virt;
map->addr = phys;
map->size = size + 1;
map->name = name;
list_add(&map->list, &sq_mapping_list);
return map;
spin_unlock_irq(&sq_mapping_lock);
}
static unsigned long __sq_get_next_addr(void)
static inline void sq_mapping_list_del(struct sq_mapping *map)
{
if (!list_empty(&sq_mapping_list)) {
struct list_head *pos, *tmp;
struct sq_mapping **p, *tmp;
/*
* Read one off the list head, as it will have the highest
* mapped allocation. Set the next one up right above it.
*
* This is somewhat sub-optimal, as we don't look at
* gaps between allocations or anything lower then the
* highest-level allocation.
*
* However, in the interest of performance and the general
* lack of desire to do constant list rebalancing, we don't
* worry about it.
*/
list_for_each_safe(pos, tmp, &sq_mapping_list) {
struct sq_mapping *entry;
spin_lock_irq(&sq_mapping_lock);
entry = list_entry(pos, typeof(*entry), list);
return entry->sq_addr + entry->size;
}
for (p = &sq_mapping_list; (tmp = *p); p = &tmp->next)
if (tmp == map) {
*p = tmp->next;
break;
}
return P4SEG_STORE_QUE;
spin_unlock_irq(&sq_mapping_lock);
}
/**
* __sq_remap - Perform a translation from the SQ to a phys addr
* @map: sq mapping containing phys and store queue addresses.
*
* Maps the store queue address specified in the mapping to the physical
* address specified in the mapping.
*/
static struct sq_mapping *__sq_remap(struct sq_mapping *map)
static int __sq_remap(struct sq_mapping *map, unsigned long flags)
{
unsigned long flags, pteh, ptel;
#if defined(CONFIG_MMU)
struct vm_struct *vma;
pgprot_t pgprot;
vma = __get_vm_area(map->size, VM_ALLOC, map->sq_addr, SQ_ADDRMAX);
if (!vma)
return -ENOMEM;
vma->phys_addr = map->addr;
if (remap_area_pages((unsigned long)vma->addr, vma->phys_addr,
map->size, flags)) {
vunmap(vma->addr);
return -EAGAIN;
}
#else
/*
* Without an MMU (or with it turned off), this is much more
* straightforward, as we can just load up each queue's QACR with
* the physical address appropriately masked.
*/
ctrl_outl(((map->addr >> 26) << 2) & 0x1c, SQ_QACR0);
ctrl_outl(((map->addr >> 26) << 2) & 0x1c, SQ_QACR1);
#endif
#ifdef CONFIG_MMU
/*
* With an MMU on the other hand, things are slightly more involved.
* Namely, we have to have a direct mapping between the SQ addr and
* the associated physical address in the UTLB by way of setting up
* a virt<->phys translation by hand. We do this by simply specifying
* the SQ addr in UTLB.VPN and the associated physical address in
* UTLB.PPN.
*
* Notably, even though this is a special case translation, and some
* of the configuration bits are meaningless, we're still required
* to have a valid ASID context in PTEH.
*
* We could also probably get by without explicitly setting PTEA, but
* we do it here just for good measure.
*/
spin_lock_irqsave(&sq_mapping_lock, flags);
pteh = map->sq_addr;
ctrl_outl((pteh & MMU_VPN_MASK) | get_asid(), MMU_PTEH);
ptel = map->addr & PAGE_MASK;
if (cpu_data->flags & CPU_HAS_PTEA)
ctrl_outl(((ptel >> 28) & 0xe) | (ptel & 0x1), MMU_PTEA);
pgprot = pgprot_noncached(PAGE_KERNEL);
ptel &= _PAGE_FLAGS_HARDWARE_MASK;
ptel |= pgprot_val(pgprot);
ctrl_outl(ptel, MMU_PTEL);
__asm__ __volatile__ ("ldtlb" : : : "memory");
spin_unlock_irqrestore(&sq_mapping_lock, flags);
/*
* Next, we need to map ourselves in the kernel page table, so that
* future accesses after a TLB flush will be handled when we take a
* page fault.
*
* Theoretically we could just do this directly and not worry about
* setting up the translation by hand ahead of time, but for the
* cases where we want a one-shot SQ mapping followed by a quick
* writeout before we hit the TLB flush, we do it anyways. This way
* we at least save ourselves the initial page fault overhead.
*/
vma = __get_vm_area(map->size, VM_ALLOC, map->sq_addr, SQ_ADDRMAX);
if (!vma)
return ERR_PTR(-ENOMEM);
vma->phys_addr = map->addr;
if (remap_area_pages((unsigned long)vma->addr, vma->phys_addr,
map->size, pgprot_val(pgprot))) {
vunmap(vma->addr);
return NULL;
}
#endif /* CONFIG_MMU */
return map;
return 0;
}
/**
@ -212,42 +133,65 @@ static struct sq_mapping *__sq_remap(struct sq_mapping *map)
* @phys: Physical address of mapping.
* @size: Length of mapping.
* @name: User invoking mapping.
* @flags: Protection flags.
*
* Remaps the physical address @phys through the next available store queue
* address of @size length. @name is logged at boot time as well as through
* the procfs interface.
*
* A pre-allocated and filled sq_mapping pointer is returned, and must be
* cleaned up with a call to sq_unmap() when the user is done with the
* mapping.
* the sysfs interface.
*/
struct sq_mapping *sq_remap(unsigned long phys, unsigned int size, const char *name)
unsigned long sq_remap(unsigned long phys, unsigned int size,
const char *name, unsigned long flags)
{
struct sq_mapping *map;
unsigned long virt, end;
unsigned long end;
unsigned int psz;
int ret, page;
/* Don't allow wraparound or zero size */
end = phys + size - 1;
if (!size || end < phys)
return NULL;
if (unlikely(!size || end < phys))
return -EINVAL;
/* Don't allow anyone to remap normal memory.. */
if (phys < virt_to_phys(high_memory))
return NULL;
if (unlikely(phys < virt_to_phys(high_memory)))
return -EINVAL;
phys &= PAGE_MASK;
size = PAGE_ALIGN(end + 1) - phys;
virt = __sq_get_next_addr();
psz = (size + (PAGE_SIZE - 1)) / PAGE_SIZE;
map = __sq_alloc_mapping(virt, phys, size, name);
printk("sqremap: %15s [%4d page%s] va 0x%08lx pa 0x%08lx\n",
map->name ? map->name : "???",
map = kmem_cache_alloc(sq_cache, GFP_KERNEL);
if (unlikely(!map))
return -ENOMEM;
map->addr = phys;
map->size = size;
map->name = name;
page = bitmap_find_free_region(sq_bitmap, 0x04000000,
get_order(map->size));
if (unlikely(page < 0)) {
ret = -ENOSPC;
goto out;
}
map->sq_addr = P4SEG_STORE_QUE + (page << PAGE_SHIFT);
ret = __sq_remap(map, flags);
if (unlikely(ret != 0))
goto out;
psz = (size + (PAGE_SIZE - 1)) >> PAGE_SHIFT;
pr_info("sqremap: %15s [%4d page%s] va 0x%08lx pa 0x%08lx\n",
likely(map->name) ? map->name : "???",
psz, psz == 1 ? " " : "s",
map->sq_addr, map->addr);
return __sq_remap(map);
sq_mapping_list_add(map);
return map->sq_addr;
out:
kmem_cache_free(sq_cache, map);
return ret;
}
/**
@ -258,188 +202,198 @@ struct sq_mapping *sq_remap(unsigned long phys, unsigned int size, const char *n
* sq_remap(). Also frees up the pte that was previously inserted into
* the kernel page table and discards the UTLB translation.
*/
void sq_unmap(struct sq_mapping *map)
void sq_unmap(unsigned long vaddr)
{
if (map->sq_addr > (unsigned long)high_memory)
vfree((void *)(map->sq_addr & PAGE_MASK));
struct sq_mapping **p, *map;
struct vm_struct *vma;
int page;
list_del(&map->list);
kfree(map);
}
/**
* sq_clear - Clear a store queue range
* @addr: Address to start clearing from.
* @len: Length to clear.
*
* A quick zero-fill implementation for clearing out memory that has been
* remapped through the store queues.
*/
void sq_clear(unsigned long addr, unsigned int len)
{
int i;
/* Clear out both queues linearly */
for (i = 0; i < 8; i++) {
ctrl_outl(0, addr + i + 0);
ctrl_outl(0, addr + i + 8);
}
sq_flush_range(addr, len);
}
/**
* sq_vma_unmap - Unmap a VMA range
* @area: VMA containing range.
* @addr: Start of range.
* @len: Length of range.
*
* Searches the sq_mapping_list for a mapping matching the sq addr @addr,
* and subsequently frees up the entry. Further cleanup is done by generic
* code.
*/
static void sq_vma_unmap(struct vm_area_struct *area,
unsigned long addr, size_t len)
{
struct list_head *pos, *tmp;
list_for_each_safe(pos, tmp, &sq_mapping_list) {
struct sq_mapping *entry;
entry = list_entry(pos, typeof(*entry), list);
if (entry->sq_addr == addr) {
/*
* We could probably get away without doing the tlb flush
* here, as generic code should take care of most of this
* when unmapping the rest of the VMA range for us. Leave
* it in for added sanity for the time being..
*/
__flush_tlb_page(get_asid(), entry->sq_addr & PAGE_MASK);
list_del(&entry->list);
kfree(entry);
for (p = &sq_mapping_list; (map = *p); p = &map->next)
if (map->sq_addr == vaddr)
break;
if (unlikely(!map)) {
printk("%s: bad store queue address 0x%08lx\n",
__FUNCTION__, vaddr);
return;
}
}
}
/**
* sq_vma_sync - Sync a VMA range
* @area: VMA containing range.
* @start: Start of range.
* @len: Length of range.
* @flags: Additional flags.
*
* Synchronizes an sq mapped range by flushing the store queue cache for
* the duration of the mapping.
*
* Used internally for user mappings, which must use msync() to prefetch
* the store queue cache.
*/
static int sq_vma_sync(struct vm_area_struct *area,
unsigned long start, size_t len, unsigned int flags)
{
sq_flush_range(start, len);
page = (map->sq_addr - P4SEG_STORE_QUE) >> PAGE_SHIFT;
bitmap_release_region(sq_bitmap, page, get_order(map->size));
return 0;
}
static struct vm_operations_struct sq_vma_ops = {
.unmap = sq_vma_unmap,
.sync = sq_vma_sync,
};
/**
* sq_mmap - mmap() for /dev/cpu/sq
* @file: unused.
* @vma: VMA to remap.
*
* Remap the specified vma @vma through the store queues, and setup associated
* information for the new mapping. Also build up the page tables for the new
* area.
*/
static int sq_mmap(struct file *file, struct vm_area_struct *vma)
{
unsigned long offset = vma->vm_pgoff << PAGE_SHIFT;
unsigned long size = vma->vm_end - vma->vm_start;
struct sq_mapping *map;
/*
* We're not interested in any arbitrary virtual address that has
* been stuck in the VMA, as we already know what addresses we
* want. Save off the size, and reposition the VMA to begin at
* the next available sq address.
*/
vma->vm_start = __sq_get_next_addr();
vma->vm_end = vma->vm_start + size;
vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
vma->vm_flags |= VM_IO | VM_RESERVED;
map = __sq_alloc_mapping(vma->vm_start, offset, size, "Userspace");
if (io_remap_pfn_range(vma, map->sq_addr, map->addr >> PAGE_SHIFT,
size, vma->vm_page_prot))
return -EAGAIN;
vma->vm_ops = &sq_vma_ops;
return 0;
}
#ifdef CONFIG_PROC_FS
static int sq_mapping_read_proc(char *buf, char **start, off_t off,
int len, int *eof, void *data)
{
struct list_head *pos;
char *p = buf;
list_for_each_prev(pos, &sq_mapping_list) {
struct sq_mapping *entry;
entry = list_entry(pos, typeof(*entry), list);
p += sprintf(p, "%08lx-%08lx [%08lx]: %s\n", entry->sq_addr,
entry->sq_addr + entry->size - 1, entry->addr,
entry->name);
}
return p - buf;
#ifdef CONFIG_MMU
vma = remove_vm_area((void *)(map->sq_addr & PAGE_MASK));
if (!vma) {
printk(KERN_ERR "%s: bad address 0x%08lx\n",
__FUNCTION__, map->sq_addr);
return;
}
#endif
static struct file_operations sq_fops = {
.owner = THIS_MODULE,
.mmap = sq_mmap,
sq_mapping_list_del(map);
kmem_cache_free(sq_cache, map);
}
/*
* Needlessly complex sysfs interface. Unfortunately it doesn't seem like
* there is any other easy way to add things on a per-cpu basis without
* putting the directory entries somewhere stupid and having to create
* links in sysfs by hand back in to the per-cpu directories.
*
* Some day we may want to have an additional abstraction per store
* queue, but considering the kobject hell we already have to deal with,
* it's simply not worth the trouble.
*/
static struct kobject *sq_kobject[NR_CPUS];
struct sq_sysfs_attr {
struct attribute attr;
ssize_t (*show)(char *buf);
ssize_t (*store)(const char *buf, size_t count);
};
static struct miscdevice sq_dev = {
.minor = STORE_QUEUE_MINOR,
.name = "sq",
.fops = &sq_fops,
#define to_sq_sysfs_attr(attr) container_of(attr, struct sq_sysfs_attr, attr)
static ssize_t sq_sysfs_show(struct kobject *kobj, struct attribute *attr,
char *buf)
{
struct sq_sysfs_attr *sattr = to_sq_sysfs_attr(attr);
if (likely(sattr->show))
return sattr->show(buf);
return -EIO;
}
static ssize_t sq_sysfs_store(struct kobject *kobj, struct attribute *attr,
const char *buf, size_t count)
{
struct sq_sysfs_attr *sattr = to_sq_sysfs_attr(attr);
if (likely(sattr->store))
return sattr->store(buf, count);
return -EIO;
}
static ssize_t mapping_show(char *buf)
{
struct sq_mapping **list, *entry;
char *p = buf;
for (list = &sq_mapping_list; (entry = *list); list = &entry->next)
p += sprintf(p, "%08lx-%08lx [%08lx]: %s\n",
entry->sq_addr, entry->sq_addr + entry->size,
entry->addr, entry->name);
return p - buf;
}
static ssize_t mapping_store(const char *buf, size_t count)
{
unsigned long base = 0, len = 0;
sscanf(buf, "%lx %lx", &base, &len);
if (!base)
return -EIO;
if (likely(len)) {
int ret = sq_remap(base, len, "Userspace",
pgprot_val(PAGE_SHARED));
if (ret < 0)
return ret;
} else
sq_unmap(base);
return count;
}
static struct sq_sysfs_attr mapping_attr =
__ATTR(mapping, 0644, mapping_show, mapping_store);
static struct attribute *sq_sysfs_attrs[] = {
&mapping_attr.attr,
NULL,
};
static struct sysfs_ops sq_sysfs_ops = {
.show = sq_sysfs_show,
.store = sq_sysfs_store,
};
static struct kobj_type ktype_percpu_entry = {
.sysfs_ops = &sq_sysfs_ops,
.default_attrs = sq_sysfs_attrs,
};
static int __devinit sq_sysdev_add(struct sys_device *sysdev)
{
unsigned int cpu = sysdev->id;
struct kobject *kobj;
sq_kobject[cpu] = kzalloc(sizeof(struct kobject), GFP_KERNEL);
if (unlikely(!sq_kobject[cpu]))
return -ENOMEM;
kobj = sq_kobject[cpu];
kobj->parent = &sysdev->kobj;
kobject_set_name(kobj, "%s", "sq");
kobj->ktype = &ktype_percpu_entry;
return kobject_register(kobj);
}
static int __devexit sq_sysdev_remove(struct sys_device *sysdev)
{
unsigned int cpu = sysdev->id;
struct kobject *kobj = sq_kobject[cpu];
kobject_unregister(kobj);
return 0;
}
static struct sysdev_driver sq_sysdev_driver = {
.add = sq_sysdev_add,
.remove = __devexit_p(sq_sysdev_remove),
};
static int __init sq_api_init(void)
{
int ret;
unsigned int nr_pages = 0x04000000 >> PAGE_SHIFT;
unsigned int size = (nr_pages + (BITS_PER_LONG - 1)) / BITS_PER_LONG;
int ret = -ENOMEM;
printk(KERN_NOTICE "sq: Registering store queue API.\n");
create_proc_read_entry("sq_mapping", 0, 0, sq_mapping_read_proc, 0);
sq_cache = kmem_cache_create("store_queue_cache",
sizeof(struct sq_mapping), 0, 0,
NULL, NULL);
if (unlikely(!sq_cache))
return ret;
ret = misc_register(&sq_dev);
if (ret)
remove_proc_entry("sq_mapping", NULL);
sq_bitmap = kzalloc(size, GFP_KERNEL);
if (unlikely(!sq_bitmap))
goto out;
ret = sysdev_driver_register(&cpu_sysdev_class, &sq_sysdev_driver);
if (unlikely(ret != 0))
goto out;
return 0;
out:
kfree(sq_bitmap);
kmem_cache_destroy(sq_cache);
return ret;
}
static void __exit sq_api_exit(void)
{
misc_deregister(&sq_dev);
remove_proc_entry("sq_mapping", NULL);
sysdev_driver_unregister(&cpu_sysdev_class, &sq_sysdev_driver);
kfree(sq_bitmap);
kmem_cache_destroy(sq_cache);
}
module_init(sq_api_init);
@ -448,11 +402,7 @@ module_exit(sq_api_exit);
MODULE_AUTHOR("Paul Mundt <lethal@linux-sh.org>, M. R. Brown <mrbrown@0xd6.org>");
MODULE_DESCRIPTION("Simple API for SH-4 integrated Store Queues");
MODULE_LICENSE("GPL");
MODULE_ALIAS_MISCDEV(STORE_QUEUE_MINOR);
EXPORT_SYMBOL(sq_remap);
EXPORT_SYMBOL(sq_unmap);
EXPORT_SYMBOL(sq_clear);
EXPORT_SYMBOL(sq_flush);
EXPORT_SYMBOL(sq_flush_range);

View File

@ -26,23 +26,10 @@
#define SQ_QACR1 (P4SEG_REG_BASE + 0x3c)
#define SQ_ADDRMAX (P4SEG_STORE_QUE + 0x04000000)
struct sq_mapping {
const char *name;
unsigned long sq_addr;
unsigned long addr;
unsigned int size;
struct list_head list;
};
/* arch/sh/kernel/cpu/sh4/sq.c */
extern struct sq_mapping *sq_remap(unsigned long phys, unsigned int size, const char *name);
extern void sq_unmap(struct sq_mapping *map);
extern void sq_clear(unsigned long addr, unsigned int len);
extern void sq_flush(void *addr);
extern void sq_flush_range(unsigned long start, unsigned int len);
unsigned long sq_remap(unsigned long phys, unsigned int size,
const char *name, unsigned long flags);
void sq_unmap(unsigned long vaddr);
void sq_flush_range(unsigned long start, unsigned int len);
#endif /* __ASM_CPU_SH4_SQ_H */