sh: Store Queue API rework.

/*

 * 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);

struct sq_mapping;

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;

/**

 * 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");

}

#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

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;

	/*

	 * 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);

#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);

		return -ENOMEM;

	vma->phys_addr = map->addr;

	if (remap_area_pages((unsigned long)vma->addr, vma->phys_addr,

			     map->size, pgprot_val(pgprot))) {

			     map->size, flags)) {

		vunmap(vma->addr);

		return NULL;

		return -EAGAIN;

	}

#endif /* CONFIG_MMU */

#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

	return map;

	return 0;

}

/**

 * @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;

}

/**

 * 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;

	for (p = &sq_mapping_list; (map = *p); p = &map->next)

		if (map->sq_addr == vaddr)

			break;

	list_del(&map->list);

	kfree(map);

	if (unlikely(!map)) {

		printk("%s: bad store queue address 0x%08lx\n",

		       __FUNCTION__, vaddr);

		return;

	}

/**

 * 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;

	page = (map->sq_addr - P4SEG_STORE_QUE) >> PAGE_SHIFT;

	bitmap_release_region(sq_bitmap, page, get_order(map->size));

	/* Clear out both queues linearly */

	for (i = 0; i < 8; i++) {

		ctrl_outl(0, addr + i + 0);

		ctrl_outl(0, addr + i + 8);

#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

	sq_flush_range(addr, len);

	sq_mapping_list_del(map);

	kmem_cache_free(sq_cache, map);

}

/**

 * sq_vma_unmap - Unmap a VMA range

 * @area: VMA containing range.

 * @addr: Start of range.

 * @len: Length of range.

/*

 * 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.

 *

 * 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.

 * 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 void sq_vma_unmap(struct vm_area_struct *area,

			 unsigned long addr, size_t len)

{

	struct list_head *pos, *tmp;

static struct kobject *sq_kobject[NR_CPUS];

	list_for_each_safe(pos, tmp, &sq_mapping_list) {

		struct sq_mapping *entry;

struct sq_sysfs_attr {

	struct attribute attr;

	ssize_t (*show)(char *buf);

	ssize_t (*store)(const char *buf, size_t count);

};

		entry = list_entry(pos, typeof(*entry), list);

#define to_sq_sysfs_attr(attr)	container_of(attr, struct sq_sysfs_attr, attr)

		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);

static ssize_t sq_sysfs_show(struct kobject *kobj, struct attribute *attr,

			     char *buf)

{

	struct sq_sysfs_attr *sattr = to_sq_sysfs_attr(attr);

			list_del(&entry->list);

			kfree(entry);

	if (likely(sattr->show))

		return sattr->show(buf);

			return;

		}

	}

	return -EIO;

}

/**

 * 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)

static ssize_t sq_sysfs_store(struct kobject *kobj, struct attribute *attr,

			      const char *buf, size_t count)

{

	sq_flush_range(start, len);

	struct sq_sysfs_attr *sattr = to_sq_sysfs_attr(attr);

	return 0;

	if (likely(sattr->store))

		return sattr->store(buf, count);

	return -EIO;

}

static struct vm_operations_struct sq_vma_ops = {

	.unmap	= sq_vma_unmap,

	.sync	= sq_vma_sync,

};

static ssize_t mapping_show(char *buf)

{

	struct sq_mapping **list, *entry;

	char *p = buf;

/**

 * 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)

	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 offset = vma->vm_pgoff << PAGE_SHIFT;

	unsigned long size = vma->vm_end - vma->vm_start;

	struct sq_mapping *map;

	unsigned long base = 0, len = 0;

	/*

	 * 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;

	sscanf(buf, "%lx %lx", &base, &len);

	if (!base)

		return -EIO;

	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);

	if (likely(len)) {

		int ret = sq_remap(base, len, "Userspace",

				   pgprot_val(PAGE_SHARED));

		if (ret < 0)

			return ret;

	} else

		sq_unmap(base);

	vma->vm_flags |= VM_IO | VM_RESERVED;

	return count;

}

	map = __sq_alloc_mapping(vma->vm_start, offset, size, "Userspace");

static struct sq_sysfs_attr mapping_attr =

	__ATTR(mapping, 0644, mapping_show, mapping_store);

	if (io_remap_pfn_range(vma, map->sq_addr, map->addr >> PAGE_SHIFT,

				size, vma->vm_page_prot))

		return -EAGAIN;

static struct attribute *sq_sysfs_attrs[] = {

	&mapping_attr.attr,

	NULL,

};

	vma->vm_ops = &sq_vma_ops;

static struct sysfs_ops sq_sysfs_ops = {

	.show	= sq_sysfs_show,

	.store	= sq_sysfs_store,

};

	return 0;

}

static struct kobj_type ktype_percpu_entry = {

	.sysfs_ops	= &sq_sysfs_ops,

	.default_attrs	= sq_sysfs_attrs,

};

#ifdef CONFIG_PROC_FS

static int sq_mapping_read_proc(char *buf, char **start, off_t off,

				int len, int *eof, void *data)

static int __devinit sq_sysdev_add(struct sys_device *sysdev)

{

	struct list_head *pos;

	char *p = buf;

	unsigned int cpu = sysdev->id;

	struct kobject *kobj;

	list_for_each_prev(pos, &sq_mapping_list) {

		struct sq_mapping *entry;

	sq_kobject[cpu] = kzalloc(sizeof(struct kobject), GFP_KERNEL);

	if (unlikely(!sq_kobject[cpu]))

		return -ENOMEM;

		entry = list_entry(pos, typeof(*entry), list);

	kobj = sq_kobject[cpu];

	kobj->parent = &sysdev->kobj;

	kobject_set_name(kobj, "%s", "sq");

	kobj->ktype = &ktype_percpu_entry;

		p += sprintf(p, "%08lx-%08lx [%08lx]: %s\n", entry->sq_addr,

			     entry->sq_addr + entry->size - 1, entry->addr,

			     entry->name);

	return kobject_register(kobj);

}

	return p - buf;