xpc_channel.c

linux 内核源代码

/*
 * 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.
 *
 * Copyright (c) 2004-2006 Silicon Graphics, Inc.  All Rights Reserved.
 */


/*
 * Cross Partition Communication (XPC) channel support.
 *
 *	This is the part of XPC that manages the channels and
 *	sends/receives messages across them to/from other partitions.
 *
 */


#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/sched.h>
#include <linux/cache.h>
#include <linux/interrupt.h>
#include <linux/mutex.h>
#include <linux/completion.h>
#include <asm/sn/bte.h>
#include <asm/sn/sn_sal.h>
#include <asm/sn/xpc.h>


/*
 * Guarantee that the kzalloc'd memory is cacheline aligned.
 */
static void *
xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base)
{
	/* see if kzalloc will give us cachline aligned memory by default */
	*base = kzalloc(size, flags);
	if (*base == NULL) {
		return NULL;
	}
	if ((u64) *base == L1_CACHE_ALIGN((u64) *base)) {
		return *base;
	}
	kfree(*base);

	/* nope, we'll have to do it ourselves */
	*base = kzalloc(size + L1_CACHE_BYTES, flags);
	if (*base == NULL) {
		return NULL;
	}
	return (void *) L1_CACHE_ALIGN((u64) *base);
}


/*
 * Set up the initial values for the XPartition Communication channels.
 */
static void
xpc_initialize_channels(struct xpc_partition *part, partid_t partid)
{
	int ch_number;
	struct xpc_channel *ch;


	for (ch_number = 0; ch_number < part->nchannels; ch_number++) {
		ch = &part->channels[ch_number];

		ch->partid = partid;
		ch->number = ch_number;
		ch->flags = XPC_C_DISCONNECTED;

		ch->local_GP = &part->local_GPs[ch_number];
		ch->local_openclose_args =
					&part->local_openclose_args[ch_number];

		atomic_set(&ch->kthreads_assigned, 0);
		atomic_set(&ch->kthreads_idle, 0);
		atomic_set(&ch->kthreads_active, 0);

		atomic_set(&ch->references, 0);
		atomic_set(&ch->n_to_notify, 0);

		spin_lock_init(&ch->lock);
		mutex_init(&ch->msg_to_pull_mutex);
		init_completion(&ch->wdisconnect_wait);

		atomic_set(&ch->n_on_msg_allocate_wq, 0);
		init_waitqueue_head(&ch->msg_allocate_wq);
		init_waitqueue_head(&ch->idle_wq);
	}
}


/*
 * Setup the infrastructure necessary to support XPartition Communication
 * between the specified remote partition and the local one.
 */
enum xpc_retval
xpc_setup_infrastructure(struct xpc_partition *part)
{
	int ret, cpuid;
	struct timer_list *timer;
	partid_t partid = XPC_PARTID(part);


	/*
	 * Zero out MOST of the entry for this partition. Only the fields
	 * starting with `nchannels' will be zeroed. The preceding fields must
	 * remain `viable' across partition ups and downs, since they may be
	 * referenced during this memset() operation.
	 */
	memset(&part->nchannels, 0, sizeof(struct xpc_partition) -
				offsetof(struct xpc_partition, nchannels));

	/*
	 * Allocate all of the channel structures as a contiguous chunk of
	 * memory.
	 */
	part->channels = kzalloc(sizeof(struct xpc_channel) * XPC_NCHANNELS,
								GFP_KERNEL);
	if (part->channels == NULL) {
		dev_err(xpc_chan, "can't get memory for channels\n");
		return xpcNoMemory;
	}

	part->nchannels = XPC_NCHANNELS;


	/* allocate all the required GET/PUT values */

	part->local_GPs = xpc_kzalloc_cacheline_aligned(XPC_GP_SIZE,
					GFP_KERNEL, &part->local_GPs_base);
	if (part->local_GPs == NULL) {
		kfree(part->channels);
		part->channels = NULL;
		dev_err(xpc_chan, "can't get memory for local get/put "
			"values\n");
		return xpcNoMemory;
	}

	part->remote_GPs = xpc_kzalloc_cacheline_aligned(XPC_GP_SIZE,
					GFP_KERNEL, &part->remote_GPs_base);
	if (part->remote_GPs == NULL) {
		dev_err(xpc_chan, "can't get memory for remote get/put "
			"values\n");
		kfree(part->local_GPs_base);
		part->local_GPs = NULL;
		kfree(part->channels);
		part->channels = NULL;
		return xpcNoMemory;
	}


	/* allocate all the required open and close args */

	part->local_openclose_args = xpc_kzalloc_cacheline_aligned(
					XPC_OPENCLOSE_ARGS_SIZE, GFP_KERNEL,
					&part->local_openclose_args_base);
	if (part->local_openclose_args == NULL) {
		dev_err(xpc_chan, "can't get memory for local connect args\n");
		kfree(part->remote_GPs_base);
		part->remote_GPs = NULL;
		kfree(part->local_GPs_base);
		part->local_GPs = NULL;
		kfree(part->channels);
		part->channels = NULL;
		return xpcNoMemory;
	}

	part->remote_openclose_args = xpc_kzalloc_cacheline_aligned(
					XPC_OPENCLOSE_ARGS_SIZE, GFP_KERNEL,
					&part->remote_openclose_args_base);
	if (part->remote_openclose_args == NULL) {
		dev_err(xpc_chan, "can't get memory for remote connect args\n");
		kfree(part->local_openclose_args_base);
		part->local_openclose_args = NULL;
		kfree(part->remote_GPs_base);
		part->remote_GPs = NULL;
		kfree(part->local_GPs_base);
		part->local_GPs = NULL;
		kfree(part->channels);
		part->channels = NULL;
		return xpcNoMemory;
	}


	xpc_initialize_channels(part, partid);

	atomic_set(&part->nchannels_active, 0);
	atomic_set(&part->nchannels_engaged, 0);


	/* local_IPI_amo were set to 0 by an earlier memset() */

	/* Initialize this partitions AMO_t structure */
	part->local_IPI_amo_va = xpc_IPI_init(partid);

	spin_lock_init(&part->IPI_lock);

	atomic_set(&part->channel_mgr_requests, 1);
	init_waitqueue_head(&part->channel_mgr_wq);

	sprintf(part->IPI_owner, "xpc%02d", partid);
	ret = request_irq(SGI_XPC_NOTIFY, xpc_notify_IRQ_handler, IRQF_SHARED,
				part->IPI_owner, (void *) (u64) partid);
	if (ret != 0) {
		dev_err(xpc_chan, "can't register NOTIFY IRQ handler, "
			"errno=%d\n", -ret);
		kfree(part->remote_openclose_args_base);
		part->remote_openclose_args = NULL;
		kfree(part->local_openclose_args_base);
		part->local_openclose_args = NULL;
		kfree(part->remote_GPs_base);
		part->remote_GPs = NULL;
		kfree(part->local_GPs_base);
		part->local_GPs = NULL;
		kfree(part->channels);
		part->channels = NULL;
		return xpcLackOfResources;
	}

	/* Setup a timer to check for dropped IPIs */
	timer = &part->dropped_IPI_timer;
	init_timer(timer);
	timer->function = (void (*)(unsigned long)) xpc_dropped_IPI_check;
	timer->data = (unsigned long) part;
	timer->expires = jiffies + XPC_P_DROPPED_IPI_WAIT;
	add_timer(timer);

	/*
	 * With the setting of the partition setup_state to XPC_P_SETUP, we're
	 * declaring that this partition is ready to go.
	 */
	part->setup_state = XPC_P_SETUP;


	/*
	 * Setup the per partition specific variables required by the
	 * remote partition to establish channel connections with us.
	 *
	 * The setting of the magic # indicates that these per partition
	 * specific variables are ready to be used.
	 */
	xpc_vars_part[partid].GPs_pa = __pa(part->local_GPs);
	xpc_vars_part[partid].openclose_args_pa =
					__pa(part->local_openclose_args);
	xpc_vars_part[partid].IPI_amo_pa = __pa(part->local_IPI_amo_va);
	cpuid = raw_smp_processor_id();	/* any CPU in this partition will do */
	xpc_vars_part[partid].IPI_nasid = cpuid_to_nasid(cpuid);
	xpc_vars_part[partid].IPI_phys_cpuid = cpu_physical_id(cpuid);
	xpc_vars_part[partid].nchannels = part->nchannels;
	xpc_vars_part[partid].magic = XPC_VP_MAGIC1;

	return xpcSuccess;
}


/*
 * Create a wrapper that hides the underlying mechanism for pulling a cacheline
 * (or multiple cachelines) from a remote partition.
 *
 * src must be a cacheline aligned physical address on the remote partition.
 * dst must be a cacheline aligned virtual address on this partition.
 * cnt must be an cacheline sized
 */
static enum xpc_retval
xpc_pull_remote_cachelines(struct xpc_partition *part, void *dst,
				const void *src, size_t cnt)
{
	bte_result_t bte_ret;


	DBUG_ON((u64) src != L1_CACHE_ALIGN((u64) src));
	DBUG_ON((u64) dst != L1_CACHE_ALIGN((u64) dst));
	DBUG_ON(cnt != L1_CACHE_ALIGN(cnt));

	if (part->act_state == XPC_P_DEACTIVATING) {
		return part->reason;
	}

	bte_ret = xp_bte_copy((u64) src, (u64) dst, (u64) cnt,
					(BTE_NORMAL | BTE_WACQUIRE), NULL);
	if (bte_ret == BTE_SUCCESS) {
		return xpcSuccess;
	}

	dev_dbg(xpc_chan, "xp_bte_copy() from partition %d failed, ret=%d\n",
		XPC_PARTID(part), bte_ret);

	return xpc_map_bte_errors(bte_ret);
}


/*
 * Pull the remote per partition specific variables from the specified
 * partition.
 */
enum xpc_retval
xpc_pull_remote_vars_part(struct xpc_partition *part)
{
	u8 buffer[L1_CACHE_BYTES * 2];
	struct xpc_vars_part *pulled_entry_cacheline =
			(struct xpc_vars_part *) L1_CACHE_ALIGN((u64) buffer);
	struct xpc_vars_part *pulled_entry;
	u64 remote_entry_cacheline_pa, remote_entry_pa;
	partid_t partid = XPC_PARTID(part);
	enum xpc_retval ret;


	/* pull the cacheline that contains the variables we're interested in */

	DBUG_ON(part->remote_vars_part_pa !=
				L1_CACHE_ALIGN(part->remote_vars_part_pa));
	DBUG_ON(sizeof(struct xpc_vars_part) != L1_CACHE_BYTES / 2);

	remote_entry_pa = part->remote_vars_part_pa +
			sn_partition_id * sizeof(struct xpc_vars_part);

	remote_entry_cacheline_pa = (remote_entry_pa & ~(L1_CACHE_BYTES - 1));

	pulled_entry = (struct xpc_vars_part *) ((u64) pulled_entry_cacheline +
				(remote_entry_pa & (L1_CACHE_BYTES - 1)));

	ret = xpc_pull_remote_cachelines(part, pulled_entry_cacheline,
					(void *) remote_entry_cacheline_pa,
					L1_CACHE_BYTES);
	if (ret != xpcSuccess) {
		dev_dbg(xpc_chan, "failed to pull XPC vars_part from "
			"partition %d, ret=%d\n", partid, ret);
		return ret;
	}


	/* see if they've been set up yet */

	if (pulled_entry->magic != XPC_VP_MAGIC1 &&
				pulled_entry->magic != XPC_VP_MAGIC2) {

		if (pulled_entry->magic != 0) {
			dev_dbg(xpc_chan, "partition %d's XPC vars_part for "
				"partition %d has bad magic value (=0x%lx)\n",
				partid, sn_partition_id, pulled_entry->magic);
			return xpcBadMagic;
		}

		/* they've not been initialized yet */
		return xpcRetry;
	}

	if (xpc_vars_part[partid].magic == XPC_VP_MAGIC1) {

		/* validate the variables */

		if (pulled_entry->GPs_pa == 0 ||
				pulled_entry->openclose_args_pa == 0 ||
					pulled_entry->IPI_amo_pa == 0) {

			dev_err(xpc_chan, "partition %d's XPC vars_part for "
				"partition %d are not valid\n", partid,
				sn_partition_id);
			return xpcInvalidAddress;
		}

		/* the variables we imported look to be valid */

		part->remote_GPs_pa = pulled_entry->GPs_pa;
		part->remote_openclose_args_pa =
					pulled_entry->openclose_args_pa;
		part->remote_IPI_amo_va =
				      (AMO_t *) __va(pulled_entry->IPI_amo_pa);
		part->remote_IPI_nasid = pulled_entry->IPI_nasid;
		part->remote_IPI_phys_cpuid = pulled_entry->IPI_phys_cpuid;

		if (part->nchannels > pulled_entry->nchannels) {
			part->nchannels = pulled_entry->nchannels;
		}

		/* let the other side know that we've pulled their variables */

		xpc_vars_part[partid].magic = XPC_VP_MAGIC2;
	}

	if (pulled_entry->magic == XPC_VP_MAGIC1) {
		return xpcRetry;
	}

	return xpcSuccess;
}


/*
 * Get the IPI flags and pull the openclose args and/or remote GPs as needed.
 */
static u64
xpc_get_IPI_flags(struct xpc_partition *part)
{
	unsigned long irq_flags;
	u64 IPI_amo;
	enum xpc_retval ret;


	/*
	 * See if there are any IPI flags to be handled.
	 */

	spin_lock_irqsave(&part->IPI_lock, irq_flags);
	if ((IPI_amo = part->local_IPI_amo) != 0) {
		part->local_IPI_amo = 0;
	}
	spin_unlock_irqrestore(&part->IPI_lock, irq_flags);


	if (XPC_ANY_OPENCLOSE_IPI_FLAGS_SET(IPI_amo)) {
		ret = xpc_pull_remote_cachelines(part,
					part->remote_openclose_args,
					(void *) part->remote_openclose_args_pa,
					XPC_OPENCLOSE_ARGS_SIZE);
		if (ret != xpcSuccess) {
			XPC_DEACTIVATE_PARTITION(part, ret);

			dev_dbg(xpc_chan, "failed to pull openclose args from "
				"partition %d, ret=%d\n", XPC_PARTID(part),
				ret);

			/* don't bother processing IPIs anymore */
			IPI_amo = 0;
		}
	}

	if (XPC_ANY_MSG_IPI_FLAGS_SET(IPI_amo)) {
		ret = xpc_pull_remote_cachelines(part, part->remote_GPs,
						(void *) part->remote_GPs_pa,
						XPC_GP_SIZE);
		if (ret != xpcSuccess) {
			XPC_DEACTIVATE_PARTITION(part, ret);

			dev_dbg(xpc_chan, "failed to pull GPs from partition "
				"%d, ret=%d\n", XPC_PARTID(part), ret);

			/* don't bother processing IPIs anymore */
			IPI_amo = 0;
		}
	}

	return IPI_amo;
}


/*
 * Allocate the local message queue and the notify queue.
 */
static enum xpc_retval
xpc_allocate_local_msgqueue(struct xpc_channel *ch)
{
	unsigned long irq_flags;
	int nentries;
	size_t nbytes;


	// >>> may want to check for ch->flags & XPC_C_DISCONNECTING between
	// >>> iterations of the for-loop, bail if set?

	// >>> should we impose a minimum #of entries? like 4 or 8?
	for (nentries = ch->local_nentries; nentries > 0; nentries--) {

		nbytes = nentries * ch->msg_size;
		ch->local_msgqueue = xpc_kzalloc_cacheline_aligned(nbytes,
						GFP_KERNEL,
						&ch->local_msgqueue_base);
		if (ch->local_msgqueue == NULL) {
			continue;
		}

		nbytes = nentries * sizeof(struct xpc_notify);
		ch->notify_queue = kzalloc(nbytes, GFP_KERNEL);