vote.c

ocfs1.2.7 源码

/* -*- mode: c; c-basic-offset: 8; -*-
 * vim: noexpandtab sw=8 ts=8 sts=0:
 *
 * vote.c
 *
 * description here
 *
 * Copyright (C) 2003, 2004 Oracle.  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; either
 * version 2 of the License, or (at your option) any later version.
 *
 * 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.  See the GNU
 * General Public License for more details.
 *
 * You should have received a copy of the GNU General Public
 * License along with this program; if not, write to the
 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 * Boston, MA 021110-1307, USA.
 */

#include <linux/types.h>
#include <linux/slab.h>
#include <linux/highmem.h>
#include <linux/smp_lock.h>
#include <linux/kthread.h>

#include <cluster/heartbeat.h>
#include <cluster/nodemanager.h>
#include <cluster/tcp.h>

#include <dlm/dlmapi.h>

#define MLOG_MASK_PREFIX ML_VOTE
#include <cluster/masklog.h>

#include "ocfs2.h"

#include "alloc.h"
#include "dlmglue.h"
#include "extent_map.h"
#include "heartbeat.h"
#include "inode.h"
#include "journal.h"
#include "slot_map.h"
#include "vote.h"

#include "buffer_head_io.h"

#define OCFS2_MESSAGE_TYPE_VOTE     (0x1)
#define OCFS2_MESSAGE_TYPE_RESPONSE (0x2)
struct ocfs2_msg_hdr
{
	__be32 h_response_id; /* used to lookup message handle on sending
			    * node. */
	__be32 h_request;
	__be64 h_blkno;
	__be32 h_generation;
	__be32 h_node_num;    /* node sending this particular message. */
};

/* OCFS2_MAX_FILENAME_LEN is 255 characters, but we want to align this
 * for the network. */
#define OCFS2_VOTE_FILENAME_LEN 256
struct ocfs2_vote_msg
{
	struct ocfs2_msg_hdr v_hdr;
	union {
		__be32 v_generic1;
		__be32 v_orphaned_slot;	/* Used during delete votes */
		__be32 v_nlink;		/* Used during unlink votes */
	} md1;				/* Message type dependant 1 */
	__be32 v_unlink_namelen;
	__be64 v_unlink_parent;
	u8  v_unlink_dirent[OCFS2_VOTE_FILENAME_LEN];
};

/* Responses are given these values to maintain backwards
 * compatibility with older ocfs2 versions */
#define OCFS2_RESPONSE_OK		(0)
#define OCFS2_RESPONSE_BUSY		(-16)
#define OCFS2_RESPONSE_BAD_MSG		(-22)

struct ocfs2_response_msg
{
	struct ocfs2_msg_hdr r_hdr;
	__be32 r_response;
	__be32 r_orphaned_slot;
};

struct ocfs2_vote_work {
	struct list_head   w_list;
	struct ocfs2_vote_msg w_msg;
};

enum ocfs2_vote_request {
	OCFS2_VOTE_REQ_INVALID = 0,
	OCFS2_VOTE_REQ_DELETE,
	OCFS2_VOTE_REQ_UNLINK,
	OCFS2_VOTE_REQ_RENAME,
	OCFS2_VOTE_REQ_MOUNT,
	OCFS2_VOTE_REQ_UMOUNT,
	OCFS2_VOTE_REQ_LAST
};

static inline int ocfs2_is_valid_vote_request(int request)
{
	return OCFS2_VOTE_REQ_INVALID < request &&
		request < OCFS2_VOTE_REQ_LAST;
}

typedef void (*ocfs2_net_response_callback)(void *priv,
					    struct ocfs2_response_msg *resp);
struct ocfs2_net_response_cb {
	ocfs2_net_response_callback	rc_cb;
	void				*rc_priv;
};

struct ocfs2_net_wait_ctxt {
	struct list_head        n_list;
	u32                     n_response_id;
	wait_queue_head_t       n_event;
	struct ocfs2_node_map   n_node_map;
	int                     n_response; /* an agreggate response. 0 if
					     * all nodes are go, < 0 on any
					     * negative response from any
					     * node or network error. */
	struct ocfs2_net_response_cb *n_callback;
};

static void ocfs2_process_mount_request(struct ocfs2_super *osb,
					unsigned int node_num)
{
	mlog(0, "MOUNT vote from node %u\n", node_num);
	/* The other node only sends us this message when he has an EX
	 * on the superblock, so our recovery threads (if having been
	 * launched) are waiting on it.*/
	ocfs2_recovery_map_clear(osb, node_num);
	ocfs2_node_map_set_bit(osb, &osb->mounted_map, node_num);

	/* We clear the umount map here because a node may have been
	 * previously mounted, safely unmounted but never stopped
	 * heartbeating - in which case we'd have a stale entry. */
	ocfs2_node_map_clear_bit(osb, &osb->umount_map, node_num);
}

static void ocfs2_process_umount_request(struct ocfs2_super *osb,
					 unsigned int node_num)
{
	mlog(0, "UMOUNT vote from node %u\n", node_num);
	ocfs2_node_map_clear_bit(osb, &osb->mounted_map, node_num);
	ocfs2_node_map_set_bit(osb, &osb->umount_map, node_num);
}

void ocfs2_mark_inode_remotely_deleted(struct inode *inode, int deleting_node)
{
	struct ocfs2_inode_info *oi = OCFS2_I(inode);

	assert_spin_locked(&oi->ip_lock);
	/* We set the SKIP_DELETE flag on the inode so we don't try to
	 * delete it in delete_inode ourselves, thus avoiding
	 * unecessary lock pinging. If the other node failed to wipe
	 * the inode as a result of a crash, then recovery will pick
	 * up the slack. */
	oi->ip_flags |= OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE;
	if (deleting_node != O2NM_INVALID_NODE_NUM)
		oi->ip_deleting_node = deleting_node;
}

static int ocfs2_process_delete_request(struct inode *inode,
					int *orphaned_slot,
					int deleting_node)
{
	int response = OCFS2_RESPONSE_BUSY;

	mlog(0, "DELETE vote on inode %lu, read lnk_cnt = %u, slot = %d\n",
	     inode->i_ino, inode->i_nlink, *orphaned_slot);

	spin_lock(&OCFS2_I(inode)->ip_lock);

	/* Whatever our vote response is, we want to make sure that
	 * the orphaned slot is recorded properly on this node *and*
	 * on the requesting node. Technically, if the requesting node
	 * did not know which slot the inode is orphaned in but we
	 * respond with BUSY he doesn't actually need the orphaned
	 * slot, but it doesn't hurt to do it here anyway. */
	if ((*orphaned_slot) != OCFS2_INVALID_SLOT) {
		mlog_bug_on_msg(OCFS2_I(inode)->ip_orphaned_slot !=
				OCFS2_INVALID_SLOT &&
				OCFS2_I(inode)->ip_orphaned_slot !=
				(*orphaned_slot),
				"Inode %"MLFu64": This node thinks it's "
				"orphaned in slot %d, messaged it's in %d\n",
				OCFS2_I(inode)->ip_blkno,
				OCFS2_I(inode)->ip_orphaned_slot,
				*orphaned_slot);

		mlog(0, "Setting orphaned slot for inode %"MLFu64" to %d\n",
		     OCFS2_I(inode)->ip_blkno, *orphaned_slot);

		OCFS2_I(inode)->ip_orphaned_slot = *orphaned_slot;
	} else {
		mlog(0, "Sending back orphaned slot %d for inode %"MLFu64"\n",
		     OCFS2_I(inode)->ip_orphaned_slot,
		     OCFS2_I(inode)->ip_blkno);

		*orphaned_slot = OCFS2_I(inode)->ip_orphaned_slot;
	}

	/* vote no if the file is still open. */
	if (OCFS2_I(inode)->ip_open_count) {
		mlog(0, "open count = %u\n",
		     OCFS2_I(inode)->ip_open_count);
		spin_unlock(&OCFS2_I(inode)->ip_lock);
		goto done;
	}
	spin_unlock(&OCFS2_I(inode)->ip_lock);

	/* directories are a bit ugly... What if someone is sitting in
	 * it? We want to make sure the inode is removed completely as
	 * a result of the iput in process_vote. */
	if (S_ISDIR(inode->i_mode) && (atomic_read(&inode->i_count) != 1)) {
		mlog(0, "i_count = %u\n", atomic_read(&inode->i_count));
		goto done;
	}

	if (filemap_fdatawrite(inode->i_mapping)) {
		mlog(ML_ERROR, "Could not sync inode %"MLFu64" for delete!\n",
		     OCFS2_I(inode)->ip_blkno);
		goto done;
	}
	sync_mapping_buffers(inode->i_mapping);
	truncate_inode_pages(inode->i_mapping, 0);
	ocfs2_extent_map_trunc(inode, 0);

	spin_lock(&OCFS2_I(inode)->ip_lock);
	/* double check open count - someone might have raced this
	 * thread into ocfs2_file_open while we were writing out
	 * data. If we're to allow a wipe of this inode now, we *must*
	 * hold the spinlock until we've marked it. */
	if (OCFS2_I(inode)->ip_open_count) {
		mlog(0, "Raced to wipe! open count = %u\n",
		     OCFS2_I(inode)->ip_open_count);
		spin_unlock(&OCFS2_I(inode)->ip_lock);
		goto done;
	}

	/* Mark the inode as being wiped from disk. */
	ocfs2_mark_inode_remotely_deleted(inode, deleting_node);
	spin_unlock(&OCFS2_I(inode)->ip_lock);

	/* Not sure this is necessary anymore. */
	d_prune_aliases(inode);

	/* If we get here, then we're voting 'yes', so commit the
	 * delete on our side. */
	response = OCFS2_RESPONSE_OK;
done:
	return response;
}

static int ocfs2_match_dentry(struct dentry *dentry,
			      u64 parent_blkno,
			      unsigned int namelen,
			      const char *name)
{
	struct inode *parent;

	if (!dentry->d_parent) {
		mlog(0, "Detached from parent.\n");
		return 0;
	}

	parent = dentry->d_parent->d_inode;
	/* Negative parent dentry? */
	if (!parent)
		return 0;

	/* Name is in a different directory. */
	if (OCFS2_I(parent)->ip_blkno != parent_blkno)
		return 0;

	if (dentry->d_name.len != namelen)
		return 0;

	/* comparison above guarantees this is safe. */
	if (memcmp(dentry->d_name.name, name, namelen))
		return 0;

	return 1;
}

static void ocfs2_process_dentry_request(struct inode *inode,
					 int rename,
					 unsigned int new_nlink,
					 u64 parent_blkno,
					 unsigned int namelen,
					 const char *name)
{
	struct dentry *dentry = NULL;
	struct list_head *p;
	struct ocfs2_inode_info *oi = OCFS2_I(inode);

	mlog(0, "parent %"MLFu64", namelen = %u, name = %.*s\n", parent_blkno,
	     namelen, namelen, name);

	spin_lock(&dcache_lock);

	/* Another node is removing this name from the system. It is
	 * up to us to find the corresponding dentry and if it exists,
	 * unhash it from the dcache. */
	list_for_each(p, &inode->i_dentry) {
		dentry = list_entry(p, struct dentry, d_alias);

		if (ocfs2_match_dentry(dentry, parent_blkno, namelen, name)) {
			mlog(0, "dentry found: %.*s\n",
			     dentry->d_name.len, dentry->d_name.name);

			dget_locked(dentry);
			break;
		}

		dentry = NULL;
	}

	spin_unlock(&dcache_lock);

	if (dentry) {
		d_delete(dentry);
		dput(dentry);
	}

	/* rename votes don't send link counts */
	if (!rename) {
		mlog(0, "new_nlink = %u\n", new_nlink);

		/* We don't have the proper locks here to directly
		 * change i_nlink and besides, the vote is sent
		 * *before* the operation so it may have failed on the
		 * other node. This passes a hint to ocfs2_drop_inode
		 * to force ocfs2_delete_inode, who will take the
		 * proper cluster locks to sort things out. */
		if (new_nlink == 0) {
			spin_lock(&oi->ip_lock);
			oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;

#ifdef OCFS2_DELETE_INODE_WORKAROUND
			/* Do a sync now as we can't be sure whether
			 * the inode will actually be orphaned or
			 * not. We condition this on the open count as
			 * otherwise, ocfs2_file_release will handle
			 * it for us. */
			if (!oi->ip_open_count) {
				spin_unlock(&oi->ip_lock);
				write_inode_now(inode, 1);
				/* strange indentation past the
				 * 'else', but I want to keep the non
				 * hack code purty :) */
			} else
#endif
			spin_unlock(&OCFS2_I(inode)->ip_lock);
		}
	}
}

static void ocfs2_process_vote(struct ocfs2_super *osb,
			       struct ocfs2_vote_msg *msg)
{
	int net_status, vote_response;
	int orphaned_slot = 0;
	int deleting_node = O2NM_INVALID_NODE_NUM;
	int rename = 0;
	unsigned int node_num, generation, new_nlink, namelen;
	u64 blkno, parent_blkno;
	enum ocfs2_vote_request request;
	struct inode *inode = NULL;
	struct ocfs2_msg_hdr *hdr = &msg->v_hdr;
	struct ocfs2_response_msg response;

	/* decode the network mumbo jumbo into local variables. */
	request = be32_to_cpu(hdr->h_request);
	blkno = be64_to_cpu(hdr->h_blkno);
	generation = be32_to_cpu(hdr->h_generation);
	node_num = be32_to_cpu(hdr->h_node_num);
	if (request == OCFS2_VOTE_REQ_DELETE) {
		orphaned_slot = be32_to_cpu(msg->md1.v_orphaned_slot);
		deleting_node = be32_to_cpu(msg->v_hdr.h_node_num);
	}

	mlog(0, "processing vote: request = %u, blkno = %"MLFu64", "
	     "generation = %u, node_num = %u, priv1 = %u\n", request,
	     blkno, generation, node_num, be32_to_cpu(msg->md1.v_generic1));

	if (!ocfs2_is_valid_vote_request(request)) {
		mlog(ML_ERROR, "Invalid vote request %d from node %u\n",
		     request, node_num);
		vote_response = OCFS2_RESPONSE_BAD_MSG;
		goto respond;
	}

	vote_response = OCFS2_RESPONSE_OK;

	switch (request) {
	case OCFS2_VOTE_REQ_UMOUNT:
		ocfs2_process_umount_request(osb, node_num);