fs/dlm/member.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /******************************************************************************
0003 *******************************************************************************
0004 **
0005 **  Copyright (C) 2005-2011 Red Hat, Inc.  All rights reserved.
0006 **
0007 **
0008 *******************************************************************************
0009 ******************************************************************************/
0010
0011 #include "dlm_internal.h"
0012 #include "lockspace.h"
0013 #include "member.h"
0014 #include "recoverd.h"
0015 #include "recover.h"
0016 #include "rcom.h"
0017 #include "config.h"
0018 #include "midcomms.h"
0019 #include "lowcomms.h"
0020
0021 int dlm_slots_version(struct dlm_header *h)
0022 {
0023     if ((le32_to_cpu(h->h_version) & 0x0000FFFF) < DLM_HEADER_SLOTS)
0024         return 0;
0025     return 1;
0026 }
0027
0028 void dlm_slot_save(struct dlm_ls *ls, struct dlm_rcom *rc,
0029            struct dlm_member *memb)
0030 {
0031     struct rcom_config *rf = (struct rcom_config *)rc->rc_buf;
0032
0033     if (!dlm_slots_version(&rc->rc_header))
0034         return;
0035
0036     memb->slot = le16_to_cpu(rf->rf_our_slot);
0037     memb->generation = le32_to_cpu(rf->rf_generation);
0038 }
0039
0040 void dlm_slots_copy_out(struct dlm_ls *ls, struct dlm_rcom *rc)
0041 {
0042     struct dlm_slot *slot;
0043     struct rcom_slot *ro;
0044     int i;
0045
0046     ro = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config));
0047
0048     /* ls_slots array is sparse, but not rcom_slots */
0049
0050     for (i = 0; i < ls->ls_slots_size; i++) {
0051         slot = &ls->ls_slots[i];
0052         if (!slot->nodeid)
0053             continue;
0054         ro->ro_nodeid = cpu_to_le32(slot->nodeid);
0055         ro->ro_slot = cpu_to_le16(slot->slot);
0056         ro++;
0057     }
0058 }
0059
0060 #define SLOT_DEBUG_LINE 128
0061
0062 static void log_slots(struct dlm_ls *ls, uint32_t gen, int num_slots,
0063               struct rcom_slot *ro0, struct dlm_slot *array,
0064               int array_size)
0065 {
0066     char line[SLOT_DEBUG_LINE];
0067     int len = SLOT_DEBUG_LINE - 1;
0068     int pos = 0;
0069     int ret, i;
0070
0071     memset(line, 0, sizeof(line));
0072
0073     if (array) {
0074         for (i = 0; i < array_size; i++) {
0075             if (!array[i].nodeid)
0076                 continue;
0077
0078             ret = snprintf(line + pos, len - pos, " %d:%d",
0079                        array[i].slot, array[i].nodeid);
0080             if (ret >= len - pos)
0081                 break;
0082             pos += ret;
0083         }
0084     } else if (ro0) {
0085         for (i = 0; i < num_slots; i++) {
0086             ret = snprintf(line + pos, len - pos, " %d:%d",
0087                        ro0[i].ro_slot, ro0[i].ro_nodeid);
0088             if (ret >= len - pos)
0089                 break;
0090             pos += ret;
0091         }
0092     }
0093
0094     log_rinfo(ls, "generation %u slots %d%s", gen, num_slots, line);
0095 }
0096
0097 int dlm_slots_copy_in(struct dlm_ls *ls)
0098 {
0099     struct dlm_member *memb;
0100     struct dlm_rcom *rc = ls->ls_recover_buf;
0101     struct rcom_config *rf = (struct rcom_config *)rc->rc_buf;
0102     struct rcom_slot *ro0, *ro;
0103     int our_nodeid = dlm_our_nodeid();
0104     int i, num_slots;
0105     uint32_t gen;
0106
0107     if (!dlm_slots_version(&rc->rc_header))
0108         return -1;
0109
0110     gen = le32_to_cpu(rf->rf_generation);
0111     if (gen <= ls->ls_generation) {
0112         log_error(ls, "dlm_slots_copy_in gen %u old %u",
0113               gen, ls->ls_generation);
0114     }
0115     ls->ls_generation = gen;
0116
0117     num_slots = le16_to_cpu(rf->rf_num_slots);
0118     if (!num_slots)
0119         return -1;
0120
0121     ro0 = (struct rcom_slot *)(rc->rc_buf + sizeof(struct rcom_config));
0122
0123     log_slots(ls, gen, num_slots, ro0, NULL, 0);
0124
0125     list_for_each_entry(memb, &ls->ls_nodes, list) {
0126         for (i = 0, ro = ro0; i < num_slots; i++, ro++) {
0127             if (le32_to_cpu(ro->ro_nodeid) != memb->nodeid)
0128                 continue;
0129             memb->slot = le16_to_cpu(ro->ro_slot);
0130             memb->slot_prev = memb->slot;
0131             break;
0132         }
0133
0134         if (memb->nodeid == our_nodeid) {
0135             if (ls->ls_slot && ls->ls_slot != memb->slot) {
0136                 log_error(ls, "dlm_slots_copy_in our slot "
0137                       "changed %d %d", ls->ls_slot,
0138                       memb->slot);
0139                 return -1;
0140             }
0141
0142             if (!ls->ls_slot)
0143                 ls->ls_slot = memb->slot;
0144         }
0145
0146         if (!memb->slot) {
0147             log_error(ls, "dlm_slots_copy_in nodeid %d no slot",
0148                    memb->nodeid);
0149             return -1;
0150         }
0151     }
0152
0153     return 0;
0154 }
0155
0156 /* for any nodes that do not support slots, we will not have set memb->slot
0157    in wait_status_all(), so memb->slot will remain -1, and we will not
0158    assign slots or set ls_num_slots here */
0159
0160 int dlm_slots_assign(struct dlm_ls *ls, int *num_slots, int *slots_size,
0161              struct dlm_slot **slots_out, uint32_t *gen_out)
0162 {
0163     struct dlm_member *memb;
0164     struct dlm_slot *array;
0165     int our_nodeid = dlm_our_nodeid();
0166     int array_size, max_slots, i;
0167     int need = 0;
0168     int max = 0;
0169     int num = 0;
0170     uint32_t gen = 0;
0171
0172     /* our own memb struct will have slot -1 gen 0 */
0173
0174     list_for_each_entry(memb, &ls->ls_nodes, list) {
0175         if (memb->nodeid == our_nodeid) {
0176             memb->slot = ls->ls_slot;
0177             memb->generation = ls->ls_generation;
0178             break;
0179         }
0180     }
0181
0182     list_for_each_entry(memb, &ls->ls_nodes, list) {
0183         if (memb->generation > gen)
0184             gen = memb->generation;
0185
0186         /* node doesn't support slots */
0187
0188         if (memb->slot == -1)
0189             return -1;
0190
0191         /* node needs a slot assigned */
0192
0193         if (!memb->slot)
0194             need++;
0195
0196         /* node has a slot assigned */
0197
0198         num++;
0199
0200         if (!max || max < memb->slot)
0201             max = memb->slot;
0202
0203         /* sanity check, once slot is assigned it shouldn't change */
0204
0205         if (memb->slot_prev && memb->slot && memb->slot_prev != memb->slot) {
0206             log_error(ls, "nodeid %d slot changed %d %d",
0207                   memb->nodeid, memb->slot_prev, memb->slot);
0208             return -1;
0209         }
0210         memb->slot_prev = memb->slot;
0211     }
0212
0213     array_size = max + need;
0214     array = kcalloc(array_size, sizeof(*array), GFP_NOFS);
0215     if (!array)
0216         return -ENOMEM;
0217
0218     num = 0;
0219
0220     /* fill in slots (offsets) that are used */
0221
0222     list_for_each_entry(memb, &ls->ls_nodes, list) {
0223         if (!memb->slot)
0224             continue;
0225
0226         if (memb->slot > array_size) {
0227             log_error(ls, "invalid slot number %d", memb->slot);
0228             kfree(array);
0229             return -1;
0230         }
0231
0232         array[memb->slot - 1].nodeid = memb->nodeid;
0233         array[memb->slot - 1].slot = memb->slot;
0234         num++;
0235     }
0236
0237     /* assign new slots from unused offsets */
0238
0239     list_for_each_entry(memb, &ls->ls_nodes, list) {
0240         if (memb->slot)
0241             continue;
0242
0243         for (i = 0; i < array_size; i++) {
0244             if (array[i].nodeid)
0245                 continue;
0246
0247             memb->slot = i + 1;
0248             memb->slot_prev = memb->slot;
0249             array[i].nodeid = memb->nodeid;
0250             array[i].slot = memb->slot;
0251             num++;
0252
0253             if (!ls->ls_slot && memb->nodeid == our_nodeid)
0254                 ls->ls_slot = memb->slot;
0255             break;
0256         }
0257
0258         if (!memb->slot) {
0259             log_error(ls, "no free slot found");
0260             kfree(array);
0261             return -1;
0262         }
0263     }
0264
0265     gen++;
0266
0267     log_slots(ls, gen, num, NULL, array, array_size);
0268
0269     max_slots = (DLM_MAX_APP_BUFSIZE - sizeof(struct dlm_rcom) -
0270              sizeof(struct rcom_config)) / sizeof(struct rcom_slot);
0271
0272     if (num > max_slots) {
0273         log_error(ls, "num_slots %d exceeds max_slots %d",
0274               num, max_slots);
0275         kfree(array);
0276         return -1;
0277     }
0278
0279     *gen_out = gen;
0280     *slots_out = array;
0281     *slots_size = array_size;
0282     *num_slots = num;
0283     return 0;
0284 }
0285
0286 static void add_ordered_member(struct dlm_ls *ls, struct dlm_member *new)
0287 {
0288     struct dlm_member *memb = NULL;
0289     struct list_head *tmp;
0290     struct list_head *newlist = &new->list;
0291     struct list_head *head = &ls->ls_nodes;
0292
0293     list_for_each(tmp, head) {
0294         memb = list_entry(tmp, struct dlm_member, list);
0295         if (new->nodeid < memb->nodeid)
0296             break;
0297     }
0298
0299     if (!memb)
0300         list_add_tail(newlist, head);
0301     else {
0302         /* FIXME: can use list macro here */
0303         newlist->prev = tmp->prev;
0304         newlist->next = tmp;
0305         tmp->prev->next = newlist;
0306         tmp->prev = newlist;
0307     }
0308 }
0309
0310 static int dlm_add_member(struct dlm_ls *ls, struct dlm_config_node *node)
0311 {
0312     struct dlm_member *memb;
0313     int error;
0314
0315     memb = kzalloc(sizeof(*memb), GFP_NOFS);
0316     if (!memb)
0317         return -ENOMEM;
0318
0319     error = dlm_lowcomms_connect_node(node->nodeid);
0320     if (error < 0) {
0321         kfree(memb);
0322         return error;
0323     }
0324
0325     memb->nodeid = node->nodeid;
0326     memb->weight = node->weight;
0327     memb->comm_seq = node->comm_seq;
0328     dlm_midcomms_add_member(node->nodeid);
0329     add_ordered_member(ls, memb);
0330     ls->ls_num_nodes++;
0331     return 0;
0332 }
0333
0334 static struct dlm_member *find_memb(struct list_head *head, int nodeid)
0335 {
0336     struct dlm_member *memb;
0337
0338     list_for_each_entry(memb, head, list) {
0339         if (memb->nodeid == nodeid)
0340             return memb;
0341     }
0342     return NULL;
0343 }
0344
0345 int dlm_is_member(struct dlm_ls *ls, int nodeid)
0346 {
0347     if (find_memb(&ls->ls_nodes, nodeid))
0348         return 1;
0349     return 0;
0350 }
0351
0352 int dlm_is_removed(struct dlm_ls *ls, int nodeid)
0353 {
0354     if (find_memb(&ls->ls_nodes_gone, nodeid))
0355         return 1;
0356     return 0;
0357 }
0358
0359 static void clear_memb_list(struct list_head *head,
0360                 void (*after_del)(int nodeid))
0361 {
0362     struct dlm_member *memb;
0363
0364     while (!list_empty(head)) {
0365         memb = list_entry(head->next, struct dlm_member, list);
0366         list_del(&memb->list);
0367         if (after_del)
0368             after_del(memb->nodeid);
0369         kfree(memb);
0370     }
0371 }
0372
0373 static void clear_members_cb(int nodeid)
0374 {
0375     dlm_midcomms_remove_member(nodeid);
0376 }
0377
0378 void dlm_clear_members(struct dlm_ls *ls)
0379 {
0380     clear_memb_list(&ls->ls_nodes, clear_members_cb);
0381     ls->ls_num_nodes = 0;
0382 }
0383
0384 void dlm_clear_members_gone(struct dlm_ls *ls)
0385 {
0386     clear_memb_list(&ls->ls_nodes_gone, NULL);
0387 }
0388
0389 static void make_member_array(struct dlm_ls *ls)
0390 {
0391     struct dlm_member *memb;
0392     int i, w, x = 0, total = 0, all_zero = 0, *array;
0393
0394     kfree(ls->ls_node_array);
0395     ls->ls_node_array = NULL;
0396
0397     list_for_each_entry(memb, &ls->ls_nodes, list) {
0398         if (memb->weight)
0399             total += memb->weight;
0400     }
0401
0402     /* all nodes revert to weight of 1 if all have weight 0 */
0403
0404     if (!total) {
0405         total = ls->ls_num_nodes;
0406         all_zero = 1;
0407     }
0408
0409     ls->ls_total_weight = total;
0410     array = kmalloc_array(total, sizeof(*array), GFP_NOFS);
0411     if (!array)
0412         return;
0413
0414     list_for_each_entry(memb, &ls->ls_nodes, list) {
0415         if (!all_zero && !memb->weight)
0416             continue;
0417
0418         if (all_zero)
0419             w = 1;
0420         else
0421             w = memb->weight;
0422
0423         DLM_ASSERT(x < total, printk("total %d x %d\n", total, x););
0424
0425         for (i = 0; i < w; i++)
0426             array[x++] = memb->nodeid;
0427     }
0428
0429     ls->ls_node_array = array;
0430 }
0431
0432 /* send a status request to all members just to establish comms connections */
0433
0434 static int ping_members(struct dlm_ls *ls)
0435 {
0436     struct dlm_member *memb;
0437     int error = 0;
0438
0439     list_for_each_entry(memb, &ls->ls_nodes, list) {
0440         if (dlm_recovery_stopped(ls)) {
0441             error = -EINTR;
0442             break;
0443         }
0444         error = dlm_rcom_status(ls, memb->nodeid, 0);
0445         if (error)
0446             break;
0447     }
0448     if (error)
0449         log_rinfo(ls, "ping_members aborted %d last nodeid %d",
0450               error, ls->ls_recover_nodeid);
0451     return error;
0452 }
0453
0454 static void dlm_lsop_recover_prep(struct dlm_ls *ls)
0455 {
0456     if (!ls->ls_ops || !ls->ls_ops->recover_prep)
0457         return;
0458     ls->ls_ops->recover_prep(ls->ls_ops_arg);
0459 }
0460
0461 static void dlm_lsop_recover_slot(struct dlm_ls *ls, struct dlm_member *memb)
0462 {
0463     struct dlm_slot slot;
0464     uint32_t seq;
0465     int error;
0466
0467     if (!ls->ls_ops || !ls->ls_ops->recover_slot)
0468         return;
0469
0470     /* if there is no comms connection with this node
0471        or the present comms connection is newer
0472        than the one when this member was added, then
0473        we consider the node to have failed (versus
0474        being removed due to dlm_release_lockspace) */
0475
0476     error = dlm_comm_seq(memb->nodeid, &seq);
0477
0478     if (!error && seq == memb->comm_seq)
0479         return;
0480
0481     slot.nodeid = memb->nodeid;
0482     slot.slot = memb->slot;
0483
0484     ls->ls_ops->recover_slot(ls->ls_ops_arg, &slot);
0485 }
0486
0487 void dlm_lsop_recover_done(struct dlm_ls *ls)
0488 {
0489     struct dlm_member *memb;
0490     struct dlm_slot *slots;
0491     int i, num;
0492
0493     if (!ls->ls_ops || !ls->ls_ops->recover_done)
0494         return;
0495
0496     num = ls->ls_num_nodes;
0497     slots = kcalloc(num, sizeof(*slots), GFP_KERNEL);
0498     if (!slots)
0499         return;
0500
0501     i = 0;
0502     list_for_each_entry(memb, &ls->ls_nodes, list) {
0503         if (i == num) {
0504             log_error(ls, "dlm_lsop_recover_done bad num %d", num);
0505             goto out;
0506         }
0507         slots[i].nodeid = memb->nodeid;
0508         slots[i].slot = memb->slot;
0509         i++;
0510     }
0511
0512     ls->ls_ops->recover_done(ls->ls_ops_arg, slots, num,
0513                  ls->ls_slot, ls->ls_generation);
0514  out:
0515     kfree(slots);
0516 }
0517
0518 static struct dlm_config_node *find_config_node(struct dlm_recover *rv,
0519                         int nodeid)
0520 {
0521     int i;
0522
0523     for (i = 0; i < rv->nodes_count; i++) {
0524         if (rv->nodes[i].nodeid == nodeid)
0525             return &rv->nodes[i];
0526     }
0527     return NULL;
0528 }
0529
0530 int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out)
0531 {
0532     struct dlm_member *memb, *safe;
0533     struct dlm_config_node *node;
0534     int i, error, neg = 0, low = -1;
0535
0536     /* previously removed members that we've not finished removing need to
0537      * count as a negative change so the "neg" recovery steps will happen
0538      *
0539      * This functionality must report all member changes to lsops or
0540      * midcomms layer and must never return before.
0541      */
0542
0543     list_for_each_entry(memb, &ls->ls_nodes_gone, list) {
0544         log_rinfo(ls, "prev removed member %d", memb->nodeid);
0545         neg++;
0546     }
0547
0548     /* move departed members from ls_nodes to ls_nodes_gone */
0549
0550     list_for_each_entry_safe(memb, safe, &ls->ls_nodes, list) {
0551         node = find_config_node(rv, memb->nodeid);
0552         if (node && !node->new)
0553             continue;
0554
0555         if (!node) {
0556             log_rinfo(ls, "remove member %d", memb->nodeid);
0557         } else {
0558             /* removed and re-added */
0559             log_rinfo(ls, "remove member %d comm_seq %u %u",
0560                   memb->nodeid, memb->comm_seq, node->comm_seq);
0561         }
0562
0563         neg++;
0564         list_move(&memb->list, &ls->ls_nodes_gone);
0565         dlm_midcomms_remove_member(memb->nodeid);
0566         ls->ls_num_nodes--;
0567         dlm_lsop_recover_slot(ls, memb);
0568     }
0569
0570     /* add new members to ls_nodes */
0571
0572     for (i = 0; i < rv->nodes_count; i++) {
0573         node = &rv->nodes[i];
0574         if (dlm_is_member(ls, node->nodeid))
0575             continue;
0576         dlm_add_member(ls, node);
0577         log_rinfo(ls, "add member %d", node->nodeid);
0578     }
0579
0580     list_for_each_entry(memb, &ls->ls_nodes, list) {
0581         if (low == -1 || memb->nodeid < low)
0582             low = memb->nodeid;
0583     }
0584     ls->ls_low_nodeid = low;
0585
0586     make_member_array(ls);
0587     *neg_out = neg;
0588
0589     error = ping_members(ls);
0590     log_rinfo(ls, "dlm_recover_members %d nodes", ls->ls_num_nodes);
0591     return error;
0592 }
0593
0594 /* Userspace guarantees that dlm_ls_stop() has completed on all nodes before
0595    dlm_ls_start() is called on any of them to start the new recovery. */
0596
0597 int dlm_ls_stop(struct dlm_ls *ls)
0598 {
0599     int new;
0600
0601     /*
0602      * Prevent dlm_recv from being in the middle of something when we do
0603      * the stop.  This includes ensuring dlm_recv isn't processing a
0604      * recovery message (rcom), while dlm_recoverd is aborting and
0605      * resetting things from an in-progress recovery.  i.e. we want
0606      * dlm_recoverd to abort its recovery without worrying about dlm_recv
0607      * processing an rcom at the same time.  Stopping dlm_recv also makes
0608      * it easy for dlm_receive_message() to check locking stopped and add a
0609      * message to the requestqueue without races.
0610      */
0611
0612     down_write(&ls->ls_recv_active);
0613
0614     /*
0615      * Abort any recovery that's in progress (see RECOVER_STOP,
0616      * dlm_recovery_stopped()) and tell any other threads running in the
0617      * dlm to quit any processing (see RUNNING, dlm_locking_stopped()).
0618      */
0619
0620     spin_lock(&ls->ls_recover_lock);
0621     set_bit(LSFL_RECOVER_STOP, &ls->ls_flags);
0622     new = test_and_clear_bit(LSFL_RUNNING, &ls->ls_flags);
0623     ls->ls_recover_seq++;
0624     spin_unlock(&ls->ls_recover_lock);
0625
0626     /*
0627      * Let dlm_recv run again, now any normal messages will be saved on the
0628      * requestqueue for later.
0629      */
0630
0631     up_write(&ls->ls_recv_active);
0632
0633     /*
0634      * This in_recovery lock does two things:
0635      * 1) Keeps this function from returning until all threads are out
0636      *    of locking routines and locking is truly stopped.
0637      * 2) Keeps any new requests from being processed until it's unlocked
0638      *    when recovery is complete.
0639      */
0640
0641     if (new) {
0642         set_bit(LSFL_RECOVER_DOWN, &ls->ls_flags);
0643         wake_up_process(ls->ls_recoverd_task);
0644         wait_event(ls->ls_recover_lock_wait,
0645                test_bit(LSFL_RECOVER_LOCK, &ls->ls_flags));
0646     }
0647
0648     /*
0649      * The recoverd suspend/resume makes sure that dlm_recoverd (if
0650      * running) has noticed RECOVER_STOP above and quit processing the
0651      * previous recovery.
0652      */
0653
0654     dlm_recoverd_suspend(ls);
0655
0656     spin_lock(&ls->ls_recover_lock);
0657     kfree(ls->ls_slots);
0658     ls->ls_slots = NULL;
0659     ls->ls_num_slots = 0;
0660     ls->ls_slots_size = 0;
0661     ls->ls_recover_status = 0;
0662     spin_unlock(&ls->ls_recover_lock);
0663
0664     dlm_recoverd_resume(ls);
0665
0666     if (!ls->ls_recover_begin)
0667         ls->ls_recover_begin = jiffies;
0668
0669     /* call recover_prep ops only once and not multiple times
0670      * for each possible dlm_ls_stop() when recovery is already
0671      * stopped.
0672      *
0673      * If we successful was able to clear LSFL_RUNNING bit and
0674      * it was set we know it is the first dlm_ls_stop() call.
0675      */
0676     if (new)
0677         dlm_lsop_recover_prep(ls);
0678
0679     return 0;
0680 }
0681
0682 int dlm_ls_start(struct dlm_ls *ls)
0683 {
0684     struct dlm_recover *rv, *rv_old;
0685     struct dlm_config_node *nodes = NULL;
0686     int error, count;
0687
0688     rv = kzalloc(sizeof(*rv), GFP_NOFS);
0689     if (!rv)
0690         return -ENOMEM;
0691
0692     error = dlm_config_nodes(ls->ls_name, &nodes, &count);
0693     if (error < 0)
0694         goto fail_rv;
0695
0696     spin_lock(&ls->ls_recover_lock);
0697
0698     /* the lockspace needs to be stopped before it can be started */
0699
0700     if (!dlm_locking_stopped(ls)) {
0701         spin_unlock(&ls->ls_recover_lock);
0702         log_error(ls, "start ignored: lockspace running");
0703         error = -EINVAL;
0704         goto fail;
0705     }
0706
0707     rv->nodes = nodes;
0708     rv->nodes_count = count;
0709     rv->seq = ++ls->ls_recover_seq;
0710     rv_old = ls->ls_recover_args;
0711     ls->ls_recover_args = rv;
0712     spin_unlock(&ls->ls_recover_lock);
0713
0714     if (rv_old) {
0715         log_error(ls, "unused recovery %llx %d",
0716               (unsigned long long)rv_old->seq, rv_old->nodes_count);
0717         kfree(rv_old->nodes);
0718         kfree(rv_old);
0719     }
0720
0721     set_bit(LSFL_RECOVER_WORK, &ls->ls_flags);
0722     wake_up_process(ls->ls_recoverd_task);
0723     return 0;
0724
0725  fail:
0726     kfree(nodes);
0727  fail_rv:
0728     kfree(rv);
0729     return error;
0730 }
0731