Implement inode removal tool. Removes multiple objects from multiple OSDs in parallel

Makefile

@@ -2,7 +2,7 @@ BLOCKSTORE_OBJS := allocator.o blockstore.o blockstore_impl.o blockstore_init.o
 	blockstore_write.o blockstore_sync.o blockstore_stable.o blockstore_rollback.o blockstore_flush.o crc32c.o ringloop.o
 # -fsanitize=address
 CXXFLAGS := -g -O3 -Wall -Wno-sign-compare -Wno-comment -Wno-parentheses -Wno-pointer-arith -fPIC -fdiagnostics-color=always
-all: libfio_blockstore.so osd libfio_sec_osd.so libfio_cluster.so stub_osd stub_uring_osd stub_bench osd_test dump_journal qemu_driver.so nbd_proxy
+all: libfio_blockstore.so osd libfio_sec_osd.so libfio_cluster.so stub_osd stub_uring_osd stub_bench osd_test dump_journal qemu_driver.so nbd_proxy rm_inode
 clean:
 	rm -f *.o
 
@@ -44,6 +44,9 @@ libfio_cluster.so: fio_cluster.o $(FIO_CLUSTER_OBJS)
 nbd_proxy: nbd_proxy.o $(FIO_CLUSTER_OBJS)
 	g++ $(CXXFLAGS) -ltcmalloc_minimal -o $@ $< $(FIO_CLUSTER_OBJS) -luring
 
+rm_inode: rm_inode.o $(FIO_CLUSTER_OBJS)
+	g++ $(CXXFLAGS) -ltcmalloc_minimal -o $@ $< $(FIO_CLUSTER_OBJS) -luring
+
 qemu_driver.o: qemu_driver.c qemu_proxy.h
 	gcc -I qemu/b/qemu `pkg-config glib-2.0 --cflags` \
 		-I qemu/include $(CXXFLAGS) -c -o $@ $<
@@ -149,6 +152,8 @@ qemu_proxy.o: qemu_proxy.cpp cluster_client.h etcd_state_client.h http_client.h
 	g++ $(CXXFLAGS) -c -o $@ $<
 ringloop.o: ringloop.cpp ringloop.h
 	g++ $(CXXFLAGS) -c -o $@ $<
+rm_inode.o: rm_inode.cpp cluster_client.h etcd_state_client.h http_client.h json11/json11.hpp malloc_or_die.h messenger.h object_id.h osd_id.h osd_ops.h ringloop.h timerfd_manager.h
+	g++ $(CXXFLAGS) -c -o $@ $<
 rw_blocking.o: rw_blocking.cpp rw_blocking.h
 	g++ $(CXXFLAGS) -c -o $@ $<
 stub_bench.o: stub_bench.cpp object_id.h osd_id.h osd_ops.h rw_blocking.h

README.md

@@ -31,12 +31,12 @@ breaking changes in the future. However, the following is implemented:
 - QEMU driver (built out-of-tree)
 - Loadable fio engine for benchmarks (also built out-of-tree)
 - NBD proxy for kernel mounts
+- Inode removal tool (./rm_inode)
 
 ## Roadmap
 
 - Packaging for Debian and, probably, CentOS too
 - OSD creation tool (OSDs currently have to be created by hand)
-- Inode deletion tool (currently you can't delete anything :))
 - Other administrative tools
 - Per-inode I/O and space usage statistics
 - jerasure EC support with any number of data and parity drives in a group
@@ -333,19 +333,22 @@ and calculate disk offsets almost by hand. This will be fixed in near future.
     -device virtio-blk-pci,scsi=off,bus=pci.0,addr=0x5,drive=drive-virtio-disk0,id=virtio-disk0,bootindex=1,write-cache=off,physical_block_size=4096,logical_block_size=512
     -vnc 0.0.0.0:0
   ```
+- Remove inode with (for example):
+  ```
+  ./rm_inode --etcd_address 10.115.0.10:2379/v3 --pool 1 --inode 1 --parallel_osds 16 --iodepth 32
+  ```
 
 ## Known Problems
 
 - OSDs may currently crash with "can't get SQE, will fall out of sync with EPOLLET"
   if you try to load them with very long iodepths because io_uring queue (ring) is limited
   and OSDs don't check if it fills up.
-- Object deletion requests may currently lead to unfound objects on crashes because
-  proper handling of deletions in a cluster requires a "three-phase cleanup process"
-  and it's currently not implemented. In fact, even though deletion requests are
-  implemented, there's no user tool to delete anything from the cluster yet :).
-  Of course I'll create such tool, but its first implementation will be vulnerable to this issue.
-  It's not a big deal though, because you'll be able to just repeat the deletion request
-  in this case.
+- Object deletion requests may currently lead to 'incomplete' objects if your OSDs crash during
+  deletion because proper handling of object cleanup in a cluster should be "three-phase"
+  and it's currently not implemented. Inode removal tool currently can't handle unclean
+  objects, so incomplete objects become undeletable. This will be fixed in near future
+  by allowing the inode removal tool to delete unclean objects. With this problem fixed
+  you'll be able just to repeat the removal again.
 
 ## Implementation Principles
 

cluster_client.cpp

@@ -203,6 +203,12 @@ void cluster_client_t::on_load_pgs_hook(bool success)
     {
         pg_counts[pool_item.first] = pool_item.second.real_pg_count;
     }
+    pgs_loaded = true;
+    for (auto fn: on_ready_hooks)
+    {
+        fn();
+    }
+    on_ready_hooks.clear();
     for (auto op: offline_ops)
     {
         execute(op);
@@ -254,6 +260,18 @@ void cluster_client_t::on_change_osd_state_hook(uint64_t peer_osd)
     }
 }
 
+void cluster_client_t::on_ready(std::function<void(void)> fn)
+{
+    if (pgs_loaded)
+    {
+        fn();
+    }
+    else
+    {
+        on_ready_hooks.push_back(fn);
+    }
+}
+
 /**
  * How writes are synced when immediate_commit is false
  *
@@ -284,7 +302,7 @@ void cluster_client_t::on_change_osd_state_hook(uint64_t peer_osd)
 
 void cluster_client_t::execute(cluster_op_t *op)
 {
-    if (!bs_disk_alignment)
+    if (!pgs_loaded)
     {
         // We're offline
         offline_ops.push_back(op);

cluster_client.h

@@ -63,8 +63,6 @@ class cluster_client_t
     int up_wait_retry_interval = 500; // ms
 
     uint64_t op_id = 1;
-    etcd_state_client_t st_cli;
-    osd_messenger_t msgr;
     ring_consumer_t consumer;
     // operations currently in progress
     std::set<cluster_op_t*> cur_ops;
@@ -78,10 +76,17 @@ class cluster_client_t
     std::vector<cluster_op_t*> offline_ops;
     uint64_t queued_bytes = 0;
 
+    bool pgs_loaded = false;
+    std::vector<std::function<void(void)>> on_ready_hooks;
+
 public:
+    etcd_state_client_t st_cli;
+    osd_messenger_t msgr;
+
     cluster_client_t(ring_loop_t *ringloop, timerfd_manager_t *tfd, json11::Json & config);
     ~cluster_client_t();
     void execute(cluster_op_t *op);
+    void on_ready(std::function<void(void)> fn);
     void stop();
 
 protected:

nbd_proxy.cpp

@@ -107,7 +107,7 @@ public:
     {
         printf(
             "Vitastor NBD proxy\n"
-            "(c) Vitaliy Filippov, 2020 (VNPL-1.0 or GNU GPL 2.0+)\n\n"
+            "(c) Vitaliy Filippov, 2020 (VNPL-1.0)\n\n"
             "USAGE:\n"
             "  %s map --etcd_address <etcd_address> --pool <pool> --inode <inode> --size <size in bytes>\n"
             "  %s unmap /dev/nbd0\n"

rm_inode.cpp

@@ -0,0 +1,325 @@
+// Copyright (c) Vitaliy Filippov, 2019+
+// License: VNPL-1.0 (see README.md for details)
+
+/**
+ * Inode removal tool
+ * May be included into a bigger "command-line management interface" in the future
+ */
+
+#include <algorithm>
+
+#include "epoll_manager.h"
+#include "cluster_client.h"
+#include "pg_states.h"
+
+#define RM_NO_LIST 1
+#define RM_LIST_SENT 2
+#define RM_REMOVING 3
+#define RM_END 4
+
+const char *exe_name = NULL;
+
+struct rm_pg_osd_t
+{
+    pg_num_t pg_num;
+    osd_num_t osd_num;
+    int state = 0;
+    obj_ver_id *obj_list = NULL;
+    uint64_t obj_count = 0, obj_pos = 0, obj_done = 0, obj_prev_done = 0;
+    int in_flight = 0;
+};
+
+class rm_inode_t
+{
+protected:
+    uint64_t inode = 0, pool_id = 0;
+    uint64_t iodepth = 0, parallel_osds = 0;
+
+    ring_loop_t *ringloop = NULL;
+    epoll_manager_t *epmgr = NULL;
+    cluster_client_t *cli = NULL;
+    ring_consumer_t consumer;
+
+    std::vector<rm_pg_osd_t*> lists;
+    uint64_t total_count = 0, total_done = 0, total_prev_pct = 0;
+    uint64_t pgs_to_list = 0;
+    bool started = false;
+    bool progress = true;
+    int log_level = 0;
+
+public:
+    static json11::Json::object parse_args(int narg, const char *args[])
+    {
+        json11::Json::object cfg;
+        cfg["progress"] = "1";
+        for (int i = 1; i < narg; i++)
+        {
+            if (!strcmp(args[i], "-h") || !strcmp(args[i], "--help"))
+            {
+                help();
+            }
+            else if (args[i][0] == '-' && args[i][1] == '-')
+            {
+                const char *opt = args[i]+2;
+                cfg[opt] = !strcmp(opt, "json") || i == narg-1 ? "1" : args[++i];
+            }
+        }
+        return cfg;
+    }
+
+    static void help()
+    {
+        printf(
+            "Vitastor inode removal tool\n"
+            "(c) Vitaliy Filippov, 2020 (VNPL-1.0)\n\n"
+            "USAGE:\n"
+            "  %s --etcd_address <etcd_address> --pool <pool> --inode <inode>\n",
+            exe_name
+        );
+        exit(0);
+    }
+
+    void run(json11::Json cfg)
+    {
+        if (cfg["etcd_address"].string_value() == "")
+        {
+            fprintf(stderr, "etcd_address is missing\n");
+            exit(1);
+        }
+        inode = cfg["inode"].uint64_value();
+        pool_id = cfg["pool"].uint64_value();
+        if (pool_id)
+            inode = (inode & ((1l << (64-POOL_ID_BITS)) - 1)) | (pool_id << (64-POOL_ID_BITS));
+        pool_id = INODE_POOL(inode);
+        if (!pool_id)
+        {
+            fprintf(stderr, "pool is missing");
+            exit(1);
+        }
+        iodepth = cfg["iodepth"].uint64_value();
+        if (!iodepth)
+            iodepth = 32;
+        parallel_osds = cfg["parallel_osds"].uint64_value();
+        if (!parallel_osds)
+            parallel_osds = 4;
+        log_level = cfg["log_level"].uint64_value();
+        progress = cfg["progress"].uint64_value() ? true : false;
+        // Create client
+        ringloop = new ring_loop_t(512);
+        epmgr = new epoll_manager_t(ringloop);
+        cli = new cluster_client_t(ringloop, epmgr->tfd, cfg);
+        cli->on_ready([this]() { start_delete(); });
+        // Initialize job
+        consumer.loop = [this]()
+        {
+            if (started)
+                continue_delete();
+            ringloop->submit();
+        };
+        ringloop->register_consumer(&consumer);
+        // Loop until it completes
+        while (1)
+        {
+            ringloop->loop();
+            ringloop->wait();
+        }
+    }
+
+    void start_delete()
+    {
+        if (cli->st_cli.pool_config.find(pool_id) == cli->st_cli.pool_config.end())
+        {
+            fprintf(stderr, "Pool %lu does not exist\n", pool_id);
+            exit(1);
+        }
+        auto pool_cfg = cli->st_cli.pool_config[pool_id];
+        for (auto & pg_item: pool_cfg.pg_config)
+        {
+            auto & pg = pg_item.second;
+            if (pg.pause || !pg.cur_primary || pg.cur_state != PG_ACTIVE)
+            {
+                // FIXME Support deletion in non-clean active PGs by introducing a "primary-list" command
+                fprintf(stderr, "PG %u is not active+clean, skipping\n", pg_item.first);
+                continue;
+            }
+            rm_pg_osd_t *r = new rm_pg_osd_t();
+            r->pg_num = pg_item.first;
+            r->osd_num = pg.cur_primary;
+            r->state = RM_NO_LIST;
+            lists.push_back(r);
+        }
+        std::sort(lists.begin(), lists.end(), [](rm_pg_osd_t *a, rm_pg_osd_t *b)
+        {
+            return a->osd_num < b->osd_num ? true : false;
+        });
+        pgs_to_list = lists.size();
+        started = true;
+        continue_delete();
+    }
+
+    void send_list(rm_pg_osd_t *cur_list)
+    {
+        if (cli->msgr.osd_peer_fds.find(cur_list->osd_num) ==
+            cli->msgr.osd_peer_fds.end())
+        {
+            // Initiate connection
+            cli->msgr.connect_peer(cur_list->osd_num, cli->st_cli.peer_states[cur_list->osd_num]);
+            return;
+        }
+        osd_op_t *op = new osd_op_t();
+        op->op_type = OSD_OP_OUT;
+        op->peer_fd = cli->msgr.osd_peer_fds[cur_list->osd_num];
+        op->req = {
+            .sec_list = {
+                .header = {
+                    .magic = SECONDARY_OSD_OP_MAGIC,
+                    .id = cli->msgr.next_subop_id++,
+                    .opcode = OSD_OP_SEC_LIST,
+                },
+                .list_pg = cur_list->pg_num,
+                .pg_count = (pg_num_t)cli->st_cli.pool_config[pool_id].real_pg_count,
+                .pg_stripe_size = cli->st_cli.pool_config[pool_id].pg_stripe_size,
+                .min_inode = inode,
+                .max_inode = inode,
+            },
+        };
+        op->callback = [this, cur_list](osd_op_t *op)
+        {
+            pgs_to_list--;
+            if (op->reply.hdr.retval < 0)
+            {
+                fprintf(stderr, "Failed to get object list from OSD %lu (retval=%ld), skipping the PG\n",
+                    cur_list->osd_num, op->reply.hdr.retval);
+                cli->msgr.stop_client(cur_list->osd_num);
+                delete op;
+                cur_list->state = RM_END;
+                continue_delete();
+                return;
+            }
+            if (log_level > 0)
+            {
+                printf(
+                    "[PG %u] Got inode object list from OSD %lu: %ld object versions\n",
+                    cur_list->pg_num, cur_list->osd_num, op->reply.hdr.retval
+                );
+            }
+            cur_list->obj_list = (obj_ver_id*)op->buf;
+            cur_list->obj_count = (uint64_t)op->reply.hdr.retval;
+            cur_list->obj_done = cur_list->obj_prev_done = cur_list->obj_pos = 0;
+            total_count += cur_list->obj_count;
+            total_prev_pct = 0;
+            // set op->buf to NULL so it doesn't get freed
+            op->buf = NULL;
+            delete op;
+            cur_list->state = RM_REMOVING;
+            continue_delete();
+        };
+        cur_list->state = RM_LIST_SENT;
+        cli->msgr.outbox_push(op);
+    }
+
+    void send_ops(rm_pg_osd_t *cur_list)
+    {
+        if (cli->msgr.osd_peer_fds.find(cur_list->osd_num) ==
+            cli->msgr.osd_peer_fds.end())
+        {
+            // Initiate connection
+            cli->msgr.connect_peer(cur_list->osd_num, cli->st_cli.peer_states[cur_list->osd_num]);
+            return;
+        }
+        while (cur_list->in_flight < iodepth && cur_list->obj_pos < cur_list->obj_count)
+        {
+            osd_op_t *op = new osd_op_t();
+            op->op_type = OSD_OP_OUT;
+            op->peer_fd = cli->msgr.osd_peer_fds[cur_list->osd_num];
+            op->req = {
+                .rw = {
+                    .header = {
+                        .magic = SECONDARY_OSD_OP_MAGIC,
+                        .id = cli->msgr.next_subop_id++,
+                        .opcode = OSD_OP_DELETE,
+                    },
+                    .inode = cur_list->obj_list[cur_list->obj_pos].oid.inode,
+                    .offset = (cur_list->obj_list[cur_list->obj_pos].oid.stripe & ~STRIPE_MASK),
+                    .len = 0,
+                },
+            };
+            op->callback = [this, cur_list](osd_op_t *op)
+            {
+                cur_list->in_flight--;
+                if (op->reply.hdr.retval < 0)
+                {
+                    fprintf(stderr, "Failed to remove object from PG %u (OSD %lu) (retval=%ld)\n",
+                        cur_list->pg_num, cur_list->osd_num, op->reply.hdr.retval);
+                }
+                delete op;
+                cur_list->obj_done++;
+                total_done++;
+                continue_delete();
+            };
+            cli->msgr.outbox_push(op);
+            cur_list->obj_pos++;
+            cur_list->in_flight++;
+        }
+        if (!cur_list->in_flight && cur_list->obj_pos >= cur_list->obj_count)
+        {
+            free(cur_list->obj_list);
+            cur_list->obj_list = NULL;
+            cur_list->obj_count = 0;
+            cur_list->obj_done = cur_list->obj_prev_done = cur_list->obj_pos = 0;
+            cur_list->state = RM_END;
+        }
+    }
+
+    void continue_delete()
+    {
+        int par_osd = 0;
+        osd_num_t max_seen_osd = 0;
+        for (int i = 0; i < lists.size(); i++)
+        {
+            if (lists[i]->state == RM_END)
+            {
+                delete lists[i];
+                lists.erase(lists.begin()+i, lists.begin()+i+1);
+                i--;
+            }
+            else if (lists[i]->osd_num > max_seen_osd)
+            {
+                if (lists[i]->state == RM_NO_LIST)
+                {
+                    send_list(lists[i]);
+                }
+                else if (lists[i]->state == RM_REMOVING)
+                {
+                    send_ops(lists[i]);
+                }
+                par_osd++;
+                max_seen_osd = lists[i]->osd_num;
+                if (par_osd >= parallel_osds)
+                {
+                    break;
+                }
+            }
+        }
+        if (progress && total_count > 0 && total_done*1000/total_count != total_prev_pct)
+        {
+            printf("\rRemoved %lu/%lu objects, %lu more PGs to list...", total_done, total_count, pgs_to_list);
+            total_prev_pct = total_done*1000/total_count;
+        }
+        if (!lists.size())
+        {
+            printf("Done, inode %lu in pool %lu removed\n", (inode & ((1l << (64-POOL_ID_BITS)) - 1)), pool_id);
+            exit(0);
+        }
+    }
+};
+
+int main(int narg, const char *args[])
+{
+    setvbuf(stdout, NULL, _IONBF, 0);
+    setvbuf(stderr, NULL, _IONBF, 0);
+    exe_name = args[0];
+    rm_inode_t *p = new rm_inode_t();
+    p->run(rm_inode_t::parse_args(narg, args));
+    return 0;
+}