vitastor/osd_primary.cpp

#include "osd.h"
#include "osd_rmw.h"

// FIXME: Split into more files

#define SUBMIT_READ 0
#define SUBMIT_RMW_READ 1
#define SUBMIT_WRITE 2

// read: read directly or read paired stripe(s), reconstruct, return
// write: read paired stripe(s), reconstruct, modify, calculate parity, write
//
// nuance: take care to read the same version from paired stripes!
// to do so, we remember "last readable" version until a write request completes
// and we postpone other write requests to the same stripe until completion of previous ones
//
// sync: sync peers, get unstable versions, stabilize them

struct unstable_osd_num_t
{
    osd_num_t osd_num;
    int start, len;
};

struct osd_primary_op_data_t
{
    int st = 0;
    pg_num_t pg_num;
    object_id oid;
    uint64_t target_ver;
    uint64_t fact_ver = 0;
    int n_subops = 0, done = 0, errors = 0, epipe = 0;
    int degraded = 0, pg_size, pg_minsize;
    osd_rmw_stripe_t *stripes;
    osd_op_t *subops = NULL;
    uint64_t *prev_set = NULL;
    pg_osd_set_state_t *object_state = NULL;

    // for sync. oops, requires freeing
    std::vector<unstable_osd_num_t> *unstable_write_osds = NULL;
    pg_num_t *dirty_pgs = NULL;
    int dirty_pg_count = 0;
    obj_ver_id *unstable_writes = NULL;
};

void osd_t::finish_op(osd_op_t *cur_op, int retval)
{
    inflight_ops--;
    if (cur_op->op_data && cur_op->op_data->pg_num > 0)
    {
        auto & pg = pgs[cur_op->op_data->pg_num];
        int n = --pg.inflight;
        assert(n >= 0);
        if (n == 0 && (pg.state & PG_STOPPING))
        {
            finish_stop_pg(pg);
        }
    }
    if (!cur_op->peer_fd)
    {
        // Copy lambda to be unaffected by `delete op`
        std::function<void(osd_op_t*)>(cur_op->callback)(cur_op);
    }
    else
    {
        // FIXME add separate magic number
        auto cl_it = clients.find(cur_op->peer_fd);
        if (cl_it != clients.end())
        {
            cur_op->reply.hdr.magic = SECONDARY_OSD_REPLY_MAGIC;
            cur_op->reply.hdr.id = cur_op->req.hdr.id;
            cur_op->reply.hdr.opcode = cur_op->req.hdr.opcode;
            cur_op->reply.hdr.retval = retval;
            outbox_push(cl_it->second, cur_op);
        }
        else
        {
            delete cur_op;
        }
    }
}

bool osd_t::prepare_primary_rw(osd_op_t *cur_op)
{
    // PG number is calculated from the offset
    // Our EC scheme stores data in fixed chunks equal to (K*block size)
    // But we must not use K in the process of calculating the PG number
    // So we calculate the PG number using a separate setting which should be per-inode (FIXME)
    pg_num_t pg_num = (cur_op->req.rw.inode + cur_op->req.rw.offset / pg_stripe_size) % pg_count + 1;
    auto pg_it = pgs.find(pg_num);
    if (pg_it == pgs.end() || !(pg_it->second.state & PG_ACTIVE))
    {
        // This OSD is not primary for this PG or the PG is inactive
        finish_op(cur_op, -EPIPE);
        return false;
    }
    uint64_t pg_block_size = bs_block_size * pg_it->second.pg_minsize;
    object_id oid = {
        .inode = cur_op->req.rw.inode,
        // oid.stripe = starting offset of the parity stripe, so it can be mapped back to the PG
        .stripe = (cur_op->req.rw.offset / pg_stripe_size) * pg_stripe_size +
            ((cur_op->req.rw.offset % pg_stripe_size) / pg_block_size) * pg_block_size
    };
    if ((cur_op->req.rw.offset + cur_op->req.rw.len) > (oid.stripe + pg_block_size) ||
        (cur_op->req.rw.offset % bs_disk_alignment) != 0 ||
        (cur_op->req.rw.len % bs_disk_alignment) != 0)
    {
        finish_op(cur_op, -EINVAL);
        return false;
    }
    osd_primary_op_data_t *op_data = (osd_primary_op_data_t*)calloc(
        sizeof(osd_primary_op_data_t) + sizeof(osd_rmw_stripe_t) * pg_it->second.pg_size, 1
    );
    op_data->pg_num = pg_num;
    op_data->oid = oid;
    op_data->stripes = ((osd_rmw_stripe_t*)(op_data+1));
    cur_op->op_data = op_data;
    split_stripes(pg_it->second.pg_minsize, bs_block_size, (uint32_t)(cur_op->req.rw.offset - oid.stripe), cur_op->req.rw.len, op_data->stripes);
    pg_it->second.inflight++;
    return true;
}

uint64_t* get_object_osd_set(pg_t &pg, object_id &oid, uint64_t *def, pg_osd_set_state_t **object_state)
{
    if (!(pg.state & (PG_HAS_INCOMPLETE | PG_HAS_DEGRADED | PG_HAS_MISPLACED)))
    {
        *object_state = NULL;
        return def;
    }
    auto st_it = pg.incomplete_objects.find(oid);
    if (st_it != pg.incomplete_objects.end())
    {
        *object_state = st_it->second;
        return st_it->second->read_target.data();
    }
    st_it = pg.degraded_objects.find(oid);
    if (st_it != pg.degraded_objects.end())
    {
        *object_state = st_it->second;
        return st_it->second->read_target.data();
    }
    st_it = pg.misplaced_objects.find(oid);
    if (st_it != pg.misplaced_objects.end())
    {
        *object_state = st_it->second;
        return st_it->second->read_target.data();
    }
    *object_state = NULL;
    return def;
}

void osd_t::continue_primary_read(osd_op_t *cur_op)
{
    if (!cur_op->op_data && !prepare_primary_rw(cur_op))
    {
        return;
    }
    osd_primary_op_data_t *op_data = cur_op->op_data;
    if (op_data->st == 1)      goto resume_1;
    else if (op_data->st == 2) goto resume_2;
    {
        auto & pg = pgs[op_data->pg_num];
        for (int role = 0; role < pg.pg_minsize; role++)
        {
            op_data->stripes[role].read_start = op_data->stripes[role].req_start;
            op_data->stripes[role].read_end = op_data->stripes[role].req_end;
        }
        // Determine version
        auto vo_it = pg.ver_override.find(op_data->oid);
        op_data->target_ver = vo_it != pg.ver_override.end() ? vo_it->second : UINT64_MAX;
        if (pg.state == PG_ACTIVE)
        {
            // Fast happy-path
            cur_op->buf = alloc_read_buffer(op_data->stripes, pg.pg_minsize, 0);
            submit_primary_subops(SUBMIT_READ, pg.pg_minsize, pg.cur_set.data(), cur_op);
            cur_op->send_list.push_back(cur_op->buf, cur_op->req.rw.len);
            op_data->st = 1;
        }
        else
        {
            // PG may be degraded or have misplaced objects
            uint64_t* cur_set = get_object_osd_set(pg, op_data->oid, pg.cur_set.data(), &op_data->object_state);
            if (extend_missing_stripes(op_data->stripes, cur_set, pg.pg_minsize, pg.pg_size) < 0)
            {
                finish_op(cur_op, -EIO);
                return;
            }
            // Submit reads
            op_data->pg_minsize = pg.pg_minsize;
            op_data->pg_size = pg.pg_size;
            op_data->degraded = 1;
            cur_op->buf = alloc_read_buffer(op_data->stripes, pg.pg_size, 0);
            submit_primary_subops(SUBMIT_READ, pg.pg_size, cur_set, cur_op);
            op_data->st = 1;
        }
    }
resume_1:
    return;
resume_2:
    if (op_data->errors > 0)
    {
        finish_op(cur_op, op_data->epipe > 0 ? -EPIPE : -EIO);
        return;
    }
    if (op_data->degraded)
    {
        // Reconstruct missing stripes
        // FIXME: Always EC(k+1) by now. Add different coding schemes
        osd_rmw_stripe_t *stripes = op_data->stripes;
        for (int role = 0; role < op_data->pg_minsize; role++)
        {
            if (stripes[role].read_end != 0 && stripes[role].missing)
            {
                reconstruct_stripe(stripes, op_data->pg_size, role);
            }
            if (stripes[role].req_end != 0)
            {
                // Send buffer in parts to avoid copying
                cur_op->send_list.push_back(
                    stripes[role].read_buf + (stripes[role].req_start - stripes[role].read_start),
                    stripes[role].req_end - stripes[role].req_start
                );
            }
        }
    }
    finish_op(cur_op, cur_op->req.rw.len);
}

void osd_t::submit_primary_subops(int submit_type, int pg_size, const uint64_t* osd_set, osd_op_t *cur_op)
{
    bool w = submit_type == SUBMIT_WRITE;
    osd_primary_op_data_t *op_data = cur_op->op_data;
    osd_rmw_stripe_t *stripes = op_data->stripes;
    // Allocate subops
    int n_subops = 0, zero_read = -1;
    for (int role = 0; role < pg_size; role++)
    {
        if (osd_set[role] == this->osd_num || osd_set[role] != 0 && zero_read == -1)
        {
            zero_read = role;
        }
        if (osd_set[role] != 0 && (w || stripes[role].read_end != 0))
        {
            n_subops++;
        }
    }
    if (!n_subops && submit_type == SUBMIT_RMW_READ)
    {
        n_subops = 1;
    }
    else
    {
        zero_read = -1;
    }
    uint64_t op_version = w ? op_data->fact_ver+1 : (submit_type == SUBMIT_RMW_READ ? UINT64_MAX : op_data->target_ver);
    osd_op_t *subops = new osd_op_t[n_subops];
    op_data->fact_ver = 0;
    op_data->done = op_data->errors = 0;
    op_data->n_subops = n_subops;
    op_data->subops = subops;
    int subop = 0;
    for (int role = 0; role < pg_size; role++)
    {
        // We always submit zero-length writes to all replicas, even if the stripe is not modified
        if (!(w || stripes[role].read_end != 0 || zero_read == role))
        {
            continue;
        }
        osd_num_t role_osd_num = osd_set[role];
        if (role_osd_num != 0)
        {
            if (role_osd_num == this->osd_num)
            {
                subops[subop].bs_op = new blockstore_op_t({
                    .opcode = (uint64_t)(w ? BS_OP_WRITE : BS_OP_READ),
                    .callback = [cur_op, this](blockstore_op_t *subop)
                    {
                        if (subop->opcode == BS_OP_WRITE && subop->retval != subop->len)
                        {
                            // die
                            throw std::runtime_error("local write operation failed (retval = "+std::to_string(subop->retval)+")");
                        }
                        handle_primary_subop(
                            subop->opcode == BS_OP_WRITE ? OSD_OP_SECONDARY_WRITE : OSD_OP_SECONDARY_READ,
                            cur_op, subop->retval, subop->len, subop->version
                        );
                    },
                    .oid = {
                        .inode = op_data->oid.inode,
                        .stripe = op_data->oid.stripe | role,
                    },
                    .version = op_version,
                    .offset = w ? stripes[role].write_start : stripes[role].read_start,
                    .len = w ? stripes[role].write_end - stripes[role].write_start : stripes[role].read_end - stripes[role].read_start,
                    .buf = w ? stripes[role].write_buf : stripes[role].read_buf,
                });
                bs->enqueue_op(subops[subop].bs_op);
            }
            else
            {
                subops[subop].op_type = OSD_OP_OUT;
                subops[subop].send_list.push_back(subops[subop].req.buf, OSD_PACKET_SIZE);
                subops[subop].peer_fd = this->osd_peer_fds.at(role_osd_num);
                subops[subop].req.sec_rw = {
                    .header = {
                        .magic = SECONDARY_OSD_OP_MAGIC,
                        .id = this->next_subop_id++,
                        .opcode = (uint64_t)(w ? OSD_OP_SECONDARY_WRITE : OSD_OP_SECONDARY_READ),
                    },
                    .oid = {
                        .inode = op_data->oid.inode,
                        .stripe = op_data->oid.stripe | role,
                    },
                    .version = op_version,
                    .offset = w ? stripes[role].write_start : stripes[role].read_start,
                    .len = w ? stripes[role].write_end - stripes[role].write_start : stripes[role].read_end - stripes[role].read_start,
                };
                subops[subop].buf = w ? stripes[role].write_buf : stripes[role].read_buf;
                if (w && stripes[role].write_end > 0)
                {
                    subops[subop].send_list.push_back(stripes[role].write_buf, stripes[role].write_end - stripes[role].write_start);
                }
                subops[subop].callback = [cur_op, this](osd_op_t *subop)
                {
                    int fail_fd = subop->req.hdr.opcode == OSD_OP_SECONDARY_WRITE &&
                        subop->reply.hdr.retval != subop->req.sec_rw.len ? subop->peer_fd : -1;
                    // so it doesn't get freed
                    subop->buf = NULL;
                    handle_primary_subop(
                        subop->req.hdr.opcode, cur_op, subop->reply.hdr.retval,
                        subop->req.sec_rw.len, subop->reply.sec_rw.version
                    );
                    if (fail_fd >= 0)
                    {
                        // write operation failed, drop the connection
                        stop_client(fail_fd);
                    }
                };
                outbox_push(clients[subops[subop].peer_fd], &subops[subop]);
            }
            subop++;
        }
    }
}

void osd_t::handle_primary_subop(uint64_t opcode, osd_op_t *cur_op, int retval, int expected, uint64_t version)
{
    osd_primary_op_data_t *op_data = cur_op->op_data;
    if (retval != expected)
    {
        printf("%s subop failed: retval = %d (expected %d)\n", osd_op_names[opcode], retval, expected);
        if (retval == -EPIPE)
        {
            op_data->epipe++;
        }
        op_data->errors++;
    }
    else
    {
        op_data->done++;
        if (opcode == OSD_OP_SECONDARY_READ || opcode == OSD_OP_SECONDARY_WRITE)
        {
            if (op_data->fact_ver != 0 && op_data->fact_ver != version)
            {
                throw std::runtime_error("different fact_versions returned from subops: "+std::to_string(version)+" vs "+std::to_string(op_data->fact_ver));
            }
            op_data->fact_ver = version;
        }
    }
    if ((op_data->errors + op_data->done) >= op_data->n_subops)
    {
        delete[] op_data->subops;
        op_data->subops = NULL;
        op_data->st++;
        if (cur_op->req.hdr.opcode == OSD_OP_READ)
        {
            continue_primary_read(cur_op);
        }
        else if (cur_op->req.hdr.opcode == OSD_OP_WRITE)
        {
            continue_primary_write(cur_op);
        }
        else if (cur_op->req.hdr.opcode == OSD_OP_SYNC)
        {
            continue_primary_sync(cur_op);
        }
        else
        {
            throw std::runtime_error("BUG: unknown opcode");
        }
    }
}

void osd_t::pg_cancel_write_queue(pg_t & pg, object_id oid, int retval)
{
    auto st_it = pg.write_queue.find(oid), it = st_it;
    while (it != pg.write_queue.end() && it->first == oid)
    {
        finish_op(it->second, retval);
        it++;
    }
    if (st_it != it)
    {
        pg.write_queue.erase(st_it, it);
    }
}

void osd_t::continue_primary_write(osd_op_t *cur_op)
{
    if (!cur_op->op_data && !prepare_primary_rw(cur_op))
    {
        return;
    }
    osd_primary_op_data_t *op_data = cur_op->op_data;
    auto & pg = pgs[op_data->pg_num];
    if (op_data->st == 1)      goto resume_1;
    else if (op_data->st == 2) goto resume_2;
    else if (op_data->st == 3) goto resume_3;
    else if (op_data->st == 4) goto resume_4;
    else if (op_data->st == 5) goto resume_5;
    else if (op_data->st == 6) goto resume_6;
    else if (op_data->st == 7) goto resume_7;
    else if (op_data->st == 8) goto resume_8;
    assert(op_data->st == 0);
    // Check if actions are pending for this object
    {
        auto act_it = pg.flush_actions.lower_bound((obj_piece_id_t){
            .oid = op_data->oid,
            .osd_num = 0,
        });
        if (act_it != pg.flush_actions.end() &&
            act_it->first.oid.inode == op_data->oid.inode &&
            (act_it->first.oid.stripe & ~STRIPE_MASK) == op_data->oid.stripe)
        {
            pg.write_queue.emplace(op_data->oid, cur_op);
            return;
        }
    }
    // Check if there are other write requests to the same object
    {
        auto vo_it = pg.write_queue.find(op_data->oid);
        if (vo_it != pg.write_queue.end())
        {
            op_data->st = 1;
            pg.write_queue.emplace(op_data->oid, cur_op);
            return;
        }
        pg.write_queue.emplace(op_data->oid, cur_op);
    }
resume_1:
    // Determine blocks to read and write
    // Missing chunks are allowed to be overwritten even in incomplete objects
    // FIXME: Allow to do small writes to the old (degraded/misplaced) OSD set for the lower performance impact
    op_data->prev_set = get_object_osd_set(pg, op_data->oid, pg.cur_set.data(), &op_data->object_state);
    cur_op->rmw_buf = calc_rmw(cur_op->buf, op_data->stripes, op_data->prev_set,
        pg.pg_size, pg.pg_minsize, pg.pg_cursize, pg.cur_set.data(), bs_block_size);
    // Read required blocks
    submit_primary_subops(SUBMIT_RMW_READ, pg.pg_size, pg.cur_set.data(), cur_op);
resume_2:
    op_data->st = 2;
    return;
resume_3:
    if (op_data->errors > 0)
    {
        pg_cancel_write_queue(pg, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
        return;
    }
    // Save version override for parallel reads
    pg.ver_override[op_data->oid] = op_data->fact_ver;
    // Recover missing stripes, calculate parity
    calc_rmw_parity(op_data->stripes, pg.pg_size, op_data->prev_set, pg.cur_set.data(), bs_block_size);
    // Send writes
    submit_primary_subops(SUBMIT_WRITE, pg.pg_size, pg.cur_set.data(), cur_op);
resume_4:
    op_data->st = 4;
    return;
resume_5:
    if (op_data->errors > 0)
    {
        pg_cancel_write_queue(pg, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
        return;
    }
    if (op_data->object_state)
    {
        if (op_data->object_state->state & OBJ_MISPLACED)
        {
            // Remove extra chunks
            submit_primary_del_subops(cur_op, pg.cur_set.data(), op_data->object_state);
            if (op_data->n_subops > 0)
            {
                op_data->st = 8;
                return;
resume_8:
                if (op_data->errors > 0)
                {
                    pg_cancel_write_queue(pg, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
                    return;
                }
            }
        }
        // Clear object state
        if (op_data->object_state->state & OBJ_INCOMPLETE)
        {
            // Successful write means that object is not incomplete anymore
            incomplete_objects--;
            pg.incomplete_objects.erase(op_data->oid);
            if (!pg.incomplete_objects.size())
            {
                pg.state = pg.state & ~PG_HAS_INCOMPLETE;
                pg.print_state();
            }
        }
        else if (op_data->object_state->state & OBJ_DEGRADED)
        {
            degraded_objects--;
            pg.degraded_objects.erase(op_data->oid);
            if (!pg.degraded_objects.size())
            {
                pg.state = pg.state & ~PG_HAS_DEGRADED;
                pg.print_state();
            }
        }
        else if (op_data->object_state->state & OBJ_MISPLACED)
        {
            misplaced_objects--;
            pg.misplaced_objects.erase(op_data->oid);
            if (!pg.misplaced_objects.size())
            {
                pg.state = pg.state & ~PG_HAS_MISPLACED;
                pg.print_state();
            }
        }
        else
        {
            throw std::runtime_error("Invalid object state during recovery: "+std::to_string(op_data->object_state->state));
        }
        // FIXME: Track object count during normal writes, too
        pg.clean_count++;
        op_data->object_state->object_count--;
        if (!op_data->object_state->object_count)
        {
            pg.state_dict.erase(op_data->object_state->osd_set);
        }
    }
    // FIXME: Check for immediate_commit == IMMEDIATE_SMALL
    if (immediate_commit == IMMEDIATE_ALL)
    {
        op_data->unstable_write_osds = new std::vector<unstable_osd_num_t>();
        op_data->unstable_writes = new obj_ver_id[pg.pg_cursize];
        {
            int last_start = 0;
            osd_num_t *osd_set = pg.cur_set.data();
            for (int role = 0; role < pg.pg_size; role++)
            {
                if (osd_set[role] != 0)
                {
                    op_data->unstable_writes[last_start] = (obj_ver_id){
                        .oid = {
                            .inode = op_data->oid.inode,
                            .stripe = op_data->oid.stripe | role,
                        },
                        .version = op_data->fact_ver,
                    };
                    op_data->unstable_write_osds->push_back((unstable_osd_num_t){
                        .osd_num = osd_set[role],
                        .start = last_start,
                        .len = 1,
                    });
                    last_start++;
                }
            }
        }
        // Stabilize version sets
        submit_primary_stab_subops(cur_op);
resume_6:
        op_data->st = 6;
        return;
resume_7:
        // FIXME: Free those in the destructor?
        delete op_data->unstable_write_osds;
        delete[] op_data->unstable_writes;
        op_data->unstable_writes = NULL;
        op_data->unstable_write_osds = NULL;
        if (op_data->errors > 0)
        {
            pg_cancel_write_queue(pg, op_data->oid, op_data->epipe > 0 ? -EPIPE : -EIO);
            return;
        }
    }
    else
    {
        // Remember version as unstable
        osd_num_t *osd_set = pg.cur_set.data();
        for (int role = 0; role < pg.pg_size; role++)
        {
            if (osd_set[role] != 0)
            {
                this->unstable_writes[(osd_object_id_t){
                    .osd_num = osd_set[role],
                    .oid = {
                        .inode = op_data->oid.inode,
                        .stripe = op_data->oid.stripe | role,
                    },
                }] = op_data->fact_ver;
            }
        }
        // Remember PG as dirty to drop the connection when PG goes offline
        // (this is required because of the "lazy sync")
        this->clients[cur_op->peer_fd].dirty_pgs.insert(op_data->pg_num);
        dirty_pgs.insert(op_data->pg_num);
    }
    // Remove version override
    object_id oid = op_data->oid;
    pg.ver_override.erase(oid);
    finish_op(cur_op, cur_op->req.rw.len);
    // Continue other write operations to the same object
    auto next_it = pg.write_queue.find(oid);
    auto this_it = next_it;
    next_it++;
    pg.write_queue.erase(this_it);
    if (next_it != pg.write_queue.end() &&
        next_it->first == oid)
    {
        osd_op_t *next_op = next_it->second;
        continue_primary_write(next_op);
    }
}

void osd_t::autosync()
{
    if (immediate_commit != IMMEDIATE_ALL && !autosync_op)
    {
        autosync_op = new osd_op_t();
        autosync_op->op_type = OSD_OP_IN;
        autosync_op->req = {
            .sync = {
                .header = {
                    .magic = SECONDARY_OSD_OP_MAGIC,
                    .id = 1,
                    .opcode = OSD_OP_SYNC,
                },
            },
        };
        autosync_op->callback = [this](osd_op_t *op)
        {
            if (op->reply.hdr.retval < 0)
            {
                printf("Warning: automatic sync resulted in an error: %ld (%s)\n", -op->reply.hdr.retval, strerror(-op->reply.hdr.retval));
            }
            delete autosync_op;
            autosync_op = NULL;
        };
        exec_op(autosync_op);
    }
}

// Save and clear unstable_writes -> SYNC all -> STABLE all
void osd_t::continue_primary_sync(osd_op_t *cur_op)
{
    if (!cur_op->op_data)
    {
        cur_op->op_data = (osd_primary_op_data_t*)calloc(sizeof(osd_primary_op_data_t), 1);
    }
    osd_primary_op_data_t *op_data = cur_op->op_data;
    if (op_data->st == 1)      goto resume_1;
    else if (op_data->st == 2) goto resume_2;
    else if (op_data->st == 3) goto resume_3;
    else if (op_data->st == 4) goto resume_4;
    else if (op_data->st == 5) goto resume_5;
    else if (op_data->st == 6) goto resume_6;
    assert(op_data->st == 0);
    if (syncs_in_progress.size() > 0)
    {
        // Wait for previous syncs, if any
        // FIXME: We may try to execute the current one in parallel, like in Blockstore, but I'm not sure if it matters at all
        syncs_in_progress.push_back(cur_op);
        op_data->st = 1;
resume_1:
        return;
    }
    else
    {
        syncs_in_progress.push_back(cur_op);
    }
resume_2:
    if (unstable_writes.size() == 0)
    {
        // Nothing to sync
        goto finish;
    }
    // Save and clear unstable_writes
    // In theory it is possible to do in on a per-client basis, but this seems to be an unnecessary complication
    // It would be cool not to copy these here at all, but someone has to deduplicate them by object IDs anyway
    {
        op_data->unstable_write_osds = new std::vector<unstable_osd_num_t>();
        op_data->unstable_writes = new obj_ver_id[this->unstable_writes.size()];
        op_data->dirty_pgs = new pg_num_t[dirty_pgs.size()];
        op_data->dirty_pg_count = dirty_pgs.size();
        osd_num_t last_osd = 0;
        int last_start = 0, last_end = 0;
        for (auto it = this->unstable_writes.begin(); it != this->unstable_writes.end(); it++)
        {
            if (last_osd != it->first.osd_num)
            {
                if (last_osd != 0)
                {
                    op_data->unstable_write_osds->push_back((unstable_osd_num_t){
                        .osd_num = last_osd,
                        .start = last_start,
                        .len = last_end - last_start,
                    });
                }
                last_osd = it->first.osd_num;
                last_start = last_end;
            }
            op_data->unstable_writes[last_end] = (obj_ver_id){
                .oid = it->first.oid,
                .version = it->second,
            };
            last_end++;
        }
        if (last_osd != 0)
        {
            op_data->unstable_write_osds->push_back((unstable_osd_num_t){
                .osd_num = last_osd,
                .start = last_start,
                .len = last_end - last_start,
            });
        }
        int dpg = 0;
        for (auto dirty_pg_num: dirty_pgs)
        {
            pgs[dirty_pg_num].inflight++;
            op_data->dirty_pgs[dpg++] = dirty_pg_num;
        }
        dirty_pgs.clear();
        this->unstable_writes.clear();
    }
    if (immediate_commit != IMMEDIATE_ALL)
    {
        // SYNC
        submit_primary_sync_subops(cur_op);
resume_3:
        op_data->st = 3;
        return;
resume_4:
        if (op_data->errors > 0)
        {
            goto resume_6;
        }
    }
    // Stabilize version sets
    submit_primary_stab_subops(cur_op);
resume_5:
    op_data->st = 5;
    return;
resume_6:
    for (int i = 0; i < op_data->dirty_pg_count; i++)
    {
        pgs[op_data->dirty_pgs[i]].inflight--;
    }
    if (op_data->errors > 0)
    {
        // Return objects back into the unstable write set
        for (auto unstable_osd: *(op_data->unstable_write_osds))
        {
            for (int i = 0; i < unstable_osd.len; i++)
            {
                // Except those from peered PGs
                auto & w = op_data->unstable_writes[i];
                pg_num_t wpg = map_to_pg(w.oid);
                if (pgs[wpg].state & PG_ACTIVE)
                {
                    uint64_t & dest = this->unstable_writes[(osd_object_id_t){
                        .osd_num = unstable_osd.osd_num,
                        .oid = w.oid,
                    }];
                    dest = dest < w.version ? w.version : dest;
                    dirty_pgs.insert(wpg);
                }
            }
        }
    }
    // FIXME: Free those in the destructor?
    delete op_data->dirty_pgs;
    delete op_data->unstable_write_osds;
    delete[] op_data->unstable_writes;
    op_data->unstable_writes = NULL;
    op_data->unstable_write_osds = NULL;
    if (op_data->errors > 0)
    {
        finish_op(cur_op, op_data->epipe > 0 ? -EPIPE : -EIO);
    }
    else
    {
finish:
        if (cur_op->peer_fd)
        {
            auto it = clients.find(cur_op->peer_fd);
            if (it != clients.end())
                it->second.dirty_pgs.clear();
        }
        finish_op(cur_op, 0);
    }
    assert(syncs_in_progress.front() == cur_op);
    syncs_in_progress.pop_front();
    if (syncs_in_progress.size() > 0)
    {
        cur_op = syncs_in_progress.front();
        op_data->st++;
        goto resume_2;
    }
}

void osd_t::submit_primary_del_subops(osd_op_t *cur_op, uint64_t *cur_set, pg_osd_set_state_t *object_state)
{
    osd_primary_op_data_t *op_data = cur_op->op_data;
    int extra_chunks = 0;
    for (auto chunk: object_state->osd_set)
    {
        if (chunk.osd_num != cur_set[chunk.role])
        {
            extra_chunks++;
        }
    }
    op_data->n_subops = extra_chunks;
    op_data->done = op_data->errors = 0;
    if (!extra_chunks)
    {
        return;
    }
    osd_op_t *subops = new osd_op_t[extra_chunks];
    op_data->subops = subops;
    int i = 0;
    for (auto chunk: object_state->osd_set)
    {
        if (chunk.osd_num != cur_set[chunk.role])
        {
            if (chunk.osd_num == this->osd_num)
            {
                subops[i].bs_op = new blockstore_op_t({
                    .opcode = BS_OP_DELETE,
                    .callback = [cur_op, this](blockstore_op_t *subop)
                    {
                        if (subop->retval != 0)
                        {
                            // die
                            throw std::runtime_error("local delete operation failed");
                        }
                        handle_primary_subop(OSD_OP_SECONDARY_DELETE, cur_op, subop->retval, 0, 0);
                    },
                    .oid = {
                        .inode = op_data->oid.inode,
                        .stripe = op_data->oid.stripe | chunk.role,
                    },
                    // Same version as write
                    .version = op_data->fact_ver,
                });
                bs->enqueue_op(subops[i].bs_op);
            }
            else
            {
                subops[i].op_type = OSD_OP_OUT;
                subops[i].send_list.push_back(subops[i].req.buf, OSD_PACKET_SIZE);
                subops[i].peer_fd = osd_peer_fds.at(chunk.osd_num);
                subops[i].req.sec_del = {
                    .header = {
                        .magic = SECONDARY_OSD_OP_MAGIC,
                        .id = this->next_subop_id++,
                        .opcode = OSD_OP_SECONDARY_DELETE,
                    },
                    .oid = {
                        .inode = op_data->oid.inode,
                        .stripe = op_data->oid.stripe | chunk.role,
                    },
                    // Same version as write
                    .version = op_data->fact_ver,
                };
                subops[i].callback = [cur_op, this](osd_op_t *subop)
                {
                    int fail_fd = subop->reply.hdr.retval != 0 ? subop->peer_fd : 0;
                    handle_primary_subop(OSD_OP_SECONDARY_DELETE, cur_op, subop->reply.hdr.retval, 0, 0);
                    if (fail_fd >= 0)
                    {
                        // delete operation failed, drop the connection
                        stop_client(fail_fd);
                    }
                };
                outbox_push(clients[subops[i].peer_fd], &subops[i]);
            }
            i++;
        }
    }
}

void osd_t::submit_primary_sync_subops(osd_op_t *cur_op)
{
    osd_primary_op_data_t *op_data = cur_op->op_data;
    int n_osds = op_data->unstable_write_osds->size();
    osd_op_t *subops = new osd_op_t[n_osds];
    op_data->done = op_data->errors = 0;
    op_data->n_subops = n_osds;
    op_data->subops = subops;
    for (int i = 0; i < n_osds; i++)
    {
        osd_num_t sync_osd = (*(op_data->unstable_write_osds))[i].osd_num;
        if (sync_osd == this->osd_num)
        {
            subops[i].bs_op = new blockstore_op_t({
                .opcode = BS_OP_SYNC,
                .callback = [cur_op, this](blockstore_op_t *subop)
                {
                    if (subop->retval != 0)
                    {
                        // die
                        throw std::runtime_error("local sync operation failed");
                    }
                    handle_primary_subop(OSD_OP_SECONDARY_SYNC, cur_op, subop->retval, 0, 0);
                },
            });
            bs->enqueue_op(subops[i].bs_op);
        }
        else
        {
            subops[i].op_type = OSD_OP_OUT;
            subops[i].send_list.push_back(subops[i].req.buf, OSD_PACKET_SIZE);
            subops[i].peer_fd = osd_peer_fds.at(sync_osd);
            subops[i].req.sec_sync = {
                .header = {
                    .magic = SECONDARY_OSD_OP_MAGIC,
                    .id = this->next_subop_id++,
                    .opcode = OSD_OP_SECONDARY_SYNC,
                },
            };
            subops[i].callback = [cur_op, this](osd_op_t *subop)
            {
                int fail_fd = subop->reply.hdr.retval != 0 ? subop->peer_fd : 0;
                handle_primary_subop(OSD_OP_SECONDARY_SYNC, cur_op, subop->reply.hdr.retval, 0, 0);
                if (fail_fd >= 0)
                {
                    // sync operation failed, drop the connection
                    stop_client(fail_fd);
                }
            };
            outbox_push(clients[subops[i].peer_fd], &subops[i]);
        }
    }
}

void osd_t::submit_primary_stab_subops(osd_op_t *cur_op)
{
    osd_primary_op_data_t *op_data = cur_op->op_data;
    int n_osds = op_data->unstable_write_osds->size();
    osd_op_t *subops = new osd_op_t[n_osds];
    op_data->done = op_data->errors = 0;
    op_data->n_subops = n_osds;
    op_data->subops = subops;
    for (int i = 0; i < n_osds; i++)
    {
        auto & stab_osd = (*(op_data->unstable_write_osds))[i];
        if (stab_osd.osd_num == this->osd_num)
        {
            subops[i].bs_op = new blockstore_op_t({
                .opcode = BS_OP_STABLE,
                .callback = [cur_op, this](blockstore_op_t *subop)
                {
                    if (subop->retval != 0)
                    {
                        // die
                        throw std::runtime_error("local stabilize operation failed");
                    }
                    handle_primary_subop(OSD_OP_SECONDARY_STABILIZE, cur_op, subop->retval, 0, 0);
                },
                .len = (uint32_t)stab_osd.len,
                .buf = (void*)(op_data->unstable_writes + stab_osd.start),
            });
            bs->enqueue_op(subops[i].bs_op);
        }
        else
        {
            subops[i].op_type = OSD_OP_OUT;
            subops[i].send_list.push_back(subops[i].req.buf, OSD_PACKET_SIZE);
            subops[i].peer_fd = osd_peer_fds.at(stab_osd.osd_num);
            subops[i].req.sec_stab = {
                .header = {
                    .magic = SECONDARY_OSD_OP_MAGIC,
                    .id = this->next_subop_id++,
                    .opcode = OSD_OP_SECONDARY_STABILIZE,
                },
                .len = (uint64_t)(stab_osd.len * sizeof(obj_ver_id)),
            };
            subops[i].send_list.push_back(op_data->unstable_writes + stab_osd.start, stab_osd.len * sizeof(obj_ver_id));
            subops[i].callback = [cur_op, this](osd_op_t *subop)
            {
                int fail_fd = subop->reply.hdr.retval != 0 ? subop->peer_fd : 0;
                handle_primary_subop(OSD_OP_SECONDARY_STABILIZE, cur_op, subop->reply.hdr.retval, 0, 0);
                if (fail_fd >= 0)
                {
                    // sync operation failed, drop the connection
                    stop_client(fail_fd);
                }
            };
            outbox_push(clients[subops[i].peer_fd], &subops[i]);
        }
    }
}