Quelle  flush.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0-only
/*
 * Copyright 2023 Red Hat
 */

#include "flush.h"

#include <linux/mempool.h>
#include <linux/spinlock.h>

#include "logger.h"
#include "memory-alloc.h"
#include "permassert.h"

#include "admin-state.h"
#include "completion.h"
#include "io-submitter.h"
#include "logical-zone.h"
#include "slab-depot.h"
#include "types.h"
#include "vdo.h"

struct flusher {
 struct vdo_completion completion;
 /* The vdo to which this flusher belongs */
 struct vdo *vdo;
 /* The administrative state of the flusher */
 struct admin_state state;
 /* The current flush generation of the vdo */
 sequence_number_t flush_generation;
 /* The first unacknowledged flush generation */
 sequence_number_t first_unacknowledged_generation;
 /* The queue of flush requests waiting to notify other threads */
 struct vdo_wait_queue notifiers;
 /* The queue of flush requests waiting for VIOs to complete */
 struct vdo_wait_queue pending_flushes;
 /* The flush generation for which notifications are being sent */
 sequence_number_t notify_generation;
 /* The logical zone to notify next */
 struct logical_zone *logical_zone_to_notify;
 /* The ID of the thread on which flush requests should be made */
 thread_id_t thread_id;
 /* The pool of flush requests */
 mempool_t *flush_pool;
 /* Bios waiting for a flush request to become available */
 struct bio_list waiting_flush_bios;
 /* The lock to protect the previous fields */
 spinlock_t lock;
 /* The rotor for selecting the bio queue for submitting flush bios */
 zone_count_t bio_queue_rotor;
 /* The number of flushes submitted to the current bio queue */
 int flush_count;
};

/**
 * assert_on_flusher_thread() - Check that we are on the flusher thread.
 * @flusher: The flusher.
 * @caller: The function which is asserting.
 */
static inline void assert_on_flusher_thread(struct flusher *flusher, const char *caller)
{
 VDO_ASSERT_LOG_ONLY((vdo_get_callback_thread_id() == flusher->thread_id),
       "%s() called from flusher thread", caller);
}

/**
 * as_flusher() - Convert a generic vdo_completion to a flusher.
 * @completion: The completion to convert.
 *
 * Return: The completion as a flusher.
 */
static struct flusher *as_flusher(struct vdo_completion *completion)
{
 vdo_assert_completion_type(completion, VDO_FLUSH_NOTIFICATION_COMPLETION);
 return container_of(completion, struct flusher, completion);
}

/**
 * completion_as_vdo_flush() - Convert a generic vdo_completion to a vdo_flush.
 * @completion: The completion to convert.
 *
 * Return: The completion as a vdo_flush.
 */
static inline struct vdo_flush *completion_as_vdo_flush(struct vdo_completion *completion)
{
 vdo_assert_completion_type(completion, VDO_FLUSH_COMPLETION);
 return container_of(completion, struct vdo_flush, completion);
}

/**
 * vdo_waiter_as_flush() - Convert a vdo_flush's generic wait queue entry back to the vdo_flush.
 * @waiter: The wait queue entry to convert.
 *
 * Return: The wait queue entry as a vdo_flush.
 */
static struct vdo_flush *vdo_waiter_as_flush(struct vdo_waiter *waiter)
{
 return container_of(waiter, struct vdo_flush, waiter);
}

static void *allocate_flush(gfp_t gfp_mask, void *pool_data)
{
 struct vdo_flush *flush = NULL;

 if ((gfp_mask & GFP_NOWAIT) == GFP_NOWAIT) {
  flush = vdo_allocate_memory_nowait(sizeof(struct vdo_flush), __func__);
 } else {
  int result = vdo_allocate(1, struct vdo_flush, __func__, &flush);

  if (result != VDO_SUCCESS)
   vdo_log_error_strerror(result, "failed to allocate spare flush");
 }

 if (flush != NULL) {
  struct flusher *flusher = pool_data;

  vdo_initialize_completion(&flush->completion, flusher->vdo,
       VDO_FLUSH_COMPLETION);
 }

 return flush;
}

static void free_flush(void *element, void *pool_data __always_unused)
{
 vdo_free(element);
}

/**
 * vdo_make_flusher() - Make a flusher for a vdo.
 * @vdo: The vdo which owns the flusher.
 *
 * Return: VDO_SUCCESS or an error.
 */
int vdo_make_flusher(struct vdo *vdo)
{
 int result = vdo_allocate(1, struct flusher, __func__, &vdo->flusher);

 if (result != VDO_SUCCESS)
  return result;

 vdo->flusher->vdo = vdo;
 vdo->flusher->thread_id = vdo->thread_config.packer_thread;
 vdo_set_admin_state_code(&vdo->flusher->state, VDO_ADMIN_STATE_NORMAL_OPERATION);
 vdo_initialize_completion(&vdo->flusher->completion, vdo,
      VDO_FLUSH_NOTIFICATION_COMPLETION);

 spin_lock_init(&vdo->flusher->lock);
 bio_list_init(&vdo->flusher->waiting_flush_bios);
 vdo->flusher->flush_pool = mempool_create(1, allocate_flush, free_flush,
        vdo->flusher);
 return ((vdo->flusher->flush_pool == NULL) ? -ENOMEM : VDO_SUCCESS);
}

/**
 * vdo_free_flusher() - Free a flusher.
 * @flusher: The flusher to free.
 */
void vdo_free_flusher(struct flusher *flusher)
{
 if (flusher == NULL)
  return;

 if (flusher->flush_pool != NULL)
  mempool_destroy(vdo_forget(flusher->flush_pool));
 vdo_free(flusher);
}

/**
 * vdo_get_flusher_thread_id() - Get the ID of the thread on which flusher functions should be
 *                               called.
 * @flusher: The flusher to query.
 *
 * Return: The ID of the thread which handles the flusher.
 */
thread_id_t vdo_get_flusher_thread_id(struct flusher *flusher)
{
 return flusher->thread_id;
}

static void notify_flush(struct flusher *flusher);
static void vdo_complete_flush(struct vdo_flush *flush);

/**
 * finish_notification() - Finish the notification process.
 * @completion: The flusher completion.
 *
 * Finishes the notification process by checking if any flushes have completed and then starting
 * the notification of the next flush request if one came in while the current notification was in
 * progress. This callback is registered in flush_packer_callback().
 */
static void finish_notification(struct vdo_completion *completion)
{
 struct flusher *flusher = as_flusher(completion);

 assert_on_flusher_thread(flusher, __func__);

 vdo_waitq_enqueue_waiter(&flusher->pending_flushes,
     vdo_waitq_dequeue_waiter(&flusher->notifiers));
 vdo_complete_flushes(flusher);
 if (vdo_waitq_has_waiters(&flusher->notifiers))
  notify_flush(flusher);
}

/**
 * flush_packer_callback() - Flush the packer.
 * @completion: The flusher completion.
 *
 * Flushes the packer now that all of the logical and physical zones have been notified of the new
 * flush request. This callback is registered in increment_generation().
 */
static void flush_packer_callback(struct vdo_completion *completion)
{
 struct flusher *flusher = as_flusher(completion);

 vdo_increment_packer_flush_generation(flusher->vdo->packer);
 vdo_launch_completion_callback(completion, finish_notification,
           flusher->thread_id);
}

/**
 * increment_generation() - Increment the flush generation in a logical zone.
 * @completion: The flusher as a completion.
 *
 * If there are more logical zones, go on to the next one, otherwise, prepare the physical zones.
 * This callback is registered both in notify_flush() and in itself.
 */
static void increment_generation(struct vdo_completion *completion)
{
 struct flusher *flusher = as_flusher(completion);
 struct logical_zone *zone = flusher->logical_zone_to_notify;

 vdo_increment_logical_zone_flush_generation(zone, flusher->notify_generation);
 if (zone->next == NULL) {
  vdo_launch_completion_callback(completion, flush_packer_callback,
            flusher->thread_id);
  return;
 }

 flusher->logical_zone_to_notify = zone->next;
 vdo_launch_completion_callback(completion, increment_generation,
           flusher->logical_zone_to_notify->thread_id);
}

/**
 * notify_flush() - Launch a flush notification.
 * @flusher: The flusher doing the notification.
 */
static void notify_flush(struct flusher *flusher)
{
 struct vdo_flush *flush =
  vdo_waiter_as_flush(vdo_waitq_get_first_waiter(&flusher->notifiers));

 flusher->notify_generation = flush->flush_generation;
 flusher->logical_zone_to_notify = &flusher->vdo->logical_zones->zones[0];
 flusher->completion.requeue = true;
 vdo_launch_completion_callback(&flusher->completion, increment_generation,
           flusher->logical_zone_to_notify->thread_id);
}

/**
 * flush_vdo() - Start processing a flush request.
 * @completion: A flush request (as a vdo_completion)
 *
 * This callback is registered in launch_flush().
 */
static void flush_vdo(struct vdo_completion *completion)
{
 struct vdo_flush *flush = completion_as_vdo_flush(completion);
 struct flusher *flusher = completion->vdo->flusher;
 bool may_notify;
 int result;

 assert_on_flusher_thread(flusher, __func__);
 result = VDO_ASSERT(vdo_is_state_normal(&flusher->state),
       "flusher is in normal operation");
 if (result != VDO_SUCCESS) {
  vdo_enter_read_only_mode(flusher->vdo, result);
  vdo_complete_flush(flush);
  return;
 }

 flush->flush_generation = flusher->flush_generation++;
 may_notify = !vdo_waitq_has_waiters(&flusher->notifiers);
 vdo_waitq_enqueue_waiter(&flusher->notifiers, &flush->waiter);
 if (may_notify)
  notify_flush(flusher);
}

/**
 * check_for_drain_complete() - Check whether the flusher has drained.
 * @flusher: The flusher.
 */
static void check_for_drain_complete(struct flusher *flusher)
{
 bool drained;

 if (!vdo_is_state_draining(&flusher->state) ||
     vdo_waitq_has_waiters(&flusher->pending_flushes))
  return;

 spin_lock(&flusher->lock);
 drained = bio_list_empty(&flusher->waiting_flush_bios);
 spin_unlock(&flusher->lock);

 if (drained)
  vdo_finish_draining(&flusher->state);
}

/**
 * vdo_complete_flushes() - Attempt to complete any flushes which might have finished.
 * @flusher: The flusher.
 */
void vdo_complete_flushes(struct flusher *flusher)
{
 sequence_number_t oldest_active_generation = U64_MAX;
 struct logical_zone *zone;

 assert_on_flusher_thread(flusher, __func__);

 for (zone = &flusher->vdo->logical_zones->zones[0]; zone != NULL; zone = zone->next)
  oldest_active_generation =
   min(oldest_active_generation,
       READ_ONCE(zone->oldest_active_generation));

 while (vdo_waitq_has_waiters(&flusher->pending_flushes)) {
  struct vdo_flush *flush =
   vdo_waiter_as_flush(vdo_waitq_get_first_waiter(&flusher->pending_flushes));

  if (flush->flush_generation >= oldest_active_generation)
   return;

  VDO_ASSERT_LOG_ONLY((flush->flush_generation ==
         flusher->first_unacknowledged_generation),
        "acknowledged next expected flush, %llu, was: %llu",
        (unsigned long long) flusher->first_unacknowledged_generation,
        (unsigned long long) flush->flush_generation);
  vdo_waitq_dequeue_waiter(&flusher->pending_flushes);
  vdo_complete_flush(flush);
  flusher->first_unacknowledged_generation++;
 }

 check_for_drain_complete(flusher);
}

/**
 * vdo_dump_flusher() - Dump the flusher, in a thread-unsafe fashion.
 * @flusher: The flusher.
 */
void vdo_dump_flusher(const struct flusher *flusher)
{
 vdo_log_info("struct flusher");
 vdo_log_info(" flush_generation=%llu first_unacknowledged_generation=%llu",
       (unsigned long long) flusher->flush_generation,
       (unsigned long long) flusher->first_unacknowledged_generation);
 vdo_log_info(" notifiers queue is %s; pending_flushes queue is %s",
       (vdo_waitq_has_waiters(&flusher->notifiers) ? "not empty" : "empty"),
       (vdo_waitq_has_waiters(&flusher->pending_flushes) ? "not empty" : "empty"));
}

/**
 * initialize_flush() - Initialize a vdo_flush structure.
 * @flush: The flush to initialize.
 * @vdo: The vdo being flushed.
 *
 * Initializes a vdo_flush structure, transferring all the bios in the flusher's waiting_flush_bios
 * list to it. The caller MUST already hold the lock.
 */
static void initialize_flush(struct vdo_flush *flush, struct vdo *vdo)
{
 bio_list_init(&flush->bios);
 bio_list_merge_init(&flush->bios, &vdo->flusher->waiting_flush_bios);
}

static void launch_flush(struct vdo_flush *flush)
{
 struct vdo_completion *completion = &flush->completion;

 vdo_prepare_completion(completion, flush_vdo, flush_vdo,
          completion->vdo->thread_config.packer_thread, NULL);
 vdo_enqueue_completion(completion, VDO_DEFAULT_Q_FLUSH_PRIORITY);
}

/**
 * vdo_launch_flush() - Function called to start processing a flush request.
 * @vdo: The vdo.
 * @bio: The bio containing an empty flush request.
 *
 * This is called when we receive an empty flush bio from the block layer, and before acknowledging
 * a non-empty bio with the FUA flag set.
 */
void vdo_launch_flush(struct vdo *vdo, struct bio *bio)
{
 /*
 * Try to allocate a vdo_flush to represent the flush request. If the allocation fails,
 * we'll deal with it later.
 */
 struct vdo_flush *flush = mempool_alloc(vdo->flusher->flush_pool, GFP_NOWAIT);
 struct flusher *flusher = vdo->flusher;
 const struct admin_state_code *code = vdo_get_admin_state_code(&flusher->state);

 VDO_ASSERT_LOG_ONLY(!code->quiescent, "Flushing not allowed in state %s",
       code->name);

 spin_lock(&flusher->lock);

 /* We have a new bio to start. Add it to the list. */
 bio_list_add(&flusher->waiting_flush_bios, bio);

 if (flush == NULL) {
  spin_unlock(&flusher->lock);
  return;
 }

 /* We have flushes to start. Capture them in the vdo_flush structure. */
 initialize_flush(flush, vdo);
 spin_unlock(&flusher->lock);

 /* Finish launching the flushes. */
 launch_flush(flush);
}

/**
 * release_flush() - Release a vdo_flush structure that has completed its work.
 * @flush: The completed flush structure to re-use or free.
 *
 * If there are any pending flush requests whose vdo_flush allocation failed, they will be launched
 * by immediately re-using the released vdo_flush. If there is no spare vdo_flush, the released
 * structure will become the spare. Otherwise, the vdo_flush will be freed.
 */
static void release_flush(struct vdo_flush *flush)
{
 bool relaunch_flush;
 struct flusher *flusher = flush->completion.vdo->flusher;

 spin_lock(&flusher->lock);
 if (bio_list_empty(&flusher->waiting_flush_bios)) {
  relaunch_flush = false;
 } else {
  /* We have flushes to start. Capture them in a flush request. */
  initialize_flush(flush, flusher->vdo);
  relaunch_flush = true;
 }
 spin_unlock(&flusher->lock);

 if (relaunch_flush) {
  /* Finish launching the flushes. */
  launch_flush(flush);
  return;
 }

 mempool_free(flush, flusher->flush_pool);
}

/**
 * vdo_complete_flush_callback() - Function called to complete and free a flush request, registered
 *                                 in vdo_complete_flush().
 * @completion: The flush request.
 */
static void vdo_complete_flush_callback(struct vdo_completion *completion)
{
 struct vdo_flush *flush = completion_as_vdo_flush(completion);
 struct vdo *vdo = completion->vdo;
 struct bio *bio;

 while ((bio = bio_list_pop(&flush->bios)) != NULL) {
  /*
 * We're not acknowledging this bio now, but we'll never touch it again, so this is
 * the last chance to account for it.
 */
  vdo_count_bios(&vdo->stats.bios_acknowledged, bio);

  /* Update the device, and send it on down... */
  bio_set_dev(bio, vdo_get_backing_device(vdo));
  atomic64_inc(&vdo->stats.flush_out);
  submit_bio_noacct(bio);
 }


 /*
 * Release the flush structure, freeing it, re-using it as the spare, or using it to launch
 * any flushes that had to wait when allocations failed.
 */
 release_flush(flush);
}

/**
 * select_bio_queue() - Select the bio queue on which to finish a flush request.
 * @flusher: The flusher finishing the request.
 */
static thread_id_t select_bio_queue(struct flusher *flusher)
{
 struct vdo *vdo = flusher->vdo;
 zone_count_t bio_threads = flusher->vdo->thread_config.bio_thread_count;
 int interval;

 if (bio_threads == 1)
  return vdo->thread_config.bio_threads[0];

 interval = vdo->device_config->thread_counts.bio_rotation_interval;
 if (flusher->flush_count == interval) {
  flusher->flush_count = 1;
  flusher->bio_queue_rotor = ((flusher->bio_queue_rotor + 1) % bio_threads);
 } else {
  flusher->flush_count++;
 }

 return vdo->thread_config.bio_threads[flusher->bio_queue_rotor];
}

/**
 * vdo_complete_flush() - Complete and free a vdo flush request.
 * @flush: The flush request.
 */
static void vdo_complete_flush(struct vdo_flush *flush)
{
 struct vdo_completion *completion = &flush->completion;

 vdo_prepare_completion(completion, vdo_complete_flush_callback,
          vdo_complete_flush_callback,
          select_bio_queue(completion->vdo->flusher), NULL);
 vdo_enqueue_completion(completion, BIO_Q_FLUSH_PRIORITY);
}

/**
 * initiate_drain() - Initiate a drain.
 *
 * Implements vdo_admin_initiator_fn.
 */
static void initiate_drain(struct admin_state *state)
{
 check_for_drain_complete(container_of(state, struct flusher, state));
}

/**
 * vdo_drain_flusher() - Drain the flusher.
 * @flusher: The flusher to drain.
 * @completion: The completion to finish when the flusher has drained.
 *
 * Drains the flusher by preventing any more VIOs from entering the flusher and then flushing. The
 * flusher will be left in the suspended state.
 */
void vdo_drain_flusher(struct flusher *flusher, struct vdo_completion *completion)
{
 assert_on_flusher_thread(flusher, __func__);
 vdo_start_draining(&flusher->state, VDO_ADMIN_STATE_SUSPENDING, completion,
      initiate_drain);
}

/**
 * vdo_resume_flusher() - Resume a flusher which has been suspended.
 * @flusher: The flusher to resume.
 * @parent: The completion to finish when the flusher has resumed.
 */
void vdo_resume_flusher(struct flusher *flusher, struct vdo_completion *parent)
{
 assert_on_flusher_thread(flusher, __func__);
 vdo_continue_completion(parent, vdo_resume_if_quiescent(&flusher->state));
}