Quelle  vmci_transport_notify.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0-only
/*
 * VMware vSockets Driver
 *
 * Copyright (C) 2009-2013 VMware, Inc. All rights reserved.
 */

#include <linux/types.h>
#include <linux/socket.h>
#include <linux/stddef.h>
#include <net/sock.h>

#include "vmci_transport_notify.h"

#define PKT_FIELD(vsk, field_name) (vmci_trans(vsk)->notify.pkt.field_name)

static bool vmci_transport_notify_waiting_write(struct vsock_sock *vsk)
{
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 bool retval;
 u64 notify_limit;

 if (!PKT_FIELD(vsk, peer_waiting_write))
  return false;

#ifdef VSOCK_OPTIMIZATION_FLOW_CONTROL
 /* When the sender blocks, we take that as a sign that the sender is
 * faster than the receiver. To reduce the transmit rate of the sender,
 * we delay the sending of the read notification by decreasing the
 * write_notify_window. The notification is delayed until the number of
 * bytes used in the queue drops below the write_notify_window.
 */

 if (!PKT_FIELD(vsk, peer_waiting_write_detected)) {
  PKT_FIELD(vsk, peer_waiting_write_detected) = true;
  if (PKT_FIELD(vsk, write_notify_window) < PAGE_SIZE) {
   PKT_FIELD(vsk, write_notify_window) =
       PKT_FIELD(vsk, write_notify_min_window);
  } else {
   PKT_FIELD(vsk, write_notify_window) -= PAGE_SIZE;
   if (PKT_FIELD(vsk, write_notify_window) <
       PKT_FIELD(vsk, write_notify_min_window))
    PKT_FIELD(vsk, write_notify_window) =
        PKT_FIELD(vsk, write_notify_min_window);

  }
 }
 notify_limit = vmci_trans(vsk)->consume_size -
  PKT_FIELD(vsk, write_notify_window);
#else
 notify_limit = 0;
#endif

 /* For now we ignore the wait information and just see if the free
 * space exceeds the notify limit.  Note that improving this function
 * to be more intelligent will not require a protocol change and will
 * retain compatibility between endpoints with mixed versions of this
 * function.
 *
 * The notify_limit is used to delay notifications in the case where
 * flow control is enabled. Below the test is expressed in terms of
 * free space in the queue: if free_space > ConsumeSize -
 * write_notify_window then notify An alternate way of expressing this
 * is to rewrite the expression to use the data ready in the receive
 * queue: if write_notify_window > bufferReady then notify as
 * free_space == ConsumeSize - bufferReady.
 */
 retval = vmci_qpair_consume_free_space(vmci_trans(vsk)->qpair) >
  notify_limit;
#ifdef VSOCK_OPTIMIZATION_FLOW_CONTROL
 if (retval) {
  /*
 * Once we notify the peer, we reset the detected flag so the
 * next wait will again cause a decrease in the window size.
 */

  PKT_FIELD(vsk, peer_waiting_write_detected) = false;
 }
#endif
 return retval;
#else
 return true;
#endif
}

static bool vmci_transport_notify_waiting_read(struct vsock_sock *vsk)
{
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 if (!PKT_FIELD(vsk, peer_waiting_read))
  return false;

 /* For now we ignore the wait information and just see if there is any
 * data for our peer to read.  Note that improving this function to be
 * more intelligent will not require a protocol change and will retain
 * compatibility between endpoints with mixed versions of this
 * function.
 */
 return vmci_qpair_produce_buf_ready(vmci_trans(vsk)->qpair) > 0;
#else
 return true;
#endif
}

static void
vmci_transport_handle_waiting_read(struct sock *sk,
       struct vmci_transport_packet *pkt,
       bool bottom_half,
       struct sockaddr_vm *dst,
       struct sockaddr_vm *src)
{
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 struct vsock_sock *vsk;

 vsk = vsock_sk(sk);

 PKT_FIELD(vsk, peer_waiting_read) = true;
 memcpy(&PKT_FIELD(vsk, peer_waiting_read_info), &pkt->u.wait,
        sizeof(PKT_FIELD(vsk, peer_waiting_read_info)));

 if (vmci_transport_notify_waiting_read(vsk)) {
  bool sent;

  if (bottom_half)
   sent = vmci_transport_send_wrote_bh(dst, src) > 0;
  else
   sent = vmci_transport_send_wrote(sk) > 0;

  if (sent)
   PKT_FIELD(vsk, peer_waiting_read) = false;
 }
#endif
}

static void
vmci_transport_handle_waiting_write(struct sock *sk,
        struct vmci_transport_packet *pkt,
        bool bottom_half,
        struct sockaddr_vm *dst,
        struct sockaddr_vm *src)
{
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 struct vsock_sock *vsk;

 vsk = vsock_sk(sk);

 PKT_FIELD(vsk, peer_waiting_write) = true;
 memcpy(&PKT_FIELD(vsk, peer_waiting_write_info), &pkt->u.wait,
        sizeof(PKT_FIELD(vsk, peer_waiting_write_info)));

 if (vmci_transport_notify_waiting_write(vsk)) {
  bool sent;

  if (bottom_half)
   sent = vmci_transport_send_read_bh(dst, src) > 0;
  else
   sent = vmci_transport_send_read(sk) > 0;

  if (sent)
   PKT_FIELD(vsk, peer_waiting_write) = false;
 }
#endif
}

static void
vmci_transport_handle_read(struct sock *sk,
      struct vmci_transport_packet *pkt,
      bool bottom_half,
      struct sockaddr_vm *dst, struct sockaddr_vm *src)
{
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 struct vsock_sock *vsk;

 vsk = vsock_sk(sk);
 PKT_FIELD(vsk, sent_waiting_write) = false;
#endif

 sk->sk_write_space(sk);
}

static bool send_waiting_read(struct sock *sk, u64 room_needed)
{
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 struct vsock_sock *vsk;
 struct vmci_transport_waiting_info waiting_info;
 u64 tail;
 u64 head;
 u64 room_left;
 bool ret;

 vsk = vsock_sk(sk);

 if (PKT_FIELD(vsk, sent_waiting_read))
  return true;

 if (PKT_FIELD(vsk, write_notify_window) <
   vmci_trans(vsk)->consume_size)
  PKT_FIELD(vsk, write_notify_window) =
      min(PKT_FIELD(vsk, write_notify_window) + PAGE_SIZE,
   vmci_trans(vsk)->consume_size);

 vmci_qpair_get_consume_indexes(vmci_trans(vsk)->qpair, &tail, &head);
 room_left = vmci_trans(vsk)->consume_size - head;
 if (room_needed >= room_left) {
  waiting_info.offset = room_needed - room_left;
  waiting_info.generation =
      PKT_FIELD(vsk, consume_q_generation) + 1;
 } else {
  waiting_info.offset = head + room_needed;
  waiting_info.generation = PKT_FIELD(vsk, consume_q_generation);
 }

 ret = vmci_transport_send_waiting_read(sk, &waiting_info) > 0;
 if (ret)
  PKT_FIELD(vsk, sent_waiting_read) = true;

 return ret;
#else
 return true;
#endif
}

static bool send_waiting_write(struct sock *sk, u64 room_needed)
{
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 struct vsock_sock *vsk;
 struct vmci_transport_waiting_info waiting_info;
 u64 tail;
 u64 head;
 u64 room_left;
 bool ret;

 vsk = vsock_sk(sk);

 if (PKT_FIELD(vsk, sent_waiting_write))
  return true;

 vmci_qpair_get_produce_indexes(vmci_trans(vsk)->qpair, &tail, &head);
 room_left = vmci_trans(vsk)->produce_size - tail;
 if (room_needed + 1 >= room_left) {
  /* Wraps around to current generation. */
  waiting_info.offset = room_needed + 1 - room_left;
  waiting_info.generation = PKT_FIELD(vsk, produce_q_generation);
 } else {
  waiting_info.offset = tail + room_needed + 1;
  waiting_info.generation =
      PKT_FIELD(vsk, produce_q_generation) - 1;
 }

 ret = vmci_transport_send_waiting_write(sk, &waiting_info) > 0;
 if (ret)
  PKT_FIELD(vsk, sent_waiting_write) = true;

 return ret;
#else
 return true;
#endif
}

static int vmci_transport_send_read_notification(struct sock *sk)
{
 struct vsock_sock *vsk;
 bool sent_read;
 unsigned int retries;
 int err;

 vsk = vsock_sk(sk);
 sent_read = false;
 retries = 0;
 err = 0;

 if (vmci_transport_notify_waiting_write(vsk)) {
  /* Notify the peer that we have read, retrying the send on
 * failure up to our maximum value.  XXX For now we just log
 * the failure, but later we should schedule a work item to
 * handle the resend until it succeeds.  That would require
 * keeping track of work items in the vsk and cleaning them up
 * upon socket close.
 */
  while (!(vsk->peer_shutdown & RCV_SHUTDOWN) &&
         !sent_read &&
         retries < VMCI_TRANSPORT_MAX_DGRAM_RESENDS) {
   err = vmci_transport_send_read(sk);
   if (err >= 0)
    sent_read = true;

   retries++;
  }

  if (retries >= VMCI_TRANSPORT_MAX_DGRAM_RESENDS)
   pr_err("%p unable to send read notify to peer\n", sk);
  else
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
   PKT_FIELD(vsk, peer_waiting_write) = false;
#endif

 }
 return err;
}

static void
vmci_transport_handle_wrote(struct sock *sk,
       struct vmci_transport_packet *pkt,
       bool bottom_half,
       struct sockaddr_vm *dst, struct sockaddr_vm *src)
{
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 struct vsock_sock *vsk = vsock_sk(sk);
 PKT_FIELD(vsk, sent_waiting_read) = false;
#endif
 vsock_data_ready(sk);
}

static void vmci_transport_notify_pkt_socket_init(struct sock *sk)
{
 struct vsock_sock *vsk = vsock_sk(sk);

 PKT_FIELD(vsk, write_notify_window) = PAGE_SIZE;
 PKT_FIELD(vsk, write_notify_min_window) = PAGE_SIZE;
 PKT_FIELD(vsk, peer_waiting_read) = false;
 PKT_FIELD(vsk, peer_waiting_write) = false;
 PKT_FIELD(vsk, peer_waiting_write_detected) = false;
 PKT_FIELD(vsk, sent_waiting_read) = false;
 PKT_FIELD(vsk, sent_waiting_write) = false;
 PKT_FIELD(vsk, produce_q_generation) = 0;
 PKT_FIELD(vsk, consume_q_generation) = 0;

 memset(&PKT_FIELD(vsk, peer_waiting_read_info), 0,
        sizeof(PKT_FIELD(vsk, peer_waiting_read_info)));
 memset(&PKT_FIELD(vsk, peer_waiting_write_info), 0,
        sizeof(PKT_FIELD(vsk, peer_waiting_write_info)));
}

static void vmci_transport_notify_pkt_socket_destruct(struct vsock_sock *vsk)
{
}

static int
vmci_transport_notify_pkt_poll_in(struct sock *sk,
      size_t target, bool *data_ready_now)
{
 struct vsock_sock *vsk = vsock_sk(sk);

 if (vsock_stream_has_data(vsk) >= target) {
  *data_ready_now = true;
 } else {
  /* We can't read right now because there is not enough data
 * in the queue. Ask for notifications when there is something
 * to read.
 */
  if (sk->sk_state == TCP_ESTABLISHED) {
   if (!send_waiting_read(sk, 1))
    return -1;

  }
  *data_ready_now = false;
 }

 return 0;
}

static int
vmci_transport_notify_pkt_poll_out(struct sock *sk,
       size_t target, bool *space_avail_now)
{
 s64 produce_q_free_space;
 struct vsock_sock *vsk = vsock_sk(sk);

 produce_q_free_space = vsock_stream_has_space(vsk);
 if (produce_q_free_space > 0) {
  *space_avail_now = true;
  return 0;
 } else if (produce_q_free_space == 0) {
  /* This is a connected socket but we can't currently send data.
 * Notify the peer that we are waiting if the queue is full. We
 * only send a waiting write if the queue is full because
 * otherwise we end up in an infinite WAITING_WRITE, READ,
 * WAITING_WRITE, READ, etc. loop. Treat failing to send the
 * notification as a socket error, passing that back through
 * the mask.
 */
  if (!send_waiting_write(sk, 1))
   return -1;

  *space_avail_now = false;
 }

 return 0;
}

static int
vmci_transport_notify_pkt_recv_init(
   struct sock *sk,
   size_t target,
   struct vmci_transport_recv_notify_data *data)
{
 struct vsock_sock *vsk = vsock_sk(sk);

#ifdef VSOCK_OPTIMIZATION_WAITING_NOTIFY
 data->consume_head = 0;
 data->produce_tail = 0;
#ifdef VSOCK_OPTIMIZATION_FLOW_CONTROL
 data->notify_on_block = false;

 if (PKT_FIELD(vsk, write_notify_min_window) < target + 1) {
  PKT_FIELD(vsk, write_notify_min_window) = target + 1;
  if (PKT_FIELD(vsk, write_notify_window) <
      PKT_FIELD(vsk, write_notify_min_window)) {
   /* If the current window is smaller than the new
 * minimal window size, we need to reevaluate whether
 * we need to notify the sender. If the number of ready
 * bytes are smaller than the new window, we need to
 * send a notification to the sender before we block.
 */

   PKT_FIELD(vsk, write_notify_window) =
       PKT_FIELD(vsk, write_notify_min_window);
   data->notify_on_block = true;
  }
 }
#endif
#endif

 return 0;
}

static int
vmci_transport_notify_pkt_recv_pre_block(
    struct sock *sk,
    size_t target,
    struct vmci_transport_recv_notify_data *data)
{
 int err = 0;

 /* Notify our peer that we are waiting for data to read. */
 if (!send_waiting_read(sk, target)) {
  err = -EHOSTUNREACH;
  return err;
 }
#ifdef VSOCK_OPTIMIZATION_FLOW_CONTROL
 if (data->notify_on_block) {
  err = vmci_transport_send_read_notification(sk);
  if (err < 0)
   return err;

  data->notify_on_block = false;
 }
#endif

 return err;
}

static int
vmci_transport_notify_pkt_recv_pre_dequeue(
    struct sock *sk,
    size_t target,
    struct vmci_transport_recv_notify_data *data)
{
 struct vsock_sock *vsk = vsock_sk(sk);

 /* Now consume up to len bytes from the queue.  Note that since we have
 * the socket locked we should copy at least ready bytes.
 */
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 vmci_qpair_get_consume_indexes(vmci_trans(vsk)->qpair,
           &data->produce_tail,
           &data->consume_head);
#endif

 return 0;
}

static int
vmci_transport_notify_pkt_recv_post_dequeue(
    struct sock *sk,
    size_t target,
    ssize_t copied,
    bool data_read,
    struct vmci_transport_recv_notify_data *data)
{
 struct vsock_sock *vsk;
 int err;

 vsk = vsock_sk(sk);
 err = 0;

 if (data_read) {
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
  /* Detect a wrap-around to maintain queue generation.  Note
 * that this is safe since we hold the socket lock across the
 * two queue pair operations.
 */
  if (copied >=
   vmci_trans(vsk)->consume_size - data->consume_head)
   PKT_FIELD(vsk, consume_q_generation)++;
#endif

  err = vmci_transport_send_read_notification(sk);
  if (err < 0)
   return err;

 }
 return err;
}

static int
vmci_transport_notify_pkt_send_init(
   struct sock *sk,
   struct vmci_transport_send_notify_data *data)
{
#ifdef VSOCK_OPTIMIZATION_WAITING_NOTIFY
 data->consume_head = 0;
 data->produce_tail = 0;
#endif

 return 0;
}

static int
vmci_transport_notify_pkt_send_pre_block(
    struct sock *sk,
    struct vmci_transport_send_notify_data *data)
{
 /* Notify our peer that we are waiting for room to write. */
 if (!send_waiting_write(sk, 1))
  return -EHOSTUNREACH;

 return 0;
}

static int
vmci_transport_notify_pkt_send_pre_enqueue(
    struct sock *sk,
    struct vmci_transport_send_notify_data *data)
{
 struct vsock_sock *vsk = vsock_sk(sk);

#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 vmci_qpair_get_produce_indexes(vmci_trans(vsk)->qpair,
           &data->produce_tail,
           &data->consume_head);
#endif

 return 0;
}

static int
vmci_transport_notify_pkt_send_post_enqueue(
    struct sock *sk,
    ssize_t written,
    struct vmci_transport_send_notify_data *data)
{
 int err = 0;
 struct vsock_sock *vsk;
 bool sent_wrote = false;
 int retries = 0;

 vsk = vsock_sk(sk);

#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
 /* Detect a wrap-around to maintain queue generation.  Note that this
 * is safe since we hold the socket lock across the two queue pair
 * operations.
 */
 if (written >= vmci_trans(vsk)->produce_size - data->produce_tail)
  PKT_FIELD(vsk, produce_q_generation)++;

#endif

 if (vmci_transport_notify_waiting_read(vsk)) {
  /* Notify the peer that we have written, retrying the send on
 * failure up to our maximum value. See the XXX comment for the
 * corresponding piece of code in StreamRecvmsg() for potential
 * improvements.
 */
  while (!(vsk->peer_shutdown & RCV_SHUTDOWN) &&
         !sent_wrote &&
         retries < VMCI_TRANSPORT_MAX_DGRAM_RESENDS) {
   err = vmci_transport_send_wrote(sk);
   if (err >= 0)
    sent_wrote = true;

   retries++;
  }

  if (retries >= VMCI_TRANSPORT_MAX_DGRAM_RESENDS) {
   pr_err("%p unable to send wrote notify to peer\n", sk);
   return err;
  } else {
#if defined(VSOCK_OPTIMIZATION_WAITING_NOTIFY)
   PKT_FIELD(vsk, peer_waiting_read) = false;
#endif
  }
 }
 return err;
}

static void
vmci_transport_notify_pkt_handle_pkt(
   struct sock *sk,
   struct vmci_transport_packet *pkt,
   bool bottom_half,
   struct sockaddr_vm *dst,
   struct sockaddr_vm *src, bool *pkt_processed)
{
 bool processed = false;

 switch (pkt->type) {
 case VMCI_TRANSPORT_PACKET_TYPE_WROTE:
  vmci_transport_handle_wrote(sk, pkt, bottom_half, dst, src);
  processed = true;
  break;
 case VMCI_TRANSPORT_PACKET_TYPE_READ:
  vmci_transport_handle_read(sk, pkt, bottom_half, dst, src);
  processed = true;
  break;
 case VMCI_TRANSPORT_PACKET_TYPE_WAITING_WRITE:
  vmci_transport_handle_waiting_write(sk, pkt, bottom_half,
          dst, src);
  processed = true;
  break;

 case VMCI_TRANSPORT_PACKET_TYPE_WAITING_READ:
  vmci_transport_handle_waiting_read(sk, pkt, bottom_half,
         dst, src);
  processed = true;
  break;
 }

 if (pkt_processed)
  *pkt_processed = processed;
}

static void vmci_transport_notify_pkt_process_request(struct sock *sk)
{
 struct vsock_sock *vsk = vsock_sk(sk);

 PKT_FIELD(vsk, write_notify_window) = vmci_trans(vsk)->consume_size;
 if (vmci_trans(vsk)->consume_size <
  PKT_FIELD(vsk, write_notify_min_window))
  PKT_FIELD(vsk, write_notify_min_window) =
   vmci_trans(vsk)->consume_size;
}

static void vmci_transport_notify_pkt_process_negotiate(struct sock *sk)
{
 struct vsock_sock *vsk = vsock_sk(sk);

 PKT_FIELD(vsk, write_notify_window) = vmci_trans(vsk)->consume_size;
 if (vmci_trans(vsk)->consume_size <
  PKT_FIELD(vsk, write_notify_min_window))
  PKT_FIELD(vsk, write_notify_min_window) =
   vmci_trans(vsk)->consume_size;
}

/* Socket control packet based operations. */
const struct vmci_transport_notify_ops vmci_transport_notify_pkt_ops = {
 .socket_init = vmci_transport_notify_pkt_socket_init,
 .socket_destruct = vmci_transport_notify_pkt_socket_destruct,
 .poll_in = vmci_transport_notify_pkt_poll_in,
 .poll_out = vmci_transport_notify_pkt_poll_out,
 .handle_notify_pkt = vmci_transport_notify_pkt_handle_pkt,
 .recv_init = vmci_transport_notify_pkt_recv_init,
 .recv_pre_block = vmci_transport_notify_pkt_recv_pre_block,
 .recv_pre_dequeue = vmci_transport_notify_pkt_recv_pre_dequeue,
 .recv_post_dequeue = vmci_transport_notify_pkt_recv_post_dequeue,
 .send_init = vmci_transport_notify_pkt_send_init,
 .send_pre_block = vmci_transport_notify_pkt_send_pre_block,
 .send_pre_enqueue = vmci_transport_notify_pkt_send_pre_enqueue,
 .send_post_enqueue = vmci_transport_notify_pkt_send_post_enqueue,
 .process_request = vmci_transport_notify_pkt_process_request,
 .process_negotiate = vmci_transport_notify_pkt_process_negotiate,
};