| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * ACRN HSM eventfd - use eventfd objects to signal expected I/O requests |
| 4 | * |
| 5 | * Copyright (C) 2020 Intel Corporation. All rights reserved. |
| 6 | * |
| 7 | * Authors: |
| 8 | * Shuo Liu <shuo.a.liu@intel.com> |
| 9 | * Yakui Zhao <yakui.zhao@intel.com> |
| 10 | */ |
| 11 | |
| 12 | #include <linux/eventfd.h> |
| 13 | #include <linux/slab.h> |
| 14 | |
| 15 | #include "acrn_drv.h" |
| 16 | |
| 17 | /** |
| 18 | * struct hsm_ioeventfd - Properties of HSM ioeventfd |
| 19 | * @list: Entry within &acrn_vm.ioeventfds of ioeventfds of a VM |
| 20 | * @eventfd: Eventfd of the HSM ioeventfd |
| 21 | * @addr: Address of I/O range |
| 22 | * @data: Data for matching |
| 23 | * @length: Length of I/O range |
| 24 | * @type: Type of I/O range (ACRN_IOREQ_TYPE_MMIO/ACRN_IOREQ_TYPE_PORTIO) |
| 25 | * @wildcard: Data matching or not |
| 26 | */ |
| 27 | struct hsm_ioeventfd { |
| 28 | struct list_head list; |
| 29 | struct eventfd_ctx *eventfd; |
| 30 | u64 addr; |
| 31 | u64 data; |
| 32 | int length; |
| 33 | int type; |
| 34 | bool wildcard; |
| 35 | }; |
| 36 | |
| 37 | static inline int ioreq_type_from_flags(int flags) |
| 38 | { |
| 39 | return flags & ACRN_IOEVENTFD_FLAG_PIO ? |
| 40 | ACRN_IOREQ_TYPE_PORTIO : ACRN_IOREQ_TYPE_MMIO; |
| 41 | } |
| 42 | |
| 43 | static void acrn_ioeventfd_shutdown(struct acrn_vm *vm, struct hsm_ioeventfd *p) |
| 44 | { |
| 45 | lockdep_assert_held(&vm->ioeventfds_lock); |
| 46 | |
| 47 | eventfd_ctx_put(ctx: p->eventfd); |
| 48 | list_del(entry: &p->list); |
| 49 | kfree(objp: p); |
| 50 | } |
| 51 | |
| 52 | static bool hsm_ioeventfd_is_conflict(struct acrn_vm *vm, |
| 53 | struct hsm_ioeventfd *ioeventfd) |
| 54 | { |
| 55 | struct hsm_ioeventfd *p; |
| 56 | |
| 57 | lockdep_assert_held(&vm->ioeventfds_lock); |
| 58 | |
| 59 | /* Either one is wildcard, the data matching will be skipped. */ |
| 60 | list_for_each_entry(p, &vm->ioeventfds, list) |
| 61 | if (p->eventfd == ioeventfd->eventfd && |
| 62 | p->addr == ioeventfd->addr && |
| 63 | p->type == ioeventfd->type && |
| 64 | (p->wildcard || ioeventfd->wildcard || |
| 65 | p->data == ioeventfd->data)) |
| 66 | return true; |
| 67 | |
| 68 | return false; |
| 69 | } |
| 70 | |
| 71 | /* |
| 72 | * Assign an eventfd to a VM and create a HSM ioeventfd associated with the |
| 73 | * eventfd. The properties of the HSM ioeventfd are built from a &struct |
| 74 | * acrn_ioeventfd. |
| 75 | */ |
| 76 | static int acrn_ioeventfd_assign(struct acrn_vm *vm, |
| 77 | struct acrn_ioeventfd *args) |
| 78 | { |
| 79 | struct eventfd_ctx *eventfd; |
| 80 | struct hsm_ioeventfd *p; |
| 81 | int ret; |
| 82 | |
| 83 | /* Check for range overflow */ |
| 84 | if (args->addr + args->len < args->addr) |
| 85 | return -EINVAL; |
| 86 | |
| 87 | /* |
| 88 | * Currently, acrn_ioeventfd is used to support vhost. 1,2,4,8 width |
| 89 | * accesses can cover vhost's requirements. |
| 90 | */ |
| 91 | if (!(args->len == 1 || args->len == 2 || |
| 92 | args->len == 4 || args->len == 8)) |
| 93 | return -EINVAL; |
| 94 | |
| 95 | eventfd = eventfd_ctx_fdget(fd: args->fd); |
| 96 | if (IS_ERR(ptr: eventfd)) |
| 97 | return PTR_ERR(ptr: eventfd); |
| 98 | |
| 99 | p = kzalloc(sizeof(*p), GFP_KERNEL); |
| 100 | if (!p) { |
| 101 | ret = -ENOMEM; |
| 102 | goto fail; |
| 103 | } |
| 104 | |
| 105 | INIT_LIST_HEAD(list: &p->list); |
| 106 | p->addr = args->addr; |
| 107 | p->length = args->len; |
| 108 | p->eventfd = eventfd; |
| 109 | p->type = ioreq_type_from_flags(flags: args->flags); |
| 110 | |
| 111 | /* |
| 112 | * ACRN_IOEVENTFD_FLAG_DATAMATCH flag is set in virtio 1.0 support, the |
| 113 | * writing of notification register of each virtqueue may trigger the |
| 114 | * notification. There is no data matching requirement. |
| 115 | */ |
| 116 | if (args->flags & ACRN_IOEVENTFD_FLAG_DATAMATCH) |
| 117 | p->data = args->data; |
| 118 | else |
| 119 | p->wildcard = true; |
| 120 | |
| 121 | mutex_lock(&vm->ioeventfds_lock); |
| 122 | |
| 123 | if (hsm_ioeventfd_is_conflict(vm, ioeventfd: p)) { |
| 124 | ret = -EEXIST; |
| 125 | goto unlock_fail; |
| 126 | } |
| 127 | |
| 128 | /* register the I/O range into ioreq client */ |
| 129 | ret = acrn_ioreq_range_add(client: vm->ioeventfd_client, type: p->type, |
| 130 | start: p->addr, end: p->addr + p->length - 1); |
| 131 | if (ret < 0) |
| 132 | goto unlock_fail; |
| 133 | |
| 134 | list_add_tail(new: &p->list, head: &vm->ioeventfds); |
| 135 | mutex_unlock(lock: &vm->ioeventfds_lock); |
| 136 | |
| 137 | return 0; |
| 138 | |
| 139 | unlock_fail: |
| 140 | mutex_unlock(lock: &vm->ioeventfds_lock); |
| 141 | kfree(objp: p); |
| 142 | fail: |
| 143 | eventfd_ctx_put(ctx: eventfd); |
| 144 | return ret; |
| 145 | } |
| 146 | |
| 147 | static int acrn_ioeventfd_deassign(struct acrn_vm *vm, |
| 148 | struct acrn_ioeventfd *args) |
| 149 | { |
| 150 | struct hsm_ioeventfd *p; |
| 151 | struct eventfd_ctx *eventfd; |
| 152 | |
| 153 | eventfd = eventfd_ctx_fdget(fd: args->fd); |
| 154 | if (IS_ERR(ptr: eventfd)) |
| 155 | return PTR_ERR(ptr: eventfd); |
| 156 | |
| 157 | mutex_lock(&vm->ioeventfds_lock); |
| 158 | list_for_each_entry(p, &vm->ioeventfds, list) { |
| 159 | if (p->eventfd != eventfd) |
| 160 | continue; |
| 161 | |
| 162 | acrn_ioreq_range_del(client: vm->ioeventfd_client, type: p->type, |
| 163 | start: p->addr, end: p->addr + p->length - 1); |
| 164 | acrn_ioeventfd_shutdown(vm, p); |
| 165 | break; |
| 166 | } |
| 167 | mutex_unlock(lock: &vm->ioeventfds_lock); |
| 168 | |
| 169 | eventfd_ctx_put(ctx: eventfd); |
| 170 | return 0; |
| 171 | } |
| 172 | |
| 173 | static struct hsm_ioeventfd *hsm_ioeventfd_match(struct acrn_vm *vm, u64 addr, |
| 174 | u64 data, int len, int type) |
| 175 | { |
| 176 | struct hsm_ioeventfd *p = NULL; |
| 177 | |
| 178 | lockdep_assert_held(&vm->ioeventfds_lock); |
| 179 | |
| 180 | list_for_each_entry(p, &vm->ioeventfds, list) { |
| 181 | if (p->type == type && p->addr == addr && p->length >= len && |
| 182 | (p->wildcard || p->data == data)) |
| 183 | return p; |
| 184 | } |
| 185 | |
| 186 | return NULL; |
| 187 | } |
| 188 | |
| 189 | static int acrn_ioeventfd_handler(struct acrn_ioreq_client *client, |
| 190 | struct acrn_io_request *req) |
| 191 | { |
| 192 | struct hsm_ioeventfd *p; |
| 193 | u64 addr, val; |
| 194 | int size; |
| 195 | |
| 196 | if (req->type == ACRN_IOREQ_TYPE_MMIO) { |
| 197 | /* |
| 198 | * I/O requests are dispatched by range check only, so a |
| 199 | * acrn_ioreq_client need process both READ and WRITE accesses |
| 200 | * of same range. READ accesses are safe to be ignored here |
| 201 | * because virtio PCI devices write the notify registers for |
| 202 | * notification. |
| 203 | */ |
| 204 | if (req->reqs.mmio_request.direction == ACRN_IOREQ_DIR_READ) { |
| 205 | /* reading does nothing and return 0 */ |
| 206 | req->reqs.mmio_request.value = 0; |
| 207 | return 0; |
| 208 | } |
| 209 | addr = req->reqs.mmio_request.address; |
| 210 | size = req->reqs.mmio_request.size; |
| 211 | val = req->reqs.mmio_request.value; |
| 212 | } else { |
| 213 | if (req->reqs.pio_request.direction == ACRN_IOREQ_DIR_READ) { |
| 214 | /* reading does nothing and return 0 */ |
| 215 | req->reqs.pio_request.value = 0; |
| 216 | return 0; |
| 217 | } |
| 218 | addr = req->reqs.pio_request.address; |
| 219 | size = req->reqs.pio_request.size; |
| 220 | val = req->reqs.pio_request.value; |
| 221 | } |
| 222 | |
| 223 | mutex_lock(&client->vm->ioeventfds_lock); |
| 224 | p = hsm_ioeventfd_match(vm: client->vm, addr, data: val, len: size, type: req->type); |
| 225 | if (p) |
| 226 | eventfd_signal(ctx: p->eventfd); |
| 227 | mutex_unlock(lock: &client->vm->ioeventfds_lock); |
| 228 | |
| 229 | return 0; |
| 230 | } |
| 231 | |
| 232 | int acrn_ioeventfd_config(struct acrn_vm *vm, struct acrn_ioeventfd *args) |
| 233 | { |
| 234 | int ret; |
| 235 | |
| 236 | if (args->flags & ACRN_IOEVENTFD_FLAG_DEASSIGN) |
| 237 | ret = acrn_ioeventfd_deassign(vm, args); |
| 238 | else |
| 239 | ret = acrn_ioeventfd_assign(vm, args); |
| 240 | |
| 241 | return ret; |
| 242 | } |
| 243 | |
| 244 | int acrn_ioeventfd_init(struct acrn_vm *vm) |
| 245 | { |
| 246 | char name[ACRN_NAME_LEN]; |
| 247 | |
| 248 | mutex_init(&vm->ioeventfds_lock); |
| 249 | INIT_LIST_HEAD(list: &vm->ioeventfds); |
| 250 | snprintf(buf: name, size: sizeof(name), fmt: "ioeventfd-%u" , vm->vmid); |
| 251 | vm->ioeventfd_client = acrn_ioreq_client_create(vm, |
| 252 | handler: acrn_ioeventfd_handler, |
| 253 | NULL, is_default: false, name); |
| 254 | if (!vm->ioeventfd_client) { |
| 255 | dev_err(acrn_dev.this_device, "Failed to create ioeventfd ioreq client!\n" ); |
| 256 | return -EINVAL; |
| 257 | } |
| 258 | |
| 259 | dev_dbg(acrn_dev.this_device, "VM %u ioeventfd init.\n" , vm->vmid); |
| 260 | return 0; |
| 261 | } |
| 262 | |
| 263 | void acrn_ioeventfd_deinit(struct acrn_vm *vm) |
| 264 | { |
| 265 | struct hsm_ioeventfd *p, *next; |
| 266 | |
| 267 | dev_dbg(acrn_dev.this_device, "VM %u ioeventfd deinit.\n" , vm->vmid); |
| 268 | acrn_ioreq_client_destroy(client: vm->ioeventfd_client); |
| 269 | mutex_lock(&vm->ioeventfds_lock); |
| 270 | list_for_each_entry_safe(p, next, &vm->ioeventfds, list) |
| 271 | acrn_ioeventfd_shutdown(vm, p); |
| 272 | mutex_unlock(lock: &vm->ioeventfds_lock); |
| 273 | } |
| 274 | |