1 /* 2 * Intel I/OAT DMA Linux driver 3 * Copyright(c) 2004 - 2015 Intel Corporation. 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms and conditions of the GNU General Public License, 7 * version 2, as published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 12 * more details. 13 * 14 * The full GNU General Public License is included in this distribution in 15 * the file called "COPYING". 16 * 17 */ 18 19 #include <linux/init.h> 20 #include <linux/module.h> 21 #include <linux/slab.h> 22 #include <linux/pci.h> 23 #include <linux/interrupt.h> 24 #include <linux/dmaengine.h> 25 #include <linux/delay.h> 26 #include <linux/dma-mapping.h> 27 #include <linux/workqueue.h> 28 #include <linux/prefetch.h> 29 #include <linux/dca.h> 30 #include <linux/aer.h> 31 #include <linux/sizes.h> 32 #include "dma.h" 33 #include "registers.h" 34 #include "hw.h" 35 36 #include "../dmaengine.h" 37 38 MODULE_VERSION(IOAT_DMA_VERSION); 39 MODULE_LICENSE("Dual BSD/GPL"); 40 MODULE_AUTHOR("Intel Corporation"); 41 42 static const struct pci_device_id ioat_pci_tbl[] = { 43 /* I/OAT v3 platforms */ 44 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_TBG0) }, 45 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_TBG1) }, 46 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_TBG2) }, 47 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_TBG3) }, 48 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_TBG4) }, 49 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_TBG5) }, 50 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_TBG6) }, 51 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_TBG7) }, 52 53 /* I/OAT v3.2 platforms */ 54 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF0) }, 55 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF1) }, 56 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF2) }, 57 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF3) }, 58 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF4) }, 59 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF5) }, 60 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF6) }, 61 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF7) }, 62 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF8) }, 63 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_JSF9) }, 64 65 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB0) }, 66 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB1) }, 67 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB2) }, 68 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB3) }, 69 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB4) }, 70 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB5) }, 71 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB6) }, 72 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB7) }, 73 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB8) }, 74 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB9) }, 75 76 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB0) }, 77 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB1) }, 78 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB2) }, 79 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB3) }, 80 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB4) }, 81 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB5) }, 82 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB6) }, 83 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB7) }, 84 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB8) }, 85 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_IVB9) }, 86 87 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW0) }, 88 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW1) }, 89 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW2) }, 90 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW3) }, 91 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW4) }, 92 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW5) }, 93 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW6) }, 94 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW7) }, 95 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW8) }, 96 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_HSW9) }, 97 98 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX0) }, 99 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX1) }, 100 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX2) }, 101 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX3) }, 102 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX4) }, 103 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX5) }, 104 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX6) }, 105 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX7) }, 106 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX8) }, 107 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDX9) }, 108 109 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_SKX) }, 110 111 /* I/OAT v3.3 platforms */ 112 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BWD0) }, 113 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BWD1) }, 114 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BWD2) }, 115 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BWD3) }, 116 117 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDXDE0) }, 118 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDXDE1) }, 119 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDXDE2) }, 120 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_IOAT_BDXDE3) }, 121 122 { 0, } 123 }; 124 MODULE_DEVICE_TABLE(pci, ioat_pci_tbl); 125 126 static int ioat_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id); 127 static void ioat_remove(struct pci_dev *pdev); 128 static void 129 ioat_init_channel(struct ioatdma_device *ioat_dma, 130 struct ioatdma_chan *ioat_chan, int idx); 131 static void ioat_intr_quirk(struct ioatdma_device *ioat_dma); 132 static int ioat_enumerate_channels(struct ioatdma_device *ioat_dma); 133 static int ioat3_dma_self_test(struct ioatdma_device *ioat_dma); 134 135 static int ioat_dca_enabled = 1; 136 module_param(ioat_dca_enabled, int, 0644); 137 MODULE_PARM_DESC(ioat_dca_enabled, "control support of dca service (default: 1)"); 138 int ioat_pending_level = 4; 139 module_param(ioat_pending_level, int, 0644); 140 MODULE_PARM_DESC(ioat_pending_level, 141 "high-water mark for pushing ioat descriptors (default: 4)"); 142 static char ioat_interrupt_style[32] = "msix"; 143 module_param_string(ioat_interrupt_style, ioat_interrupt_style, 144 sizeof(ioat_interrupt_style), 0644); 145 MODULE_PARM_DESC(ioat_interrupt_style, 146 "set ioat interrupt style: msix (default), msi, intx"); 147 148 struct kmem_cache *ioat_cache; 149 struct kmem_cache *ioat_sed_cache; 150 151 static bool is_jf_ioat(struct pci_dev *pdev) 152 { 153 switch (pdev->device) { 154 case PCI_DEVICE_ID_INTEL_IOAT_JSF0: 155 case PCI_DEVICE_ID_INTEL_IOAT_JSF1: 156 case PCI_DEVICE_ID_INTEL_IOAT_JSF2: 157 case PCI_DEVICE_ID_INTEL_IOAT_JSF3: 158 case PCI_DEVICE_ID_INTEL_IOAT_JSF4: 159 case PCI_DEVICE_ID_INTEL_IOAT_JSF5: 160 case PCI_DEVICE_ID_INTEL_IOAT_JSF6: 161 case PCI_DEVICE_ID_INTEL_IOAT_JSF7: 162 case PCI_DEVICE_ID_INTEL_IOAT_JSF8: 163 case PCI_DEVICE_ID_INTEL_IOAT_JSF9: 164 return true; 165 default: 166 return false; 167 } 168 } 169 170 static bool is_snb_ioat(struct pci_dev *pdev) 171 { 172 switch (pdev->device) { 173 case PCI_DEVICE_ID_INTEL_IOAT_SNB0: 174 case PCI_DEVICE_ID_INTEL_IOAT_SNB1: 175 case PCI_DEVICE_ID_INTEL_IOAT_SNB2: 176 case PCI_DEVICE_ID_INTEL_IOAT_SNB3: 177 case PCI_DEVICE_ID_INTEL_IOAT_SNB4: 178 case PCI_DEVICE_ID_INTEL_IOAT_SNB5: 179 case PCI_DEVICE_ID_INTEL_IOAT_SNB6: 180 case PCI_DEVICE_ID_INTEL_IOAT_SNB7: 181 case PCI_DEVICE_ID_INTEL_IOAT_SNB8: 182 case PCI_DEVICE_ID_INTEL_IOAT_SNB9: 183 return true; 184 default: 185 return false; 186 } 187 } 188 189 static bool is_ivb_ioat(struct pci_dev *pdev) 190 { 191 switch (pdev->device) { 192 case PCI_DEVICE_ID_INTEL_IOAT_IVB0: 193 case PCI_DEVICE_ID_INTEL_IOAT_IVB1: 194 case PCI_DEVICE_ID_INTEL_IOAT_IVB2: 195 case PCI_DEVICE_ID_INTEL_IOAT_IVB3: 196 case PCI_DEVICE_ID_INTEL_IOAT_IVB4: 197 case PCI_DEVICE_ID_INTEL_IOAT_IVB5: 198 case PCI_DEVICE_ID_INTEL_IOAT_IVB6: 199 case PCI_DEVICE_ID_INTEL_IOAT_IVB7: 200 case PCI_DEVICE_ID_INTEL_IOAT_IVB8: 201 case PCI_DEVICE_ID_INTEL_IOAT_IVB9: 202 return true; 203 default: 204 return false; 205 } 206 207 } 208 209 static bool is_hsw_ioat(struct pci_dev *pdev) 210 { 211 switch (pdev->device) { 212 case PCI_DEVICE_ID_INTEL_IOAT_HSW0: 213 case PCI_DEVICE_ID_INTEL_IOAT_HSW1: 214 case PCI_DEVICE_ID_INTEL_IOAT_HSW2: 215 case PCI_DEVICE_ID_INTEL_IOAT_HSW3: 216 case PCI_DEVICE_ID_INTEL_IOAT_HSW4: 217 case PCI_DEVICE_ID_INTEL_IOAT_HSW5: 218 case PCI_DEVICE_ID_INTEL_IOAT_HSW6: 219 case PCI_DEVICE_ID_INTEL_IOAT_HSW7: 220 case PCI_DEVICE_ID_INTEL_IOAT_HSW8: 221 case PCI_DEVICE_ID_INTEL_IOAT_HSW9: 222 return true; 223 default: 224 return false; 225 } 226 227 } 228 229 static bool is_bdx_ioat(struct pci_dev *pdev) 230 { 231 switch (pdev->device) { 232 case PCI_DEVICE_ID_INTEL_IOAT_BDX0: 233 case PCI_DEVICE_ID_INTEL_IOAT_BDX1: 234 case PCI_DEVICE_ID_INTEL_IOAT_BDX2: 235 case PCI_DEVICE_ID_INTEL_IOAT_BDX3: 236 case PCI_DEVICE_ID_INTEL_IOAT_BDX4: 237 case PCI_DEVICE_ID_INTEL_IOAT_BDX5: 238 case PCI_DEVICE_ID_INTEL_IOAT_BDX6: 239 case PCI_DEVICE_ID_INTEL_IOAT_BDX7: 240 case PCI_DEVICE_ID_INTEL_IOAT_BDX8: 241 case PCI_DEVICE_ID_INTEL_IOAT_BDX9: 242 return true; 243 default: 244 return false; 245 } 246 } 247 248 static inline bool is_skx_ioat(struct pci_dev *pdev) 249 { 250 return (pdev->device == PCI_DEVICE_ID_INTEL_IOAT_SKX) ? true : false; 251 } 252 253 static bool is_xeon_cb32(struct pci_dev *pdev) 254 { 255 return is_jf_ioat(pdev) || is_snb_ioat(pdev) || is_ivb_ioat(pdev) || 256 is_hsw_ioat(pdev) || is_bdx_ioat(pdev) || is_skx_ioat(pdev); 257 } 258 259 bool is_bwd_ioat(struct pci_dev *pdev) 260 { 261 switch (pdev->device) { 262 case PCI_DEVICE_ID_INTEL_IOAT_BWD0: 263 case PCI_DEVICE_ID_INTEL_IOAT_BWD1: 264 case PCI_DEVICE_ID_INTEL_IOAT_BWD2: 265 case PCI_DEVICE_ID_INTEL_IOAT_BWD3: 266 /* even though not Atom, BDX-DE has same DMA silicon */ 267 case PCI_DEVICE_ID_INTEL_IOAT_BDXDE0: 268 case PCI_DEVICE_ID_INTEL_IOAT_BDXDE1: 269 case PCI_DEVICE_ID_INTEL_IOAT_BDXDE2: 270 case PCI_DEVICE_ID_INTEL_IOAT_BDXDE3: 271 return true; 272 default: 273 return false; 274 } 275 } 276 277 static bool is_bwd_noraid(struct pci_dev *pdev) 278 { 279 switch (pdev->device) { 280 case PCI_DEVICE_ID_INTEL_IOAT_BWD2: 281 case PCI_DEVICE_ID_INTEL_IOAT_BWD3: 282 case PCI_DEVICE_ID_INTEL_IOAT_BDXDE0: 283 case PCI_DEVICE_ID_INTEL_IOAT_BDXDE1: 284 case PCI_DEVICE_ID_INTEL_IOAT_BDXDE2: 285 case PCI_DEVICE_ID_INTEL_IOAT_BDXDE3: 286 return true; 287 default: 288 return false; 289 } 290 291 } 292 293 /* 294 * Perform a IOAT transaction to verify the HW works. 295 */ 296 #define IOAT_TEST_SIZE 2000 297 298 static void ioat_dma_test_callback(void *dma_async_param) 299 { 300 struct completion *cmp = dma_async_param; 301 302 complete(cmp); 303 } 304 305 /** 306 * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works. 307 * @ioat_dma: dma device to be tested 308 */ 309 static int ioat_dma_self_test(struct ioatdma_device *ioat_dma) 310 { 311 int i; 312 u8 *src; 313 u8 *dest; 314 struct dma_device *dma = &ioat_dma->dma_dev; 315 struct device *dev = &ioat_dma->pdev->dev; 316 struct dma_chan *dma_chan; 317 struct dma_async_tx_descriptor *tx; 318 dma_addr_t dma_dest, dma_src; 319 dma_cookie_t cookie; 320 int err = 0; 321 struct completion cmp; 322 unsigned long tmo; 323 unsigned long flags; 324 325 src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL); 326 if (!src) 327 return -ENOMEM; 328 dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL); 329 if (!dest) { 330 kfree(src); 331 return -ENOMEM; 332 } 333 334 /* Fill in src buffer */ 335 for (i = 0; i < IOAT_TEST_SIZE; i++) 336 src[i] = (u8)i; 337 338 /* Start copy, using first DMA channel */ 339 dma_chan = container_of(dma->channels.next, struct dma_chan, 340 device_node); 341 if (dma->device_alloc_chan_resources(dma_chan) < 1) { 342 dev_err(dev, "selftest cannot allocate chan resource\n"); 343 err = -ENODEV; 344 goto out; 345 } 346 347 dma_src = dma_map_single(dev, src, IOAT_TEST_SIZE, DMA_TO_DEVICE); 348 if (dma_mapping_error(dev, dma_src)) { 349 dev_err(dev, "mapping src buffer failed\n"); 350 err = -ENOMEM; 351 goto free_resources; 352 } 353 dma_dest = dma_map_single(dev, dest, IOAT_TEST_SIZE, DMA_FROM_DEVICE); 354 if (dma_mapping_error(dev, dma_dest)) { 355 dev_err(dev, "mapping dest buffer failed\n"); 356 err = -ENOMEM; 357 goto unmap_src; 358 } 359 flags = DMA_PREP_INTERRUPT; 360 tx = ioat_dma->dma_dev.device_prep_dma_memcpy(dma_chan, dma_dest, 361 dma_src, IOAT_TEST_SIZE, 362 flags); 363 if (!tx) { 364 dev_err(dev, "Self-test prep failed, disabling\n"); 365 err = -ENODEV; 366 goto unmap_dma; 367 } 368 369 async_tx_ack(tx); 370 init_completion(&cmp); 371 tx->callback = ioat_dma_test_callback; 372 tx->callback_param = &cmp; 373 cookie = tx->tx_submit(tx); 374 if (cookie < 0) { 375 dev_err(dev, "Self-test setup failed, disabling\n"); 376 err = -ENODEV; 377 goto unmap_dma; 378 } 379 dma->device_issue_pending(dma_chan); 380 381 tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000)); 382 383 if (tmo == 0 || 384 dma->device_tx_status(dma_chan, cookie, NULL) 385 != DMA_COMPLETE) { 386 dev_err(dev, "Self-test copy timed out, disabling\n"); 387 err = -ENODEV; 388 goto unmap_dma; 389 } 390 if (memcmp(src, dest, IOAT_TEST_SIZE)) { 391 dev_err(dev, "Self-test copy failed compare, disabling\n"); 392 err = -ENODEV; 393 goto free_resources; 394 } 395 396 unmap_dma: 397 dma_unmap_single(dev, dma_dest, IOAT_TEST_SIZE, DMA_FROM_DEVICE); 398 unmap_src: 399 dma_unmap_single(dev, dma_src, IOAT_TEST_SIZE, DMA_TO_DEVICE); 400 free_resources: 401 dma->device_free_chan_resources(dma_chan); 402 out: 403 kfree(src); 404 kfree(dest); 405 return err; 406 } 407 408 /** 409 * ioat_dma_setup_interrupts - setup interrupt handler 410 * @ioat_dma: ioat dma device 411 */ 412 int ioat_dma_setup_interrupts(struct ioatdma_device *ioat_dma) 413 { 414 struct ioatdma_chan *ioat_chan; 415 struct pci_dev *pdev = ioat_dma->pdev; 416 struct device *dev = &pdev->dev; 417 struct msix_entry *msix; 418 int i, j, msixcnt; 419 int err = -EINVAL; 420 u8 intrctrl = 0; 421 422 if (!strcmp(ioat_interrupt_style, "msix")) 423 goto msix; 424 if (!strcmp(ioat_interrupt_style, "msi")) 425 goto msi; 426 if (!strcmp(ioat_interrupt_style, "intx")) 427 goto intx; 428 dev_err(dev, "invalid ioat_interrupt_style %s\n", ioat_interrupt_style); 429 goto err_no_irq; 430 431 msix: 432 /* The number of MSI-X vectors should equal the number of channels */ 433 msixcnt = ioat_dma->dma_dev.chancnt; 434 for (i = 0; i < msixcnt; i++) 435 ioat_dma->msix_entries[i].entry = i; 436 437 err = pci_enable_msix_exact(pdev, ioat_dma->msix_entries, msixcnt); 438 if (err) 439 goto msi; 440 441 for (i = 0; i < msixcnt; i++) { 442 msix = &ioat_dma->msix_entries[i]; 443 ioat_chan = ioat_chan_by_index(ioat_dma, i); 444 err = devm_request_irq(dev, msix->vector, 445 ioat_dma_do_interrupt_msix, 0, 446 "ioat-msix", ioat_chan); 447 if (err) { 448 for (j = 0; j < i; j++) { 449 msix = &ioat_dma->msix_entries[j]; 450 ioat_chan = ioat_chan_by_index(ioat_dma, j); 451 devm_free_irq(dev, msix->vector, ioat_chan); 452 } 453 goto msi; 454 } 455 } 456 intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL; 457 ioat_dma->irq_mode = IOAT_MSIX; 458 goto done; 459 460 msi: 461 err = pci_enable_msi(pdev); 462 if (err) 463 goto intx; 464 465 err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt, 0, 466 "ioat-msi", ioat_dma); 467 if (err) { 468 pci_disable_msi(pdev); 469 goto intx; 470 } 471 ioat_dma->irq_mode = IOAT_MSI; 472 goto done; 473 474 intx: 475 err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt, 476 IRQF_SHARED, "ioat-intx", ioat_dma); 477 if (err) 478 goto err_no_irq; 479 480 ioat_dma->irq_mode = IOAT_INTX; 481 done: 482 if (is_bwd_ioat(pdev)) 483 ioat_intr_quirk(ioat_dma); 484 intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN; 485 writeb(intrctrl, ioat_dma->reg_base + IOAT_INTRCTRL_OFFSET); 486 return 0; 487 488 err_no_irq: 489 /* Disable all interrupt generation */ 490 writeb(0, ioat_dma->reg_base + IOAT_INTRCTRL_OFFSET); 491 ioat_dma->irq_mode = IOAT_NOIRQ; 492 dev_err(dev, "no usable interrupts\n"); 493 return err; 494 } 495 496 static void ioat_disable_interrupts(struct ioatdma_device *ioat_dma) 497 { 498 /* Disable all interrupt generation */ 499 writeb(0, ioat_dma->reg_base + IOAT_INTRCTRL_OFFSET); 500 } 501 502 static int ioat_probe(struct ioatdma_device *ioat_dma) 503 { 504 int err = -ENODEV; 505 struct dma_device *dma = &ioat_dma->dma_dev; 506 struct pci_dev *pdev = ioat_dma->pdev; 507 struct device *dev = &pdev->dev; 508 509 ioat_dma->completion_pool = dma_pool_create("completion_pool", dev, 510 sizeof(u64), 511 SMP_CACHE_BYTES, 512 SMP_CACHE_BYTES); 513 514 if (!ioat_dma->completion_pool) { 515 err = -ENOMEM; 516 goto err_out; 517 } 518 519 ioat_enumerate_channels(ioat_dma); 520 521 dma_cap_set(DMA_MEMCPY, dma->cap_mask); 522 dma->dev = &pdev->dev; 523 524 if (!dma->chancnt) { 525 dev_err(dev, "channel enumeration error\n"); 526 goto err_setup_interrupts; 527 } 528 529 err = ioat_dma_setup_interrupts(ioat_dma); 530 if (err) 531 goto err_setup_interrupts; 532 533 err = ioat3_dma_self_test(ioat_dma); 534 if (err) 535 goto err_self_test; 536 537 return 0; 538 539 err_self_test: 540 ioat_disable_interrupts(ioat_dma); 541 err_setup_interrupts: 542 dma_pool_destroy(ioat_dma->completion_pool); 543 err_out: 544 return err; 545 } 546 547 static int ioat_register(struct ioatdma_device *ioat_dma) 548 { 549 int err = dma_async_device_register(&ioat_dma->dma_dev); 550 551 if (err) { 552 ioat_disable_interrupts(ioat_dma); 553 dma_pool_destroy(ioat_dma->completion_pool); 554 } 555 556 return err; 557 } 558 559 static void ioat_dma_remove(struct ioatdma_device *ioat_dma) 560 { 561 struct dma_device *dma = &ioat_dma->dma_dev; 562 563 ioat_disable_interrupts(ioat_dma); 564 565 ioat_kobject_del(ioat_dma); 566 567 dma_async_device_unregister(dma); 568 569 dma_pool_destroy(ioat_dma->completion_pool); 570 571 INIT_LIST_HEAD(&dma->channels); 572 } 573 574 /** 575 * ioat_enumerate_channels - find and initialize the device's channels 576 * @ioat_dma: the ioat dma device to be enumerated 577 */ 578 static int ioat_enumerate_channels(struct ioatdma_device *ioat_dma) 579 { 580 struct ioatdma_chan *ioat_chan; 581 struct device *dev = &ioat_dma->pdev->dev; 582 struct dma_device *dma = &ioat_dma->dma_dev; 583 u8 xfercap_log; 584 int i; 585 586 INIT_LIST_HEAD(&dma->channels); 587 dma->chancnt = readb(ioat_dma->reg_base + IOAT_CHANCNT_OFFSET); 588 dma->chancnt &= 0x1f; /* bits [4:0] valid */ 589 if (dma->chancnt > ARRAY_SIZE(ioat_dma->idx)) { 590 dev_warn(dev, "(%d) exceeds max supported channels (%zu)\n", 591 dma->chancnt, ARRAY_SIZE(ioat_dma->idx)); 592 dma->chancnt = ARRAY_SIZE(ioat_dma->idx); 593 } 594 xfercap_log = readb(ioat_dma->reg_base + IOAT_XFERCAP_OFFSET); 595 xfercap_log &= 0x1f; /* bits [4:0] valid */ 596 if (xfercap_log == 0) 597 return 0; 598 dev_dbg(dev, "%s: xfercap = %d\n", __func__, 1 << xfercap_log); 599 600 for (i = 0; i < dma->chancnt; i++) { 601 ioat_chan = devm_kzalloc(dev, sizeof(*ioat_chan), GFP_KERNEL); 602 if (!ioat_chan) 603 break; 604 605 ioat_init_channel(ioat_dma, ioat_chan, i); 606 ioat_chan->xfercap_log = xfercap_log; 607 spin_lock_init(&ioat_chan->prep_lock); 608 if (ioat_reset_hw(ioat_chan)) { 609 i = 0; 610 break; 611 } 612 } 613 dma->chancnt = i; 614 return i; 615 } 616 617 /** 618 * ioat_free_chan_resources - release all the descriptors 619 * @chan: the channel to be cleaned 620 */ 621 static void ioat_free_chan_resources(struct dma_chan *c) 622 { 623 struct ioatdma_chan *ioat_chan = to_ioat_chan(c); 624 struct ioatdma_device *ioat_dma = ioat_chan->ioat_dma; 625 struct ioat_ring_ent *desc; 626 const int total_descs = 1 << ioat_chan->alloc_order; 627 int descs; 628 int i; 629 630 /* Before freeing channel resources first check 631 * if they have been previously allocated for this channel. 632 */ 633 if (!ioat_chan->ring) 634 return; 635 636 ioat_stop(ioat_chan); 637 ioat_reset_hw(ioat_chan); 638 639 spin_lock_bh(&ioat_chan->cleanup_lock); 640 spin_lock_bh(&ioat_chan->prep_lock); 641 descs = ioat_ring_space(ioat_chan); 642 dev_dbg(to_dev(ioat_chan), "freeing %d idle descriptors\n", descs); 643 for (i = 0; i < descs; i++) { 644 desc = ioat_get_ring_ent(ioat_chan, ioat_chan->head + i); 645 ioat_free_ring_ent(desc, c); 646 } 647 648 if (descs < total_descs) 649 dev_err(to_dev(ioat_chan), "Freeing %d in use descriptors!\n", 650 total_descs - descs); 651 652 for (i = 0; i < total_descs - descs; i++) { 653 desc = ioat_get_ring_ent(ioat_chan, ioat_chan->tail + i); 654 dump_desc_dbg(ioat_chan, desc); 655 ioat_free_ring_ent(desc, c); 656 } 657 658 for (i = 0; i < ioat_chan->desc_chunks; i++) { 659 dma_free_coherent(to_dev(ioat_chan), SZ_2M, 660 ioat_chan->descs[i].virt, 661 ioat_chan->descs[i].hw); 662 ioat_chan->descs[i].virt = NULL; 663 ioat_chan->descs[i].hw = 0; 664 } 665 ioat_chan->desc_chunks = 0; 666 667 kfree(ioat_chan->ring); 668 ioat_chan->ring = NULL; 669 ioat_chan->alloc_order = 0; 670 dma_pool_free(ioat_dma->completion_pool, ioat_chan->completion, 671 ioat_chan->completion_dma); 672 spin_unlock_bh(&ioat_chan->prep_lock); 673 spin_unlock_bh(&ioat_chan->cleanup_lock); 674 675 ioat_chan->last_completion = 0; 676 ioat_chan->completion_dma = 0; 677 ioat_chan->dmacount = 0; 678 } 679 680 /* ioat_alloc_chan_resources - allocate/initialize ioat descriptor ring 681 * @chan: channel to be initialized 682 */ 683 static int ioat_alloc_chan_resources(struct dma_chan *c) 684 { 685 struct ioatdma_chan *ioat_chan = to_ioat_chan(c); 686 struct ioat_ring_ent **ring; 687 u64 status; 688 int order; 689 int i = 0; 690 u32 chanerr; 691 692 /* have we already been set up? */ 693 if (ioat_chan->ring) 694 return 1 << ioat_chan->alloc_order; 695 696 /* Setup register to interrupt and write completion status on error */ 697 writew(IOAT_CHANCTRL_RUN, ioat_chan->reg_base + IOAT_CHANCTRL_OFFSET); 698 699 /* allocate a completion writeback area */ 700 /* doing 2 32bit writes to mmio since 1 64b write doesn't work */ 701 ioat_chan->completion = 702 dma_pool_zalloc(ioat_chan->ioat_dma->completion_pool, 703 GFP_NOWAIT, &ioat_chan->completion_dma); 704 if (!ioat_chan->completion) 705 return -ENOMEM; 706 707 writel(((u64)ioat_chan->completion_dma) & 0x00000000FFFFFFFF, 708 ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_LOW); 709 writel(((u64)ioat_chan->completion_dma) >> 32, 710 ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH); 711 712 order = IOAT_MAX_ORDER; 713 ring = ioat_alloc_ring(c, order, GFP_NOWAIT); 714 if (!ring) 715 return -ENOMEM; 716 717 spin_lock_bh(&ioat_chan->cleanup_lock); 718 spin_lock_bh(&ioat_chan->prep_lock); 719 ioat_chan->ring = ring; 720 ioat_chan->head = 0; 721 ioat_chan->issued = 0; 722 ioat_chan->tail = 0; 723 ioat_chan->alloc_order = order; 724 set_bit(IOAT_RUN, &ioat_chan->state); 725 spin_unlock_bh(&ioat_chan->prep_lock); 726 spin_unlock_bh(&ioat_chan->cleanup_lock); 727 728 ioat_start_null_desc(ioat_chan); 729 730 /* check that we got off the ground */ 731 do { 732 udelay(1); 733 status = ioat_chansts(ioat_chan); 734 } while (i++ < 20 && !is_ioat_active(status) && !is_ioat_idle(status)); 735 736 if (is_ioat_active(status) || is_ioat_idle(status)) 737 return 1 << ioat_chan->alloc_order; 738 739 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET); 740 741 dev_WARN(to_dev(ioat_chan), 742 "failed to start channel chanerr: %#x\n", chanerr); 743 ioat_free_chan_resources(c); 744 return -EFAULT; 745 } 746 747 /* common channel initialization */ 748 static void 749 ioat_init_channel(struct ioatdma_device *ioat_dma, 750 struct ioatdma_chan *ioat_chan, int idx) 751 { 752 struct dma_device *dma = &ioat_dma->dma_dev; 753 struct dma_chan *c = &ioat_chan->dma_chan; 754 unsigned long data = (unsigned long) c; 755 756 ioat_chan->ioat_dma = ioat_dma; 757 ioat_chan->reg_base = ioat_dma->reg_base + (0x80 * (idx + 1)); 758 spin_lock_init(&ioat_chan->cleanup_lock); 759 ioat_chan->dma_chan.device = dma; 760 dma_cookie_init(&ioat_chan->dma_chan); 761 list_add_tail(&ioat_chan->dma_chan.device_node, &dma->channels); 762 ioat_dma->idx[idx] = ioat_chan; 763 timer_setup(&ioat_chan->timer, ioat_timer_event, 0); 764 tasklet_init(&ioat_chan->cleanup_task, ioat_cleanup_event, data); 765 } 766 767 #define IOAT_NUM_SRC_TEST 6 /* must be <= 8 */ 768 static int ioat_xor_val_self_test(struct ioatdma_device *ioat_dma) 769 { 770 int i, src_idx; 771 struct page *dest; 772 struct page *xor_srcs[IOAT_NUM_SRC_TEST]; 773 struct page *xor_val_srcs[IOAT_NUM_SRC_TEST + 1]; 774 dma_addr_t dma_srcs[IOAT_NUM_SRC_TEST + 1]; 775 dma_addr_t dest_dma; 776 struct dma_async_tx_descriptor *tx; 777 struct dma_chan *dma_chan; 778 dma_cookie_t cookie; 779 u8 cmp_byte = 0; 780 u32 cmp_word; 781 u32 xor_val_result; 782 int err = 0; 783 struct completion cmp; 784 unsigned long tmo; 785 struct device *dev = &ioat_dma->pdev->dev; 786 struct dma_device *dma = &ioat_dma->dma_dev; 787 u8 op = 0; 788 789 dev_dbg(dev, "%s\n", __func__); 790 791 if (!dma_has_cap(DMA_XOR, dma->cap_mask)) 792 return 0; 793 794 for (src_idx = 0; src_idx < IOAT_NUM_SRC_TEST; src_idx++) { 795 xor_srcs[src_idx] = alloc_page(GFP_KERNEL); 796 if (!xor_srcs[src_idx]) { 797 while (src_idx--) 798 __free_page(xor_srcs[src_idx]); 799 return -ENOMEM; 800 } 801 } 802 803 dest = alloc_page(GFP_KERNEL); 804 if (!dest) { 805 while (src_idx--) 806 __free_page(xor_srcs[src_idx]); 807 return -ENOMEM; 808 } 809 810 /* Fill in src buffers */ 811 for (src_idx = 0; src_idx < IOAT_NUM_SRC_TEST; src_idx++) { 812 u8 *ptr = page_address(xor_srcs[src_idx]); 813 814 for (i = 0; i < PAGE_SIZE; i++) 815 ptr[i] = (1 << src_idx); 816 } 817 818 for (src_idx = 0; src_idx < IOAT_NUM_SRC_TEST; src_idx++) 819 cmp_byte ^= (u8) (1 << src_idx); 820 821 cmp_word = (cmp_byte << 24) | (cmp_byte << 16) | 822 (cmp_byte << 8) | cmp_byte; 823 824 memset(page_address(dest), 0, PAGE_SIZE); 825 826 dma_chan = container_of(dma->channels.next, struct dma_chan, 827 device_node); 828 if (dma->device_alloc_chan_resources(dma_chan) < 1) { 829 err = -ENODEV; 830 goto out; 831 } 832 833 /* test xor */ 834 op = IOAT_OP_XOR; 835 836 dest_dma = dma_map_page(dev, dest, 0, PAGE_SIZE, DMA_FROM_DEVICE); 837 if (dma_mapping_error(dev, dest_dma)) { 838 err = -ENOMEM; 839 goto free_resources; 840 } 841 842 for (i = 0; i < IOAT_NUM_SRC_TEST; i++) { 843 dma_srcs[i] = dma_map_page(dev, xor_srcs[i], 0, PAGE_SIZE, 844 DMA_TO_DEVICE); 845 if (dma_mapping_error(dev, dma_srcs[i])) { 846 err = -ENOMEM; 847 goto dma_unmap; 848 } 849 } 850 tx = dma->device_prep_dma_xor(dma_chan, dest_dma, dma_srcs, 851 IOAT_NUM_SRC_TEST, PAGE_SIZE, 852 DMA_PREP_INTERRUPT); 853 854 if (!tx) { 855 dev_err(dev, "Self-test xor prep failed\n"); 856 err = -ENODEV; 857 goto dma_unmap; 858 } 859 860 async_tx_ack(tx); 861 init_completion(&cmp); 862 tx->callback = ioat_dma_test_callback; 863 tx->callback_param = &cmp; 864 cookie = tx->tx_submit(tx); 865 if (cookie < 0) { 866 dev_err(dev, "Self-test xor setup failed\n"); 867 err = -ENODEV; 868 goto dma_unmap; 869 } 870 dma->device_issue_pending(dma_chan); 871 872 tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000)); 873 874 if (tmo == 0 || 875 dma->device_tx_status(dma_chan, cookie, NULL) != DMA_COMPLETE) { 876 dev_err(dev, "Self-test xor timed out\n"); 877 err = -ENODEV; 878 goto dma_unmap; 879 } 880 881 for (i = 0; i < IOAT_NUM_SRC_TEST; i++) 882 dma_unmap_page(dev, dma_srcs[i], PAGE_SIZE, DMA_TO_DEVICE); 883 884 dma_sync_single_for_cpu(dev, dest_dma, PAGE_SIZE, DMA_FROM_DEVICE); 885 for (i = 0; i < (PAGE_SIZE / sizeof(u32)); i++) { 886 u32 *ptr = page_address(dest); 887 888 if (ptr[i] != cmp_word) { 889 dev_err(dev, "Self-test xor failed compare\n"); 890 err = -ENODEV; 891 goto free_resources; 892 } 893 } 894 dma_sync_single_for_device(dev, dest_dma, PAGE_SIZE, DMA_FROM_DEVICE); 895 896 dma_unmap_page(dev, dest_dma, PAGE_SIZE, DMA_FROM_DEVICE); 897 898 /* skip validate if the capability is not present */ 899 if (!dma_has_cap(DMA_XOR_VAL, dma_chan->device->cap_mask)) 900 goto free_resources; 901 902 op = IOAT_OP_XOR_VAL; 903 904 /* validate the sources with the destintation page */ 905 for (i = 0; i < IOAT_NUM_SRC_TEST; i++) 906 xor_val_srcs[i] = xor_srcs[i]; 907 xor_val_srcs[i] = dest; 908 909 xor_val_result = 1; 910 911 for (i = 0; i < IOAT_NUM_SRC_TEST + 1; i++) { 912 dma_srcs[i] = dma_map_page(dev, xor_val_srcs[i], 0, PAGE_SIZE, 913 DMA_TO_DEVICE); 914 if (dma_mapping_error(dev, dma_srcs[i])) { 915 err = -ENOMEM; 916 goto dma_unmap; 917 } 918 } 919 tx = dma->device_prep_dma_xor_val(dma_chan, dma_srcs, 920 IOAT_NUM_SRC_TEST + 1, PAGE_SIZE, 921 &xor_val_result, DMA_PREP_INTERRUPT); 922 if (!tx) { 923 dev_err(dev, "Self-test zero prep failed\n"); 924 err = -ENODEV; 925 goto dma_unmap; 926 } 927 928 async_tx_ack(tx); 929 init_completion(&cmp); 930 tx->callback = ioat_dma_test_callback; 931 tx->callback_param = &cmp; 932 cookie = tx->tx_submit(tx); 933 if (cookie < 0) { 934 dev_err(dev, "Self-test zero setup failed\n"); 935 err = -ENODEV; 936 goto dma_unmap; 937 } 938 dma->device_issue_pending(dma_chan); 939 940 tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000)); 941 942 if (tmo == 0 || 943 dma->device_tx_status(dma_chan, cookie, NULL) != DMA_COMPLETE) { 944 dev_err(dev, "Self-test validate timed out\n"); 945 err = -ENODEV; 946 goto dma_unmap; 947 } 948 949 for (i = 0; i < IOAT_NUM_SRC_TEST + 1; i++) 950 dma_unmap_page(dev, dma_srcs[i], PAGE_SIZE, DMA_TO_DEVICE); 951 952 if (xor_val_result != 0) { 953 dev_err(dev, "Self-test validate failed compare\n"); 954 err = -ENODEV; 955 goto free_resources; 956 } 957 958 memset(page_address(dest), 0, PAGE_SIZE); 959 960 /* test for non-zero parity sum */ 961 op = IOAT_OP_XOR_VAL; 962 963 xor_val_result = 0; 964 for (i = 0; i < IOAT_NUM_SRC_TEST + 1; i++) { 965 dma_srcs[i] = dma_map_page(dev, xor_val_srcs[i], 0, PAGE_SIZE, 966 DMA_TO_DEVICE); 967 if (dma_mapping_error(dev, dma_srcs[i])) { 968 err = -ENOMEM; 969 goto dma_unmap; 970 } 971 } 972 tx = dma->device_prep_dma_xor_val(dma_chan, dma_srcs, 973 IOAT_NUM_SRC_TEST + 1, PAGE_SIZE, 974 &xor_val_result, DMA_PREP_INTERRUPT); 975 if (!tx) { 976 dev_err(dev, "Self-test 2nd zero prep failed\n"); 977 err = -ENODEV; 978 goto dma_unmap; 979 } 980 981 async_tx_ack(tx); 982 init_completion(&cmp); 983 tx->callback = ioat_dma_test_callback; 984 tx->callback_param = &cmp; 985 cookie = tx->tx_submit(tx); 986 if (cookie < 0) { 987 dev_err(dev, "Self-test 2nd zero setup failed\n"); 988 err = -ENODEV; 989 goto dma_unmap; 990 } 991 dma->device_issue_pending(dma_chan); 992 993 tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000)); 994 995 if (tmo == 0 || 996 dma->device_tx_status(dma_chan, cookie, NULL) != DMA_COMPLETE) { 997 dev_err(dev, "Self-test 2nd validate timed out\n"); 998 err = -ENODEV; 999 goto dma_unmap; 1000 } 1001 1002 if (xor_val_result != SUM_CHECK_P_RESULT) { 1003 dev_err(dev, "Self-test validate failed compare\n"); 1004 err = -ENODEV; 1005 goto dma_unmap; 1006 } 1007 1008 for (i = 0; i < IOAT_NUM_SRC_TEST + 1; i++) 1009 dma_unmap_page(dev, dma_srcs[i], PAGE_SIZE, DMA_TO_DEVICE); 1010 1011 goto free_resources; 1012 dma_unmap: 1013 if (op == IOAT_OP_XOR) { 1014 while (--i >= 0) 1015 dma_unmap_page(dev, dma_srcs[i], PAGE_SIZE, 1016 DMA_TO_DEVICE); 1017 dma_unmap_page(dev, dest_dma, PAGE_SIZE, DMA_FROM_DEVICE); 1018 } else if (op == IOAT_OP_XOR_VAL) { 1019 while (--i >= 0) 1020 dma_unmap_page(dev, dma_srcs[i], PAGE_SIZE, 1021 DMA_TO_DEVICE); 1022 } 1023 free_resources: 1024 dma->device_free_chan_resources(dma_chan); 1025 out: 1026 src_idx = IOAT_NUM_SRC_TEST; 1027 while (src_idx--) 1028 __free_page(xor_srcs[src_idx]); 1029 __free_page(dest); 1030 return err; 1031 } 1032 1033 static int ioat3_dma_self_test(struct ioatdma_device *ioat_dma) 1034 { 1035 int rc; 1036 1037 rc = ioat_dma_self_test(ioat_dma); 1038 if (rc) 1039 return rc; 1040 1041 rc = ioat_xor_val_self_test(ioat_dma); 1042 1043 return rc; 1044 } 1045 1046 static void ioat_intr_quirk(struct ioatdma_device *ioat_dma) 1047 { 1048 struct dma_device *dma; 1049 struct dma_chan *c; 1050 struct ioatdma_chan *ioat_chan; 1051 u32 errmask; 1052 1053 dma = &ioat_dma->dma_dev; 1054 1055 /* 1056 * if we have descriptor write back error status, we mask the 1057 * error interrupts 1058 */ 1059 if (ioat_dma->cap & IOAT_CAP_DWBES) { 1060 list_for_each_entry(c, &dma->channels, device_node) { 1061 ioat_chan = to_ioat_chan(c); 1062 errmask = readl(ioat_chan->reg_base + 1063 IOAT_CHANERR_MASK_OFFSET); 1064 errmask |= IOAT_CHANERR_XOR_P_OR_CRC_ERR | 1065 IOAT_CHANERR_XOR_Q_ERR; 1066 writel(errmask, ioat_chan->reg_base + 1067 IOAT_CHANERR_MASK_OFFSET); 1068 } 1069 } 1070 } 1071 1072 static int ioat3_dma_probe(struct ioatdma_device *ioat_dma, int dca) 1073 { 1074 struct pci_dev *pdev = ioat_dma->pdev; 1075 int dca_en = system_has_dca_enabled(pdev); 1076 struct dma_device *dma; 1077 struct dma_chan *c; 1078 struct ioatdma_chan *ioat_chan; 1079 int err; 1080 u16 val16; 1081 1082 dma = &ioat_dma->dma_dev; 1083 dma->device_prep_dma_memcpy = ioat_dma_prep_memcpy_lock; 1084 dma->device_issue_pending = ioat_issue_pending; 1085 dma->device_alloc_chan_resources = ioat_alloc_chan_resources; 1086 dma->device_free_chan_resources = ioat_free_chan_resources; 1087 1088 dma_cap_set(DMA_INTERRUPT, dma->cap_mask); 1089 dma->device_prep_dma_interrupt = ioat_prep_interrupt_lock; 1090 1091 ioat_dma->cap = readl(ioat_dma->reg_base + IOAT_DMA_CAP_OFFSET); 1092 1093 if (is_xeon_cb32(pdev) || is_bwd_noraid(pdev)) 1094 ioat_dma->cap &= 1095 ~(IOAT_CAP_XOR | IOAT_CAP_PQ | IOAT_CAP_RAID16SS); 1096 1097 /* dca is incompatible with raid operations */ 1098 if (dca_en && (ioat_dma->cap & (IOAT_CAP_XOR|IOAT_CAP_PQ))) 1099 ioat_dma->cap &= ~(IOAT_CAP_XOR|IOAT_CAP_PQ); 1100 1101 if (ioat_dma->cap & IOAT_CAP_XOR) { 1102 dma->max_xor = 8; 1103 1104 dma_cap_set(DMA_XOR, dma->cap_mask); 1105 dma->device_prep_dma_xor = ioat_prep_xor; 1106 1107 dma_cap_set(DMA_XOR_VAL, dma->cap_mask); 1108 dma->device_prep_dma_xor_val = ioat_prep_xor_val; 1109 } 1110 1111 if (ioat_dma->cap & IOAT_CAP_PQ) { 1112 1113 dma->device_prep_dma_pq = ioat_prep_pq; 1114 dma->device_prep_dma_pq_val = ioat_prep_pq_val; 1115 dma_cap_set(DMA_PQ, dma->cap_mask); 1116 dma_cap_set(DMA_PQ_VAL, dma->cap_mask); 1117 1118 if (ioat_dma->cap & IOAT_CAP_RAID16SS) 1119 dma_set_maxpq(dma, 16, 0); 1120 else 1121 dma_set_maxpq(dma, 8, 0); 1122 1123 if (!(ioat_dma->cap & IOAT_CAP_XOR)) { 1124 dma->device_prep_dma_xor = ioat_prep_pqxor; 1125 dma->device_prep_dma_xor_val = ioat_prep_pqxor_val; 1126 dma_cap_set(DMA_XOR, dma->cap_mask); 1127 dma_cap_set(DMA_XOR_VAL, dma->cap_mask); 1128 1129 if (ioat_dma->cap & IOAT_CAP_RAID16SS) 1130 dma->max_xor = 16; 1131 else 1132 dma->max_xor = 8; 1133 } 1134 } 1135 1136 dma->device_tx_status = ioat_tx_status; 1137 1138 /* starting with CB3.3 super extended descriptors are supported */ 1139 if (ioat_dma->cap & IOAT_CAP_RAID16SS) { 1140 char pool_name[14]; 1141 int i; 1142 1143 for (i = 0; i < MAX_SED_POOLS; i++) { 1144 snprintf(pool_name, 14, "ioat_hw%d_sed", i); 1145 1146 /* allocate SED DMA pool */ 1147 ioat_dma->sed_hw_pool[i] = dmam_pool_create(pool_name, 1148 &pdev->dev, 1149 SED_SIZE * (i + 1), 64, 0); 1150 if (!ioat_dma->sed_hw_pool[i]) 1151 return -ENOMEM; 1152 1153 } 1154 } 1155 1156 if (!(ioat_dma->cap & (IOAT_CAP_XOR | IOAT_CAP_PQ))) 1157 dma_cap_set(DMA_PRIVATE, dma->cap_mask); 1158 1159 err = ioat_probe(ioat_dma); 1160 if (err) 1161 return err; 1162 1163 list_for_each_entry(c, &dma->channels, device_node) { 1164 ioat_chan = to_ioat_chan(c); 1165 writel(IOAT_DMA_DCA_ANY_CPU, 1166 ioat_chan->reg_base + IOAT_DCACTRL_OFFSET); 1167 } 1168 1169 err = ioat_register(ioat_dma); 1170 if (err) 1171 return err; 1172 1173 ioat_kobject_add(ioat_dma, &ioat_ktype); 1174 1175 if (dca) 1176 ioat_dma->dca = ioat_dca_init(pdev, ioat_dma->reg_base); 1177 1178 /* disable relaxed ordering */ 1179 err = pcie_capability_read_word(pdev, IOAT_DEVCTRL_OFFSET, &val16); 1180 if (err) 1181 return err; 1182 1183 /* clear relaxed ordering enable */ 1184 val16 &= ~IOAT_DEVCTRL_ROE; 1185 err = pcie_capability_write_word(pdev, IOAT_DEVCTRL_OFFSET, val16); 1186 if (err) 1187 return err; 1188 1189 return 0; 1190 } 1191 1192 static void ioat_shutdown(struct pci_dev *pdev) 1193 { 1194 struct ioatdma_device *ioat_dma = pci_get_drvdata(pdev); 1195 struct ioatdma_chan *ioat_chan; 1196 int i; 1197 1198 if (!ioat_dma) 1199 return; 1200 1201 for (i = 0; i < IOAT_MAX_CHANS; i++) { 1202 ioat_chan = ioat_dma->idx[i]; 1203 if (!ioat_chan) 1204 continue; 1205 1206 spin_lock_bh(&ioat_chan->prep_lock); 1207 set_bit(IOAT_CHAN_DOWN, &ioat_chan->state); 1208 del_timer_sync(&ioat_chan->timer); 1209 spin_unlock_bh(&ioat_chan->prep_lock); 1210 /* this should quiesce then reset */ 1211 ioat_reset_hw(ioat_chan); 1212 } 1213 1214 ioat_disable_interrupts(ioat_dma); 1215 } 1216 1217 static void ioat_resume(struct ioatdma_device *ioat_dma) 1218 { 1219 struct ioatdma_chan *ioat_chan; 1220 u32 chanerr; 1221 int i; 1222 1223 for (i = 0; i < IOAT_MAX_CHANS; i++) { 1224 ioat_chan = ioat_dma->idx[i]; 1225 if (!ioat_chan) 1226 continue; 1227 1228 spin_lock_bh(&ioat_chan->prep_lock); 1229 clear_bit(IOAT_CHAN_DOWN, &ioat_chan->state); 1230 spin_unlock_bh(&ioat_chan->prep_lock); 1231 1232 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET); 1233 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET); 1234 1235 /* no need to reset as shutdown already did that */ 1236 } 1237 } 1238 1239 #define DRV_NAME "ioatdma" 1240 1241 static pci_ers_result_t ioat_pcie_error_detected(struct pci_dev *pdev, 1242 enum pci_channel_state error) 1243 { 1244 dev_dbg(&pdev->dev, "%s: PCIe AER error %d\n", DRV_NAME, error); 1245 1246 /* quiesce and block I/O */ 1247 ioat_shutdown(pdev); 1248 1249 return PCI_ERS_RESULT_NEED_RESET; 1250 } 1251 1252 static pci_ers_result_t ioat_pcie_error_slot_reset(struct pci_dev *pdev) 1253 { 1254 pci_ers_result_t result = PCI_ERS_RESULT_RECOVERED; 1255 int err; 1256 1257 dev_dbg(&pdev->dev, "%s post reset handling\n", DRV_NAME); 1258 1259 if (pci_enable_device_mem(pdev) < 0) { 1260 dev_err(&pdev->dev, 1261 "Failed to enable PCIe device after reset.\n"); 1262 result = PCI_ERS_RESULT_DISCONNECT; 1263 } else { 1264 pci_set_master(pdev); 1265 pci_restore_state(pdev); 1266 pci_save_state(pdev); 1267 pci_wake_from_d3(pdev, false); 1268 } 1269 1270 err = pci_cleanup_aer_uncorrect_error_status(pdev); 1271 if (err) { 1272 dev_err(&pdev->dev, 1273 "AER uncorrect error status clear failed: %#x\n", err); 1274 } 1275 1276 return result; 1277 } 1278 1279 static void ioat_pcie_error_resume(struct pci_dev *pdev) 1280 { 1281 struct ioatdma_device *ioat_dma = pci_get_drvdata(pdev); 1282 1283 dev_dbg(&pdev->dev, "%s: AER handling resuming\n", DRV_NAME); 1284 1285 /* initialize and bring everything back */ 1286 ioat_resume(ioat_dma); 1287 } 1288 1289 static const struct pci_error_handlers ioat_err_handler = { 1290 .error_detected = ioat_pcie_error_detected, 1291 .slot_reset = ioat_pcie_error_slot_reset, 1292 .resume = ioat_pcie_error_resume, 1293 }; 1294 1295 static struct pci_driver ioat_pci_driver = { 1296 .name = DRV_NAME, 1297 .id_table = ioat_pci_tbl, 1298 .probe = ioat_pci_probe, 1299 .remove = ioat_remove, 1300 .shutdown = ioat_shutdown, 1301 .err_handler = &ioat_err_handler, 1302 }; 1303 1304 static struct ioatdma_device * 1305 alloc_ioatdma(struct pci_dev *pdev, void __iomem *iobase) 1306 { 1307 struct device *dev = &pdev->dev; 1308 struct ioatdma_device *d = devm_kzalloc(dev, sizeof(*d), GFP_KERNEL); 1309 1310 if (!d) 1311 return NULL; 1312 d->pdev = pdev; 1313 d->reg_base = iobase; 1314 return d; 1315 } 1316 1317 static int ioat_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id) 1318 { 1319 void __iomem * const *iomap; 1320 struct device *dev = &pdev->dev; 1321 struct ioatdma_device *device; 1322 int err; 1323 1324 err = pcim_enable_device(pdev); 1325 if (err) 1326 return err; 1327 1328 err = pcim_iomap_regions(pdev, 1 << IOAT_MMIO_BAR, DRV_NAME); 1329 if (err) 1330 return err; 1331 iomap = pcim_iomap_table(pdev); 1332 if (!iomap) 1333 return -ENOMEM; 1334 1335 err = pci_set_dma_mask(pdev, DMA_BIT_MASK(64)); 1336 if (err) 1337 err = pci_set_dma_mask(pdev, DMA_BIT_MASK(32)); 1338 if (err) 1339 return err; 1340 1341 err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64)); 1342 if (err) 1343 err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32)); 1344 if (err) 1345 return err; 1346 1347 device = alloc_ioatdma(pdev, iomap[IOAT_MMIO_BAR]); 1348 if (!device) 1349 return -ENOMEM; 1350 pci_set_master(pdev); 1351 pci_set_drvdata(pdev, device); 1352 1353 device->version = readb(device->reg_base + IOAT_VER_OFFSET); 1354 if (device->version >= IOAT_VER_3_0) { 1355 if (is_skx_ioat(pdev)) 1356 device->version = IOAT_VER_3_2; 1357 err = ioat3_dma_probe(device, ioat_dca_enabled); 1358 1359 if (device->version >= IOAT_VER_3_3) 1360 pci_enable_pcie_error_reporting(pdev); 1361 } else 1362 return -ENODEV; 1363 1364 if (err) { 1365 dev_err(dev, "Intel(R) I/OAT DMA Engine init failed\n"); 1366 pci_disable_pcie_error_reporting(pdev); 1367 return -ENODEV; 1368 } 1369 1370 return 0; 1371 } 1372 1373 static void ioat_remove(struct pci_dev *pdev) 1374 { 1375 struct ioatdma_device *device = pci_get_drvdata(pdev); 1376 1377 if (!device) 1378 return; 1379 1380 dev_err(&pdev->dev, "Removing dma and dca services\n"); 1381 if (device->dca) { 1382 unregister_dca_provider(device->dca, &pdev->dev); 1383 free_dca_provider(device->dca); 1384 device->dca = NULL; 1385 } 1386 1387 pci_disable_pcie_error_reporting(pdev); 1388 ioat_dma_remove(device); 1389 } 1390 1391 static int __init ioat_init_module(void) 1392 { 1393 int err = -ENOMEM; 1394 1395 pr_info("%s: Intel(R) QuickData Technology Driver %s\n", 1396 DRV_NAME, IOAT_DMA_VERSION); 1397 1398 ioat_cache = kmem_cache_create("ioat", sizeof(struct ioat_ring_ent), 1399 0, SLAB_HWCACHE_ALIGN, NULL); 1400 if (!ioat_cache) 1401 return -ENOMEM; 1402 1403 ioat_sed_cache = KMEM_CACHE(ioat_sed_ent, 0); 1404 if (!ioat_sed_cache) 1405 goto err_ioat_cache; 1406 1407 err = pci_register_driver(&ioat_pci_driver); 1408 if (err) 1409 goto err_ioat3_cache; 1410 1411 return 0; 1412 1413 err_ioat3_cache: 1414 kmem_cache_destroy(ioat_sed_cache); 1415 1416 err_ioat_cache: 1417 kmem_cache_destroy(ioat_cache); 1418 1419 return err; 1420 } 1421 module_init(ioat_init_module); 1422 1423 static void __exit ioat_exit_module(void) 1424 { 1425 pci_unregister_driver(&ioat_pci_driver); 1426 kmem_cache_destroy(ioat_cache); 1427 } 1428 module_exit(ioat_exit_module); 1429