Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2// Copyright (C) 2017 Broadcom
 
 
 
 
 
 
 
 
 
 
   3
   4/*
   5 * Broadcom SBA RAID Driver
   6 *
   7 * The Broadcom stream buffer accelerator (SBA) provides offloading
   8 * capabilities for RAID operations. The SBA offload engine is accessible
   9 * via Broadcom SoC specific ring manager. Two or more offload engines
  10 * can share same Broadcom SoC specific ring manager due to this Broadcom
  11 * SoC specific ring manager driver is implemented as a mailbox controller
  12 * driver and offload engine drivers are implemented as mallbox clients.
  13 *
  14 * Typically, Broadcom SoC specific ring manager will implement larger
  15 * number of hardware rings over one or more SBA hardware devices. By
  16 * design, the internal buffer size of SBA hardware device is limited
  17 * but all offload operations supported by SBA can be broken down into
  18 * multiple small size requests and executed parallely on multiple SBA
  19 * hardware devices for achieving high through-put.
  20 *
  21 * The Broadcom SBA RAID driver does not require any register programming
  22 * except submitting request to SBA hardware device via mailbox channels.
  23 * This driver implements a DMA device with one DMA channel using a single
  24 * mailbox channel provided by Broadcom SoC specific ring manager driver.
  25 * For having more SBA DMA channels, we can create more SBA device nodes
  26 * in Broadcom SoC specific DTS based on number of hardware rings supported
  27 * by Broadcom SoC ring manager.
  28 */
  29
  30#include <linux/bitops.h>
  31#include <linux/debugfs.h>
  32#include <linux/dma-mapping.h>
  33#include <linux/dmaengine.h>
  34#include <linux/list.h>
  35#include <linux/mailbox_client.h>
  36#include <linux/mailbox/brcm-message.h>
  37#include <linux/module.h>
  38#include <linux/of.h>
  39#include <linux/of_platform.h>
  40#include <linux/platform_device.h>
  41#include <linux/slab.h>
  42#include <linux/raid/pq.h>
  43
  44#include "dmaengine.h"
  45
  46/* ====== Driver macros and defines ===== */
  47
  48#define SBA_TYPE_SHIFT					48
  49#define SBA_TYPE_MASK					GENMASK(1, 0)
  50#define SBA_TYPE_A					0x0
  51#define SBA_TYPE_B					0x2
  52#define SBA_TYPE_C					0x3
  53#define SBA_USER_DEF_SHIFT				32
  54#define SBA_USER_DEF_MASK				GENMASK(15, 0)
  55#define SBA_R_MDATA_SHIFT				24
  56#define SBA_R_MDATA_MASK				GENMASK(7, 0)
  57#define SBA_C_MDATA_MS_SHIFT				18
  58#define SBA_C_MDATA_MS_MASK				GENMASK(1, 0)
  59#define SBA_INT_SHIFT					17
  60#define SBA_INT_MASK					BIT(0)
  61#define SBA_RESP_SHIFT					16
  62#define SBA_RESP_MASK					BIT(0)
  63#define SBA_C_MDATA_SHIFT				8
  64#define SBA_C_MDATA_MASK				GENMASK(7, 0)
  65#define SBA_C_MDATA_BNUMx_SHIFT(__bnum)			(2 * (__bnum))
  66#define SBA_C_MDATA_BNUMx_MASK				GENMASK(1, 0)
  67#define SBA_C_MDATA_DNUM_SHIFT				5
  68#define SBA_C_MDATA_DNUM_MASK				GENMASK(4, 0)
  69#define SBA_C_MDATA_LS(__v)				((__v) & 0xff)
  70#define SBA_C_MDATA_MS(__v)				(((__v) >> 8) & 0x3)
  71#define SBA_CMD_SHIFT					0
  72#define SBA_CMD_MASK					GENMASK(3, 0)
  73#define SBA_CMD_ZERO_BUFFER				0x4
  74#define SBA_CMD_ZERO_ALL_BUFFERS			0x8
  75#define SBA_CMD_LOAD_BUFFER				0x9
  76#define SBA_CMD_XOR					0xa
  77#define SBA_CMD_GALOIS_XOR				0xb
  78#define SBA_CMD_WRITE_BUFFER				0xc
  79#define SBA_CMD_GALOIS					0xe
  80
  81#define SBA_MAX_REQ_PER_MBOX_CHANNEL			8192
  82#define SBA_MAX_MSG_SEND_PER_MBOX_CHANNEL		8
  83
  84/* Driver helper macros */
  85#define to_sba_request(tx)		\
  86	container_of(tx, struct sba_request, tx)
  87#define to_sba_device(dchan)		\
  88	container_of(dchan, struct sba_device, dma_chan)
  89
  90/* ===== Driver data structures ===== */
  91
  92enum sba_request_flags {
  93	SBA_REQUEST_STATE_FREE		= 0x001,
  94	SBA_REQUEST_STATE_ALLOCED	= 0x002,
  95	SBA_REQUEST_STATE_PENDING	= 0x004,
  96	SBA_REQUEST_STATE_ACTIVE	= 0x008,
  97	SBA_REQUEST_STATE_ABORTED	= 0x010,
  98	SBA_REQUEST_STATE_MASK		= 0x0ff,
  99	SBA_REQUEST_FENCE		= 0x100,
 100};
 101
 102struct sba_request {
 103	/* Global state */
 104	struct list_head node;
 105	struct sba_device *sba;
 106	u32 flags;
 107	/* Chained requests management */
 108	struct sba_request *first;
 109	struct list_head next;
 110	atomic_t next_pending_count;
 111	/* BRCM message data */
 112	struct brcm_message msg;
 113	struct dma_async_tx_descriptor tx;
 114	/* SBA commands */
 115	struct brcm_sba_command cmds[];
 116};
 117
 118enum sba_version {
 119	SBA_VER_1 = 0,
 120	SBA_VER_2
 121};
 122
 123struct sba_device {
 124	/* Underlying device */
 125	struct device *dev;
 126	/* DT configuration parameters */
 127	enum sba_version ver;
 128	/* Derived configuration parameters */
 129	u32 max_req;
 130	u32 hw_buf_size;
 131	u32 hw_resp_size;
 132	u32 max_pq_coefs;
 133	u32 max_pq_srcs;
 134	u32 max_cmd_per_req;
 135	u32 max_xor_srcs;
 136	u32 max_resp_pool_size;
 137	u32 max_cmds_pool_size;
 138	/* Maibox client and Mailbox channels */
 139	struct mbox_client client;
 140	struct mbox_chan *mchan;
 141	struct device *mbox_dev;
 142	/* DMA device and DMA channel */
 143	struct dma_device dma_dev;
 144	struct dma_chan dma_chan;
 145	/* DMA channel resources */
 146	void *resp_base;
 147	dma_addr_t resp_dma_base;
 148	void *cmds_base;
 149	dma_addr_t cmds_dma_base;
 150	spinlock_t reqs_lock;
 151	bool reqs_fence;
 152	struct list_head reqs_alloc_list;
 153	struct list_head reqs_pending_list;
 154	struct list_head reqs_active_list;
 155	struct list_head reqs_aborted_list;
 156	struct list_head reqs_free_list;
 157	/* DebugFS directory entries */
 158	struct dentry *root;
 
 159};
 160
 161/* ====== Command helper routines ===== */
 162
 163static inline u64 __pure sba_cmd_enc(u64 cmd, u32 val, u32 shift, u32 mask)
 164{
 165	cmd &= ~((u64)mask << shift);
 166	cmd |= ((u64)(val & mask) << shift);
 167	return cmd;
 168}
 169
 170static inline u32 __pure sba_cmd_load_c_mdata(u32 b0)
 171{
 172	return b0 & SBA_C_MDATA_BNUMx_MASK;
 173}
 174
 175static inline u32 __pure sba_cmd_write_c_mdata(u32 b0)
 176{
 177	return b0 & SBA_C_MDATA_BNUMx_MASK;
 178}
 179
 180static inline u32 __pure sba_cmd_xor_c_mdata(u32 b1, u32 b0)
 181{
 182	return (b0 & SBA_C_MDATA_BNUMx_MASK) |
 183	       ((b1 & SBA_C_MDATA_BNUMx_MASK) << SBA_C_MDATA_BNUMx_SHIFT(1));
 184}
 185
 186static inline u32 __pure sba_cmd_pq_c_mdata(u32 d, u32 b1, u32 b0)
 187{
 188	return (b0 & SBA_C_MDATA_BNUMx_MASK) |
 189	       ((b1 & SBA_C_MDATA_BNUMx_MASK) << SBA_C_MDATA_BNUMx_SHIFT(1)) |
 190	       ((d & SBA_C_MDATA_DNUM_MASK) << SBA_C_MDATA_DNUM_SHIFT);
 191}
 192
 193/* ====== General helper routines ===== */
 194
 195static struct sba_request *sba_alloc_request(struct sba_device *sba)
 196{
 197	bool found = false;
 198	unsigned long flags;
 199	struct sba_request *req = NULL;
 200
 201	spin_lock_irqsave(&sba->reqs_lock, flags);
 202	list_for_each_entry(req, &sba->reqs_free_list, node) {
 203		if (async_tx_test_ack(&req->tx)) {
 204			list_move_tail(&req->node, &sba->reqs_alloc_list);
 205			found = true;
 206			break;
 207		}
 208	}
 209	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 210
 211	if (!found) {
 212		/*
 213		 * We have no more free requests so, we peek
 214		 * mailbox channels hoping few active requests
 215		 * would have completed which will create more
 216		 * room for new requests.
 217		 */
 218		mbox_client_peek_data(sba->mchan);
 219		return NULL;
 220	}
 221
 222	req->flags = SBA_REQUEST_STATE_ALLOCED;
 223	req->first = req;
 224	INIT_LIST_HEAD(&req->next);
 225	atomic_set(&req->next_pending_count, 1);
 226
 227	dma_async_tx_descriptor_init(&req->tx, &sba->dma_chan);
 228	async_tx_ack(&req->tx);
 229
 230	return req;
 231}
 232
 233/* Note: Must be called with sba->reqs_lock held */
 234static void _sba_pending_request(struct sba_device *sba,
 235				 struct sba_request *req)
 236{
 237	lockdep_assert_held(&sba->reqs_lock);
 238	req->flags &= ~SBA_REQUEST_STATE_MASK;
 239	req->flags |= SBA_REQUEST_STATE_PENDING;
 240	list_move_tail(&req->node, &sba->reqs_pending_list);
 241	if (list_empty(&sba->reqs_active_list))
 242		sba->reqs_fence = false;
 243}
 244
 245/* Note: Must be called with sba->reqs_lock held */
 246static bool _sba_active_request(struct sba_device *sba,
 247				struct sba_request *req)
 248{
 249	lockdep_assert_held(&sba->reqs_lock);
 250	if (list_empty(&sba->reqs_active_list))
 251		sba->reqs_fence = false;
 252	if (sba->reqs_fence)
 253		return false;
 254	req->flags &= ~SBA_REQUEST_STATE_MASK;
 255	req->flags |= SBA_REQUEST_STATE_ACTIVE;
 256	list_move_tail(&req->node, &sba->reqs_active_list);
 257	if (req->flags & SBA_REQUEST_FENCE)
 258		sba->reqs_fence = true;
 259	return true;
 260}
 261
 262/* Note: Must be called with sba->reqs_lock held */
 263static void _sba_abort_request(struct sba_device *sba,
 264			       struct sba_request *req)
 265{
 266	lockdep_assert_held(&sba->reqs_lock);
 267	req->flags &= ~SBA_REQUEST_STATE_MASK;
 268	req->flags |= SBA_REQUEST_STATE_ABORTED;
 269	list_move_tail(&req->node, &sba->reqs_aborted_list);
 270	if (list_empty(&sba->reqs_active_list))
 271		sba->reqs_fence = false;
 272}
 273
 274/* Note: Must be called with sba->reqs_lock held */
 275static void _sba_free_request(struct sba_device *sba,
 276			      struct sba_request *req)
 277{
 278	lockdep_assert_held(&sba->reqs_lock);
 279	req->flags &= ~SBA_REQUEST_STATE_MASK;
 280	req->flags |= SBA_REQUEST_STATE_FREE;
 281	list_move_tail(&req->node, &sba->reqs_free_list);
 282	if (list_empty(&sba->reqs_active_list))
 283		sba->reqs_fence = false;
 284}
 285
 286static void sba_free_chained_requests(struct sba_request *req)
 287{
 288	unsigned long flags;
 289	struct sba_request *nreq;
 290	struct sba_device *sba = req->sba;
 291
 292	spin_lock_irqsave(&sba->reqs_lock, flags);
 293
 294	_sba_free_request(sba, req);
 295	list_for_each_entry(nreq, &req->next, next)
 296		_sba_free_request(sba, nreq);
 297
 298	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 299}
 300
 301static void sba_chain_request(struct sba_request *first,
 302			      struct sba_request *req)
 303{
 304	unsigned long flags;
 305	struct sba_device *sba = req->sba;
 306
 307	spin_lock_irqsave(&sba->reqs_lock, flags);
 308
 309	list_add_tail(&req->next, &first->next);
 310	req->first = first;
 311	atomic_inc(&first->next_pending_count);
 312
 313	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 314}
 315
 316static void sba_cleanup_nonpending_requests(struct sba_device *sba)
 317{
 318	unsigned long flags;
 319	struct sba_request *req, *req1;
 320
 321	spin_lock_irqsave(&sba->reqs_lock, flags);
 322
 323	/* Freeup all alloced request */
 324	list_for_each_entry_safe(req, req1, &sba->reqs_alloc_list, node)
 325		_sba_free_request(sba, req);
 326
 327	/* Set all active requests as aborted */
 328	list_for_each_entry_safe(req, req1, &sba->reqs_active_list, node)
 329		_sba_abort_request(sba, req);
 330
 331	/*
 332	 * Note: We expect that aborted request will be eventually
 333	 * freed by sba_receive_message()
 334	 */
 335
 336	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 337}
 338
 339static void sba_cleanup_pending_requests(struct sba_device *sba)
 340{
 341	unsigned long flags;
 342	struct sba_request *req, *req1;
 343
 344	spin_lock_irqsave(&sba->reqs_lock, flags);
 345
 346	/* Freeup all pending request */
 347	list_for_each_entry_safe(req, req1, &sba->reqs_pending_list, node)
 348		_sba_free_request(sba, req);
 349
 350	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 351}
 352
 353static int sba_send_mbox_request(struct sba_device *sba,
 354				 struct sba_request *req)
 355{
 356	int ret = 0;
 357
 358	/* Send message for the request */
 359	req->msg.error = 0;
 360	ret = mbox_send_message(sba->mchan, &req->msg);
 361	if (ret < 0) {
 362		dev_err(sba->dev, "send message failed with error %d", ret);
 363		return ret;
 364	}
 365
 366	/* Check error returned by mailbox controller */
 367	ret = req->msg.error;
 368	if (ret < 0) {
 369		dev_err(sba->dev, "message error %d", ret);
 370	}
 371
 372	/* Signal txdone for mailbox channel */
 373	mbox_client_txdone(sba->mchan, ret);
 374
 375	return ret;
 376}
 377
 378/* Note: Must be called with sba->reqs_lock held */
 379static void _sba_process_pending_requests(struct sba_device *sba)
 380{
 381	int ret;
 382	u32 count;
 383	struct sba_request *req;
 384
 385	/* Process few pending requests */
 386	count = SBA_MAX_MSG_SEND_PER_MBOX_CHANNEL;
 387	while (!list_empty(&sba->reqs_pending_list) && count) {
 388		/* Get the first pending request */
 389		req = list_first_entry(&sba->reqs_pending_list,
 390				       struct sba_request, node);
 391
 392		/* Try to make request active */
 393		if (!_sba_active_request(sba, req))
 394			break;
 395
 396		/* Send request to mailbox channel */
 397		ret = sba_send_mbox_request(sba, req);
 398		if (ret < 0) {
 399			_sba_pending_request(sba, req);
 400			break;
 401		}
 402
 403		count--;
 404	}
 405}
 406
 407static void sba_process_received_request(struct sba_device *sba,
 408					 struct sba_request *req)
 409{
 410	unsigned long flags;
 411	struct dma_async_tx_descriptor *tx;
 412	struct sba_request *nreq, *first = req->first;
 413
 414	/* Process only after all chained requests are received */
 415	if (!atomic_dec_return(&first->next_pending_count)) {
 416		tx = &first->tx;
 417
 418		WARN_ON(tx->cookie < 0);
 419		if (tx->cookie > 0) {
 420			spin_lock_irqsave(&sba->reqs_lock, flags);
 421			dma_cookie_complete(tx);
 422			spin_unlock_irqrestore(&sba->reqs_lock, flags);
 423			dmaengine_desc_get_callback_invoke(tx, NULL);
 424			dma_descriptor_unmap(tx);
 425			tx->callback = NULL;
 426			tx->callback_result = NULL;
 427		}
 428
 429		dma_run_dependencies(tx);
 430
 431		spin_lock_irqsave(&sba->reqs_lock, flags);
 432
 433		/* Free all requests chained to first request */
 434		list_for_each_entry(nreq, &first->next, next)
 435			_sba_free_request(sba, nreq);
 436		INIT_LIST_HEAD(&first->next);
 437
 438		/* Free the first request */
 439		_sba_free_request(sba, first);
 440
 441		/* Process pending requests */
 442		_sba_process_pending_requests(sba);
 443
 444		spin_unlock_irqrestore(&sba->reqs_lock, flags);
 445	}
 446}
 447
 448static void sba_write_stats_in_seqfile(struct sba_device *sba,
 449				       struct seq_file *file)
 450{
 451	unsigned long flags;
 452	struct sba_request *req;
 453	u32 free_count = 0, alloced_count = 0;
 454	u32 pending_count = 0, active_count = 0, aborted_count = 0;
 455
 456	spin_lock_irqsave(&sba->reqs_lock, flags);
 457
 458	list_for_each_entry(req, &sba->reqs_free_list, node)
 459		if (async_tx_test_ack(&req->tx))
 460			free_count++;
 461
 462	list_for_each_entry(req, &sba->reqs_alloc_list, node)
 463		alloced_count++;
 464
 465	list_for_each_entry(req, &sba->reqs_pending_list, node)
 466		pending_count++;
 467
 468	list_for_each_entry(req, &sba->reqs_active_list, node)
 469		active_count++;
 470
 471	list_for_each_entry(req, &sba->reqs_aborted_list, node)
 472		aborted_count++;
 473
 474	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 475
 476	seq_printf(file, "maximum requests   = %d\n", sba->max_req);
 477	seq_printf(file, "free requests      = %d\n", free_count);
 478	seq_printf(file, "alloced requests   = %d\n", alloced_count);
 479	seq_printf(file, "pending requests   = %d\n", pending_count);
 480	seq_printf(file, "active requests    = %d\n", active_count);
 481	seq_printf(file, "aborted requests   = %d\n", aborted_count);
 482}
 483
 484/* ====== DMAENGINE callbacks ===== */
 485
 486static void sba_free_chan_resources(struct dma_chan *dchan)
 487{
 488	/*
 489	 * Channel resources are pre-alloced so we just free-up
 490	 * whatever we can so that we can re-use pre-alloced
 491	 * channel resources next time.
 492	 */
 493	sba_cleanup_nonpending_requests(to_sba_device(dchan));
 494}
 495
 496static int sba_device_terminate_all(struct dma_chan *dchan)
 497{
 498	/* Cleanup all pending requests */
 499	sba_cleanup_pending_requests(to_sba_device(dchan));
 500
 501	return 0;
 502}
 503
 504static void sba_issue_pending(struct dma_chan *dchan)
 505{
 506	unsigned long flags;
 507	struct sba_device *sba = to_sba_device(dchan);
 508
 509	/* Process pending requests */
 510	spin_lock_irqsave(&sba->reqs_lock, flags);
 511	_sba_process_pending_requests(sba);
 512	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 513}
 514
 515static dma_cookie_t sba_tx_submit(struct dma_async_tx_descriptor *tx)
 516{
 517	unsigned long flags;
 518	dma_cookie_t cookie;
 519	struct sba_device *sba;
 520	struct sba_request *req, *nreq;
 521
 522	if (unlikely(!tx))
 523		return -EINVAL;
 524
 525	sba = to_sba_device(tx->chan);
 526	req = to_sba_request(tx);
 527
 528	/* Assign cookie and mark all chained requests pending */
 529	spin_lock_irqsave(&sba->reqs_lock, flags);
 530	cookie = dma_cookie_assign(tx);
 531	_sba_pending_request(sba, req);
 532	list_for_each_entry(nreq, &req->next, next)
 533		_sba_pending_request(sba, nreq);
 534	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 535
 536	return cookie;
 537}
 538
 539static enum dma_status sba_tx_status(struct dma_chan *dchan,
 540				     dma_cookie_t cookie,
 541				     struct dma_tx_state *txstate)
 542{
 543	enum dma_status ret;
 544	struct sba_device *sba = to_sba_device(dchan);
 545
 546	ret = dma_cookie_status(dchan, cookie, txstate);
 547	if (ret == DMA_COMPLETE)
 548		return ret;
 549
 550	mbox_client_peek_data(sba->mchan);
 551
 552	return dma_cookie_status(dchan, cookie, txstate);
 553}
 554
 555static void sba_fillup_interrupt_msg(struct sba_request *req,
 556				     struct brcm_sba_command *cmds,
 557				     struct brcm_message *msg)
 558{
 559	u64 cmd;
 560	u32 c_mdata;
 561	dma_addr_t resp_dma = req->tx.phys;
 562	struct brcm_sba_command *cmdsp = cmds;
 563
 564	/* Type-B command to load dummy data into buf0 */
 565	cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 566			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 567	cmd = sba_cmd_enc(cmd, req->sba->hw_resp_size,
 568			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 569	c_mdata = sba_cmd_load_c_mdata(0);
 570	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 571			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 572	cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 573			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 574	cmdsp->cmd = cmd;
 575	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 576	cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 577	cmdsp->data = resp_dma;
 578	cmdsp->data_len = req->sba->hw_resp_size;
 579	cmdsp++;
 580
 581	/* Type-A command to write buf0 to dummy location */
 582	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 583			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 584	cmd = sba_cmd_enc(cmd, req->sba->hw_resp_size,
 585			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 586	cmd = sba_cmd_enc(cmd, 0x1,
 587			  SBA_RESP_SHIFT, SBA_RESP_MASK);
 588	c_mdata = sba_cmd_write_c_mdata(0);
 589	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 590			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 591	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
 592			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 593	cmdsp->cmd = cmd;
 594	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 595	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
 596	if (req->sba->hw_resp_size) {
 597		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
 598		cmdsp->resp = resp_dma;
 599		cmdsp->resp_len = req->sba->hw_resp_size;
 600	}
 601	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
 602	cmdsp->data = resp_dma;
 603	cmdsp->data_len = req->sba->hw_resp_size;
 604	cmdsp++;
 605
 606	/* Fillup brcm_message */
 607	msg->type = BRCM_MESSAGE_SBA;
 608	msg->sba.cmds = cmds;
 609	msg->sba.cmds_count = cmdsp - cmds;
 610	msg->ctx = req;
 611	msg->error = 0;
 612}
 613
 614static struct dma_async_tx_descriptor *
 615sba_prep_dma_interrupt(struct dma_chan *dchan, unsigned long flags)
 616{
 617	struct sba_request *req = NULL;
 618	struct sba_device *sba = to_sba_device(dchan);
 619
 620	/* Alloc new request */
 621	req = sba_alloc_request(sba);
 622	if (!req)
 623		return NULL;
 624
 625	/*
 626	 * Force fence so that no requests are submitted
 627	 * until DMA callback for this request is invoked.
 628	 */
 629	req->flags |= SBA_REQUEST_FENCE;
 630
 631	/* Fillup request message */
 632	sba_fillup_interrupt_msg(req, req->cmds, &req->msg);
 633
 634	/* Init async_tx descriptor */
 635	req->tx.flags = flags;
 636	req->tx.cookie = -EBUSY;
 637
 638	return &req->tx;
 639}
 640
 641static void sba_fillup_memcpy_msg(struct sba_request *req,
 642				  struct brcm_sba_command *cmds,
 643				  struct brcm_message *msg,
 644				  dma_addr_t msg_offset, size_t msg_len,
 645				  dma_addr_t dst, dma_addr_t src)
 646{
 647	u64 cmd;
 648	u32 c_mdata;
 649	dma_addr_t resp_dma = req->tx.phys;
 650	struct brcm_sba_command *cmdsp = cmds;
 651
 652	/* Type-B command to load data into buf0 */
 653	cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 654			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 655	cmd = sba_cmd_enc(cmd, msg_len,
 656			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 657	c_mdata = sba_cmd_load_c_mdata(0);
 658	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 659			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 660	cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 661			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 662	cmdsp->cmd = cmd;
 663	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 664	cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 665	cmdsp->data = src + msg_offset;
 666	cmdsp->data_len = msg_len;
 667	cmdsp++;
 668
 669	/* Type-A command to write buf0 */
 670	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 671			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 672	cmd = sba_cmd_enc(cmd, msg_len,
 673			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 674	cmd = sba_cmd_enc(cmd, 0x1,
 675			  SBA_RESP_SHIFT, SBA_RESP_MASK);
 676	c_mdata = sba_cmd_write_c_mdata(0);
 677	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 678			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 679	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
 680			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 681	cmdsp->cmd = cmd;
 682	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 683	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
 684	if (req->sba->hw_resp_size) {
 685		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
 686		cmdsp->resp = resp_dma;
 687		cmdsp->resp_len = req->sba->hw_resp_size;
 688	}
 689	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
 690	cmdsp->data = dst + msg_offset;
 691	cmdsp->data_len = msg_len;
 692	cmdsp++;
 693
 694	/* Fillup brcm_message */
 695	msg->type = BRCM_MESSAGE_SBA;
 696	msg->sba.cmds = cmds;
 697	msg->sba.cmds_count = cmdsp - cmds;
 698	msg->ctx = req;
 699	msg->error = 0;
 700}
 701
 702static struct sba_request *
 703sba_prep_dma_memcpy_req(struct sba_device *sba,
 704			dma_addr_t off, dma_addr_t dst, dma_addr_t src,
 705			size_t len, unsigned long flags)
 706{
 707	struct sba_request *req = NULL;
 708
 709	/* Alloc new request */
 710	req = sba_alloc_request(sba);
 711	if (!req)
 712		return NULL;
 713	if (flags & DMA_PREP_FENCE)
 714		req->flags |= SBA_REQUEST_FENCE;
 715
 716	/* Fillup request message */
 717	sba_fillup_memcpy_msg(req, req->cmds, &req->msg,
 718			      off, len, dst, src);
 719
 720	/* Init async_tx descriptor */
 721	req->tx.flags = flags;
 722	req->tx.cookie = -EBUSY;
 723
 724	return req;
 725}
 726
 727static struct dma_async_tx_descriptor *
 728sba_prep_dma_memcpy(struct dma_chan *dchan, dma_addr_t dst, dma_addr_t src,
 729		    size_t len, unsigned long flags)
 730{
 731	size_t req_len;
 732	dma_addr_t off = 0;
 733	struct sba_device *sba = to_sba_device(dchan);
 734	struct sba_request *first = NULL, *req;
 735
 736	/* Create chained requests where each request is upto hw_buf_size */
 737	while (len) {
 738		req_len = (len < sba->hw_buf_size) ? len : sba->hw_buf_size;
 739
 740		req = sba_prep_dma_memcpy_req(sba, off, dst, src,
 741					      req_len, flags);
 742		if (!req) {
 743			if (first)
 744				sba_free_chained_requests(first);
 745			return NULL;
 746		}
 747
 748		if (first)
 749			sba_chain_request(first, req);
 750		else
 751			first = req;
 752
 753		off += req_len;
 754		len -= req_len;
 755	}
 756
 757	return (first) ? &first->tx : NULL;
 758}
 759
 760static void sba_fillup_xor_msg(struct sba_request *req,
 761				struct brcm_sba_command *cmds,
 762				struct brcm_message *msg,
 763				dma_addr_t msg_offset, size_t msg_len,
 764				dma_addr_t dst, dma_addr_t *src, u32 src_cnt)
 765{
 766	u64 cmd;
 767	u32 c_mdata;
 768	unsigned int i;
 769	dma_addr_t resp_dma = req->tx.phys;
 770	struct brcm_sba_command *cmdsp = cmds;
 771
 772	/* Type-B command to load data into buf0 */
 773	cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 774			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 775	cmd = sba_cmd_enc(cmd, msg_len,
 776			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 777	c_mdata = sba_cmd_load_c_mdata(0);
 778	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 779			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 780	cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 781			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 782	cmdsp->cmd = cmd;
 783	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 784	cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 785	cmdsp->data = src[0] + msg_offset;
 786	cmdsp->data_len = msg_len;
 787	cmdsp++;
 788
 789	/* Type-B commands to xor data with buf0 and put it back in buf0 */
 790	for (i = 1; i < src_cnt; i++) {
 791		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 792				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 793		cmd = sba_cmd_enc(cmd, msg_len,
 794				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 795		c_mdata = sba_cmd_xor_c_mdata(0, 0);
 796		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 797				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 798		cmd = sba_cmd_enc(cmd, SBA_CMD_XOR,
 799				  SBA_CMD_SHIFT, SBA_CMD_MASK);
 800		cmdsp->cmd = cmd;
 801		*cmdsp->cmd_dma = cpu_to_le64(cmd);
 802		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 803		cmdsp->data = src[i] + msg_offset;
 804		cmdsp->data_len = msg_len;
 805		cmdsp++;
 806	}
 807
 808	/* Type-A command to write buf0 */
 809	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 810			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 811	cmd = sba_cmd_enc(cmd, msg_len,
 812			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 813	cmd = sba_cmd_enc(cmd, 0x1,
 814			  SBA_RESP_SHIFT, SBA_RESP_MASK);
 815	c_mdata = sba_cmd_write_c_mdata(0);
 816	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 817			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 818	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
 819			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 820	cmdsp->cmd = cmd;
 821	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 822	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
 823	if (req->sba->hw_resp_size) {
 824		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
 825		cmdsp->resp = resp_dma;
 826		cmdsp->resp_len = req->sba->hw_resp_size;
 827	}
 828	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
 829	cmdsp->data = dst + msg_offset;
 830	cmdsp->data_len = msg_len;
 831	cmdsp++;
 832
 833	/* Fillup brcm_message */
 834	msg->type = BRCM_MESSAGE_SBA;
 835	msg->sba.cmds = cmds;
 836	msg->sba.cmds_count = cmdsp - cmds;
 837	msg->ctx = req;
 838	msg->error = 0;
 839}
 840
 841static struct sba_request *
 842sba_prep_dma_xor_req(struct sba_device *sba,
 843		     dma_addr_t off, dma_addr_t dst, dma_addr_t *src,
 844		     u32 src_cnt, size_t len, unsigned long flags)
 845{
 846	struct sba_request *req = NULL;
 847
 848	/* Alloc new request */
 849	req = sba_alloc_request(sba);
 850	if (!req)
 851		return NULL;
 852	if (flags & DMA_PREP_FENCE)
 853		req->flags |= SBA_REQUEST_FENCE;
 854
 855	/* Fillup request message */
 856	sba_fillup_xor_msg(req, req->cmds, &req->msg,
 857			   off, len, dst, src, src_cnt);
 858
 859	/* Init async_tx descriptor */
 860	req->tx.flags = flags;
 861	req->tx.cookie = -EBUSY;
 862
 863	return req;
 864}
 865
 866static struct dma_async_tx_descriptor *
 867sba_prep_dma_xor(struct dma_chan *dchan, dma_addr_t dst, dma_addr_t *src,
 868		 u32 src_cnt, size_t len, unsigned long flags)
 869{
 870	size_t req_len;
 871	dma_addr_t off = 0;
 872	struct sba_device *sba = to_sba_device(dchan);
 873	struct sba_request *first = NULL, *req;
 874
 875	/* Sanity checks */
 876	if (unlikely(src_cnt > sba->max_xor_srcs))
 877		return NULL;
 878
 879	/* Create chained requests where each request is upto hw_buf_size */
 880	while (len) {
 881		req_len = (len < sba->hw_buf_size) ? len : sba->hw_buf_size;
 882
 883		req = sba_prep_dma_xor_req(sba, off, dst, src, src_cnt,
 884					   req_len, flags);
 885		if (!req) {
 886			if (first)
 887				sba_free_chained_requests(first);
 888			return NULL;
 889		}
 890
 891		if (first)
 892			sba_chain_request(first, req);
 893		else
 894			first = req;
 895
 896		off += req_len;
 897		len -= req_len;
 898	}
 899
 900	return (first) ? &first->tx : NULL;
 901}
 902
 903static void sba_fillup_pq_msg(struct sba_request *req,
 904				bool pq_continue,
 905				struct brcm_sba_command *cmds,
 906				struct brcm_message *msg,
 907				dma_addr_t msg_offset, size_t msg_len,
 908				dma_addr_t *dst_p, dma_addr_t *dst_q,
 909				const u8 *scf, dma_addr_t *src, u32 src_cnt)
 910{
 911	u64 cmd;
 912	u32 c_mdata;
 913	unsigned int i;
 914	dma_addr_t resp_dma = req->tx.phys;
 915	struct brcm_sba_command *cmdsp = cmds;
 916
 917	if (pq_continue) {
 918		/* Type-B command to load old P into buf0 */
 919		if (dst_p) {
 920			cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 921				SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 922			cmd = sba_cmd_enc(cmd, msg_len,
 923				SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 924			c_mdata = sba_cmd_load_c_mdata(0);
 925			cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 926				SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 927			cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 928				SBA_CMD_SHIFT, SBA_CMD_MASK);
 929			cmdsp->cmd = cmd;
 930			*cmdsp->cmd_dma = cpu_to_le64(cmd);
 931			cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 932			cmdsp->data = *dst_p + msg_offset;
 933			cmdsp->data_len = msg_len;
 934			cmdsp++;
 935		}
 936
 937		/* Type-B command to load old Q into buf1 */
 938		if (dst_q) {
 939			cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 940				SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 941			cmd = sba_cmd_enc(cmd, msg_len,
 942				SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 943			c_mdata = sba_cmd_load_c_mdata(1);
 944			cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 945				SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 946			cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 947				SBA_CMD_SHIFT, SBA_CMD_MASK);
 948			cmdsp->cmd = cmd;
 949			*cmdsp->cmd_dma = cpu_to_le64(cmd);
 950			cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 951			cmdsp->data = *dst_q + msg_offset;
 952			cmdsp->data_len = msg_len;
 953			cmdsp++;
 954		}
 955	} else {
 956		/* Type-A command to zero all buffers */
 957		cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 958				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 959		cmd = sba_cmd_enc(cmd, msg_len,
 960				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 961		cmd = sba_cmd_enc(cmd, SBA_CMD_ZERO_ALL_BUFFERS,
 962				  SBA_CMD_SHIFT, SBA_CMD_MASK);
 963		cmdsp->cmd = cmd;
 964		*cmdsp->cmd_dma = cpu_to_le64(cmd);
 965		cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
 966		cmdsp++;
 967	}
 968
 969	/* Type-B commands for generate P onto buf0 and Q onto buf1 */
 970	for (i = 0; i < src_cnt; i++) {
 971		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 972				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 973		cmd = sba_cmd_enc(cmd, msg_len,
 974				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 975		c_mdata = sba_cmd_pq_c_mdata(raid6_gflog[scf[i]], 1, 0);
 976		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 977				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 978		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_MS(c_mdata),
 979				  SBA_C_MDATA_MS_SHIFT, SBA_C_MDATA_MS_MASK);
 980		cmd = sba_cmd_enc(cmd, SBA_CMD_GALOIS_XOR,
 981				  SBA_CMD_SHIFT, SBA_CMD_MASK);
 982		cmdsp->cmd = cmd;
 983		*cmdsp->cmd_dma = cpu_to_le64(cmd);
 984		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 985		cmdsp->data = src[i] + msg_offset;
 986		cmdsp->data_len = msg_len;
 987		cmdsp++;
 988	}
 989
 990	/* Type-A command to write buf0 */
 991	if (dst_p) {
 992		cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 993				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 994		cmd = sba_cmd_enc(cmd, msg_len,
 995				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 996		cmd = sba_cmd_enc(cmd, 0x1,
 997				  SBA_RESP_SHIFT, SBA_RESP_MASK);
 998		c_mdata = sba_cmd_write_c_mdata(0);
 999		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1000				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1001		cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
1002				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1003		cmdsp->cmd = cmd;
1004		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1005		cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1006		if (req->sba->hw_resp_size) {
1007			cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
1008			cmdsp->resp = resp_dma;
1009			cmdsp->resp_len = req->sba->hw_resp_size;
1010		}
1011		cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
1012		cmdsp->data = *dst_p + msg_offset;
1013		cmdsp->data_len = msg_len;
1014		cmdsp++;
1015	}
1016
1017	/* Type-A command to write buf1 */
1018	if (dst_q) {
1019		cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1020				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1021		cmd = sba_cmd_enc(cmd, msg_len,
1022				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1023		cmd = sba_cmd_enc(cmd, 0x1,
1024				  SBA_RESP_SHIFT, SBA_RESP_MASK);
1025		c_mdata = sba_cmd_write_c_mdata(1);
1026		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1027				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1028		cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
1029				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1030		cmdsp->cmd = cmd;
1031		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1032		cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1033		if (req->sba->hw_resp_size) {
1034			cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
1035			cmdsp->resp = resp_dma;
1036			cmdsp->resp_len = req->sba->hw_resp_size;
1037		}
1038		cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
1039		cmdsp->data = *dst_q + msg_offset;
1040		cmdsp->data_len = msg_len;
1041		cmdsp++;
1042	}
1043
1044	/* Fillup brcm_message */
1045	msg->type = BRCM_MESSAGE_SBA;
1046	msg->sba.cmds = cmds;
1047	msg->sba.cmds_count = cmdsp - cmds;
1048	msg->ctx = req;
1049	msg->error = 0;
1050}
1051
1052static struct sba_request *
1053sba_prep_dma_pq_req(struct sba_device *sba, dma_addr_t off,
1054		    dma_addr_t *dst_p, dma_addr_t *dst_q, dma_addr_t *src,
1055		    u32 src_cnt, const u8 *scf, size_t len, unsigned long flags)
1056{
1057	struct sba_request *req = NULL;
1058
1059	/* Alloc new request */
1060	req = sba_alloc_request(sba);
1061	if (!req)
1062		return NULL;
1063	if (flags & DMA_PREP_FENCE)
1064		req->flags |= SBA_REQUEST_FENCE;
1065
1066	/* Fillup request messages */
1067	sba_fillup_pq_msg(req, dmaf_continue(flags),
1068			  req->cmds, &req->msg,
1069			  off, len, dst_p, dst_q, scf, src, src_cnt);
1070
1071	/* Init async_tx descriptor */
1072	req->tx.flags = flags;
1073	req->tx.cookie = -EBUSY;
1074
1075	return req;
1076}
1077
1078static void sba_fillup_pq_single_msg(struct sba_request *req,
1079				bool pq_continue,
1080				struct brcm_sba_command *cmds,
1081				struct brcm_message *msg,
1082				dma_addr_t msg_offset, size_t msg_len,
1083				dma_addr_t *dst_p, dma_addr_t *dst_q,
1084				dma_addr_t src, u8 scf)
1085{
1086	u64 cmd;
1087	u32 c_mdata;
1088	u8 pos, dpos = raid6_gflog[scf];
1089	dma_addr_t resp_dma = req->tx.phys;
1090	struct brcm_sba_command *cmdsp = cmds;
1091
1092	if (!dst_p)
1093		goto skip_p;
1094
1095	if (pq_continue) {
1096		/* Type-B command to load old P into buf0 */
1097		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1098				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1099		cmd = sba_cmd_enc(cmd, msg_len,
1100				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1101		c_mdata = sba_cmd_load_c_mdata(0);
1102		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1103				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1104		cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
1105				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1106		cmdsp->cmd = cmd;
1107		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1108		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1109		cmdsp->data = *dst_p + msg_offset;
1110		cmdsp->data_len = msg_len;
1111		cmdsp++;
1112
1113		/*
1114		 * Type-B commands to xor data with buf0 and put it
1115		 * back in buf0
1116		 */
1117		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1118				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1119		cmd = sba_cmd_enc(cmd, msg_len,
1120				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1121		c_mdata = sba_cmd_xor_c_mdata(0, 0);
1122		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1123				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1124		cmd = sba_cmd_enc(cmd, SBA_CMD_XOR,
1125				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1126		cmdsp->cmd = cmd;
1127		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1128		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1129		cmdsp->data = src + msg_offset;
1130		cmdsp->data_len = msg_len;
1131		cmdsp++;
1132	} else {
1133		/* Type-B command to load old P into buf0 */
1134		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1135				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1136		cmd = sba_cmd_enc(cmd, msg_len,
1137				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1138		c_mdata = sba_cmd_load_c_mdata(0);
1139		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1140				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1141		cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
1142				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1143		cmdsp->cmd = cmd;
1144		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1145		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1146		cmdsp->data = src + msg_offset;
1147		cmdsp->data_len = msg_len;
1148		cmdsp++;
1149	}
1150
1151	/* Type-A command to write buf0 */
1152	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1153			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1154	cmd = sba_cmd_enc(cmd, msg_len,
1155			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1156	cmd = sba_cmd_enc(cmd, 0x1,
1157			  SBA_RESP_SHIFT, SBA_RESP_MASK);
1158	c_mdata = sba_cmd_write_c_mdata(0);
1159	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1160			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1161	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
1162			  SBA_CMD_SHIFT, SBA_CMD_MASK);
1163	cmdsp->cmd = cmd;
1164	*cmdsp->cmd_dma = cpu_to_le64(cmd);
1165	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1166	if (req->sba->hw_resp_size) {
1167		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
1168		cmdsp->resp = resp_dma;
1169		cmdsp->resp_len = req->sba->hw_resp_size;
1170	}
1171	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
1172	cmdsp->data = *dst_p + msg_offset;
1173	cmdsp->data_len = msg_len;
1174	cmdsp++;
1175
1176skip_p:
1177	if (!dst_q)
1178		goto skip_q;
1179
1180	/* Type-A command to zero all buffers */
1181	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1182			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1183	cmd = sba_cmd_enc(cmd, msg_len,
1184			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1185	cmd = sba_cmd_enc(cmd, SBA_CMD_ZERO_ALL_BUFFERS,
1186			  SBA_CMD_SHIFT, SBA_CMD_MASK);
1187	cmdsp->cmd = cmd;
1188	*cmdsp->cmd_dma = cpu_to_le64(cmd);
1189	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1190	cmdsp++;
1191
1192	if (dpos == 255)
1193		goto skip_q_computation;
1194	pos = (dpos < req->sba->max_pq_coefs) ?
1195		dpos : (req->sba->max_pq_coefs - 1);
1196
1197	/*
1198	 * Type-B command to generate initial Q from data
1199	 * and store output into buf0
1200	 */
1201	cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1202			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1203	cmd = sba_cmd_enc(cmd, msg_len,
1204			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1205	c_mdata = sba_cmd_pq_c_mdata(pos, 0, 0);
1206	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1207			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1208	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_MS(c_mdata),
1209			  SBA_C_MDATA_MS_SHIFT, SBA_C_MDATA_MS_MASK);
1210	cmd = sba_cmd_enc(cmd, SBA_CMD_GALOIS,
1211			  SBA_CMD_SHIFT, SBA_CMD_MASK);
1212	cmdsp->cmd = cmd;
1213	*cmdsp->cmd_dma = cpu_to_le64(cmd);
1214	cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1215	cmdsp->data = src + msg_offset;
1216	cmdsp->data_len = msg_len;
1217	cmdsp++;
1218
1219	dpos -= pos;
1220
1221	/* Multiple Type-A command to generate final Q */
1222	while (dpos) {
1223		pos = (dpos < req->sba->max_pq_coefs) ?
1224			dpos : (req->sba->max_pq_coefs - 1);
1225
1226		/*
1227		 * Type-A command to generate Q with buf0 and
1228		 * buf1 store result in buf0
1229		 */
1230		cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1231				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1232		cmd = sba_cmd_enc(cmd, msg_len,
1233				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1234		c_mdata = sba_cmd_pq_c_mdata(pos, 0, 1);
1235		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1236				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1237		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_MS(c_mdata),
1238				  SBA_C_MDATA_MS_SHIFT, SBA_C_MDATA_MS_MASK);
1239		cmd = sba_cmd_enc(cmd, SBA_CMD_GALOIS,
1240				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1241		cmdsp->cmd = cmd;
1242		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1243		cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1244		cmdsp++;
1245
1246		dpos -= pos;
1247	}
1248
1249skip_q_computation:
1250	if (pq_continue) {
1251		/*
1252		 * Type-B command to XOR previous output with
1253		 * buf0 and write it into buf0
1254		 */
1255		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1256				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1257		cmd = sba_cmd_enc(cmd, msg_len,
1258				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1259		c_mdata = sba_cmd_xor_c_mdata(0, 0);
1260		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1261				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1262		cmd = sba_cmd_enc(cmd, SBA_CMD_XOR,
1263				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1264		cmdsp->cmd = cmd;
1265		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1266		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1267		cmdsp->data = *dst_q + msg_offset;
1268		cmdsp->data_len = msg_len;
1269		cmdsp++;
1270	}
1271
1272	/* Type-A command to write buf0 */
1273	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1274			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1275	cmd = sba_cmd_enc(cmd, msg_len,
1276			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1277	cmd = sba_cmd_enc(cmd, 0x1,
1278			  SBA_RESP_SHIFT, SBA_RESP_MASK);
1279	c_mdata = sba_cmd_write_c_mdata(0);
1280	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1281			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1282	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
1283			  SBA_CMD_SHIFT, SBA_CMD_MASK);
1284	cmdsp->cmd = cmd;
1285	*cmdsp->cmd_dma = cpu_to_le64(cmd);
1286	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1287	if (req->sba->hw_resp_size) {
1288		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
1289		cmdsp->resp = resp_dma;
1290		cmdsp->resp_len = req->sba->hw_resp_size;
1291	}
1292	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
1293	cmdsp->data = *dst_q + msg_offset;
1294	cmdsp->data_len = msg_len;
1295	cmdsp++;
1296
1297skip_q:
1298	/* Fillup brcm_message */
1299	msg->type = BRCM_MESSAGE_SBA;
1300	msg->sba.cmds = cmds;
1301	msg->sba.cmds_count = cmdsp - cmds;
1302	msg->ctx = req;
1303	msg->error = 0;
1304}
1305
1306static struct sba_request *
1307sba_prep_dma_pq_single_req(struct sba_device *sba, dma_addr_t off,
1308			   dma_addr_t *dst_p, dma_addr_t *dst_q,
1309			   dma_addr_t src, u8 scf, size_t len,
1310			   unsigned long flags)
1311{
1312	struct sba_request *req = NULL;
1313
1314	/* Alloc new request */
1315	req = sba_alloc_request(sba);
1316	if (!req)
1317		return NULL;
1318	if (flags & DMA_PREP_FENCE)
1319		req->flags |= SBA_REQUEST_FENCE;
1320
1321	/* Fillup request messages */
1322	sba_fillup_pq_single_msg(req,  dmaf_continue(flags),
1323				 req->cmds, &req->msg, off, len,
1324				 dst_p, dst_q, src, scf);
1325
1326	/* Init async_tx descriptor */
1327	req->tx.flags = flags;
1328	req->tx.cookie = -EBUSY;
1329
1330	return req;
1331}
1332
1333static struct dma_async_tx_descriptor *
1334sba_prep_dma_pq(struct dma_chan *dchan, dma_addr_t *dst, dma_addr_t *src,
1335		u32 src_cnt, const u8 *scf, size_t len, unsigned long flags)
1336{
1337	u32 i, dst_q_index;
1338	size_t req_len;
1339	bool slow = false;
1340	dma_addr_t off = 0;
1341	dma_addr_t *dst_p = NULL, *dst_q = NULL;
1342	struct sba_device *sba = to_sba_device(dchan);
1343	struct sba_request *first = NULL, *req;
1344
1345	/* Sanity checks */
1346	if (unlikely(src_cnt > sba->max_pq_srcs))
1347		return NULL;
1348	for (i = 0; i < src_cnt; i++)
1349		if (sba->max_pq_coefs <= raid6_gflog[scf[i]])
1350			slow = true;
1351
1352	/* Figure-out P and Q destination addresses */
1353	if (!(flags & DMA_PREP_PQ_DISABLE_P))
1354		dst_p = &dst[0];
1355	if (!(flags & DMA_PREP_PQ_DISABLE_Q))
1356		dst_q = &dst[1];
1357
1358	/* Create chained requests where each request is upto hw_buf_size */
1359	while (len) {
1360		req_len = (len < sba->hw_buf_size) ? len : sba->hw_buf_size;
1361
1362		if (slow) {
1363			dst_q_index = src_cnt;
1364
1365			if (dst_q) {
1366				for (i = 0; i < src_cnt; i++) {
1367					if (*dst_q == src[i]) {
1368						dst_q_index = i;
1369						break;
1370					}
1371				}
1372			}
1373
1374			if (dst_q_index < src_cnt) {
1375				i = dst_q_index;
1376				req = sba_prep_dma_pq_single_req(sba,
1377					off, dst_p, dst_q, src[i], scf[i],
1378					req_len, flags | DMA_PREP_FENCE);
1379				if (!req)
1380					goto fail;
1381
1382				if (first)
1383					sba_chain_request(first, req);
1384				else
1385					first = req;
1386
1387				flags |= DMA_PREP_CONTINUE;
1388			}
1389
1390			for (i = 0; i < src_cnt; i++) {
1391				if (dst_q_index == i)
1392					continue;
1393
1394				req = sba_prep_dma_pq_single_req(sba,
1395					off, dst_p, dst_q, src[i], scf[i],
1396					req_len, flags | DMA_PREP_FENCE);
1397				if (!req)
1398					goto fail;
1399
1400				if (first)
1401					sba_chain_request(first, req);
1402				else
1403					first = req;
1404
1405				flags |= DMA_PREP_CONTINUE;
1406			}
1407		} else {
1408			req = sba_prep_dma_pq_req(sba, off,
1409						  dst_p, dst_q, src, src_cnt,
1410						  scf, req_len, flags);
1411			if (!req)
1412				goto fail;
1413
1414			if (first)
1415				sba_chain_request(first, req);
1416			else
1417				first = req;
1418		}
1419
1420		off += req_len;
1421		len -= req_len;
1422	}
1423
1424	return (first) ? &first->tx : NULL;
1425
1426fail:
1427	if (first)
1428		sba_free_chained_requests(first);
1429	return NULL;
1430}
1431
1432/* ====== Mailbox callbacks ===== */
1433
1434static void sba_receive_message(struct mbox_client *cl, void *msg)
1435{
1436	struct brcm_message *m = msg;
1437	struct sba_request *req = m->ctx;
1438	struct sba_device *sba = req->sba;
1439
1440	/* Error count if message has error */
1441	if (m->error < 0)
1442		dev_err(sba->dev, "%s got message with error %d",
1443			dma_chan_name(&sba->dma_chan), m->error);
1444
1445	/* Process received request */
1446	sba_process_received_request(sba, req);
1447}
1448
1449/* ====== Debugfs callbacks ====== */
1450
1451static int sba_debugfs_stats_show(struct seq_file *file, void *offset)
1452{
1453	struct sba_device *sba = dev_get_drvdata(file->private);
 
1454
1455	/* Write stats in file */
1456	sba_write_stats_in_seqfile(sba, file);
1457
1458	return 0;
1459}
1460
1461/* ====== Platform driver routines ===== */
1462
1463static int sba_prealloc_channel_resources(struct sba_device *sba)
1464{
1465	int i, j, ret = 0;
1466	struct sba_request *req = NULL;
1467
1468	sba->resp_base = dma_alloc_coherent(sba->mbox_dev,
1469					    sba->max_resp_pool_size,
1470					    &sba->resp_dma_base, GFP_KERNEL);
1471	if (!sba->resp_base)
1472		return -ENOMEM;
1473
1474	sba->cmds_base = dma_alloc_coherent(sba->mbox_dev,
1475					    sba->max_cmds_pool_size,
1476					    &sba->cmds_dma_base, GFP_KERNEL);
1477	if (!sba->cmds_base) {
1478		ret = -ENOMEM;
1479		goto fail_free_resp_pool;
1480	}
1481
1482	spin_lock_init(&sba->reqs_lock);
1483	sba->reqs_fence = false;
1484	INIT_LIST_HEAD(&sba->reqs_alloc_list);
1485	INIT_LIST_HEAD(&sba->reqs_pending_list);
1486	INIT_LIST_HEAD(&sba->reqs_active_list);
1487	INIT_LIST_HEAD(&sba->reqs_aborted_list);
1488	INIT_LIST_HEAD(&sba->reqs_free_list);
1489
1490	for (i = 0; i < sba->max_req; i++) {
1491		req = devm_kzalloc(sba->dev,
1492				   struct_size(req, cmds, sba->max_cmd_per_req),
1493				   GFP_KERNEL);
 
1494		if (!req) {
1495			ret = -ENOMEM;
1496			goto fail_free_cmds_pool;
1497		}
1498		INIT_LIST_HEAD(&req->node);
1499		req->sba = sba;
1500		req->flags = SBA_REQUEST_STATE_FREE;
1501		INIT_LIST_HEAD(&req->next);
1502		atomic_set(&req->next_pending_count, 0);
1503		for (j = 0; j < sba->max_cmd_per_req; j++) {
1504			req->cmds[j].cmd = 0;
1505			req->cmds[j].cmd_dma = sba->cmds_base +
1506				(i * sba->max_cmd_per_req + j) * sizeof(u64);
1507			req->cmds[j].cmd_dma_addr = sba->cmds_dma_base +
1508				(i * sba->max_cmd_per_req + j) * sizeof(u64);
1509			req->cmds[j].flags = 0;
1510		}
1511		memset(&req->msg, 0, sizeof(req->msg));
1512		dma_async_tx_descriptor_init(&req->tx, &sba->dma_chan);
1513		async_tx_ack(&req->tx);
1514		req->tx.tx_submit = sba_tx_submit;
1515		req->tx.phys = sba->resp_dma_base + i * sba->hw_resp_size;
1516		list_add_tail(&req->node, &sba->reqs_free_list);
1517	}
1518
1519	return 0;
1520
1521fail_free_cmds_pool:
1522	dma_free_coherent(sba->mbox_dev,
1523			  sba->max_cmds_pool_size,
1524			  sba->cmds_base, sba->cmds_dma_base);
1525fail_free_resp_pool:
1526	dma_free_coherent(sba->mbox_dev,
1527			  sba->max_resp_pool_size,
1528			  sba->resp_base, sba->resp_dma_base);
1529	return ret;
1530}
1531
1532static void sba_freeup_channel_resources(struct sba_device *sba)
1533{
1534	dmaengine_terminate_all(&sba->dma_chan);
1535	dma_free_coherent(sba->mbox_dev, sba->max_cmds_pool_size,
1536			  sba->cmds_base, sba->cmds_dma_base);
1537	dma_free_coherent(sba->mbox_dev, sba->max_resp_pool_size,
1538			  sba->resp_base, sba->resp_dma_base);
1539	sba->resp_base = NULL;
1540	sba->resp_dma_base = 0;
1541}
1542
1543static int sba_async_register(struct sba_device *sba)
1544{
1545	int ret;
1546	struct dma_device *dma_dev = &sba->dma_dev;
1547
1548	/* Initialize DMA channel cookie */
1549	sba->dma_chan.device = dma_dev;
1550	dma_cookie_init(&sba->dma_chan);
1551
1552	/* Initialize DMA device capability mask */
1553	dma_cap_zero(dma_dev->cap_mask);
1554	dma_cap_set(DMA_INTERRUPT, dma_dev->cap_mask);
1555	dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
1556	dma_cap_set(DMA_XOR, dma_dev->cap_mask);
1557	dma_cap_set(DMA_PQ, dma_dev->cap_mask);
1558
1559	/*
1560	 * Set mailbox channel device as the base device of
1561	 * our dma_device because the actual memory accesses
1562	 * will be done by mailbox controller
1563	 */
1564	dma_dev->dev = sba->mbox_dev;
1565
1566	/* Set base prep routines */
1567	dma_dev->device_free_chan_resources = sba_free_chan_resources;
1568	dma_dev->device_terminate_all = sba_device_terminate_all;
1569	dma_dev->device_issue_pending = sba_issue_pending;
1570	dma_dev->device_tx_status = sba_tx_status;
1571
1572	/* Set interrupt routine */
1573	if (dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask))
1574		dma_dev->device_prep_dma_interrupt = sba_prep_dma_interrupt;
1575
1576	/* Set memcpy routine */
1577	if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask))
1578		dma_dev->device_prep_dma_memcpy = sba_prep_dma_memcpy;
1579
1580	/* Set xor routine and capability */
1581	if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1582		dma_dev->device_prep_dma_xor = sba_prep_dma_xor;
1583		dma_dev->max_xor = sba->max_xor_srcs;
1584	}
1585
1586	/* Set pq routine and capability */
1587	if (dma_has_cap(DMA_PQ, dma_dev->cap_mask)) {
1588		dma_dev->device_prep_dma_pq = sba_prep_dma_pq;
1589		dma_set_maxpq(dma_dev, sba->max_pq_srcs, 0);
1590	}
1591
1592	/* Initialize DMA device channel list */
1593	INIT_LIST_HEAD(&dma_dev->channels);
1594	list_add_tail(&sba->dma_chan.device_node, &dma_dev->channels);
1595
1596	/* Register with Linux async DMA framework*/
1597	ret = dma_async_device_register(dma_dev);
1598	if (ret) {
1599		dev_err(sba->dev, "async device register error %d", ret);
1600		return ret;
1601	}
1602
1603	dev_info(sba->dev, "%s capabilities: %s%s%s%s\n",
1604	dma_chan_name(&sba->dma_chan),
1605	dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask) ? "interrupt " : "",
1606	dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask) ? "memcpy " : "",
1607	dma_has_cap(DMA_XOR, dma_dev->cap_mask) ? "xor " : "",
1608	dma_has_cap(DMA_PQ, dma_dev->cap_mask) ? "pq " : "");
1609
1610	return 0;
1611}
1612
1613static int sba_probe(struct platform_device *pdev)
1614{
1615	int ret = 0;
1616	struct sba_device *sba;
1617	struct platform_device *mbox_pdev;
1618	struct of_phandle_args args;
1619
1620	/* Allocate main SBA struct */
1621	sba = devm_kzalloc(&pdev->dev, sizeof(*sba), GFP_KERNEL);
1622	if (!sba)
1623		return -ENOMEM;
1624
1625	sba->dev = &pdev->dev;
1626	platform_set_drvdata(pdev, sba);
1627
1628	/* Number of mailbox channels should be atleast 1 */
1629	ret = of_count_phandle_with_args(pdev->dev.of_node,
1630					 "mboxes", "#mbox-cells");
1631	if (ret <= 0)
1632		return -ENODEV;
1633
1634	/* Determine SBA version from DT compatible string */
1635	if (of_device_is_compatible(sba->dev->of_node, "brcm,iproc-sba"))
1636		sba->ver = SBA_VER_1;
1637	else if (of_device_is_compatible(sba->dev->of_node,
1638					 "brcm,iproc-sba-v2"))
1639		sba->ver = SBA_VER_2;
1640	else
1641		return -ENODEV;
1642
1643	/* Derived Configuration parameters */
1644	switch (sba->ver) {
1645	case SBA_VER_1:
1646		sba->hw_buf_size = 4096;
1647		sba->hw_resp_size = 8;
1648		sba->max_pq_coefs = 6;
1649		sba->max_pq_srcs = 6;
1650		break;
1651	case SBA_VER_2:
1652		sba->hw_buf_size = 4096;
1653		sba->hw_resp_size = 8;
1654		sba->max_pq_coefs = 30;
1655		/*
1656		 * We can support max_pq_srcs == max_pq_coefs because
1657		 * we are limited by number of SBA commands that we can
1658		 * fit in one message for underlying ring manager HW.
1659		 */
1660		sba->max_pq_srcs = 12;
1661		break;
1662	default:
1663		return -EINVAL;
1664	}
1665	sba->max_req = SBA_MAX_REQ_PER_MBOX_CHANNEL;
1666	sba->max_cmd_per_req = sba->max_pq_srcs + 3;
1667	sba->max_xor_srcs = sba->max_cmd_per_req - 1;
1668	sba->max_resp_pool_size = sba->max_req * sba->hw_resp_size;
1669	sba->max_cmds_pool_size = sba->max_req *
1670				  sba->max_cmd_per_req * sizeof(u64);
1671
1672	/* Setup mailbox client */
1673	sba->client.dev			= &pdev->dev;
1674	sba->client.rx_callback		= sba_receive_message;
1675	sba->client.tx_block		= false;
1676	sba->client.knows_txdone	= true;
1677	sba->client.tx_tout		= 0;
1678
1679	/* Request mailbox channel */
1680	sba->mchan = mbox_request_channel(&sba->client, 0);
1681	if (IS_ERR(sba->mchan)) {
1682		ret = PTR_ERR(sba->mchan);
1683		goto fail_free_mchan;
1684	}
1685
1686	/* Find-out underlying mailbox device */
1687	ret = of_parse_phandle_with_args(pdev->dev.of_node,
1688					 "mboxes", "#mbox-cells", 0, &args);
1689	if (ret)
1690		goto fail_free_mchan;
1691	mbox_pdev = of_find_device_by_node(args.np);
1692	of_node_put(args.np);
1693	if (!mbox_pdev) {
1694		ret = -ENODEV;
1695		goto fail_free_mchan;
1696	}
1697	sba->mbox_dev = &mbox_pdev->dev;
1698
1699	/* Prealloc channel resource */
1700	ret = sba_prealloc_channel_resources(sba);
1701	if (ret)
1702		goto fail_free_mchan;
1703
1704	/* Check availability of debugfs */
1705	if (!debugfs_initialized())
1706		goto skip_debugfs;
1707
1708	/* Create debugfs root entry */
1709	sba->root = debugfs_create_dir(dev_name(sba->dev), NULL);
 
 
 
 
 
1710
1711	/* Create debugfs stats entry */
1712	debugfs_create_devm_seqfile(sba->dev, "stats", sba->root,
1713				    sba_debugfs_stats_show);
1714
 
1715skip_debugfs:
1716
1717	/* Register DMA device with Linux async framework */
1718	ret = sba_async_register(sba);
1719	if (ret)
1720		goto fail_free_resources;
1721
1722	/* Print device info */
1723	dev_info(sba->dev, "%s using SBAv%d mailbox channel from %s",
1724		 dma_chan_name(&sba->dma_chan), sba->ver+1,
1725		 dev_name(sba->mbox_dev));
1726
1727	return 0;
1728
1729fail_free_resources:
1730	debugfs_remove_recursive(sba->root);
1731	sba_freeup_channel_resources(sba);
1732fail_free_mchan:
1733	mbox_free_channel(sba->mchan);
1734	return ret;
1735}
1736
1737static void sba_remove(struct platform_device *pdev)
1738{
1739	struct sba_device *sba = platform_get_drvdata(pdev);
1740
1741	dma_async_device_unregister(&sba->dma_dev);
1742
1743	debugfs_remove_recursive(sba->root);
1744
1745	sba_freeup_channel_resources(sba);
1746
1747	mbox_free_channel(sba->mchan);
 
 
1748}
1749
1750static const struct of_device_id sba_of_match[] = {
1751	{ .compatible = "brcm,iproc-sba", },
1752	{ .compatible = "brcm,iproc-sba-v2", },
1753	{},
1754};
1755MODULE_DEVICE_TABLE(of, sba_of_match);
1756
1757static struct platform_driver sba_driver = {
1758	.probe = sba_probe,
1759	.remove_new = sba_remove,
1760	.driver = {
1761		.name = "bcm-sba-raid",
1762		.of_match_table = sba_of_match,
1763	},
1764};
1765module_platform_driver(sba_driver);
1766
1767MODULE_DESCRIPTION("Broadcom SBA RAID driver");
1768MODULE_AUTHOR("Anup Patel <anup.patel@broadcom.com>");
1769MODULE_LICENSE("GPL v2");
v4.17
   1/*
   2 * Copyright (C) 2017 Broadcom
   3 *
   4 * This program is free software; you can redistribute it and/or
   5 * modify it under the terms of the GNU General Public License as
   6 * published by the Free Software Foundation version 2.
   7 *
   8 * This program is distributed "as is" WITHOUT ANY WARRANTY of any
   9 * kind, whether express or implied; without even the implied warranty
  10 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  11 * GNU General Public License for more details.
  12 */
  13
  14/*
  15 * Broadcom SBA RAID Driver
  16 *
  17 * The Broadcom stream buffer accelerator (SBA) provides offloading
  18 * capabilities for RAID operations. The SBA offload engine is accessible
  19 * via Broadcom SoC specific ring manager. Two or more offload engines
  20 * can share same Broadcom SoC specific ring manager due to this Broadcom
  21 * SoC specific ring manager driver is implemented as a mailbox controller
  22 * driver and offload engine drivers are implemented as mallbox clients.
  23 *
  24 * Typically, Broadcom SoC specific ring manager will implement larger
  25 * number of hardware rings over one or more SBA hardware devices. By
  26 * design, the internal buffer size of SBA hardware device is limited
  27 * but all offload operations supported by SBA can be broken down into
  28 * multiple small size requests and executed parallely on multiple SBA
  29 * hardware devices for achieving high through-put.
  30 *
  31 * The Broadcom SBA RAID driver does not require any register programming
  32 * except submitting request to SBA hardware device via mailbox channels.
  33 * This driver implements a DMA device with one DMA channel using a single
  34 * mailbox channel provided by Broadcom SoC specific ring manager driver.
  35 * For having more SBA DMA channels, we can create more SBA device nodes
  36 * in Broadcom SoC specific DTS based on number of hardware rings supported
  37 * by Broadcom SoC ring manager.
  38 */
  39
  40#include <linux/bitops.h>
  41#include <linux/debugfs.h>
  42#include <linux/dma-mapping.h>
  43#include <linux/dmaengine.h>
  44#include <linux/list.h>
  45#include <linux/mailbox_client.h>
  46#include <linux/mailbox/brcm-message.h>
  47#include <linux/module.h>
  48#include <linux/of_device.h>
 
 
  49#include <linux/slab.h>
  50#include <linux/raid/pq.h>
  51
  52#include "dmaengine.h"
  53
  54/* ====== Driver macros and defines ===== */
  55
  56#define SBA_TYPE_SHIFT					48
  57#define SBA_TYPE_MASK					GENMASK(1, 0)
  58#define SBA_TYPE_A					0x0
  59#define SBA_TYPE_B					0x2
  60#define SBA_TYPE_C					0x3
  61#define SBA_USER_DEF_SHIFT				32
  62#define SBA_USER_DEF_MASK				GENMASK(15, 0)
  63#define SBA_R_MDATA_SHIFT				24
  64#define SBA_R_MDATA_MASK				GENMASK(7, 0)
  65#define SBA_C_MDATA_MS_SHIFT				18
  66#define SBA_C_MDATA_MS_MASK				GENMASK(1, 0)
  67#define SBA_INT_SHIFT					17
  68#define SBA_INT_MASK					BIT(0)
  69#define SBA_RESP_SHIFT					16
  70#define SBA_RESP_MASK					BIT(0)
  71#define SBA_C_MDATA_SHIFT				8
  72#define SBA_C_MDATA_MASK				GENMASK(7, 0)
  73#define SBA_C_MDATA_BNUMx_SHIFT(__bnum)			(2 * (__bnum))
  74#define SBA_C_MDATA_BNUMx_MASK				GENMASK(1, 0)
  75#define SBA_C_MDATA_DNUM_SHIFT				5
  76#define SBA_C_MDATA_DNUM_MASK				GENMASK(4, 0)
  77#define SBA_C_MDATA_LS(__v)				((__v) & 0xff)
  78#define SBA_C_MDATA_MS(__v)				(((__v) >> 8) & 0x3)
  79#define SBA_CMD_SHIFT					0
  80#define SBA_CMD_MASK					GENMASK(3, 0)
  81#define SBA_CMD_ZERO_BUFFER				0x4
  82#define SBA_CMD_ZERO_ALL_BUFFERS			0x8
  83#define SBA_CMD_LOAD_BUFFER				0x9
  84#define SBA_CMD_XOR					0xa
  85#define SBA_CMD_GALOIS_XOR				0xb
  86#define SBA_CMD_WRITE_BUFFER				0xc
  87#define SBA_CMD_GALOIS					0xe
  88
  89#define SBA_MAX_REQ_PER_MBOX_CHANNEL			8192
  90#define SBA_MAX_MSG_SEND_PER_MBOX_CHANNEL		8
  91
  92/* Driver helper macros */
  93#define to_sba_request(tx)		\
  94	container_of(tx, struct sba_request, tx)
  95#define to_sba_device(dchan)		\
  96	container_of(dchan, struct sba_device, dma_chan)
  97
  98/* ===== Driver data structures ===== */
  99
 100enum sba_request_flags {
 101	SBA_REQUEST_STATE_FREE		= 0x001,
 102	SBA_REQUEST_STATE_ALLOCED	= 0x002,
 103	SBA_REQUEST_STATE_PENDING	= 0x004,
 104	SBA_REQUEST_STATE_ACTIVE	= 0x008,
 105	SBA_REQUEST_STATE_ABORTED	= 0x010,
 106	SBA_REQUEST_STATE_MASK		= 0x0ff,
 107	SBA_REQUEST_FENCE		= 0x100,
 108};
 109
 110struct sba_request {
 111	/* Global state */
 112	struct list_head node;
 113	struct sba_device *sba;
 114	u32 flags;
 115	/* Chained requests management */
 116	struct sba_request *first;
 117	struct list_head next;
 118	atomic_t next_pending_count;
 119	/* BRCM message data */
 120	struct brcm_message msg;
 121	struct dma_async_tx_descriptor tx;
 122	/* SBA commands */
 123	struct brcm_sba_command cmds[0];
 124};
 125
 126enum sba_version {
 127	SBA_VER_1 = 0,
 128	SBA_VER_2
 129};
 130
 131struct sba_device {
 132	/* Underlying device */
 133	struct device *dev;
 134	/* DT configuration parameters */
 135	enum sba_version ver;
 136	/* Derived configuration parameters */
 137	u32 max_req;
 138	u32 hw_buf_size;
 139	u32 hw_resp_size;
 140	u32 max_pq_coefs;
 141	u32 max_pq_srcs;
 142	u32 max_cmd_per_req;
 143	u32 max_xor_srcs;
 144	u32 max_resp_pool_size;
 145	u32 max_cmds_pool_size;
 146	/* Maibox client and Mailbox channels */
 147	struct mbox_client client;
 148	struct mbox_chan *mchan;
 149	struct device *mbox_dev;
 150	/* DMA device and DMA channel */
 151	struct dma_device dma_dev;
 152	struct dma_chan dma_chan;
 153	/* DMA channel resources */
 154	void *resp_base;
 155	dma_addr_t resp_dma_base;
 156	void *cmds_base;
 157	dma_addr_t cmds_dma_base;
 158	spinlock_t reqs_lock;
 159	bool reqs_fence;
 160	struct list_head reqs_alloc_list;
 161	struct list_head reqs_pending_list;
 162	struct list_head reqs_active_list;
 163	struct list_head reqs_aborted_list;
 164	struct list_head reqs_free_list;
 165	/* DebugFS directory entries */
 166	struct dentry *root;
 167	struct dentry *stats;
 168};
 169
 170/* ====== Command helper routines ===== */
 171
 172static inline u64 __pure sba_cmd_enc(u64 cmd, u32 val, u32 shift, u32 mask)
 173{
 174	cmd &= ~((u64)mask << shift);
 175	cmd |= ((u64)(val & mask) << shift);
 176	return cmd;
 177}
 178
 179static inline u32 __pure sba_cmd_load_c_mdata(u32 b0)
 180{
 181	return b0 & SBA_C_MDATA_BNUMx_MASK;
 182}
 183
 184static inline u32 __pure sba_cmd_write_c_mdata(u32 b0)
 185{
 186	return b0 & SBA_C_MDATA_BNUMx_MASK;
 187}
 188
 189static inline u32 __pure sba_cmd_xor_c_mdata(u32 b1, u32 b0)
 190{
 191	return (b0 & SBA_C_MDATA_BNUMx_MASK) |
 192	       ((b1 & SBA_C_MDATA_BNUMx_MASK) << SBA_C_MDATA_BNUMx_SHIFT(1));
 193}
 194
 195static inline u32 __pure sba_cmd_pq_c_mdata(u32 d, u32 b1, u32 b0)
 196{
 197	return (b0 & SBA_C_MDATA_BNUMx_MASK) |
 198	       ((b1 & SBA_C_MDATA_BNUMx_MASK) << SBA_C_MDATA_BNUMx_SHIFT(1)) |
 199	       ((d & SBA_C_MDATA_DNUM_MASK) << SBA_C_MDATA_DNUM_SHIFT);
 200}
 201
 202/* ====== General helper routines ===== */
 203
 204static struct sba_request *sba_alloc_request(struct sba_device *sba)
 205{
 206	bool found = false;
 207	unsigned long flags;
 208	struct sba_request *req = NULL;
 209
 210	spin_lock_irqsave(&sba->reqs_lock, flags);
 211	list_for_each_entry(req, &sba->reqs_free_list, node) {
 212		if (async_tx_test_ack(&req->tx)) {
 213			list_move_tail(&req->node, &sba->reqs_alloc_list);
 214			found = true;
 215			break;
 216		}
 217	}
 218	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 219
 220	if (!found) {
 221		/*
 222		 * We have no more free requests so, we peek
 223		 * mailbox channels hoping few active requests
 224		 * would have completed which will create more
 225		 * room for new requests.
 226		 */
 227		mbox_client_peek_data(sba->mchan);
 228		return NULL;
 229	}
 230
 231	req->flags = SBA_REQUEST_STATE_ALLOCED;
 232	req->first = req;
 233	INIT_LIST_HEAD(&req->next);
 234	atomic_set(&req->next_pending_count, 1);
 235
 236	dma_async_tx_descriptor_init(&req->tx, &sba->dma_chan);
 237	async_tx_ack(&req->tx);
 238
 239	return req;
 240}
 241
 242/* Note: Must be called with sba->reqs_lock held */
 243static void _sba_pending_request(struct sba_device *sba,
 244				 struct sba_request *req)
 245{
 246	lockdep_assert_held(&sba->reqs_lock);
 247	req->flags &= ~SBA_REQUEST_STATE_MASK;
 248	req->flags |= SBA_REQUEST_STATE_PENDING;
 249	list_move_tail(&req->node, &sba->reqs_pending_list);
 250	if (list_empty(&sba->reqs_active_list))
 251		sba->reqs_fence = false;
 252}
 253
 254/* Note: Must be called with sba->reqs_lock held */
 255static bool _sba_active_request(struct sba_device *sba,
 256				struct sba_request *req)
 257{
 258	lockdep_assert_held(&sba->reqs_lock);
 259	if (list_empty(&sba->reqs_active_list))
 260		sba->reqs_fence = false;
 261	if (sba->reqs_fence)
 262		return false;
 263	req->flags &= ~SBA_REQUEST_STATE_MASK;
 264	req->flags |= SBA_REQUEST_STATE_ACTIVE;
 265	list_move_tail(&req->node, &sba->reqs_active_list);
 266	if (req->flags & SBA_REQUEST_FENCE)
 267		sba->reqs_fence = true;
 268	return true;
 269}
 270
 271/* Note: Must be called with sba->reqs_lock held */
 272static void _sba_abort_request(struct sba_device *sba,
 273			       struct sba_request *req)
 274{
 275	lockdep_assert_held(&sba->reqs_lock);
 276	req->flags &= ~SBA_REQUEST_STATE_MASK;
 277	req->flags |= SBA_REQUEST_STATE_ABORTED;
 278	list_move_tail(&req->node, &sba->reqs_aborted_list);
 279	if (list_empty(&sba->reqs_active_list))
 280		sba->reqs_fence = false;
 281}
 282
 283/* Note: Must be called with sba->reqs_lock held */
 284static void _sba_free_request(struct sba_device *sba,
 285			      struct sba_request *req)
 286{
 287	lockdep_assert_held(&sba->reqs_lock);
 288	req->flags &= ~SBA_REQUEST_STATE_MASK;
 289	req->flags |= SBA_REQUEST_STATE_FREE;
 290	list_move_tail(&req->node, &sba->reqs_free_list);
 291	if (list_empty(&sba->reqs_active_list))
 292		sba->reqs_fence = false;
 293}
 294
 295static void sba_free_chained_requests(struct sba_request *req)
 296{
 297	unsigned long flags;
 298	struct sba_request *nreq;
 299	struct sba_device *sba = req->sba;
 300
 301	spin_lock_irqsave(&sba->reqs_lock, flags);
 302
 303	_sba_free_request(sba, req);
 304	list_for_each_entry(nreq, &req->next, next)
 305		_sba_free_request(sba, nreq);
 306
 307	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 308}
 309
 310static void sba_chain_request(struct sba_request *first,
 311			      struct sba_request *req)
 312{
 313	unsigned long flags;
 314	struct sba_device *sba = req->sba;
 315
 316	spin_lock_irqsave(&sba->reqs_lock, flags);
 317
 318	list_add_tail(&req->next, &first->next);
 319	req->first = first;
 320	atomic_inc(&first->next_pending_count);
 321
 322	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 323}
 324
 325static void sba_cleanup_nonpending_requests(struct sba_device *sba)
 326{
 327	unsigned long flags;
 328	struct sba_request *req, *req1;
 329
 330	spin_lock_irqsave(&sba->reqs_lock, flags);
 331
 332	/* Freeup all alloced request */
 333	list_for_each_entry_safe(req, req1, &sba->reqs_alloc_list, node)
 334		_sba_free_request(sba, req);
 335
 336	/* Set all active requests as aborted */
 337	list_for_each_entry_safe(req, req1, &sba->reqs_active_list, node)
 338		_sba_abort_request(sba, req);
 339
 340	/*
 341	 * Note: We expect that aborted request will be eventually
 342	 * freed by sba_receive_message()
 343	 */
 344
 345	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 346}
 347
 348static void sba_cleanup_pending_requests(struct sba_device *sba)
 349{
 350	unsigned long flags;
 351	struct sba_request *req, *req1;
 352
 353	spin_lock_irqsave(&sba->reqs_lock, flags);
 354
 355	/* Freeup all pending request */
 356	list_for_each_entry_safe(req, req1, &sba->reqs_pending_list, node)
 357		_sba_free_request(sba, req);
 358
 359	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 360}
 361
 362static int sba_send_mbox_request(struct sba_device *sba,
 363				 struct sba_request *req)
 364{
 365	int ret = 0;
 366
 367	/* Send message for the request */
 368	req->msg.error = 0;
 369	ret = mbox_send_message(sba->mchan, &req->msg);
 370	if (ret < 0) {
 371		dev_err(sba->dev, "send message failed with error %d", ret);
 372		return ret;
 373	}
 374
 375	/* Check error returned by mailbox controller */
 376	ret = req->msg.error;
 377	if (ret < 0) {
 378		dev_err(sba->dev, "message error %d", ret);
 379	}
 380
 381	/* Signal txdone for mailbox channel */
 382	mbox_client_txdone(sba->mchan, ret);
 383
 384	return ret;
 385}
 386
 387/* Note: Must be called with sba->reqs_lock held */
 388static void _sba_process_pending_requests(struct sba_device *sba)
 389{
 390	int ret;
 391	u32 count;
 392	struct sba_request *req;
 393
 394	/* Process few pending requests */
 395	count = SBA_MAX_MSG_SEND_PER_MBOX_CHANNEL;
 396	while (!list_empty(&sba->reqs_pending_list) && count) {
 397		/* Get the first pending request */
 398		req = list_first_entry(&sba->reqs_pending_list,
 399				       struct sba_request, node);
 400
 401		/* Try to make request active */
 402		if (!_sba_active_request(sba, req))
 403			break;
 404
 405		/* Send request to mailbox channel */
 406		ret = sba_send_mbox_request(sba, req);
 407		if (ret < 0) {
 408			_sba_pending_request(sba, req);
 409			break;
 410		}
 411
 412		count--;
 413	}
 414}
 415
 416static void sba_process_received_request(struct sba_device *sba,
 417					 struct sba_request *req)
 418{
 419	unsigned long flags;
 420	struct dma_async_tx_descriptor *tx;
 421	struct sba_request *nreq, *first = req->first;
 422
 423	/* Process only after all chained requests are received */
 424	if (!atomic_dec_return(&first->next_pending_count)) {
 425		tx = &first->tx;
 426
 427		WARN_ON(tx->cookie < 0);
 428		if (tx->cookie > 0) {
 429			spin_lock_irqsave(&sba->reqs_lock, flags);
 430			dma_cookie_complete(tx);
 431			spin_unlock_irqrestore(&sba->reqs_lock, flags);
 432			dmaengine_desc_get_callback_invoke(tx, NULL);
 433			dma_descriptor_unmap(tx);
 434			tx->callback = NULL;
 435			tx->callback_result = NULL;
 436		}
 437
 438		dma_run_dependencies(tx);
 439
 440		spin_lock_irqsave(&sba->reqs_lock, flags);
 441
 442		/* Free all requests chained to first request */
 443		list_for_each_entry(nreq, &first->next, next)
 444			_sba_free_request(sba, nreq);
 445		INIT_LIST_HEAD(&first->next);
 446
 447		/* Free the first request */
 448		_sba_free_request(sba, first);
 449
 450		/* Process pending requests */
 451		_sba_process_pending_requests(sba);
 452
 453		spin_unlock_irqrestore(&sba->reqs_lock, flags);
 454	}
 455}
 456
 457static void sba_write_stats_in_seqfile(struct sba_device *sba,
 458				       struct seq_file *file)
 459{
 460	unsigned long flags;
 461	struct sba_request *req;
 462	u32 free_count = 0, alloced_count = 0;
 463	u32 pending_count = 0, active_count = 0, aborted_count = 0;
 464
 465	spin_lock_irqsave(&sba->reqs_lock, flags);
 466
 467	list_for_each_entry(req, &sba->reqs_free_list, node)
 468		if (async_tx_test_ack(&req->tx))
 469			free_count++;
 470
 471	list_for_each_entry(req, &sba->reqs_alloc_list, node)
 472		alloced_count++;
 473
 474	list_for_each_entry(req, &sba->reqs_pending_list, node)
 475		pending_count++;
 476
 477	list_for_each_entry(req, &sba->reqs_active_list, node)
 478		active_count++;
 479
 480	list_for_each_entry(req, &sba->reqs_aborted_list, node)
 481		aborted_count++;
 482
 483	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 484
 485	seq_printf(file, "maximum requests   = %d\n", sba->max_req);
 486	seq_printf(file, "free requests      = %d\n", free_count);
 487	seq_printf(file, "alloced requests   = %d\n", alloced_count);
 488	seq_printf(file, "pending requests   = %d\n", pending_count);
 489	seq_printf(file, "active requests    = %d\n", active_count);
 490	seq_printf(file, "aborted requests   = %d\n", aborted_count);
 491}
 492
 493/* ====== DMAENGINE callbacks ===== */
 494
 495static void sba_free_chan_resources(struct dma_chan *dchan)
 496{
 497	/*
 498	 * Channel resources are pre-alloced so we just free-up
 499	 * whatever we can so that we can re-use pre-alloced
 500	 * channel resources next time.
 501	 */
 502	sba_cleanup_nonpending_requests(to_sba_device(dchan));
 503}
 504
 505static int sba_device_terminate_all(struct dma_chan *dchan)
 506{
 507	/* Cleanup all pending requests */
 508	sba_cleanup_pending_requests(to_sba_device(dchan));
 509
 510	return 0;
 511}
 512
 513static void sba_issue_pending(struct dma_chan *dchan)
 514{
 515	unsigned long flags;
 516	struct sba_device *sba = to_sba_device(dchan);
 517
 518	/* Process pending requests */
 519	spin_lock_irqsave(&sba->reqs_lock, flags);
 520	_sba_process_pending_requests(sba);
 521	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 522}
 523
 524static dma_cookie_t sba_tx_submit(struct dma_async_tx_descriptor *tx)
 525{
 526	unsigned long flags;
 527	dma_cookie_t cookie;
 528	struct sba_device *sba;
 529	struct sba_request *req, *nreq;
 530
 531	if (unlikely(!tx))
 532		return -EINVAL;
 533
 534	sba = to_sba_device(tx->chan);
 535	req = to_sba_request(tx);
 536
 537	/* Assign cookie and mark all chained requests pending */
 538	spin_lock_irqsave(&sba->reqs_lock, flags);
 539	cookie = dma_cookie_assign(tx);
 540	_sba_pending_request(sba, req);
 541	list_for_each_entry(nreq, &req->next, next)
 542		_sba_pending_request(sba, nreq);
 543	spin_unlock_irqrestore(&sba->reqs_lock, flags);
 544
 545	return cookie;
 546}
 547
 548static enum dma_status sba_tx_status(struct dma_chan *dchan,
 549				     dma_cookie_t cookie,
 550				     struct dma_tx_state *txstate)
 551{
 552	enum dma_status ret;
 553	struct sba_device *sba = to_sba_device(dchan);
 554
 555	ret = dma_cookie_status(dchan, cookie, txstate);
 556	if (ret == DMA_COMPLETE)
 557		return ret;
 558
 559	mbox_client_peek_data(sba->mchan);
 560
 561	return dma_cookie_status(dchan, cookie, txstate);
 562}
 563
 564static void sba_fillup_interrupt_msg(struct sba_request *req,
 565				     struct brcm_sba_command *cmds,
 566				     struct brcm_message *msg)
 567{
 568	u64 cmd;
 569	u32 c_mdata;
 570	dma_addr_t resp_dma = req->tx.phys;
 571	struct brcm_sba_command *cmdsp = cmds;
 572
 573	/* Type-B command to load dummy data into buf0 */
 574	cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 575			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 576	cmd = sba_cmd_enc(cmd, req->sba->hw_resp_size,
 577			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 578	c_mdata = sba_cmd_load_c_mdata(0);
 579	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 580			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 581	cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 582			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 583	cmdsp->cmd = cmd;
 584	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 585	cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 586	cmdsp->data = resp_dma;
 587	cmdsp->data_len = req->sba->hw_resp_size;
 588	cmdsp++;
 589
 590	/* Type-A command to write buf0 to dummy location */
 591	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 592			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 593	cmd = sba_cmd_enc(cmd, req->sba->hw_resp_size,
 594			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 595	cmd = sba_cmd_enc(cmd, 0x1,
 596			  SBA_RESP_SHIFT, SBA_RESP_MASK);
 597	c_mdata = sba_cmd_write_c_mdata(0);
 598	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 599			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 600	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
 601			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 602	cmdsp->cmd = cmd;
 603	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 604	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
 605	if (req->sba->hw_resp_size) {
 606		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
 607		cmdsp->resp = resp_dma;
 608		cmdsp->resp_len = req->sba->hw_resp_size;
 609	}
 610	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
 611	cmdsp->data = resp_dma;
 612	cmdsp->data_len = req->sba->hw_resp_size;
 613	cmdsp++;
 614
 615	/* Fillup brcm_message */
 616	msg->type = BRCM_MESSAGE_SBA;
 617	msg->sba.cmds = cmds;
 618	msg->sba.cmds_count = cmdsp - cmds;
 619	msg->ctx = req;
 620	msg->error = 0;
 621}
 622
 623static struct dma_async_tx_descriptor *
 624sba_prep_dma_interrupt(struct dma_chan *dchan, unsigned long flags)
 625{
 626	struct sba_request *req = NULL;
 627	struct sba_device *sba = to_sba_device(dchan);
 628
 629	/* Alloc new request */
 630	req = sba_alloc_request(sba);
 631	if (!req)
 632		return NULL;
 633
 634	/*
 635	 * Force fence so that no requests are submitted
 636	 * until DMA callback for this request is invoked.
 637	 */
 638	req->flags |= SBA_REQUEST_FENCE;
 639
 640	/* Fillup request message */
 641	sba_fillup_interrupt_msg(req, req->cmds, &req->msg);
 642
 643	/* Init async_tx descriptor */
 644	req->tx.flags = flags;
 645	req->tx.cookie = -EBUSY;
 646
 647	return &req->tx;
 648}
 649
 650static void sba_fillup_memcpy_msg(struct sba_request *req,
 651				  struct brcm_sba_command *cmds,
 652				  struct brcm_message *msg,
 653				  dma_addr_t msg_offset, size_t msg_len,
 654				  dma_addr_t dst, dma_addr_t src)
 655{
 656	u64 cmd;
 657	u32 c_mdata;
 658	dma_addr_t resp_dma = req->tx.phys;
 659	struct brcm_sba_command *cmdsp = cmds;
 660
 661	/* Type-B command to load data into buf0 */
 662	cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 663			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 664	cmd = sba_cmd_enc(cmd, msg_len,
 665			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 666	c_mdata = sba_cmd_load_c_mdata(0);
 667	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 668			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 669	cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 670			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 671	cmdsp->cmd = cmd;
 672	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 673	cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 674	cmdsp->data = src + msg_offset;
 675	cmdsp->data_len = msg_len;
 676	cmdsp++;
 677
 678	/* Type-A command to write buf0 */
 679	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 680			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 681	cmd = sba_cmd_enc(cmd, msg_len,
 682			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 683	cmd = sba_cmd_enc(cmd, 0x1,
 684			  SBA_RESP_SHIFT, SBA_RESP_MASK);
 685	c_mdata = sba_cmd_write_c_mdata(0);
 686	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 687			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 688	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
 689			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 690	cmdsp->cmd = cmd;
 691	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 692	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
 693	if (req->sba->hw_resp_size) {
 694		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
 695		cmdsp->resp = resp_dma;
 696		cmdsp->resp_len = req->sba->hw_resp_size;
 697	}
 698	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
 699	cmdsp->data = dst + msg_offset;
 700	cmdsp->data_len = msg_len;
 701	cmdsp++;
 702
 703	/* Fillup brcm_message */
 704	msg->type = BRCM_MESSAGE_SBA;
 705	msg->sba.cmds = cmds;
 706	msg->sba.cmds_count = cmdsp - cmds;
 707	msg->ctx = req;
 708	msg->error = 0;
 709}
 710
 711static struct sba_request *
 712sba_prep_dma_memcpy_req(struct sba_device *sba,
 713			dma_addr_t off, dma_addr_t dst, dma_addr_t src,
 714			size_t len, unsigned long flags)
 715{
 716	struct sba_request *req = NULL;
 717
 718	/* Alloc new request */
 719	req = sba_alloc_request(sba);
 720	if (!req)
 721		return NULL;
 722	if (flags & DMA_PREP_FENCE)
 723		req->flags |= SBA_REQUEST_FENCE;
 724
 725	/* Fillup request message */
 726	sba_fillup_memcpy_msg(req, req->cmds, &req->msg,
 727			      off, len, dst, src);
 728
 729	/* Init async_tx descriptor */
 730	req->tx.flags = flags;
 731	req->tx.cookie = -EBUSY;
 732
 733	return req;
 734}
 735
 736static struct dma_async_tx_descriptor *
 737sba_prep_dma_memcpy(struct dma_chan *dchan, dma_addr_t dst, dma_addr_t src,
 738		    size_t len, unsigned long flags)
 739{
 740	size_t req_len;
 741	dma_addr_t off = 0;
 742	struct sba_device *sba = to_sba_device(dchan);
 743	struct sba_request *first = NULL, *req;
 744
 745	/* Create chained requests where each request is upto hw_buf_size */
 746	while (len) {
 747		req_len = (len < sba->hw_buf_size) ? len : sba->hw_buf_size;
 748
 749		req = sba_prep_dma_memcpy_req(sba, off, dst, src,
 750					      req_len, flags);
 751		if (!req) {
 752			if (first)
 753				sba_free_chained_requests(first);
 754			return NULL;
 755		}
 756
 757		if (first)
 758			sba_chain_request(first, req);
 759		else
 760			first = req;
 761
 762		off += req_len;
 763		len -= req_len;
 764	}
 765
 766	return (first) ? &first->tx : NULL;
 767}
 768
 769static void sba_fillup_xor_msg(struct sba_request *req,
 770				struct brcm_sba_command *cmds,
 771				struct brcm_message *msg,
 772				dma_addr_t msg_offset, size_t msg_len,
 773				dma_addr_t dst, dma_addr_t *src, u32 src_cnt)
 774{
 775	u64 cmd;
 776	u32 c_mdata;
 777	unsigned int i;
 778	dma_addr_t resp_dma = req->tx.phys;
 779	struct brcm_sba_command *cmdsp = cmds;
 780
 781	/* Type-B command to load data into buf0 */
 782	cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 783			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 784	cmd = sba_cmd_enc(cmd, msg_len,
 785			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 786	c_mdata = sba_cmd_load_c_mdata(0);
 787	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 788			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 789	cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 790			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 791	cmdsp->cmd = cmd;
 792	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 793	cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 794	cmdsp->data = src[0] + msg_offset;
 795	cmdsp->data_len = msg_len;
 796	cmdsp++;
 797
 798	/* Type-B commands to xor data with buf0 and put it back in buf0 */
 799	for (i = 1; i < src_cnt; i++) {
 800		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 801				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 802		cmd = sba_cmd_enc(cmd, msg_len,
 803				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 804		c_mdata = sba_cmd_xor_c_mdata(0, 0);
 805		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 806				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 807		cmd = sba_cmd_enc(cmd, SBA_CMD_XOR,
 808				  SBA_CMD_SHIFT, SBA_CMD_MASK);
 809		cmdsp->cmd = cmd;
 810		*cmdsp->cmd_dma = cpu_to_le64(cmd);
 811		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 812		cmdsp->data = src[i] + msg_offset;
 813		cmdsp->data_len = msg_len;
 814		cmdsp++;
 815	}
 816
 817	/* Type-A command to write buf0 */
 818	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 819			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 820	cmd = sba_cmd_enc(cmd, msg_len,
 821			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 822	cmd = sba_cmd_enc(cmd, 0x1,
 823			  SBA_RESP_SHIFT, SBA_RESP_MASK);
 824	c_mdata = sba_cmd_write_c_mdata(0);
 825	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 826			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 827	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
 828			  SBA_CMD_SHIFT, SBA_CMD_MASK);
 829	cmdsp->cmd = cmd;
 830	*cmdsp->cmd_dma = cpu_to_le64(cmd);
 831	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
 832	if (req->sba->hw_resp_size) {
 833		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
 834		cmdsp->resp = resp_dma;
 835		cmdsp->resp_len = req->sba->hw_resp_size;
 836	}
 837	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
 838	cmdsp->data = dst + msg_offset;
 839	cmdsp->data_len = msg_len;
 840	cmdsp++;
 841
 842	/* Fillup brcm_message */
 843	msg->type = BRCM_MESSAGE_SBA;
 844	msg->sba.cmds = cmds;
 845	msg->sba.cmds_count = cmdsp - cmds;
 846	msg->ctx = req;
 847	msg->error = 0;
 848}
 849
 850static struct sba_request *
 851sba_prep_dma_xor_req(struct sba_device *sba,
 852		     dma_addr_t off, dma_addr_t dst, dma_addr_t *src,
 853		     u32 src_cnt, size_t len, unsigned long flags)
 854{
 855	struct sba_request *req = NULL;
 856
 857	/* Alloc new request */
 858	req = sba_alloc_request(sba);
 859	if (!req)
 860		return NULL;
 861	if (flags & DMA_PREP_FENCE)
 862		req->flags |= SBA_REQUEST_FENCE;
 863
 864	/* Fillup request message */
 865	sba_fillup_xor_msg(req, req->cmds, &req->msg,
 866			   off, len, dst, src, src_cnt);
 867
 868	/* Init async_tx descriptor */
 869	req->tx.flags = flags;
 870	req->tx.cookie = -EBUSY;
 871
 872	return req;
 873}
 874
 875static struct dma_async_tx_descriptor *
 876sba_prep_dma_xor(struct dma_chan *dchan, dma_addr_t dst, dma_addr_t *src,
 877		 u32 src_cnt, size_t len, unsigned long flags)
 878{
 879	size_t req_len;
 880	dma_addr_t off = 0;
 881	struct sba_device *sba = to_sba_device(dchan);
 882	struct sba_request *first = NULL, *req;
 883
 884	/* Sanity checks */
 885	if (unlikely(src_cnt > sba->max_xor_srcs))
 886		return NULL;
 887
 888	/* Create chained requests where each request is upto hw_buf_size */
 889	while (len) {
 890		req_len = (len < sba->hw_buf_size) ? len : sba->hw_buf_size;
 891
 892		req = sba_prep_dma_xor_req(sba, off, dst, src, src_cnt,
 893					   req_len, flags);
 894		if (!req) {
 895			if (first)
 896				sba_free_chained_requests(first);
 897			return NULL;
 898		}
 899
 900		if (first)
 901			sba_chain_request(first, req);
 902		else
 903			first = req;
 904
 905		off += req_len;
 906		len -= req_len;
 907	}
 908
 909	return (first) ? &first->tx : NULL;
 910}
 911
 912static void sba_fillup_pq_msg(struct sba_request *req,
 913				bool pq_continue,
 914				struct brcm_sba_command *cmds,
 915				struct brcm_message *msg,
 916				dma_addr_t msg_offset, size_t msg_len,
 917				dma_addr_t *dst_p, dma_addr_t *dst_q,
 918				const u8 *scf, dma_addr_t *src, u32 src_cnt)
 919{
 920	u64 cmd;
 921	u32 c_mdata;
 922	unsigned int i;
 923	dma_addr_t resp_dma = req->tx.phys;
 924	struct brcm_sba_command *cmdsp = cmds;
 925
 926	if (pq_continue) {
 927		/* Type-B command to load old P into buf0 */
 928		if (dst_p) {
 929			cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 930				SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 931			cmd = sba_cmd_enc(cmd, msg_len,
 932				SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 933			c_mdata = sba_cmd_load_c_mdata(0);
 934			cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 935				SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 936			cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 937				SBA_CMD_SHIFT, SBA_CMD_MASK);
 938			cmdsp->cmd = cmd;
 939			*cmdsp->cmd_dma = cpu_to_le64(cmd);
 940			cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 941			cmdsp->data = *dst_p + msg_offset;
 942			cmdsp->data_len = msg_len;
 943			cmdsp++;
 944		}
 945
 946		/* Type-B command to load old Q into buf1 */
 947		if (dst_q) {
 948			cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 949				SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 950			cmd = sba_cmd_enc(cmd, msg_len,
 951				SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 952			c_mdata = sba_cmd_load_c_mdata(1);
 953			cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 954				SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 955			cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
 956				SBA_CMD_SHIFT, SBA_CMD_MASK);
 957			cmdsp->cmd = cmd;
 958			*cmdsp->cmd_dma = cpu_to_le64(cmd);
 959			cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 960			cmdsp->data = *dst_q + msg_offset;
 961			cmdsp->data_len = msg_len;
 962			cmdsp++;
 963		}
 964	} else {
 965		/* Type-A command to zero all buffers */
 966		cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
 967				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 968		cmd = sba_cmd_enc(cmd, msg_len,
 969				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 970		cmd = sba_cmd_enc(cmd, SBA_CMD_ZERO_ALL_BUFFERS,
 971				  SBA_CMD_SHIFT, SBA_CMD_MASK);
 972		cmdsp->cmd = cmd;
 973		*cmdsp->cmd_dma = cpu_to_le64(cmd);
 974		cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
 975		cmdsp++;
 976	}
 977
 978	/* Type-B commands for generate P onto buf0 and Q onto buf1 */
 979	for (i = 0; i < src_cnt; i++) {
 980		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
 981				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
 982		cmd = sba_cmd_enc(cmd, msg_len,
 983				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
 984		c_mdata = sba_cmd_pq_c_mdata(raid6_gflog[scf[i]], 1, 0);
 985		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
 986				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
 987		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_MS(c_mdata),
 988				  SBA_C_MDATA_MS_SHIFT, SBA_C_MDATA_MS_MASK);
 989		cmd = sba_cmd_enc(cmd, SBA_CMD_GALOIS_XOR,
 990				  SBA_CMD_SHIFT, SBA_CMD_MASK);
 991		cmdsp->cmd = cmd;
 992		*cmdsp->cmd_dma = cpu_to_le64(cmd);
 993		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
 994		cmdsp->data = src[i] + msg_offset;
 995		cmdsp->data_len = msg_len;
 996		cmdsp++;
 997	}
 998
 999	/* Type-A command to write buf0 */
1000	if (dst_p) {
1001		cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1002				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1003		cmd = sba_cmd_enc(cmd, msg_len,
1004				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1005		cmd = sba_cmd_enc(cmd, 0x1,
1006				  SBA_RESP_SHIFT, SBA_RESP_MASK);
1007		c_mdata = sba_cmd_write_c_mdata(0);
1008		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1009				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1010		cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
1011				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1012		cmdsp->cmd = cmd;
1013		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1014		cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1015		if (req->sba->hw_resp_size) {
1016			cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
1017			cmdsp->resp = resp_dma;
1018			cmdsp->resp_len = req->sba->hw_resp_size;
1019		}
1020		cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
1021		cmdsp->data = *dst_p + msg_offset;
1022		cmdsp->data_len = msg_len;
1023		cmdsp++;
1024	}
1025
1026	/* Type-A command to write buf1 */
1027	if (dst_q) {
1028		cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1029				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1030		cmd = sba_cmd_enc(cmd, msg_len,
1031				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1032		cmd = sba_cmd_enc(cmd, 0x1,
1033				  SBA_RESP_SHIFT, SBA_RESP_MASK);
1034		c_mdata = sba_cmd_write_c_mdata(1);
1035		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1036				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1037		cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
1038				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1039		cmdsp->cmd = cmd;
1040		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1041		cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1042		if (req->sba->hw_resp_size) {
1043			cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
1044			cmdsp->resp = resp_dma;
1045			cmdsp->resp_len = req->sba->hw_resp_size;
1046		}
1047		cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
1048		cmdsp->data = *dst_q + msg_offset;
1049		cmdsp->data_len = msg_len;
1050		cmdsp++;
1051	}
1052
1053	/* Fillup brcm_message */
1054	msg->type = BRCM_MESSAGE_SBA;
1055	msg->sba.cmds = cmds;
1056	msg->sba.cmds_count = cmdsp - cmds;
1057	msg->ctx = req;
1058	msg->error = 0;
1059}
1060
1061static struct sba_request *
1062sba_prep_dma_pq_req(struct sba_device *sba, dma_addr_t off,
1063		    dma_addr_t *dst_p, dma_addr_t *dst_q, dma_addr_t *src,
1064		    u32 src_cnt, const u8 *scf, size_t len, unsigned long flags)
1065{
1066	struct sba_request *req = NULL;
1067
1068	/* Alloc new request */
1069	req = sba_alloc_request(sba);
1070	if (!req)
1071		return NULL;
1072	if (flags & DMA_PREP_FENCE)
1073		req->flags |= SBA_REQUEST_FENCE;
1074
1075	/* Fillup request messages */
1076	sba_fillup_pq_msg(req, dmaf_continue(flags),
1077			  req->cmds, &req->msg,
1078			  off, len, dst_p, dst_q, scf, src, src_cnt);
1079
1080	/* Init async_tx descriptor */
1081	req->tx.flags = flags;
1082	req->tx.cookie = -EBUSY;
1083
1084	return req;
1085}
1086
1087static void sba_fillup_pq_single_msg(struct sba_request *req,
1088				bool pq_continue,
1089				struct brcm_sba_command *cmds,
1090				struct brcm_message *msg,
1091				dma_addr_t msg_offset, size_t msg_len,
1092				dma_addr_t *dst_p, dma_addr_t *dst_q,
1093				dma_addr_t src, u8 scf)
1094{
1095	u64 cmd;
1096	u32 c_mdata;
1097	u8 pos, dpos = raid6_gflog[scf];
1098	dma_addr_t resp_dma = req->tx.phys;
1099	struct brcm_sba_command *cmdsp = cmds;
1100
1101	if (!dst_p)
1102		goto skip_p;
1103
1104	if (pq_continue) {
1105		/* Type-B command to load old P into buf0 */
1106		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1107				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1108		cmd = sba_cmd_enc(cmd, msg_len,
1109				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1110		c_mdata = sba_cmd_load_c_mdata(0);
1111		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1112				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1113		cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
1114				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1115		cmdsp->cmd = cmd;
1116		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1117		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1118		cmdsp->data = *dst_p + msg_offset;
1119		cmdsp->data_len = msg_len;
1120		cmdsp++;
1121
1122		/*
1123		 * Type-B commands to xor data with buf0 and put it
1124		 * back in buf0
1125		 */
1126		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1127				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1128		cmd = sba_cmd_enc(cmd, msg_len,
1129				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1130		c_mdata = sba_cmd_xor_c_mdata(0, 0);
1131		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1132				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1133		cmd = sba_cmd_enc(cmd, SBA_CMD_XOR,
1134				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1135		cmdsp->cmd = cmd;
1136		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1137		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1138		cmdsp->data = src + msg_offset;
1139		cmdsp->data_len = msg_len;
1140		cmdsp++;
1141	} else {
1142		/* Type-B command to load old P into buf0 */
1143		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1144				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1145		cmd = sba_cmd_enc(cmd, msg_len,
1146				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1147		c_mdata = sba_cmd_load_c_mdata(0);
1148		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1149				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1150		cmd = sba_cmd_enc(cmd, SBA_CMD_LOAD_BUFFER,
1151				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1152		cmdsp->cmd = cmd;
1153		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1154		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1155		cmdsp->data = src + msg_offset;
1156		cmdsp->data_len = msg_len;
1157		cmdsp++;
1158	}
1159
1160	/* Type-A command to write buf0 */
1161	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1162			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1163	cmd = sba_cmd_enc(cmd, msg_len,
1164			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1165	cmd = sba_cmd_enc(cmd, 0x1,
1166			  SBA_RESP_SHIFT, SBA_RESP_MASK);
1167	c_mdata = sba_cmd_write_c_mdata(0);
1168	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1169			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1170	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
1171			  SBA_CMD_SHIFT, SBA_CMD_MASK);
1172	cmdsp->cmd = cmd;
1173	*cmdsp->cmd_dma = cpu_to_le64(cmd);
1174	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1175	if (req->sba->hw_resp_size) {
1176		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
1177		cmdsp->resp = resp_dma;
1178		cmdsp->resp_len = req->sba->hw_resp_size;
1179	}
1180	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
1181	cmdsp->data = *dst_p + msg_offset;
1182	cmdsp->data_len = msg_len;
1183	cmdsp++;
1184
1185skip_p:
1186	if (!dst_q)
1187		goto skip_q;
1188
1189	/* Type-A command to zero all buffers */
1190	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1191			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1192	cmd = sba_cmd_enc(cmd, msg_len,
1193			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1194	cmd = sba_cmd_enc(cmd, SBA_CMD_ZERO_ALL_BUFFERS,
1195			  SBA_CMD_SHIFT, SBA_CMD_MASK);
1196	cmdsp->cmd = cmd;
1197	*cmdsp->cmd_dma = cpu_to_le64(cmd);
1198	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1199	cmdsp++;
1200
1201	if (dpos == 255)
1202		goto skip_q_computation;
1203	pos = (dpos < req->sba->max_pq_coefs) ?
1204		dpos : (req->sba->max_pq_coefs - 1);
1205
1206	/*
1207	 * Type-B command to generate initial Q from data
1208	 * and store output into buf0
1209	 */
1210	cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1211			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1212	cmd = sba_cmd_enc(cmd, msg_len,
1213			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1214	c_mdata = sba_cmd_pq_c_mdata(pos, 0, 0);
1215	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1216			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1217	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_MS(c_mdata),
1218			  SBA_C_MDATA_MS_SHIFT, SBA_C_MDATA_MS_MASK);
1219	cmd = sba_cmd_enc(cmd, SBA_CMD_GALOIS,
1220			  SBA_CMD_SHIFT, SBA_CMD_MASK);
1221	cmdsp->cmd = cmd;
1222	*cmdsp->cmd_dma = cpu_to_le64(cmd);
1223	cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1224	cmdsp->data = src + msg_offset;
1225	cmdsp->data_len = msg_len;
1226	cmdsp++;
1227
1228	dpos -= pos;
1229
1230	/* Multiple Type-A command to generate final Q */
1231	while (dpos) {
1232		pos = (dpos < req->sba->max_pq_coefs) ?
1233			dpos : (req->sba->max_pq_coefs - 1);
1234
1235		/*
1236		 * Type-A command to generate Q with buf0 and
1237		 * buf1 store result in buf0
1238		 */
1239		cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1240				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1241		cmd = sba_cmd_enc(cmd, msg_len,
1242				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1243		c_mdata = sba_cmd_pq_c_mdata(pos, 0, 1);
1244		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1245				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1246		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_MS(c_mdata),
1247				  SBA_C_MDATA_MS_SHIFT, SBA_C_MDATA_MS_MASK);
1248		cmd = sba_cmd_enc(cmd, SBA_CMD_GALOIS,
1249				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1250		cmdsp->cmd = cmd;
1251		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1252		cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1253		cmdsp++;
1254
1255		dpos -= pos;
1256	}
1257
1258skip_q_computation:
1259	if (pq_continue) {
1260		/*
1261		 * Type-B command to XOR previous output with
1262		 * buf0 and write it into buf0
1263		 */
1264		cmd = sba_cmd_enc(0x0, SBA_TYPE_B,
1265				  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1266		cmd = sba_cmd_enc(cmd, msg_len,
1267				  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1268		c_mdata = sba_cmd_xor_c_mdata(0, 0);
1269		cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1270				  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1271		cmd = sba_cmd_enc(cmd, SBA_CMD_XOR,
1272				  SBA_CMD_SHIFT, SBA_CMD_MASK);
1273		cmdsp->cmd = cmd;
1274		*cmdsp->cmd_dma = cpu_to_le64(cmd);
1275		cmdsp->flags = BRCM_SBA_CMD_TYPE_B;
1276		cmdsp->data = *dst_q + msg_offset;
1277		cmdsp->data_len = msg_len;
1278		cmdsp++;
1279	}
1280
1281	/* Type-A command to write buf0 */
1282	cmd = sba_cmd_enc(0x0, SBA_TYPE_A,
1283			  SBA_TYPE_SHIFT, SBA_TYPE_MASK);
1284	cmd = sba_cmd_enc(cmd, msg_len,
1285			  SBA_USER_DEF_SHIFT, SBA_USER_DEF_MASK);
1286	cmd = sba_cmd_enc(cmd, 0x1,
1287			  SBA_RESP_SHIFT, SBA_RESP_MASK);
1288	c_mdata = sba_cmd_write_c_mdata(0);
1289	cmd = sba_cmd_enc(cmd, SBA_C_MDATA_LS(c_mdata),
1290			  SBA_C_MDATA_SHIFT, SBA_C_MDATA_MASK);
1291	cmd = sba_cmd_enc(cmd, SBA_CMD_WRITE_BUFFER,
1292			  SBA_CMD_SHIFT, SBA_CMD_MASK);
1293	cmdsp->cmd = cmd;
1294	*cmdsp->cmd_dma = cpu_to_le64(cmd);
1295	cmdsp->flags = BRCM_SBA_CMD_TYPE_A;
1296	if (req->sba->hw_resp_size) {
1297		cmdsp->flags |= BRCM_SBA_CMD_HAS_RESP;
1298		cmdsp->resp = resp_dma;
1299		cmdsp->resp_len = req->sba->hw_resp_size;
1300	}
1301	cmdsp->flags |= BRCM_SBA_CMD_HAS_OUTPUT;
1302	cmdsp->data = *dst_q + msg_offset;
1303	cmdsp->data_len = msg_len;
1304	cmdsp++;
1305
1306skip_q:
1307	/* Fillup brcm_message */
1308	msg->type = BRCM_MESSAGE_SBA;
1309	msg->sba.cmds = cmds;
1310	msg->sba.cmds_count = cmdsp - cmds;
1311	msg->ctx = req;
1312	msg->error = 0;
1313}
1314
1315static struct sba_request *
1316sba_prep_dma_pq_single_req(struct sba_device *sba, dma_addr_t off,
1317			   dma_addr_t *dst_p, dma_addr_t *dst_q,
1318			   dma_addr_t src, u8 scf, size_t len,
1319			   unsigned long flags)
1320{
1321	struct sba_request *req = NULL;
1322
1323	/* Alloc new request */
1324	req = sba_alloc_request(sba);
1325	if (!req)
1326		return NULL;
1327	if (flags & DMA_PREP_FENCE)
1328		req->flags |= SBA_REQUEST_FENCE;
1329
1330	/* Fillup request messages */
1331	sba_fillup_pq_single_msg(req,  dmaf_continue(flags),
1332				 req->cmds, &req->msg, off, len,
1333				 dst_p, dst_q, src, scf);
1334
1335	/* Init async_tx descriptor */
1336	req->tx.flags = flags;
1337	req->tx.cookie = -EBUSY;
1338
1339	return req;
1340}
1341
1342static struct dma_async_tx_descriptor *
1343sba_prep_dma_pq(struct dma_chan *dchan, dma_addr_t *dst, dma_addr_t *src,
1344		u32 src_cnt, const u8 *scf, size_t len, unsigned long flags)
1345{
1346	u32 i, dst_q_index;
1347	size_t req_len;
1348	bool slow = false;
1349	dma_addr_t off = 0;
1350	dma_addr_t *dst_p = NULL, *dst_q = NULL;
1351	struct sba_device *sba = to_sba_device(dchan);
1352	struct sba_request *first = NULL, *req;
1353
1354	/* Sanity checks */
1355	if (unlikely(src_cnt > sba->max_pq_srcs))
1356		return NULL;
1357	for (i = 0; i < src_cnt; i++)
1358		if (sba->max_pq_coefs <= raid6_gflog[scf[i]])
1359			slow = true;
1360
1361	/* Figure-out P and Q destination addresses */
1362	if (!(flags & DMA_PREP_PQ_DISABLE_P))
1363		dst_p = &dst[0];
1364	if (!(flags & DMA_PREP_PQ_DISABLE_Q))
1365		dst_q = &dst[1];
1366
1367	/* Create chained requests where each request is upto hw_buf_size */
1368	while (len) {
1369		req_len = (len < sba->hw_buf_size) ? len : sba->hw_buf_size;
1370
1371		if (slow) {
1372			dst_q_index = src_cnt;
1373
1374			if (dst_q) {
1375				for (i = 0; i < src_cnt; i++) {
1376					if (*dst_q == src[i]) {
1377						dst_q_index = i;
1378						break;
1379					}
1380				}
1381			}
1382
1383			if (dst_q_index < src_cnt) {
1384				i = dst_q_index;
1385				req = sba_prep_dma_pq_single_req(sba,
1386					off, dst_p, dst_q, src[i], scf[i],
1387					req_len, flags | DMA_PREP_FENCE);
1388				if (!req)
1389					goto fail;
1390
1391				if (first)
1392					sba_chain_request(first, req);
1393				else
1394					first = req;
1395
1396				flags |= DMA_PREP_CONTINUE;
1397			}
1398
1399			for (i = 0; i < src_cnt; i++) {
1400				if (dst_q_index == i)
1401					continue;
1402
1403				req = sba_prep_dma_pq_single_req(sba,
1404					off, dst_p, dst_q, src[i], scf[i],
1405					req_len, flags | DMA_PREP_FENCE);
1406				if (!req)
1407					goto fail;
1408
1409				if (first)
1410					sba_chain_request(first, req);
1411				else
1412					first = req;
1413
1414				flags |= DMA_PREP_CONTINUE;
1415			}
1416		} else {
1417			req = sba_prep_dma_pq_req(sba, off,
1418						  dst_p, dst_q, src, src_cnt,
1419						  scf, req_len, flags);
1420			if (!req)
1421				goto fail;
1422
1423			if (first)
1424				sba_chain_request(first, req);
1425			else
1426				first = req;
1427		}
1428
1429		off += req_len;
1430		len -= req_len;
1431	}
1432
1433	return (first) ? &first->tx : NULL;
1434
1435fail:
1436	if (first)
1437		sba_free_chained_requests(first);
1438	return NULL;
1439}
1440
1441/* ====== Mailbox callbacks ===== */
1442
1443static void sba_receive_message(struct mbox_client *cl, void *msg)
1444{
1445	struct brcm_message *m = msg;
1446	struct sba_request *req = m->ctx;
1447	struct sba_device *sba = req->sba;
1448
1449	/* Error count if message has error */
1450	if (m->error < 0)
1451		dev_err(sba->dev, "%s got message with error %d",
1452			dma_chan_name(&sba->dma_chan), m->error);
1453
1454	/* Process received request */
1455	sba_process_received_request(sba, req);
1456}
1457
1458/* ====== Debugfs callbacks ====== */
1459
1460static int sba_debugfs_stats_show(struct seq_file *file, void *offset)
1461{
1462	struct platform_device *pdev = to_platform_device(file->private);
1463	struct sba_device *sba = platform_get_drvdata(pdev);
1464
1465	/* Write stats in file */
1466	sba_write_stats_in_seqfile(sba, file);
1467
1468	return 0;
1469}
1470
1471/* ====== Platform driver routines ===== */
1472
1473static int sba_prealloc_channel_resources(struct sba_device *sba)
1474{
1475	int i, j, ret = 0;
1476	struct sba_request *req = NULL;
1477
1478	sba->resp_base = dma_alloc_coherent(sba->mbox_dev,
1479					    sba->max_resp_pool_size,
1480					    &sba->resp_dma_base, GFP_KERNEL);
1481	if (!sba->resp_base)
1482		return -ENOMEM;
1483
1484	sba->cmds_base = dma_alloc_coherent(sba->mbox_dev,
1485					    sba->max_cmds_pool_size,
1486					    &sba->cmds_dma_base, GFP_KERNEL);
1487	if (!sba->cmds_base) {
1488		ret = -ENOMEM;
1489		goto fail_free_resp_pool;
1490	}
1491
1492	spin_lock_init(&sba->reqs_lock);
1493	sba->reqs_fence = false;
1494	INIT_LIST_HEAD(&sba->reqs_alloc_list);
1495	INIT_LIST_HEAD(&sba->reqs_pending_list);
1496	INIT_LIST_HEAD(&sba->reqs_active_list);
1497	INIT_LIST_HEAD(&sba->reqs_aborted_list);
1498	INIT_LIST_HEAD(&sba->reqs_free_list);
1499
1500	for (i = 0; i < sba->max_req; i++) {
1501		req = devm_kzalloc(sba->dev,
1502				sizeof(*req) +
1503				sba->max_cmd_per_req * sizeof(req->cmds[0]),
1504				GFP_KERNEL);
1505		if (!req) {
1506			ret = -ENOMEM;
1507			goto fail_free_cmds_pool;
1508		}
1509		INIT_LIST_HEAD(&req->node);
1510		req->sba = sba;
1511		req->flags = SBA_REQUEST_STATE_FREE;
1512		INIT_LIST_HEAD(&req->next);
1513		atomic_set(&req->next_pending_count, 0);
1514		for (j = 0; j < sba->max_cmd_per_req; j++) {
1515			req->cmds[j].cmd = 0;
1516			req->cmds[j].cmd_dma = sba->cmds_base +
1517				(i * sba->max_cmd_per_req + j) * sizeof(u64);
1518			req->cmds[j].cmd_dma_addr = sba->cmds_dma_base +
1519				(i * sba->max_cmd_per_req + j) * sizeof(u64);
1520			req->cmds[j].flags = 0;
1521		}
1522		memset(&req->msg, 0, sizeof(req->msg));
1523		dma_async_tx_descriptor_init(&req->tx, &sba->dma_chan);
1524		async_tx_ack(&req->tx);
1525		req->tx.tx_submit = sba_tx_submit;
1526		req->tx.phys = sba->resp_dma_base + i * sba->hw_resp_size;
1527		list_add_tail(&req->node, &sba->reqs_free_list);
1528	}
1529
1530	return 0;
1531
1532fail_free_cmds_pool:
1533	dma_free_coherent(sba->mbox_dev,
1534			  sba->max_cmds_pool_size,
1535			  sba->cmds_base, sba->cmds_dma_base);
1536fail_free_resp_pool:
1537	dma_free_coherent(sba->mbox_dev,
1538			  sba->max_resp_pool_size,
1539			  sba->resp_base, sba->resp_dma_base);
1540	return ret;
1541}
1542
1543static void sba_freeup_channel_resources(struct sba_device *sba)
1544{
1545	dmaengine_terminate_all(&sba->dma_chan);
1546	dma_free_coherent(sba->mbox_dev, sba->max_cmds_pool_size,
1547			  sba->cmds_base, sba->cmds_dma_base);
1548	dma_free_coherent(sba->mbox_dev, sba->max_resp_pool_size,
1549			  sba->resp_base, sba->resp_dma_base);
1550	sba->resp_base = NULL;
1551	sba->resp_dma_base = 0;
1552}
1553
1554static int sba_async_register(struct sba_device *sba)
1555{
1556	int ret;
1557	struct dma_device *dma_dev = &sba->dma_dev;
1558
1559	/* Initialize DMA channel cookie */
1560	sba->dma_chan.device = dma_dev;
1561	dma_cookie_init(&sba->dma_chan);
1562
1563	/* Initialize DMA device capability mask */
1564	dma_cap_zero(dma_dev->cap_mask);
1565	dma_cap_set(DMA_INTERRUPT, dma_dev->cap_mask);
1566	dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
1567	dma_cap_set(DMA_XOR, dma_dev->cap_mask);
1568	dma_cap_set(DMA_PQ, dma_dev->cap_mask);
1569
1570	/*
1571	 * Set mailbox channel device as the base device of
1572	 * our dma_device because the actual memory accesses
1573	 * will be done by mailbox controller
1574	 */
1575	dma_dev->dev = sba->mbox_dev;
1576
1577	/* Set base prep routines */
1578	dma_dev->device_free_chan_resources = sba_free_chan_resources;
1579	dma_dev->device_terminate_all = sba_device_terminate_all;
1580	dma_dev->device_issue_pending = sba_issue_pending;
1581	dma_dev->device_tx_status = sba_tx_status;
1582
1583	/* Set interrupt routine */
1584	if (dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask))
1585		dma_dev->device_prep_dma_interrupt = sba_prep_dma_interrupt;
1586
1587	/* Set memcpy routine */
1588	if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask))
1589		dma_dev->device_prep_dma_memcpy = sba_prep_dma_memcpy;
1590
1591	/* Set xor routine and capability */
1592	if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1593		dma_dev->device_prep_dma_xor = sba_prep_dma_xor;
1594		dma_dev->max_xor = sba->max_xor_srcs;
1595	}
1596
1597	/* Set pq routine and capability */
1598	if (dma_has_cap(DMA_PQ, dma_dev->cap_mask)) {
1599		dma_dev->device_prep_dma_pq = sba_prep_dma_pq;
1600		dma_set_maxpq(dma_dev, sba->max_pq_srcs, 0);
1601	}
1602
1603	/* Initialize DMA device channel list */
1604	INIT_LIST_HEAD(&dma_dev->channels);
1605	list_add_tail(&sba->dma_chan.device_node, &dma_dev->channels);
1606
1607	/* Register with Linux async DMA framework*/
1608	ret = dma_async_device_register(dma_dev);
1609	if (ret) {
1610		dev_err(sba->dev, "async device register error %d", ret);
1611		return ret;
1612	}
1613
1614	dev_info(sba->dev, "%s capabilities: %s%s%s%s\n",
1615	dma_chan_name(&sba->dma_chan),
1616	dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask) ? "interrupt " : "",
1617	dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask) ? "memcpy " : "",
1618	dma_has_cap(DMA_XOR, dma_dev->cap_mask) ? "xor " : "",
1619	dma_has_cap(DMA_PQ, dma_dev->cap_mask) ? "pq " : "");
1620
1621	return 0;
1622}
1623
1624static int sba_probe(struct platform_device *pdev)
1625{
1626	int ret = 0;
1627	struct sba_device *sba;
1628	struct platform_device *mbox_pdev;
1629	struct of_phandle_args args;
1630
1631	/* Allocate main SBA struct */
1632	sba = devm_kzalloc(&pdev->dev, sizeof(*sba), GFP_KERNEL);
1633	if (!sba)
1634		return -ENOMEM;
1635
1636	sba->dev = &pdev->dev;
1637	platform_set_drvdata(pdev, sba);
1638
1639	/* Number of mailbox channels should be atleast 1 */
1640	ret = of_count_phandle_with_args(pdev->dev.of_node,
1641					 "mboxes", "#mbox-cells");
1642	if (ret <= 0)
1643		return -ENODEV;
1644
1645	/* Determine SBA version from DT compatible string */
1646	if (of_device_is_compatible(sba->dev->of_node, "brcm,iproc-sba"))
1647		sba->ver = SBA_VER_1;
1648	else if (of_device_is_compatible(sba->dev->of_node,
1649					 "brcm,iproc-sba-v2"))
1650		sba->ver = SBA_VER_2;
1651	else
1652		return -ENODEV;
1653
1654	/* Derived Configuration parameters */
1655	switch (sba->ver) {
1656	case SBA_VER_1:
1657		sba->hw_buf_size = 4096;
1658		sba->hw_resp_size = 8;
1659		sba->max_pq_coefs = 6;
1660		sba->max_pq_srcs = 6;
1661		break;
1662	case SBA_VER_2:
1663		sba->hw_buf_size = 4096;
1664		sba->hw_resp_size = 8;
1665		sba->max_pq_coefs = 30;
1666		/*
1667		 * We can support max_pq_srcs == max_pq_coefs because
1668		 * we are limited by number of SBA commands that we can
1669		 * fit in one message for underlying ring manager HW.
1670		 */
1671		sba->max_pq_srcs = 12;
1672		break;
1673	default:
1674		return -EINVAL;
1675	}
1676	sba->max_req = SBA_MAX_REQ_PER_MBOX_CHANNEL;
1677	sba->max_cmd_per_req = sba->max_pq_srcs + 3;
1678	sba->max_xor_srcs = sba->max_cmd_per_req - 1;
1679	sba->max_resp_pool_size = sba->max_req * sba->hw_resp_size;
1680	sba->max_cmds_pool_size = sba->max_req *
1681				  sba->max_cmd_per_req * sizeof(u64);
1682
1683	/* Setup mailbox client */
1684	sba->client.dev			= &pdev->dev;
1685	sba->client.rx_callback		= sba_receive_message;
1686	sba->client.tx_block		= false;
1687	sba->client.knows_txdone	= true;
1688	sba->client.tx_tout		= 0;
1689
1690	/* Request mailbox channel */
1691	sba->mchan = mbox_request_channel(&sba->client, 0);
1692	if (IS_ERR(sba->mchan)) {
1693		ret = PTR_ERR(sba->mchan);
1694		goto fail_free_mchan;
1695	}
1696
1697	/* Find-out underlying mailbox device */
1698	ret = of_parse_phandle_with_args(pdev->dev.of_node,
1699					 "mboxes", "#mbox-cells", 0, &args);
1700	if (ret)
1701		goto fail_free_mchan;
1702	mbox_pdev = of_find_device_by_node(args.np);
1703	of_node_put(args.np);
1704	if (!mbox_pdev) {
1705		ret = -ENODEV;
1706		goto fail_free_mchan;
1707	}
1708	sba->mbox_dev = &mbox_pdev->dev;
1709
1710	/* Prealloc channel resource */
1711	ret = sba_prealloc_channel_resources(sba);
1712	if (ret)
1713		goto fail_free_mchan;
1714
1715	/* Check availability of debugfs */
1716	if (!debugfs_initialized())
1717		goto skip_debugfs;
1718
1719	/* Create debugfs root entry */
1720	sba->root = debugfs_create_dir(dev_name(sba->dev), NULL);
1721	if (IS_ERR_OR_NULL(sba->root)) {
1722		dev_err(sba->dev, "failed to create debugfs root entry\n");
1723		sba->root = NULL;
1724		goto skip_debugfs;
1725	}
1726
1727	/* Create debugfs stats entry */
1728	sba->stats = debugfs_create_devm_seqfile(sba->dev, "stats", sba->root,
1729						 sba_debugfs_stats_show);
1730	if (IS_ERR_OR_NULL(sba->stats))
1731		dev_err(sba->dev, "failed to create debugfs stats file\n");
1732skip_debugfs:
1733
1734	/* Register DMA device with Linux async framework */
1735	ret = sba_async_register(sba);
1736	if (ret)
1737		goto fail_free_resources;
1738
1739	/* Print device info */
1740	dev_info(sba->dev, "%s using SBAv%d mailbox channel from %s",
1741		 dma_chan_name(&sba->dma_chan), sba->ver+1,
1742		 dev_name(sba->mbox_dev));
1743
1744	return 0;
1745
1746fail_free_resources:
1747	debugfs_remove_recursive(sba->root);
1748	sba_freeup_channel_resources(sba);
1749fail_free_mchan:
1750	mbox_free_channel(sba->mchan);
1751	return ret;
1752}
1753
1754static int sba_remove(struct platform_device *pdev)
1755{
1756	struct sba_device *sba = platform_get_drvdata(pdev);
1757
1758	dma_async_device_unregister(&sba->dma_dev);
1759
1760	debugfs_remove_recursive(sba->root);
1761
1762	sba_freeup_channel_resources(sba);
1763
1764	mbox_free_channel(sba->mchan);
1765
1766	return 0;
1767}
1768
1769static const struct of_device_id sba_of_match[] = {
1770	{ .compatible = "brcm,iproc-sba", },
1771	{ .compatible = "brcm,iproc-sba-v2", },
1772	{},
1773};
1774MODULE_DEVICE_TABLE(of, sba_of_match);
1775
1776static struct platform_driver sba_driver = {
1777	.probe = sba_probe,
1778	.remove = sba_remove,
1779	.driver = {
1780		.name = "bcm-sba-raid",
1781		.of_match_table = sba_of_match,
1782	},
1783};
1784module_platform_driver(sba_driver);
1785
1786MODULE_DESCRIPTION("Broadcom SBA RAID driver");
1787MODULE_AUTHOR("Anup Patel <anup.patel@broadcom.com>");
1788MODULE_LICENSE("GPL v2");