xref: /freebsd/sys/dev/qat/qat_common/adf_freebsd_uio_cleanup.c (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 /* SPDX-License-Identifier: BSD-3-Clause */
2 /* Copyright(c) 2007-2022 Intel Corporation */
3 #include "qat_freebsd.h"
4 #include "adf_cfg.h"
5 #include "adf_common_drv.h"
6 #include "adf_accel_devices.h"
7 #include "icp_qat_uclo.h"
8 #include "icp_qat_fw.h"
9 #include "icp_qat_fw_init_admin.h"
10 #include "adf_cfg_strings.h"
11 #include "adf_uio_control.h"
12 #include "adf_uio_cleanup.h"
13 #include "adf_uio.h"
14 #include "adf_transport_access_macros.h"
15 #include "adf_transport_internal.h"
16 #include <sys/param.h>
17 #include <sys/lock.h>
18 #include <sys/rwlock.h>
19 #include <sys/sglist.h>
20 #include <sys/systm.h>
21 #include <sys/proc.h>
22 #include <vm/vm.h>
23 #include <vm/pmap.h>
24 #include <vm/vm_kern.h>
25 #include <vm/vm_map.h>
26 #include <vm/vm_object.h>
27 #include <vm/vm_page.h>
28 #include <vm/vm_pager.h>
29 #include <vm/vm_param.h>
30 
31 #define TX_RINGS_DISABLE 0
32 #define TX_RINGS_ENABLE 1
33 #define PKE_REQ_SIZE 64
34 #define BASE_ADDR_SHIFT 6
35 #define PKE_RX_RING_0 0
36 #define PKE_RX_RING_1 1
37 
38 #define ADF_RING_EMPTY_RETRY_DELAY 2
39 #define ADF_RING_EMPTY_MAX_RETRY 15
40 
41 struct bundle_orphan_ring {
42 	unsigned long tx_mask;
43 	unsigned long rx_mask;
44 	unsigned long asym_mask;
45 	int bank;
46 	struct resource *csr_base;
47 	struct adf_uio_control_bundle *bundle;
48 };
49 
50 /*
51  *    if orphan->tx_mask does not match with orphan->rx_mask
52  */
53 static void
54 check_orphan_ring(struct adf_accel_dev *accel_dev,
55 		  struct bundle_orphan_ring *orphan,
56 		  struct adf_hw_device_data *hw_data)
57 {
58 	struct adf_hw_csr_ops *csr_ops = GET_CSR_OPS(accel_dev);
59 	int i;
60 	int tx_rx_gap = hw_data->tx_rx_gap;
61 	u8 num_rings_per_bank = hw_data->num_rings_per_bank;
62 	struct resource *csr_base = orphan->csr_base;
63 	int bank = orphan->bank;
64 
65 	for (i = 0; i < num_rings_per_bank; i++) {
66 		if (test_bit(i, &orphan->tx_mask)) {
67 			int rx_ring = i + tx_rx_gap;
68 
69 			if (!test_bit(rx_ring, &orphan->rx_mask)) {
70 				__clear_bit(i, &orphan->tx_mask);
71 
72 				/* clean up this tx ring  */
73 				csr_ops->write_csr_ring_config(csr_base,
74 							       bank,
75 							       i,
76 							       0);
77 				csr_ops->write_csr_ring_base(csr_base,
78 							     bank,
79 							     i,
80 							     0);
81 			}
82 
83 		} else if (test_bit(i, &orphan->rx_mask)) {
84 			int tx_ring = i - tx_rx_gap;
85 
86 			if (!test_bit(tx_ring, &orphan->tx_mask)) {
87 				__clear_bit(i, &orphan->rx_mask);
88 
89 				/* clean up this rx ring */
90 				csr_ops->write_csr_ring_config(csr_base,
91 							       bank,
92 							       i,
93 							       0);
94 				csr_ops->write_csr_ring_base(csr_base,
95 							     bank,
96 							     i,
97 							     0);
98 			}
99 		}
100 	}
101 }
102 
103 static int
104 get_orphan_bundle(int bank,
105 		  struct adf_uio_control_accel *accel,
106 		  struct bundle_orphan_ring **orphan_bundle_out)
107 {
108 	int i;
109 	int ret = 0;
110 	struct resource *csr_base;
111 	unsigned long tx_mask;
112 	unsigned long asym_mask;
113 	struct adf_accel_dev *accel_dev = accel->accel_dev;
114 	struct adf_hw_csr_ops *csr_ops = GET_CSR_OPS(accel_dev);
115 	struct adf_hw_device_data *hw_data = accel_dev->hw_device;
116 	u8 num_rings_per_bank = hw_data->num_rings_per_bank;
117 	struct bundle_orphan_ring *orphan_bundle = NULL;
118 	uint64_t base;
119 	struct list_head *entry;
120 	struct adf_uio_instance_rings *instance_rings;
121 	struct adf_uio_control_bundle *bundle;
122 	u16 ring_mask = 0;
123 
124 	orphan_bundle =
125 	    malloc(sizeof(*orphan_bundle), M_QAT, M_WAITOK | M_ZERO);
126 	csr_base = accel->bar->virt_addr;
127 	orphan_bundle->csr_base = csr_base;
128 	orphan_bundle->bank = bank;
129 
130 	orphan_bundle->tx_mask = 0;
131 	orphan_bundle->rx_mask = 0;
132 	tx_mask = accel_dev->hw_device->tx_rings_mask;
133 	asym_mask = accel_dev->hw_device->asym_rings_mask;
134 
135 	/* Get ring mask for this process. */
136 	bundle = &accel->bundle[bank];
137 	orphan_bundle->bundle = bundle;
138 	mutex_lock(&bundle->list_lock);
139 	list_for_each(entry, &bundle->list)
140 	{
141 		instance_rings =
142 		    list_entry(entry, struct adf_uio_instance_rings, list);
143 		if (instance_rings->user_pid == curproc->p_pid) {
144 			ring_mask = instance_rings->ring_mask;
145 			break;
146 		}
147 	}
148 	mutex_unlock(&bundle->list_lock);
149 
150 	for (i = 0; i < num_rings_per_bank; i++) {
151 		base = csr_ops->read_csr_ring_base(csr_base, bank, i);
152 
153 		if (!base)
154 			continue;
155 		if (!(ring_mask & 1 << i))
156 			continue; /* Not reserved for this process. */
157 
158 		if (test_bit(i, &tx_mask))
159 			__set_bit(i, &orphan_bundle->tx_mask);
160 		else
161 			__set_bit(i, &orphan_bundle->rx_mask);
162 
163 		if (test_bit(i, &asym_mask))
164 			__set_bit(i, &orphan_bundle->asym_mask);
165 	}
166 
167 	if (orphan_bundle->tx_mask || orphan_bundle->rx_mask)
168 		check_orphan_ring(accel_dev, orphan_bundle, hw_data);
169 
170 	*orphan_bundle_out = orphan_bundle;
171 	return ret;
172 }
173 
174 static void
175 put_orphan_bundle(struct bundle_orphan_ring *bundle)
176 {
177 	if (!bundle)
178 		return;
179 
180 	free(bundle, M_QAT);
181 }
182 
183 /* cleanup all ring  */
184 static void
185 cleanup_all_ring(struct adf_uio_control_accel *accel,
186 		 struct bundle_orphan_ring *orphan)
187 {
188 	int i;
189 	struct resource *csr_base = orphan->csr_base;
190 	unsigned long mask = orphan->rx_mask | orphan->tx_mask;
191 	struct adf_accel_dev *accel_dev = accel->accel_dev;
192 	struct adf_hw_csr_ops *csr_ops = GET_CSR_OPS(accel_dev);
193 	struct adf_hw_device_data *hw_data = accel_dev->hw_device;
194 	u8 num_rings_per_bank = hw_data->num_rings_per_bank;
195 	int bank = orphan->bank;
196 
197 	mutex_lock(&orphan->bundle->lock);
198 	orphan->bundle->rings_enabled &= ~mask;
199 	adf_update_uio_ring_arb(orphan->bundle);
200 	mutex_unlock(&orphan->bundle->lock);
201 
202 	for (i = 0; i < num_rings_per_bank; i++) {
203 		if (!test_bit(i, &mask))
204 			continue;
205 
206 		csr_ops->write_csr_ring_config(csr_base, bank, i, 0);
207 		csr_ops->write_csr_ring_base(csr_base, bank, i, 0);
208 	}
209 }
210 
211 /*
212  * Return true, if number of messages in tx ring is equal to number
213  * of messages in corresponding rx ring, else false.
214  */
215 static bool
216 is_all_resp_recvd(struct adf_hw_csr_ops *csr_ops,
217 		  struct bundle_orphan_ring *bundle,
218 		  const u8 num_rings_per_bank)
219 {
220 	u32 rx_tail = 0, tx_head = 0, rx_ring_msg_offset = 0,
221 	    tx_ring_msg_offset = 0, tx_rx_offset = num_rings_per_bank / 2,
222 	    idx = 0, retry = 0, delay = ADF_RING_EMPTY_RETRY_DELAY;
223 
224 	do {
225 		for_each_set_bit(idx, &bundle->tx_mask, tx_rx_offset)
226 		{
227 			rx_tail =
228 			    csr_ops->read_csr_ring_tail(bundle->csr_base,
229 							0,
230 							(idx + tx_rx_offset));
231 			tx_head = csr_ops->read_csr_ring_head(bundle->csr_base,
232 							      0,
233 							      idx);
234 
235 			/*
236 			 * Normalize messages in tx rings to match rx ring
237 			 * message size, i.e., size of response message(32).
238 			 * Asym messages are 64 bytes each, so right shift
239 			 * by 1 to normalize to 32. Sym and compression
240 			 * messages are 128 bytes each, so right shift by 2
241 			 * to normalize to 32.
242 			 */
243 			if (bundle->asym_mask & (1 << idx))
244 				tx_ring_msg_offset = (tx_head >> 1);
245 			else
246 				tx_ring_msg_offset = (tx_head >> 2);
247 
248 			rx_ring_msg_offset = rx_tail;
249 
250 			if (tx_ring_msg_offset != rx_ring_msg_offset)
251 				break;
252 		}
253 		if (idx == tx_rx_offset)
254 			/* All Tx and Rx ring message counts match */
255 			return true;
256 
257 		DELAY(delay);
258 		delay *= 2;
259 	} while (++retry < ADF_RING_EMPTY_MAX_RETRY);
260 
261 	return false;
262 }
263 
264 static int
265 bundle_need_cleanup(int bank, struct adf_uio_control_accel *accel)
266 {
267 	struct resource *csr_base = accel->bar->virt_addr;
268 	struct adf_accel_dev *accel_dev = accel->accel_dev;
269 	struct adf_hw_csr_ops *csr_ops = GET_CSR_OPS(accel_dev);
270 	struct adf_hw_device_data *hw_data = accel_dev->hw_device;
271 	u8 num_rings_per_bank = hw_data->num_rings_per_bank;
272 	int i;
273 
274 	if (!csr_base)
275 		return 0;
276 
277 	for (i = 0; i < num_rings_per_bank; i++) {
278 		if (csr_ops->read_csr_ring_base(csr_base, bank, i))
279 			return 1;
280 	}
281 
282 	return 0;
283 }
284 
285 static void
286 cleanup_orphan_ring(struct bundle_orphan_ring *orphan,
287 		    struct adf_uio_control_accel *accel)
288 {
289 	struct adf_accel_dev *accel_dev = accel->accel_dev;
290 	struct adf_hw_csr_ops *csr_ops = GET_CSR_OPS(accel_dev);
291 	struct adf_hw_device_data *hw_data = accel_dev->hw_device;
292 	u8 number_rings_per_bank = hw_data->num_rings_per_bank;
293 
294 	/* disable the interrupt */
295 	csr_ops->write_csr_int_col_en(orphan->csr_base, orphan->bank, 0);
296 
297 	/*
298 	 * wait firmware finish the in-process ring
299 	 * 1. disable all tx rings
300 	 * 2. check if all responses are received
301 	 * 3. reset all rings
302 	 */
303 	adf_disable_ring_arb(accel_dev, orphan->csr_base, 0, orphan->tx_mask);
304 
305 	if (!is_all_resp_recvd(csr_ops, orphan, number_rings_per_bank)) {
306 		device_printf(GET_DEV(accel_dev),
307 			      "Failed to clean up orphan rings");
308 		return;
309 	}
310 
311 	/*
312 	 * When the execution reaches here, it is assumed that
313 	 * there is no inflight request in the rings and that
314 	 * there is no in-process ring.
315 	 */
316 
317 	cleanup_all_ring(accel, orphan);
318 }
319 
320 void
321 adf_uio_do_cleanup_orphan(int bank, struct adf_uio_control_accel *accel)
322 {
323 	int ret;
324 	struct adf_uio_instance_rings *instance_rings, *tmp;
325 	struct adf_uio_control_bundle *bundle;
326 	/* orphan is local pointer allocated and deallocated in this function */
327 	struct bundle_orphan_ring *orphan = NULL;
328 	struct adf_accel_dev *accel_dev = accel->accel_dev;
329 	struct adf_hw_device_data *hw_data = accel_dev->hw_device;
330 
331 	if (!bundle_need_cleanup(bank, accel))
332 		goto release;
333 
334 	ret = get_orphan_bundle(bank, accel, &orphan);
335 	if (ret != 0)
336 		return;
337 
338 	/*
339 	 * If driver supports ring pair reset, no matter process
340 	 * exits normally or abnormally, just do ring pair reset.
341 	 * ring pair reset will reset all ring pair registers to
342 	 * default value. Driver only needs to reset ring mask
343 	 */
344 	if (hw_data->ring_pair_reset) {
345 		hw_data->ring_pair_reset(
346 		    accel_dev, orphan->bundle->hardware_bundle_number);
347 		/*
348 		 * If processes exit normally, rx_mask, tx_mask
349 		 * and rings_enabled are all 0, below expression
350 		 * have no impact on rings_enabled.
351 		 * If processes exit abnormally, rings_enabled
352 		 * will be set as 0 by below expression.
353 		 */
354 		orphan->bundle->rings_enabled &=
355 		    ~(orphan->rx_mask | orphan->tx_mask);
356 		goto out;
357 	}
358 
359 	if (!orphan->tx_mask && !orphan->rx_mask)
360 		goto out;
361 
362 	device_printf(GET_DEV(accel_dev),
363 		      "Process %d %s exit with orphan rings %lx:%lx\n",
364 		      curproc->p_pid,
365 		      curproc->p_comm,
366 		      orphan->tx_mask,
367 		      orphan->rx_mask);
368 
369 	if (!test_bit(ADF_STATUS_RESTARTING, &accel_dev->status)) {
370 		cleanup_orphan_ring(orphan, accel);
371 	}
372 out:
373 	put_orphan_bundle(orphan);
374 
375 release:
376 
377 	bundle = &accel->bundle[bank];
378 	/*
379 	 * If the user process died without releasing the rings
380 	 * then force a release here.
381 	 */
382 	mutex_lock(&bundle->list_lock);
383 	list_for_each_entry_safe(instance_rings, tmp, &bundle->list, list)
384 	{
385 		if (instance_rings->user_pid == curproc->p_pid) {
386 			bundle->rings_used &= ~instance_rings->ring_mask;
387 			break;
388 		}
389 	}
390 	mutex_unlock(&bundle->list_lock);
391 }
392