xref: /spdk/module/bdev/nvme/bdev_nvme.h (revision 952532af6688423ac8aab3367c4a3ab13b87842a)
1 /*   SPDX-License-Identifier: BSD-3-Clause
2  *   Copyright (C) 2016 Intel Corporation. All rights reserved.
3  *   Copyright (c) 2019 Mellanox Technologies LTD. All rights reserved.
4  *   Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
5  *   Copyright (c) 2022 Dell Inc, or its subsidiaries. All rights reserved.
6  */
7 
8 #ifndef SPDK_BDEV_NVME_H
9 #define SPDK_BDEV_NVME_H
10 
11 #include "spdk/stdinc.h"
12 
13 #include "spdk/queue.h"
14 #include "spdk/nvme.h"
15 #include "spdk/bdev_module.h"
16 #include "spdk/module/bdev/nvme.h"
17 #include "spdk/jsonrpc.h"
18 
19 TAILQ_HEAD(nvme_bdev_ctrlrs, nvme_bdev_ctrlr);
20 extern struct nvme_bdev_ctrlrs g_nvme_bdev_ctrlrs;
21 extern pthread_mutex_t g_bdev_nvme_mutex;
22 extern bool g_bdev_nvme_module_finish;
23 extern struct spdk_thread *g_bdev_nvme_init_thread;
24 
25 #define NVME_MAX_CONTROLLERS 1024
26 
27 typedef void (*spdk_bdev_nvme_start_discovery_fn)(void *ctx, int status);
28 typedef void (*spdk_bdev_nvme_stop_discovery_fn)(void *ctx);
29 
30 struct nvme_async_probe_ctx {
31 	struct spdk_nvme_probe_ctx *probe_ctx;
32 	const char *base_name;
33 	const char **names;
34 	uint32_t max_bdevs;
35 	uint32_t reported_bdevs;
36 	struct spdk_poller *poller;
37 	struct spdk_nvme_transport_id trid;
38 	struct spdk_bdev_nvme_ctrlr_opts bdev_opts;
39 	struct spdk_nvme_ctrlr_opts drv_opts;
40 	spdk_bdev_nvme_create_cb cb_fn;
41 	void *cb_ctx;
42 	uint32_t populates_in_progress;
43 	bool ctrlr_attached;
44 	bool probe_done;
45 	bool namespaces_populated;
46 };
47 
48 struct nvme_ns {
49 	uint32_t			id;
50 	struct spdk_nvme_ns		*ns;
51 	struct nvme_ctrlr		*ctrlr;
52 	struct nvme_bdev		*bdev;
53 	uint32_t			ana_group_id;
54 	enum spdk_nvme_ana_state	ana_state;
55 	bool				ana_state_updating;
56 	bool				ana_transition_timedout;
57 	struct spdk_poller		*anatt_timer;
58 	struct nvme_async_probe_ctx	*probe_ctx;
59 	TAILQ_ENTRY(nvme_ns)		tailq;
60 	RB_ENTRY(nvme_ns)		node;
61 
62 	/**
63 	 * record io path stat before destroyed. Allocation of stat is
64 	 * decided by option io_path_stat of RPC
65 	 * bdev_nvme_set_options
66 	 */
67 	struct spdk_bdev_io_stat	*stat;
68 };
69 
70 struct nvme_bdev_io;
71 struct nvme_bdev_ctrlr;
72 struct nvme_bdev;
73 struct nvme_io_path;
74 struct nvme_ctrlr_channel_iter;
75 struct nvme_bdev_channel_iter;
76 
77 struct nvme_path_id {
78 	struct spdk_nvme_transport_id		trid;
79 	struct spdk_nvme_host_id		hostid;
80 	TAILQ_ENTRY(nvme_path_id)		link;
81 	uint64_t				last_failed_tsc;
82 };
83 
84 typedef void (*bdev_nvme_ctrlr_op_cb)(void *cb_arg, int rc);
85 typedef void (*nvme_ctrlr_disconnected_cb)(struct nvme_ctrlr *nvme_ctrlr);
86 
87 struct nvme_ctrlr {
88 	/**
89 	 * points to pinned, physically contiguous memory region;
90 	 * contains 4KB IDENTIFY structure for controller which is
91 	 *  target for CONTROLLER IDENTIFY command during initialization
92 	 */
93 	struct spdk_nvme_ctrlr			*ctrlr;
94 	struct nvme_path_id			*active_path_id;
95 	int					ref;
96 
97 	uint32_t				resetting : 1;
98 	uint32_t				reconnect_is_delayed : 1;
99 	uint32_t				in_failover : 1;
100 	uint32_t				pending_failover : 1;
101 	uint32_t				fast_io_fail_timedout : 1;
102 	uint32_t				destruct : 1;
103 	uint32_t				ana_log_page_updating : 1;
104 	uint32_t				io_path_cache_clearing : 1;
105 	uint32_t				dont_retry : 1;
106 	uint32_t				disabled : 1;
107 
108 	struct spdk_bdev_nvme_ctrlr_opts	opts;
109 
110 	RB_HEAD(nvme_ns_tree, nvme_ns)		namespaces;
111 
112 	struct spdk_opal_dev			*opal_dev;
113 
114 	struct spdk_poller			*adminq_timer_poller;
115 	struct spdk_thread			*thread;
116 
117 	bdev_nvme_ctrlr_op_cb			ctrlr_op_cb_fn;
118 	void					*ctrlr_op_cb_arg;
119 	/* Poller used to check for reset/detach completion */
120 	struct spdk_poller			*reset_detach_poller;
121 	struct spdk_nvme_detach_ctx		*detach_ctx;
122 
123 	uint64_t				reset_start_tsc;
124 	struct spdk_poller			*reconnect_delay_timer;
125 
126 	nvme_ctrlr_disconnected_cb		disconnected_cb;
127 
128 	/** linked list pointer for device list */
129 	TAILQ_ENTRY(nvme_ctrlr)			tailq;
130 	struct nvme_bdev_ctrlr			*nbdev_ctrlr;
131 
132 	TAILQ_HEAD(nvme_paths, nvme_path_id)	trids;
133 
134 	uint32_t				max_ana_log_page_size;
135 	struct spdk_nvme_ana_page		*ana_log_page;
136 	struct spdk_nvme_ana_group_descriptor	*copied_ana_desc;
137 
138 	struct nvme_async_probe_ctx		*probe_ctx;
139 	struct spdk_key				*psk;
140 	struct spdk_key				*dhchap_key;
141 	struct spdk_key				*dhchap_ctrlr_key;
142 
143 	pthread_mutex_t				mutex;
144 };
145 
146 struct nvme_bdev_ctrlr {
147 	char				*name;
148 	TAILQ_HEAD(, nvme_ctrlr)	ctrlrs;
149 	TAILQ_HEAD(, nvme_bdev)		bdevs;
150 	TAILQ_ENTRY(nvme_bdev_ctrlr)	tailq;
151 };
152 
153 struct nvme_error_stat {
154 	uint32_t status_type[8];
155 	uint32_t status[4][256];
156 };
157 
158 struct nvme_bdev {
159 	struct spdk_bdev			disk;
160 	uint32_t				nsid;
161 	struct nvme_bdev_ctrlr			*nbdev_ctrlr;
162 	pthread_mutex_t				mutex;
163 	int					ref;
164 	enum spdk_bdev_nvme_multipath_policy	mp_policy;
165 	enum spdk_bdev_nvme_multipath_selector	mp_selector;
166 	uint32_t				rr_min_io;
167 	TAILQ_HEAD(, nvme_ns)			nvme_ns_list;
168 	bool					opal;
169 	TAILQ_ENTRY(nvme_bdev)			tailq;
170 	struct nvme_error_stat			*err_stat;
171 };
172 
173 struct nvme_qpair {
174 	struct nvme_ctrlr		*ctrlr;
175 	struct spdk_nvme_qpair		*qpair;
176 	struct nvme_poll_group		*group;
177 	struct nvme_ctrlr_channel	*ctrlr_ch;
178 
179 	/* The following is used to update io_path cache of nvme_bdev_channels. */
180 	TAILQ_HEAD(, nvme_io_path)	io_path_list;
181 
182 	TAILQ_ENTRY(nvme_qpair)		tailq;
183 };
184 
185 struct nvme_ctrlr_channel {
186 	struct nvme_qpair		*qpair;
187 	TAILQ_HEAD(, nvme_bdev_io)	pending_resets;
188 
189 	struct nvme_ctrlr_channel_iter	*reset_iter;
190 	struct spdk_poller		*connect_poller;
191 };
192 
193 struct nvme_io_path {
194 	struct nvme_ns			*nvme_ns;
195 	struct nvme_qpair		*qpair;
196 	STAILQ_ENTRY(nvme_io_path)	stailq;
197 
198 	/* The following are used to update io_path cache of the nvme_bdev_channel. */
199 	struct nvme_bdev_channel	*nbdev_ch;
200 	TAILQ_ENTRY(nvme_io_path)	tailq;
201 
202 	/* allocation of stat is decided by option io_path_stat of RPC bdev_nvme_set_options */
203 	struct spdk_bdev_io_stat	*stat;
204 };
205 
206 struct nvme_bdev_channel {
207 	struct nvme_io_path			*current_io_path;
208 	enum spdk_bdev_nvme_multipath_policy	mp_policy;
209 	enum spdk_bdev_nvme_multipath_selector	mp_selector;
210 	uint32_t				rr_min_io;
211 	uint32_t				rr_counter;
212 	STAILQ_HEAD(, nvme_io_path)		io_path_list;
213 	TAILQ_HEAD(retry_io_head, nvme_bdev_io)	retry_io_list;
214 	struct spdk_poller			*retry_io_poller;
215 	bool					resetting;
216 };
217 
218 struct nvme_poll_group {
219 	struct spdk_nvme_poll_group		*group;
220 	struct spdk_io_channel			*accel_channel;
221 	struct spdk_poller			*poller;
222 	bool					collect_spin_stat;
223 	uint64_t				spin_ticks;
224 	uint64_t				start_ticks;
225 	uint64_t				end_ticks;
226 	TAILQ_HEAD(, nvme_qpair)		qpair_list;
227 };
228 
229 void nvme_io_path_info_json(struct spdk_json_write_ctx *w, struct nvme_io_path *io_path);
230 
231 struct nvme_ctrlr *nvme_ctrlr_get_by_name(const char *name);
232 
233 typedef void (*nvme_ctrlr_for_each_channel_msg)(struct nvme_ctrlr_channel_iter *iter,
234 		struct nvme_ctrlr *nvme_ctrlr,
235 		struct nvme_ctrlr_channel *ctrlr_ch,
236 		void *ctx);
237 
238 typedef void (*nvme_ctrlr_for_each_channel_done)(struct nvme_ctrlr *nvme_ctrlr,
239 		void *ctx, int status);
240 
241 void nvme_ctrlr_for_each_channel(struct nvme_ctrlr *nvme_ctrlr,
242 				 nvme_ctrlr_for_each_channel_msg fn, void *ctx,
243 				 nvme_ctrlr_for_each_channel_done cpl);
244 
245 void nvme_ctrlr_for_each_channel_continue(struct nvme_ctrlr_channel_iter *iter,
246 		int status);
247 
248 
249 typedef void (*nvme_bdev_for_each_channel_msg)(struct nvme_bdev_channel_iter *iter,
250 		struct nvme_bdev *nbdev,
251 		struct nvme_bdev_channel *nbdev_ch,
252 		void *ctx);
253 
254 typedef void (*nvme_bdev_for_each_channel_done)(struct nvme_bdev *nbdev,
255 		void *ctx, int status);
256 
257 void nvme_bdev_for_each_channel(struct nvme_bdev *nbdev,
258 				nvme_bdev_for_each_channel_msg fn, void *ctx,
259 				nvme_bdev_for_each_channel_done cpl);
260 
261 void nvme_bdev_for_each_channel_continue(struct nvme_bdev_channel_iter *iter,
262 		int status);
263 
264 struct nvme_ctrlr *nvme_bdev_ctrlr_get_ctrlr_by_id(struct nvme_bdev_ctrlr *nbdev_ctrlr,
265 		uint16_t cntlid);
266 
267 struct nvme_bdev_ctrlr *nvme_bdev_ctrlr_get_by_name(const char *name);
268 
269 typedef void (*nvme_bdev_ctrlr_for_each_fn)(struct nvme_bdev_ctrlr *nbdev_ctrlr, void *ctx);
270 
271 void nvme_bdev_ctrlr_for_each(nvme_bdev_ctrlr_for_each_fn fn, void *ctx);
272 
273 void nvme_bdev_dump_trid_json(const struct spdk_nvme_transport_id *trid,
274 			      struct spdk_json_write_ctx *w);
275 
276 void nvme_ctrlr_info_json(struct spdk_json_write_ctx *w, struct nvme_ctrlr *nvme_ctrlr);
277 
278 struct nvme_ns *nvme_ctrlr_get_ns(struct nvme_ctrlr *nvme_ctrlr, uint32_t nsid);
279 struct nvme_ns *nvme_ctrlr_get_first_active_ns(struct nvme_ctrlr *nvme_ctrlr);
280 struct nvme_ns *nvme_ctrlr_get_next_active_ns(struct nvme_ctrlr *nvme_ctrlr, struct nvme_ns *ns);
281 
282 enum spdk_bdev_timeout_action {
283 	SPDK_BDEV_NVME_TIMEOUT_ACTION_NONE = 0,
284 	SPDK_BDEV_NVME_TIMEOUT_ACTION_RESET,
285 	SPDK_BDEV_NVME_TIMEOUT_ACTION_ABORT,
286 };
287 
288 struct spdk_bdev_nvme_opts {
289 	enum spdk_bdev_timeout_action action_on_timeout;
290 	uint64_t timeout_us;
291 	uint64_t timeout_admin_us;
292 	uint32_t keep_alive_timeout_ms;
293 	/* The number of attempts per I/O in the transport layer before an I/O fails. */
294 	uint32_t transport_retry_count;
295 	uint32_t arbitration_burst;
296 	uint32_t low_priority_weight;
297 	uint32_t medium_priority_weight;
298 	uint32_t high_priority_weight;
299 	uint64_t nvme_adminq_poll_period_us;
300 	uint64_t nvme_ioq_poll_period_us;
301 	uint32_t io_queue_requests;
302 	bool delay_cmd_submit;
303 	/* The number of attempts per I/O in the bdev layer before an I/O fails. */
304 	int32_t bdev_retry_count;
305 	uint8_t transport_ack_timeout;
306 	int32_t ctrlr_loss_timeout_sec;
307 	uint32_t reconnect_delay_sec;
308 	uint32_t fast_io_fail_timeout_sec;
309 	bool disable_auto_failback;
310 	bool generate_uuids;
311 	/* Type of Service - RDMA only */
312 	uint8_t transport_tos;
313 	bool nvme_error_stat;
314 	uint32_t rdma_srq_size;
315 	bool io_path_stat;
316 	bool allow_accel_sequence;
317 	uint32_t rdma_max_cq_size;
318 	uint16_t rdma_cm_event_timeout_ms;
319 	uint32_t dhchap_digests;
320 	uint32_t dhchap_dhgroups;
321 };
322 
323 struct spdk_nvme_qpair *bdev_nvme_get_io_qpair(struct spdk_io_channel *ctrlr_io_ch);
324 void bdev_nvme_get_opts(struct spdk_bdev_nvme_opts *opts);
325 int bdev_nvme_set_opts(const struct spdk_bdev_nvme_opts *opts);
326 int bdev_nvme_set_hotplug(bool enabled, uint64_t period_us, spdk_msg_fn cb, void *cb_ctx);
327 
328 int bdev_nvme_start_discovery(struct spdk_nvme_transport_id *trid, const char *base_name,
329 			      struct spdk_nvme_ctrlr_opts *drv_opts, struct spdk_bdev_nvme_ctrlr_opts *bdev_opts,
330 			      uint64_t timeout, bool from_mdns,
331 			      spdk_bdev_nvme_start_discovery_fn cb_fn, void *cb_ctx);
332 int bdev_nvme_stop_discovery(const char *name, spdk_bdev_nvme_stop_discovery_fn cb_fn,
333 			     void *cb_ctx);
334 void bdev_nvme_get_discovery_info(struct spdk_json_write_ctx *w);
335 
336 int bdev_nvme_start_mdns_discovery(const char *base_name,
337 				   const char *svcname,
338 				   struct spdk_nvme_ctrlr_opts *drv_opts,
339 				   struct spdk_bdev_nvme_ctrlr_opts *bdev_opts);
340 int bdev_nvme_stop_mdns_discovery(const char *name);
341 void bdev_nvme_get_mdns_discovery_info(struct spdk_jsonrpc_request *request);
342 void bdev_nvme_mdns_discovery_config_json(struct spdk_json_write_ctx *w);
343 
344 typedef void (*bdev_nvme_set_keys_cb)(void *ctx, int status);
345 
346 int bdev_nvme_set_keys(const char *name, const char *dhchap_key, const char *dhchap_ctrlr_key,
347 		       bdev_nvme_set_keys_cb cb_fn, void *cb_ctx);
348 
349 struct spdk_nvme_ctrlr *bdev_nvme_get_ctrlr(struct spdk_bdev *bdev);
350 
351 typedef void (*bdev_nvme_delete_done_fn)(void *ctx, int rc);
352 
353 /**
354  * Delete NVMe controller with all bdevs on top of it, or delete the specified path
355  * if there is any alternative path. Requires to pass name of NVMe controller.
356  *
357  * \param name NVMe controller name
358  * \param path_id The specified path to remove (optional)
359  * \param delete_done Callback function on delete complete (optional)
360  * \param delete_done_ctx Context passed to callback (optional)
361  * \return zero on success,
362  *		-EINVAL on wrong parameters or
363  *		-ENODEV if controller is not found or
364  *		-ENOMEM on no memory
365  */
366 int bdev_nvme_delete(const char *name, const struct nvme_path_id *path_id,
367 		     bdev_nvme_delete_done_fn delete_done, void *delete_done_ctx);
368 
369 enum nvme_ctrlr_op {
370 	NVME_CTRLR_OP_RESET = 1,
371 	NVME_CTRLR_OP_ENABLE,
372 	NVME_CTRLR_OP_DISABLE,
373 };
374 
375 /**
376  * Perform specified operation on an NVMe controller.
377  *
378  * NOTE: The callback function is always called after this function returns except for
379  * out of memory cases.
380  *
381  * \param nvme_ctrlr The specified NVMe controller to operate
382  * \param op Operation code
383  * \param cb_fn Function to be called back after operation completes
384  * \param cb_arg Argument for callback function
385  */
386 void nvme_ctrlr_op_rpc(struct nvme_ctrlr *nvme_ctrlr, enum nvme_ctrlr_op op,
387 		       bdev_nvme_ctrlr_op_cb cb_fn, void *cb_arg);
388 
389 /**
390  * Perform specified operation on all NVMe controllers in an NVMe bdev controller.
391  *
392  * NOTE: The callback function is always called after this function returns except for
393  * out of memory cases.
394  *
395  * \param nbdev_ctrlr The specified NVMe bdev controller to operate
396  * \param op Operation code
397  * \param cb_fn Function to be called back after operation completes
398  * \param cb_arg Argument for callback function
399  */
400 void nvme_bdev_ctrlr_op_rpc(struct nvme_bdev_ctrlr *nbdev_ctrlr, enum nvme_ctrlr_op op,
401 			    bdev_nvme_ctrlr_op_cb cb_fn, void *cb_arg);
402 
403 typedef void (*bdev_nvme_set_preferred_path_cb)(void *cb_arg, int rc);
404 
405 /**
406  * Set the preferred I/O path for an NVMe bdev in multipath mode.
407  *
408  * NOTE: This function does not support NVMe bdevs in failover mode.
409  *
410  * \param name NVMe bdev name
411  * \param cntlid NVMe-oF controller ID
412  * \param cb_fn Function to be called back after completion.
413  * \param cb_arg Argument for callback function.
414  */
415 void bdev_nvme_set_preferred_path(const char *name, uint16_t cntlid,
416 				  bdev_nvme_set_preferred_path_cb cb_fn, void *cb_arg);
417 
418 #endif /* SPDK_BDEV_NVME_H */
419