]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - net/sunrpc/backchannel_rqst.c
net/smc: Send directly when TCP_CORK is cleared
[mirror_ubuntu-jammy-kernel.git] / net / sunrpc / backchannel_rqst.c
1 /******************************************************************************
2
3 (c) 2007 Network Appliance, Inc. All Rights Reserved.
4 (c) 2009 NetApp. All Rights Reserved.
5
6 NetApp provides this source code under the GPL v2 License.
7 The GPL v2 license is available at
8 https://opensource.org/licenses/gpl-license.php.
9
10 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
11 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
12 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
13 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
14 CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
15 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
16 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
17 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
18 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
19 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
20 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
21
22 ******************************************************************************/
23
24 #include <linux/tcp.h>
25 #include <linux/slab.h>
26 #include <linux/sunrpc/xprt.h>
27 #include <linux/export.h>
28 #include <linux/sunrpc/bc_xprt.h>
29
30 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
31 #define RPCDBG_FACILITY RPCDBG_TRANS
32 #endif
33
34 #define BC_MAX_SLOTS 64U
35
36 unsigned int xprt_bc_max_slots(struct rpc_xprt *xprt)
37 {
38 return BC_MAX_SLOTS;
39 }
40
41 /*
42 * Helper routines that track the number of preallocation elements
43 * on the transport.
44 */
45 static inline int xprt_need_to_requeue(struct rpc_xprt *xprt)
46 {
47 return xprt->bc_alloc_count < xprt->bc_alloc_max;
48 }
49
50 /*
51 * Free the preallocated rpc_rqst structure and the memory
52 * buffers hanging off of it.
53 */
54 static void xprt_free_allocation(struct rpc_rqst *req)
55 {
56 struct xdr_buf *xbufp;
57
58 dprintk("RPC: free allocations for req= %p\n", req);
59 WARN_ON_ONCE(test_bit(RPC_BC_PA_IN_USE, &req->rq_bc_pa_state));
60 xbufp = &req->rq_rcv_buf;
61 free_page((unsigned long)xbufp->head[0].iov_base);
62 xbufp = &req->rq_snd_buf;
63 free_page((unsigned long)xbufp->head[0].iov_base);
64 kfree(req);
65 }
66
67 static int xprt_alloc_xdr_buf(struct xdr_buf *buf, gfp_t gfp_flags)
68 {
69 struct page *page;
70 /* Preallocate one XDR receive buffer */
71 page = alloc_page(gfp_flags);
72 if (page == NULL)
73 return -ENOMEM;
74 xdr_buf_init(buf, page_address(page), PAGE_SIZE);
75 return 0;
76 }
77
78 static
79 struct rpc_rqst *xprt_alloc_bc_req(struct rpc_xprt *xprt, gfp_t gfp_flags)
80 {
81 struct rpc_rqst *req;
82
83 /* Pre-allocate one backchannel rpc_rqst */
84 req = kzalloc(sizeof(*req), gfp_flags);
85 if (req == NULL)
86 return NULL;
87
88 req->rq_xprt = xprt;
89 INIT_LIST_HEAD(&req->rq_bc_list);
90
91 /* Preallocate one XDR receive buffer */
92 if (xprt_alloc_xdr_buf(&req->rq_rcv_buf, gfp_flags) < 0) {
93 printk(KERN_ERR "Failed to create bc receive xbuf\n");
94 goto out_free;
95 }
96 req->rq_rcv_buf.len = PAGE_SIZE;
97
98 /* Preallocate one XDR send buffer */
99 if (xprt_alloc_xdr_buf(&req->rq_snd_buf, gfp_flags) < 0) {
100 printk(KERN_ERR "Failed to create bc snd xbuf\n");
101 goto out_free;
102 }
103 return req;
104 out_free:
105 xprt_free_allocation(req);
106 return NULL;
107 }
108
109 /*
110 * Preallocate up to min_reqs structures and related buffers for use
111 * by the backchannel. This function can be called multiple times
112 * when creating new sessions that use the same rpc_xprt. The
113 * preallocated buffers are added to the pool of resources used by
114 * the rpc_xprt. Any one of these resources may be used by an
115 * incoming callback request. It's up to the higher levels in the
116 * stack to enforce that the maximum number of session slots is not
117 * being exceeded.
118 *
119 * Some callback arguments can be large. For example, a pNFS server
120 * using multiple deviceids. The list can be unbound, but the client
121 * has the ability to tell the server the maximum size of the callback
122 * requests. Each deviceID is 16 bytes, so allocate one page
123 * for the arguments to have enough room to receive a number of these
124 * deviceIDs. The NFS client indicates to the pNFS server that its
125 * callback requests can be up to 4096 bytes in size.
126 */
127 int xprt_setup_backchannel(struct rpc_xprt *xprt, unsigned int min_reqs)
128 {
129 if (!xprt->ops->bc_setup)
130 return 0;
131 return xprt->ops->bc_setup(xprt, min_reqs);
132 }
133 EXPORT_SYMBOL_GPL(xprt_setup_backchannel);
134
135 int xprt_setup_bc(struct rpc_xprt *xprt, unsigned int min_reqs)
136 {
137 struct rpc_rqst *req;
138 struct list_head tmp_list;
139 int i;
140
141 dprintk("RPC: setup backchannel transport\n");
142
143 if (min_reqs > BC_MAX_SLOTS)
144 min_reqs = BC_MAX_SLOTS;
145
146 /*
147 * We use a temporary list to keep track of the preallocated
148 * buffers. Once we're done building the list we splice it
149 * into the backchannel preallocation list off of the rpc_xprt
150 * struct. This helps minimize the amount of time the list
151 * lock is held on the rpc_xprt struct. It also makes cleanup
152 * easier in case of memory allocation errors.
153 */
154 INIT_LIST_HEAD(&tmp_list);
155 for (i = 0; i < min_reqs; i++) {
156 /* Pre-allocate one backchannel rpc_rqst */
157 req = xprt_alloc_bc_req(xprt, GFP_KERNEL);
158 if (req == NULL) {
159 printk(KERN_ERR "Failed to create bc rpc_rqst\n");
160 goto out_free;
161 }
162
163 /* Add the allocated buffer to the tmp list */
164 dprintk("RPC: adding req= %p\n", req);
165 list_add(&req->rq_bc_pa_list, &tmp_list);
166 }
167
168 /*
169 * Add the temporary list to the backchannel preallocation list
170 */
171 spin_lock(&xprt->bc_pa_lock);
172 list_splice(&tmp_list, &xprt->bc_pa_list);
173 xprt->bc_alloc_count += min_reqs;
174 xprt->bc_alloc_max += min_reqs;
175 atomic_add(min_reqs, &xprt->bc_slot_count);
176 spin_unlock(&xprt->bc_pa_lock);
177
178 dprintk("RPC: setup backchannel transport done\n");
179 return 0;
180
181 out_free:
182 /*
183 * Memory allocation failed, free the temporary list
184 */
185 while (!list_empty(&tmp_list)) {
186 req = list_first_entry(&tmp_list,
187 struct rpc_rqst,
188 rq_bc_pa_list);
189 list_del(&req->rq_bc_pa_list);
190 xprt_free_allocation(req);
191 }
192
193 dprintk("RPC: setup backchannel transport failed\n");
194 return -ENOMEM;
195 }
196
197 /**
198 * xprt_destroy_backchannel - Destroys the backchannel preallocated structures.
199 * @xprt: the transport holding the preallocated strucures
200 * @max_reqs: the maximum number of preallocated structures to destroy
201 *
202 * Since these structures may have been allocated by multiple calls
203 * to xprt_setup_backchannel, we only destroy up to the maximum number
204 * of reqs specified by the caller.
205 */
206 void xprt_destroy_backchannel(struct rpc_xprt *xprt, unsigned int max_reqs)
207 {
208 if (xprt->ops->bc_destroy)
209 xprt->ops->bc_destroy(xprt, max_reqs);
210 }
211 EXPORT_SYMBOL_GPL(xprt_destroy_backchannel);
212
213 void xprt_destroy_bc(struct rpc_xprt *xprt, unsigned int max_reqs)
214 {
215 struct rpc_rqst *req = NULL, *tmp = NULL;
216
217 dprintk("RPC: destroy backchannel transport\n");
218
219 if (max_reqs == 0)
220 goto out;
221
222 spin_lock_bh(&xprt->bc_pa_lock);
223 xprt->bc_alloc_max -= min(max_reqs, xprt->bc_alloc_max);
224 list_for_each_entry_safe(req, tmp, &xprt->bc_pa_list, rq_bc_pa_list) {
225 dprintk("RPC: req=%p\n", req);
226 list_del(&req->rq_bc_pa_list);
227 xprt_free_allocation(req);
228 xprt->bc_alloc_count--;
229 atomic_dec(&xprt->bc_slot_count);
230 if (--max_reqs == 0)
231 break;
232 }
233 spin_unlock_bh(&xprt->bc_pa_lock);
234
235 out:
236 dprintk("RPC: backchannel list empty= %s\n",
237 list_empty(&xprt->bc_pa_list) ? "true" : "false");
238 }
239
240 static struct rpc_rqst *xprt_get_bc_request(struct rpc_xprt *xprt, __be32 xid,
241 struct rpc_rqst *new)
242 {
243 struct rpc_rqst *req = NULL;
244
245 dprintk("RPC: allocate a backchannel request\n");
246 if (list_empty(&xprt->bc_pa_list)) {
247 if (!new)
248 goto not_found;
249 if (atomic_read(&xprt->bc_slot_count) >= BC_MAX_SLOTS)
250 goto not_found;
251 list_add_tail(&new->rq_bc_pa_list, &xprt->bc_pa_list);
252 xprt->bc_alloc_count++;
253 atomic_inc(&xprt->bc_slot_count);
254 }
255 req = list_first_entry(&xprt->bc_pa_list, struct rpc_rqst,
256 rq_bc_pa_list);
257 req->rq_reply_bytes_recvd = 0;
258 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
259 sizeof(req->rq_private_buf));
260 req->rq_xid = xid;
261 req->rq_connect_cookie = xprt->connect_cookie;
262 dprintk("RPC: backchannel req=%p\n", req);
263 not_found:
264 return req;
265 }
266
267 /*
268 * Return the preallocated rpc_rqst structure and XDR buffers
269 * associated with this rpc_task.
270 */
271 void xprt_free_bc_request(struct rpc_rqst *req)
272 {
273 struct rpc_xprt *xprt = req->rq_xprt;
274
275 xprt->ops->bc_free_rqst(req);
276 }
277
278 void xprt_free_bc_rqst(struct rpc_rqst *req)
279 {
280 struct rpc_xprt *xprt = req->rq_xprt;
281
282 dprintk("RPC: free backchannel req=%p\n", req);
283
284 req->rq_connect_cookie = xprt->connect_cookie - 1;
285 smp_mb__before_atomic();
286 clear_bit(RPC_BC_PA_IN_USE, &req->rq_bc_pa_state);
287 smp_mb__after_atomic();
288
289 /*
290 * Return it to the list of preallocations so that it
291 * may be reused by a new callback request.
292 */
293 spin_lock_bh(&xprt->bc_pa_lock);
294 if (xprt_need_to_requeue(xprt)) {
295 list_add_tail(&req->rq_bc_pa_list, &xprt->bc_pa_list);
296 xprt->bc_alloc_count++;
297 atomic_inc(&xprt->bc_slot_count);
298 req = NULL;
299 }
300 spin_unlock_bh(&xprt->bc_pa_lock);
301 if (req != NULL) {
302 /*
303 * The last remaining session was destroyed while this
304 * entry was in use. Free the entry and don't attempt
305 * to add back to the list because there is no need to
306 * have anymore preallocated entries.
307 */
308 dprintk("RPC: Last session removed req=%p\n", req);
309 xprt_free_allocation(req);
310 }
311 xprt_put(xprt);
312 }
313
314 /*
315 * One or more rpc_rqst structure have been preallocated during the
316 * backchannel setup. Buffer space for the send and private XDR buffers
317 * has been preallocated as well. Use xprt_alloc_bc_request to allocate
318 * to this request. Use xprt_free_bc_request to return it.
319 *
320 * We know that we're called in soft interrupt context, grab the spin_lock
321 * since there is no need to grab the bottom half spin_lock.
322 *
323 * Return an available rpc_rqst, otherwise NULL if non are available.
324 */
325 struct rpc_rqst *xprt_lookup_bc_request(struct rpc_xprt *xprt, __be32 xid)
326 {
327 struct rpc_rqst *req, *new = NULL;
328
329 do {
330 spin_lock(&xprt->bc_pa_lock);
331 list_for_each_entry(req, &xprt->bc_pa_list, rq_bc_pa_list) {
332 if (req->rq_connect_cookie != xprt->connect_cookie)
333 continue;
334 if (req->rq_xid == xid)
335 goto found;
336 }
337 req = xprt_get_bc_request(xprt, xid, new);
338 found:
339 spin_unlock(&xprt->bc_pa_lock);
340 if (new) {
341 if (req != new)
342 xprt_free_allocation(new);
343 break;
344 } else if (req)
345 break;
346 new = xprt_alloc_bc_req(xprt, GFP_KERNEL);
347 } while (new);
348 return req;
349 }
350
351 /*
352 * Add callback request to callback list. The callback
353 * service sleeps on the sv_cb_waitq waiting for new
354 * requests. Wake it up after adding enqueing the
355 * request.
356 */
357 void xprt_complete_bc_request(struct rpc_rqst *req, uint32_t copied)
358 {
359 struct rpc_xprt *xprt = req->rq_xprt;
360 struct svc_serv *bc_serv = xprt->bc_serv;
361
362 spin_lock(&xprt->bc_pa_lock);
363 list_del(&req->rq_bc_pa_list);
364 xprt->bc_alloc_count--;
365 spin_unlock(&xprt->bc_pa_lock);
366
367 req->rq_private_buf.len = copied;
368 set_bit(RPC_BC_PA_IN_USE, &req->rq_bc_pa_state);
369
370 dprintk("RPC: add callback request to list\n");
371 xprt_get(xprt);
372 spin_lock(&bc_serv->sv_cb_lock);
373 list_add(&req->rq_bc_list, &bc_serv->sv_cb_list);
374 wake_up(&bc_serv->sv_cb_waitq);
375 spin_unlock(&bc_serv->sv_cb_lock);
376 }