sys/kern/tty_outq.c

   1 /*-
   2  * Copyright (c) 2008 Ed Schouten <ed@FreeBSD.org>
   3  * All rights reserved.
   4  *
   5  * Portions of this software were developed under sponsorship from Snow
   6  * B.V., the Netherlands.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 #include <sys/cdefs.h>
  31 __FBSDID("$FreeBSD$");
  32
  33 #include <sys/param.h>
  34 #include <sys/kernel.h>
  35 #include <sys/lock.h>
  36 #include <sys/queue.h>
  37 #include <sys/sysctl.h>
  38 #include <sys/systm.h>
  39 #include <sys/tty.h>
  40 #include <sys/uio.h>
  41
  42 #include <vm/uma.h>
  43
  44 /*
  45  * TTY output queue buffering.
  46  *
  47  * The previous design of the TTY layer offered the so-called clists.
  48  * These clists were used for both the input queues and the output
  49  * queue. We don't use certain features on the output side, like quoting
  50  * bits for parity marking and such. This mechanism is similar to the
  51  * old clists, but only contains the features we need to buffer the
  52  * output.
  53  */
  54
  55 /* Statistics. */
  56 static unsigned long ttyoutq_nfast = 0;
  57 SYSCTL_ULONG(_kern, OID_AUTO, tty_outq_nfast, CTLFLAG_RD,
  58         &ttyoutq_nfast, 0, "Unbuffered reads to userspace on output");
  59 static unsigned long ttyoutq_nslow = 0;
  60 SYSCTL_ULONG(_kern, OID_AUTO, tty_outq_nslow, CTLFLAG_RD,
  61         &ttyoutq_nslow, 0, "Buffered reads to userspace on output");
  62
  63 struct ttyoutq_block {
  64         struct ttyoutq_block    *tob_next;
  65         char                    tob_data[TTYOUTQ_DATASIZE];
  66 };
  67
  68 static uma_zone_t ttyoutq_zone;
  69
  70 #define TTYOUTQ_INSERT_TAIL(to, tob) do {                               \
  71         if (to->to_end == 0) {                                          \
  72                 tob->tob_next = to->to_firstblock;                      \
  73                 to->to_firstblock = tob;                                \
  74         } else {                                                        \
  75                 tob->tob_next = to->to_lastblock->tob_next;             \
  76                 to->to_lastblock->tob_next = tob;                       \
  77         }                                                               \
  78         to->to_nblocks++;                                               \
  79 } while (0)
  80
  81 #define TTYOUTQ_REMOVE_HEAD(to) do {                                    \
  82         to->to_firstblock = to->to_firstblock->tob_next;                \
  83         to->to_nblocks--;                                               \
  84 } while (0)
  85
  86 #define TTYOUTQ_RECYCLE(to, tob) do {                                   \
  87         if (to->to_quota <= to->to_nblocks)                             \
  88                 uma_zfree(ttyoutq_zone, tob);                           \
  89         else                                                            \
  90                 TTYOUTQ_INSERT_TAIL(to, tob);                           \
  91 } while(0)
  92
  93 void
  94 ttyoutq_flush(struct ttyoutq *to)
  95 {
  96
  97         to->to_begin = 0;
  98         to->to_end = 0;
  99 }
 100
 101 void
 102 ttyoutq_setsize(struct ttyoutq *to, struct tty *tp, size_t size)
 103 {
 104         struct ttyoutq_block *tob;
 105
 106         to->to_quota = howmany(size, TTYOUTQ_DATASIZE);
 107
 108         while (to->to_quota > to->to_nblocks) {
 109                 /*
 110                  * List is getting bigger.
 111                  * Add new blocks to the tail of the list.
 112                  *
 113                  * We must unlock the TTY temporarily, because we need
 114                  * to allocate memory. This won't be a problem, because
 115                  * in the worst case, another thread ends up here, which
 116                  * may cause us to allocate too many blocks, but this
 117                  * will be caught by the loop below.
 118                  */
 119                 tty_unlock(tp);
 120                 tob = uma_zalloc(ttyoutq_zone, M_WAITOK);
 121                 tty_lock(tp);
 122
 123                 TTYOUTQ_INSERT_TAIL(to, tob);
 124         }
 125 }
 126
 127 void
 128 ttyoutq_free(struct ttyoutq *to)
 129 {
 130         struct ttyoutq_block *tob;
 131
 132         ttyoutq_flush(to);
 133         to->to_quota = 0;
 134
 135         while ((tob = to->to_firstblock) != NULL) {
 136                 TTYOUTQ_REMOVE_HEAD(to);
 137                 uma_zfree(ttyoutq_zone, tob);
 138         }
 139
 140         MPASS(to->to_nblocks == 0);
 141 }
 142
 143 size_t
 144 ttyoutq_read(struct ttyoutq *to, void *buf, size_t len)
 145 {
 146         char *cbuf = buf;
 147
 148         while (len > 0) {
 149                 struct ttyoutq_block *tob;
 150                 size_t cbegin, cend, clen;
 151
 152                 /* See if there still is data. */
 153                 if (to->to_begin == to->to_end)
 154                         break;
 155                 tob = to->to_firstblock;
 156                 if (tob == NULL)
 157                         break;
 158
 159                 /*
 160                  * The end address should be the lowest of these three:
 161                  * - The write pointer
 162                  * - The blocksize - we can't read beyond the block
 163                  * - The end address if we could perform the full read
 164                  */
 165                 cbegin = to->to_begin;
 166                 cend = MIN(MIN(to->to_end, to->to_begin + len),
 167                     TTYOUTQ_DATASIZE);
 168                 clen = cend - cbegin;
 169
 170                 /* Copy the data out of the buffers. */
 171                 memcpy(cbuf, tob->tob_data + cbegin, clen);
 172                 cbuf += clen;
 173                 len -= clen;
 174
 175                 if (cend == to->to_end) {
 176                         /* Read the complete queue. */
 177                         to->to_begin = 0;
 178                         to->to_end = 0;
 179                 } else if (cend == TTYOUTQ_DATASIZE) {
 180                         /* Read the block until the end. */
 181                         TTYOUTQ_REMOVE_HEAD(to);
 182                         to->to_begin = 0;
 183                         to->to_end -= TTYOUTQ_DATASIZE;
 184                         TTYOUTQ_RECYCLE(to, tob);
 185                 } else {
 186                         /* Read the block partially. */
 187                         to->to_begin += clen;
 188                 }
 189         }
 190
 191         return (cbuf - (char *)buf);
 192 }
 193
 194 /*
 195  * An optimized version of ttyoutq_read() which can be used in pseudo
 196  * TTY drivers to directly copy data from the outq to userspace, instead
 197  * of buffering it.
 198  *
 199  * We can only copy data directly if we need to read the entire block
 200  * back to the user, because we temporarily remove the block from the
 201  * queue. Otherwise we need to copy it to a temporary buffer first, to
 202  * make sure data remains in the correct order.
 203  */
 204 int
 205 ttyoutq_read_uio(struct ttyoutq *to, struct tty *tp, struct uio *uio)
 206 {
 207
 208         while (uio->uio_resid > 0) {
 209                 int error;
 210                 struct ttyoutq_block *tob;
 211                 size_t cbegin, cend, clen;
 212
 213                 /* See if there still is data. */
 214                 if (to->to_begin == to->to_end)
 215                         return (0);
 216                 tob = to->to_firstblock;
 217                 if (tob == NULL)
 218                         return (0);
 219
 220                 /*
 221                  * The end address should be the lowest of these three:
 222                  * - The write pointer
 223                  * - The blocksize - we can't read beyond the block
 224                  * - The end address if we could perform the full read
 225                  */
 226                 cbegin = to->to_begin;
 227                 cend = MIN(MIN(to->to_end, to->to_begin + uio->uio_resid),
 228                     TTYOUTQ_DATASIZE);
 229                 clen = cend - cbegin;
 230
 231                 /*
 232                  * We can prevent buffering in some cases:
 233                  * - We need to read the block until the end.
 234                  * - We don't need to read the block until the end, but
 235                  *   there is no data beyond it, which allows us to move
 236                  *   the write pointer to a new block.
 237                  */
 238                 if (cend == TTYOUTQ_DATASIZE || cend == to->to_end) {
 239                         atomic_add_long(&ttyoutq_nfast, 1);
 240
 241                         /*
 242                          * Fast path: zero copy. Remove the first block,
 243                          * so we can unlock the TTY temporarily.
 244                          */
 245                         TTYOUTQ_REMOVE_HEAD(to);
 246                         to->to_begin = 0;
 247                         if (to->to_end <= TTYOUTQ_DATASIZE)
 248                                 to->to_end = 0;
 249                         else
 250                                 to->to_end -= TTYOUTQ_DATASIZE;
 251
 252                         /* Temporary unlock and copy the data to userspace. */
 253                         tty_unlock(tp);
 254                         error = uiomove(tob->tob_data + cbegin, clen, uio);
 255                         tty_lock(tp);
 256
 257                         /* Block can now be readded to the list. */
 258                         TTYOUTQ_RECYCLE(to, tob);
 259                 } else {
 260                         char ob[TTYOUTQ_DATASIZE - 1];
 261                         atomic_add_long(&ttyoutq_nslow, 1);
 262
 263                         /*
 264                          * Slow path: store data in a temporary buffer.
 265                          */
 266                         memcpy(ob, tob->tob_data + cbegin, clen);
 267                         to->to_begin += clen;
 268                         MPASS(to->to_begin < TTYOUTQ_DATASIZE);
 269
 270                         /* Temporary unlock and copy the data to userspace. */
 271                         tty_unlock(tp);
 272                         error = uiomove(ob, clen, uio);
 273                         tty_lock(tp);
 274                 }
 275
 276                 if (error != 0)
 277                         return (error);
 278         }
 279
 280         return (0);
 281 }
 282
 283 size_t
 284 ttyoutq_write(struct ttyoutq *to, const void *buf, size_t nbytes)
 285 {
 286         const char *cbuf = buf;
 287         struct ttyoutq_block *tob;
 288         unsigned int boff;
 289         size_t l;
 290
 291         while (nbytes > 0) {
 292                 boff = to->to_end % TTYOUTQ_DATASIZE;
 293
 294                 if (to->to_end == 0) {
 295                         /* First time we're being used or drained. */
 296                         MPASS(to->to_begin == 0);
 297                         tob = to->to_firstblock;
 298                         if (tob == NULL) {
 299                                 /* Queue has no blocks. */
 300                                 break;
 301                         }
 302                         to->to_lastblock = tob;
 303                 } else if (boff == 0) {
 304                         /* We reached the end of this block on last write. */
 305                         tob = to->to_lastblock->tob_next;
 306                         if (tob == NULL) {
 307                                 /* We've reached the watermark. */
 308                                 break;
 309                         }
 310                         to->to_lastblock = tob;
 311                 } else {
 312                         tob = to->to_lastblock;
 313                 }
 314
 315                 /* Don't copy more than was requested. */
 316                 l = MIN(nbytes, TTYOUTQ_DATASIZE - boff);
 317                 MPASS(l > 0);
 318                 memcpy(tob->tob_data + boff, cbuf, l);
 319
 320                 cbuf += l;
 321                 nbytes -= l;
 322                 to->to_end += l;
 323         }
 324
 325         return (cbuf - (const char *)buf);
 326 }
 327
 328 int
 329 ttyoutq_write_nofrag(struct ttyoutq *to, const void *buf, size_t nbytes)
 330 {
 331         size_t ret;
 332
 333         if (ttyoutq_bytesleft(to) < nbytes)
 334                 return (-1);
 335
 336         /* We should always be able to write it back. */
 337         ret = ttyoutq_write(to, buf, nbytes);
 338         MPASS(ret == nbytes);
 339
 340         return (0);
 341 }
 342
 343 static void
 344 ttyoutq_startup(void *dummy)
 345 {
 346
 347         ttyoutq_zone = uma_zcreate("ttyoutq", sizeof(struct ttyoutq_block),
 348             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
 349 }
 350
 351 SYSINIT(ttyoutq, SI_SUB_DRIVERS, SI_ORDER_FIRST, ttyoutq_startup, NULL);