src/lib/ringbuffer/ring_buffer_splice.c

   1 /* SPDX-License-Identifier: GPL-2.0-only
   2  *
   3  * ring_buffer_splice.c
   4  *
   5  * Copyright (C) 2002-2005 - Tom Zanussi <zanussi@us.ibm.com>, IBM Corp
   6  * Copyright (C) 1999-2005 - Karim Yaghmour <karim@opersys.com>
   7  * Copyright (C) 2008-2012 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   8  *
   9  * Re-using code from kernel/relay.c, which is why it is licensed under
  10  * the GPL-2.0.
  11  */
  12
  13 #include <linux/module.h>
  14 #include <linux/fs.h>
  15 #include <linux/splice.h>
  16 #include <lttng/kernel-version.h>
  17
  18 #include <ringbuffer/backend.h>
  19 #include <ringbuffer/frontend.h>
  20 #include <ringbuffer/vfs.h>
  21
  22 #ifdef DEBUG
  23 #define printk_dbg(fmt, args...) printk(fmt, args)
  24 #else
  25 #define printk_dbg(fmt, args...)                        \
  26 do {                                                    \
  27         /* do nothing but check printf format */        \
  28         if (0)                                          \
  29                 printk(fmt, ## args);                   \
  30 } while (0)
  31 #endif
  32
  33 loff_t vfs_lib_ring_buffer_no_llseek(struct file *file, loff_t offset,
  34                 int origin)
  35 {
  36         return -ESPIPE;
  37 }
  38 EXPORT_SYMBOL_GPL(vfs_lib_ring_buffer_no_llseek);
  39
  40 /*
  41  * Release pages from the buffer so splice pipe_to_file can move them.
  42  * Called after the pipe has been populated with buffer pages.
  43  */
  44 static void lib_ring_buffer_pipe_buf_release(struct pipe_inode_info *pipe,
  45                                              struct pipe_buffer *pbuf)
  46 {
  47         __free_page(pbuf->page);
  48 }
  49
  50 #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(5,8,0))
  51 static const struct pipe_buf_operations ring_buffer_pipe_buf_ops = {
  52         .release = lib_ring_buffer_pipe_buf_release,
  53         .try_steal = generic_pipe_buf_try_steal,
  54         .get = generic_pipe_buf_get
  55 };
  56 #elif (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(5,1,0))
  57 static const struct pipe_buf_operations ring_buffer_pipe_buf_ops = {
  58         .confirm = generic_pipe_buf_confirm,
  59         .release = lib_ring_buffer_pipe_buf_release,
  60         .steal = generic_pipe_buf_steal,
  61         .get = generic_pipe_buf_get
  62 };
  63 #else
  64 static const struct pipe_buf_operations ring_buffer_pipe_buf_ops = {
  65         .can_merge = 0,
  66         .confirm = generic_pipe_buf_confirm,
  67         .release = lib_ring_buffer_pipe_buf_release,
  68         .steal = generic_pipe_buf_steal,
  69         .get = generic_pipe_buf_get
  70 };
  71 #endif
  72
  73 /*
  74  * Page release operation after splice pipe_to_file ends.
  75  */
  76 static void lib_ring_buffer_page_release(struct splice_pipe_desc *spd,
  77                                          unsigned int i)
  78 {
  79         __free_page(spd->pages[i]);
  80 }
  81
  82 /*
  83  *      subbuf_splice_actor - splice up to one subbuf's worth of data
  84  */
  85 static int subbuf_splice_actor(struct file *in,
  86                                loff_t *ppos,
  87                                struct pipe_inode_info *pipe,
  88                                size_t len,
  89                                unsigned int flags,
  90                                struct lttng_kernel_ring_buffer *buf)
  91 {
  92         struct lttng_kernel_ring_buffer_channel *chan = buf->backend.chan;
  93         const struct lttng_kernel_ring_buffer_config *config = &chan->backend.config;
  94         unsigned int poff, subbuf_pages, nr_pages;
  95         struct page *pages[PIPE_DEF_BUFFERS];
  96         struct partial_page partial[PIPE_DEF_BUFFERS];
  97         struct splice_pipe_desc spd = {
  98                 .pages = pages,
  99                 .nr_pages = 0,
 100                 .partial = partial,
 101 #if (LTTNG_LINUX_VERSION_CODE < LTTNG_KERNEL_VERSION(4,12,0))
 102                 .flags = flags,
 103 #endif
 104                 .ops = &ring_buffer_pipe_buf_ops,
 105                 .spd_release = lib_ring_buffer_page_release,
 106         };
 107         unsigned long consumed_old, roffset;
 108         unsigned long bytes_avail;
 109
 110         /*
 111          * Check that a GET_SUBBUF ioctl has been done before.
 112          */
 113         WARN_ON(atomic_long_read(&buf->active_readers) != 1);
 114         consumed_old = lib_ring_buffer_get_consumed(config, buf);
 115         consumed_old += *ppos;
 116
 117         /*
 118          * Adjust read len, if longer than what is available.
 119          * Max read size is 1 subbuffer due to get_subbuf/put_subbuf for
 120          * protection.
 121          */
 122         bytes_avail = chan->backend.subbuf_size;
 123         WARN_ON(bytes_avail > chan->backend.buf_size);
 124         len = min_t(size_t, len, bytes_avail);
 125         subbuf_pages = bytes_avail >> PAGE_SHIFT;
 126         nr_pages = min_t(unsigned int, subbuf_pages, PIPE_DEF_BUFFERS);
 127         roffset = consumed_old & PAGE_MASK;
 128         poff = consumed_old & ~PAGE_MASK;
 129         printk_dbg(KERN_DEBUG "LTTng: SPLICE actor len %zu pos %zd write_pos %ld\n",
 130                    len, (ssize_t)*ppos, lib_ring_buffer_get_offset(config, buf));
 131
 132         for (; spd.nr_pages < nr_pages; spd.nr_pages++) {
 133                 unsigned int this_len;
 134                 unsigned long *pfnp, new_pfn;
 135                 struct page *new_page;
 136                 void **virt;
 137
 138                 if (!len)
 139                         break;
 140                 printk_dbg(KERN_DEBUG "LTTng: SPLICE actor loop len %zu roffset %ld\n",
 141                            len, roffset);
 142
 143                 /*
 144                  * We have to replace the page we are moving into the splice
 145                  * pipe.
 146                  */
 147                 new_page = alloc_pages_node(cpu_to_node(max(buf->backend.cpu,
 148                                                             0)),
 149                                             GFP_KERNEL | __GFP_ZERO, 0);
 150                 if (!new_page)
 151                         break;
 152                 new_pfn = page_to_pfn(new_page);
 153                 this_len = PAGE_SIZE - poff;
 154                 pfnp = lib_ring_buffer_read_get_pfn(&buf->backend, roffset, &virt);
 155                 spd.pages[spd.nr_pages] = pfn_to_page(*pfnp);
 156                 *pfnp = new_pfn;
 157                 *virt = page_address(new_page);
 158                 spd.partial[spd.nr_pages].offset = poff;
 159                 spd.partial[spd.nr_pages].len = this_len;
 160
 161                 poff = 0;
 162                 roffset += PAGE_SIZE;
 163                 len -= this_len;
 164         }
 165
 166         if (!spd.nr_pages)
 167                 return 0;
 168
 169         return splice_to_pipe(pipe, &spd);
 170 }
 171
 172 ssize_t lib_ring_buffer_splice_read(struct file *in, loff_t *ppos,
 173                                     struct pipe_inode_info *pipe, size_t len,
 174                                     unsigned int flags,
 175                                     struct lttng_kernel_ring_buffer *buf)
 176 {
 177         struct lttng_kernel_ring_buffer_channel *chan = buf->backend.chan;
 178         const struct lttng_kernel_ring_buffer_config *config = &chan->backend.config;
 179         ssize_t spliced;
 180         int ret;
 181
 182         if (config->output != RING_BUFFER_SPLICE)
 183                 return -EINVAL;
 184
 185         /*
 186          * We require ppos and length to be page-aligned for performance reasons
 187          * (no page copy). Size is known using the ioctl
 188          * RING_BUFFER_GET_PADDED_SUBBUF_SIZE, which is page-size padded.
 189          * We fail when the ppos or len passed is not page-sized, because splice
 190          * is not allowed to copy more than the length passed as parameter (so
 191          * the ABI does not let us silently copy more than requested to include
 192          * padding).
 193          */
 194         if (*ppos != PAGE_ALIGN(*ppos) || len != PAGE_ALIGN(len))
 195                 return -EINVAL;
 196
 197         ret = 0;
 198         spliced = 0;
 199
 200         printk_dbg(KERN_DEBUG "LTTng: SPLICE read len %zu pos %zd\n", len,
 201                    (ssize_t)*ppos);
 202         while (len && !spliced) {
 203                 ret = subbuf_splice_actor(in, ppos, pipe, len, flags, buf);
 204                 printk_dbg(KERN_DEBUG "LTTng: SPLICE read loop ret %d\n", ret);
 205                 if (ret < 0)
 206                         break;
 207                 else if (!ret) {
 208                         if (flags & SPLICE_F_NONBLOCK)
 209                                 ret = -EAGAIN;
 210                         break;
 211                 }
 212
 213                 *ppos += ret;
 214                 if (ret > len)
 215                         len = 0;
 216                 else
 217                         len -= ret;
 218                 spliced += ret;
 219         }
 220
 221         if (spliced)
 222                 return spliced;
 223
 224         return ret;
 225 }
 226 EXPORT_SYMBOL_GPL(lib_ring_buffer_splice_read);
 227
 228 ssize_t vfs_lib_ring_buffer_splice_read(struct file *in, loff_t *ppos,
 229                                     struct pipe_inode_info *pipe, size_t len,
 230                                     unsigned int flags)
 231 {
 232         struct lttng_kernel_ring_buffer *buf = in->private_data;
 233
 234         return lib_ring_buffer_splice_read(in, ppos, pipe, len, flags, buf);
 235 }
 236 EXPORT_SYMBOL_GPL(vfs_lib_ring_buffer_splice_read);