Back to home page

OSCL-LXR
 
 

     

0001 /*
0002  * Copyright (C) 2003 Sistina Software
0003  * Copyright (C) 2006 Red Hat GmbH
0004  *
0005  * This file is released under the GPL.
0006  */
0007 
0008 #include "dm-core.h"
0009 
0010 #include <linux/device-mapper.h>
0011 
0012 #include <linux/bio.h>
0013 #include <linux/completion.h>
0014 #include <linux/mempool.h>
0015 #include <linux/module.h>
0016 #include <linux/sched.h>
0017 #include <linux/slab.h>
0018 #include <linux/dm-io.h>
0019 
0020 #define DM_MSG_PREFIX "io"
0021 
0022 #define DM_IO_MAX_REGIONS   BITS_PER_LONG
0023 
0024 struct dm_io_client {
0025     mempool_t pool;
0026     struct bio_set bios;
0027 };
0028 
0029 /*
0030  * Aligning 'struct io' reduces the number of bits required to store
0031  * its address.  Refer to store_io_and_region_in_bio() below.
0032  */
0033 struct io {
0034     unsigned long error_bits;
0035     atomic_t count;
0036     struct dm_io_client *client;
0037     io_notify_fn callback;
0038     void *context;
0039     void *vma_invalidate_address;
0040     unsigned long vma_invalidate_size;
0041 } __attribute__((aligned(DM_IO_MAX_REGIONS)));
0042 
0043 static struct kmem_cache *_dm_io_cache;
0044 
0045 /*
0046  * Create a client with mempool and bioset.
0047  */
0048 struct dm_io_client *dm_io_client_create(void)
0049 {
0050     struct dm_io_client *client;
0051     unsigned min_ios = dm_get_reserved_bio_based_ios();
0052     int ret;
0053 
0054     client = kzalloc(sizeof(*client), GFP_KERNEL);
0055     if (!client)
0056         return ERR_PTR(-ENOMEM);
0057 
0058     ret = mempool_init_slab_pool(&client->pool, min_ios, _dm_io_cache);
0059     if (ret)
0060         goto bad;
0061 
0062     ret = bioset_init(&client->bios, min_ios, 0, BIOSET_NEED_BVECS);
0063     if (ret)
0064         goto bad;
0065 
0066     return client;
0067 
0068    bad:
0069     mempool_exit(&client->pool);
0070     kfree(client);
0071     return ERR_PTR(ret);
0072 }
0073 EXPORT_SYMBOL(dm_io_client_create);
0074 
0075 void dm_io_client_destroy(struct dm_io_client *client)
0076 {
0077     mempool_exit(&client->pool);
0078     bioset_exit(&client->bios);
0079     kfree(client);
0080 }
0081 EXPORT_SYMBOL(dm_io_client_destroy);
0082 
0083 /*-----------------------------------------------------------------
0084  * We need to keep track of which region a bio is doing io for.
0085  * To avoid a memory allocation to store just 5 or 6 bits, we
0086  * ensure the 'struct io' pointer is aligned so enough low bits are
0087  * always zero and then combine it with the region number directly in
0088  * bi_private.
0089  *---------------------------------------------------------------*/
0090 static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
0091                        unsigned region)
0092 {
0093     if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
0094         DMCRIT("Unaligned struct io pointer %p", io);
0095         BUG();
0096     }
0097 
0098     bio->bi_private = (void *)((unsigned long)io | region);
0099 }
0100 
0101 static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
0102                        unsigned *region)
0103 {
0104     unsigned long val = (unsigned long)bio->bi_private;
0105 
0106     *io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
0107     *region = val & (DM_IO_MAX_REGIONS - 1);
0108 }
0109 
0110 /*-----------------------------------------------------------------
0111  * We need an io object to keep track of the number of bios that
0112  * have been dispatched for a particular io.
0113  *---------------------------------------------------------------*/
0114 static void complete_io(struct io *io)
0115 {
0116     unsigned long error_bits = io->error_bits;
0117     io_notify_fn fn = io->callback;
0118     void *context = io->context;
0119 
0120     if (io->vma_invalidate_size)
0121         invalidate_kernel_vmap_range(io->vma_invalidate_address,
0122                          io->vma_invalidate_size);
0123 
0124     mempool_free(io, &io->client->pool);
0125     fn(error_bits, context);
0126 }
0127 
0128 static void dec_count(struct io *io, unsigned int region, blk_status_t error)
0129 {
0130     if (error)
0131         set_bit(region, &io->error_bits);
0132 
0133     if (atomic_dec_and_test(&io->count))
0134         complete_io(io);
0135 }
0136 
0137 static void endio(struct bio *bio)
0138 {
0139     struct io *io;
0140     unsigned region;
0141     blk_status_t error;
0142 
0143     if (bio->bi_status && bio_data_dir(bio) == READ)
0144         zero_fill_bio(bio);
0145 
0146     /*
0147      * The bio destructor in bio_put() may use the io object.
0148      */
0149     retrieve_io_and_region_from_bio(bio, &io, &region);
0150 
0151     error = bio->bi_status;
0152     bio_put(bio);
0153 
0154     dec_count(io, region, error);
0155 }
0156 
0157 /*-----------------------------------------------------------------
0158  * These little objects provide an abstraction for getting a new
0159  * destination page for io.
0160  *---------------------------------------------------------------*/
0161 struct dpages {
0162     void (*get_page)(struct dpages *dp,
0163              struct page **p, unsigned long *len, unsigned *offset);
0164     void (*next_page)(struct dpages *dp);
0165 
0166     union {
0167         unsigned context_u;
0168         struct bvec_iter context_bi;
0169     };
0170     void *context_ptr;
0171 
0172     void *vma_invalidate_address;
0173     unsigned long vma_invalidate_size;
0174 };
0175 
0176 /*
0177  * Functions for getting the pages from a list.
0178  */
0179 static void list_get_page(struct dpages *dp,
0180           struct page **p, unsigned long *len, unsigned *offset)
0181 {
0182     unsigned o = dp->context_u;
0183     struct page_list *pl = (struct page_list *) dp->context_ptr;
0184 
0185     *p = pl->page;
0186     *len = PAGE_SIZE - o;
0187     *offset = o;
0188 }
0189 
0190 static void list_next_page(struct dpages *dp)
0191 {
0192     struct page_list *pl = (struct page_list *) dp->context_ptr;
0193     dp->context_ptr = pl->next;
0194     dp->context_u = 0;
0195 }
0196 
0197 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
0198 {
0199     dp->get_page = list_get_page;
0200     dp->next_page = list_next_page;
0201     dp->context_u = offset;
0202     dp->context_ptr = pl;
0203 }
0204 
0205 /*
0206  * Functions for getting the pages from a bvec.
0207  */
0208 static void bio_get_page(struct dpages *dp, struct page **p,
0209              unsigned long *len, unsigned *offset)
0210 {
0211     struct bio_vec bvec = bvec_iter_bvec((struct bio_vec *)dp->context_ptr,
0212                          dp->context_bi);
0213 
0214     *p = bvec.bv_page;
0215     *len = bvec.bv_len;
0216     *offset = bvec.bv_offset;
0217 
0218     /* avoid figuring it out again in bio_next_page() */
0219     dp->context_bi.bi_sector = (sector_t)bvec.bv_len;
0220 }
0221 
0222 static void bio_next_page(struct dpages *dp)
0223 {
0224     unsigned int len = (unsigned int)dp->context_bi.bi_sector;
0225 
0226     bvec_iter_advance((struct bio_vec *)dp->context_ptr,
0227               &dp->context_bi, len);
0228 }
0229 
0230 static void bio_dp_init(struct dpages *dp, struct bio *bio)
0231 {
0232     dp->get_page = bio_get_page;
0233     dp->next_page = bio_next_page;
0234 
0235     /*
0236      * We just use bvec iterator to retrieve pages, so it is ok to
0237      * access the bvec table directly here
0238      */
0239     dp->context_ptr = bio->bi_io_vec;
0240     dp->context_bi = bio->bi_iter;
0241 }
0242 
0243 /*
0244  * Functions for getting the pages from a VMA.
0245  */
0246 static void vm_get_page(struct dpages *dp,
0247          struct page **p, unsigned long *len, unsigned *offset)
0248 {
0249     *p = vmalloc_to_page(dp->context_ptr);
0250     *offset = dp->context_u;
0251     *len = PAGE_SIZE - dp->context_u;
0252 }
0253 
0254 static void vm_next_page(struct dpages *dp)
0255 {
0256     dp->context_ptr += PAGE_SIZE - dp->context_u;
0257     dp->context_u = 0;
0258 }
0259 
0260 static void vm_dp_init(struct dpages *dp, void *data)
0261 {
0262     dp->get_page = vm_get_page;
0263     dp->next_page = vm_next_page;
0264     dp->context_u = offset_in_page(data);
0265     dp->context_ptr = data;
0266 }
0267 
0268 /*
0269  * Functions for getting the pages from kernel memory.
0270  */
0271 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
0272             unsigned *offset)
0273 {
0274     *p = virt_to_page(dp->context_ptr);
0275     *offset = dp->context_u;
0276     *len = PAGE_SIZE - dp->context_u;
0277 }
0278 
0279 static void km_next_page(struct dpages *dp)
0280 {
0281     dp->context_ptr += PAGE_SIZE - dp->context_u;
0282     dp->context_u = 0;
0283 }
0284 
0285 static void km_dp_init(struct dpages *dp, void *data)
0286 {
0287     dp->get_page = km_get_page;
0288     dp->next_page = km_next_page;
0289     dp->context_u = offset_in_page(data);
0290     dp->context_ptr = data;
0291 }
0292 
0293 /*-----------------------------------------------------------------
0294  * IO routines that accept a list of pages.
0295  *---------------------------------------------------------------*/
0296 static void do_region(const blk_opf_t opf, unsigned region,
0297               struct dm_io_region *where, struct dpages *dp,
0298               struct io *io)
0299 {
0300     struct bio *bio;
0301     struct page *page;
0302     unsigned long len;
0303     unsigned offset;
0304     unsigned num_bvecs;
0305     sector_t remaining = where->count;
0306     struct request_queue *q = bdev_get_queue(where->bdev);
0307     sector_t num_sectors;
0308     unsigned int special_cmd_max_sectors;
0309     const enum req_op op = opf & REQ_OP_MASK;
0310 
0311     /*
0312      * Reject unsupported discard and write same requests.
0313      */
0314     if (op == REQ_OP_DISCARD)
0315         special_cmd_max_sectors = bdev_max_discard_sectors(where->bdev);
0316     else if (op == REQ_OP_WRITE_ZEROES)
0317         special_cmd_max_sectors = q->limits.max_write_zeroes_sectors;
0318     if ((op == REQ_OP_DISCARD || op == REQ_OP_WRITE_ZEROES) &&
0319         special_cmd_max_sectors == 0) {
0320         atomic_inc(&io->count);
0321         dec_count(io, region, BLK_STS_NOTSUPP);
0322         return;
0323     }
0324 
0325     /*
0326      * where->count may be zero if op holds a flush and we need to
0327      * send a zero-sized flush.
0328      */
0329     do {
0330         /*
0331          * Allocate a suitably sized-bio.
0332          */
0333         switch (op) {
0334         case REQ_OP_DISCARD:
0335         case REQ_OP_WRITE_ZEROES:
0336             num_bvecs = 0;
0337             break;
0338         default:
0339             num_bvecs = bio_max_segs(dm_sector_div_up(remaining,
0340                         (PAGE_SIZE >> SECTOR_SHIFT)));
0341         }
0342 
0343         bio = bio_alloc_bioset(where->bdev, num_bvecs, opf, GFP_NOIO,
0344                        &io->client->bios);
0345         bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
0346         bio->bi_end_io = endio;
0347         store_io_and_region_in_bio(bio, io, region);
0348 
0349         if (op == REQ_OP_DISCARD || op == REQ_OP_WRITE_ZEROES) {
0350             num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
0351             bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
0352             remaining -= num_sectors;
0353         } else while (remaining) {
0354             /*
0355              * Try and add as many pages as possible.
0356              */
0357             dp->get_page(dp, &page, &len, &offset);
0358             len = min(len, to_bytes(remaining));
0359             if (!bio_add_page(bio, page, len, offset))
0360                 break;
0361 
0362             offset = 0;
0363             remaining -= to_sector(len);
0364             dp->next_page(dp);
0365         }
0366 
0367         atomic_inc(&io->count);
0368         submit_bio(bio);
0369     } while (remaining);
0370 }
0371 
0372 static void dispatch_io(blk_opf_t opf, unsigned int num_regions,
0373             struct dm_io_region *where, struct dpages *dp,
0374             struct io *io, int sync)
0375 {
0376     int i;
0377     struct dpages old_pages = *dp;
0378 
0379     BUG_ON(num_regions > DM_IO_MAX_REGIONS);
0380 
0381     if (sync)
0382         opf |= REQ_SYNC;
0383 
0384     /*
0385      * For multiple regions we need to be careful to rewind
0386      * the dp object for each call to do_region.
0387      */
0388     for (i = 0; i < num_regions; i++) {
0389         *dp = old_pages;
0390         if (where[i].count || (opf & REQ_PREFLUSH))
0391             do_region(opf, i, where + i, dp, io);
0392     }
0393 
0394     /*
0395      * Drop the extra reference that we were holding to avoid
0396      * the io being completed too early.
0397      */
0398     dec_count(io, 0, 0);
0399 }
0400 
0401 struct sync_io {
0402     unsigned long error_bits;
0403     struct completion wait;
0404 };
0405 
0406 static void sync_io_complete(unsigned long error, void *context)
0407 {
0408     struct sync_io *sio = context;
0409 
0410     sio->error_bits = error;
0411     complete(&sio->wait);
0412 }
0413 
0414 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
0415            struct dm_io_region *where, blk_opf_t opf, struct dpages *dp,
0416            unsigned long *error_bits)
0417 {
0418     struct io *io;
0419     struct sync_io sio;
0420 
0421     if (num_regions > 1 && !op_is_write(opf)) {
0422         WARN_ON(1);
0423         return -EIO;
0424     }
0425 
0426     init_completion(&sio.wait);
0427 
0428     io = mempool_alloc(&client->pool, GFP_NOIO);
0429     io->error_bits = 0;
0430     atomic_set(&io->count, 1); /* see dispatch_io() */
0431     io->client = client;
0432     io->callback = sync_io_complete;
0433     io->context = &sio;
0434 
0435     io->vma_invalidate_address = dp->vma_invalidate_address;
0436     io->vma_invalidate_size = dp->vma_invalidate_size;
0437 
0438     dispatch_io(opf, num_regions, where, dp, io, 1);
0439 
0440     wait_for_completion_io(&sio.wait);
0441 
0442     if (error_bits)
0443         *error_bits = sio.error_bits;
0444 
0445     return sio.error_bits ? -EIO : 0;
0446 }
0447 
0448 static int async_io(struct dm_io_client *client, unsigned int num_regions,
0449             struct dm_io_region *where, blk_opf_t opf,
0450             struct dpages *dp, io_notify_fn fn, void *context)
0451 {
0452     struct io *io;
0453 
0454     if (num_regions > 1 && !op_is_write(opf)) {
0455         WARN_ON(1);
0456         fn(1, context);
0457         return -EIO;
0458     }
0459 
0460     io = mempool_alloc(&client->pool, GFP_NOIO);
0461     io->error_bits = 0;
0462     atomic_set(&io->count, 1); /* see dispatch_io() */
0463     io->client = client;
0464     io->callback = fn;
0465     io->context = context;
0466 
0467     io->vma_invalidate_address = dp->vma_invalidate_address;
0468     io->vma_invalidate_size = dp->vma_invalidate_size;
0469 
0470     dispatch_io(opf, num_regions, where, dp, io, 0);
0471     return 0;
0472 }
0473 
0474 static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
0475            unsigned long size)
0476 {
0477     /* Set up dpages based on memory type */
0478 
0479     dp->vma_invalidate_address = NULL;
0480     dp->vma_invalidate_size = 0;
0481 
0482     switch (io_req->mem.type) {
0483     case DM_IO_PAGE_LIST:
0484         list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
0485         break;
0486 
0487     case DM_IO_BIO:
0488         bio_dp_init(dp, io_req->mem.ptr.bio);
0489         break;
0490 
0491     case DM_IO_VMA:
0492         flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
0493         if ((io_req->bi_opf & REQ_OP_MASK) == REQ_OP_READ) {
0494             dp->vma_invalidate_address = io_req->mem.ptr.vma;
0495             dp->vma_invalidate_size = size;
0496         }
0497         vm_dp_init(dp, io_req->mem.ptr.vma);
0498         break;
0499 
0500     case DM_IO_KMEM:
0501         km_dp_init(dp, io_req->mem.ptr.addr);
0502         break;
0503 
0504     default:
0505         return -EINVAL;
0506     }
0507 
0508     return 0;
0509 }
0510 
0511 int dm_io(struct dm_io_request *io_req, unsigned num_regions,
0512       struct dm_io_region *where, unsigned long *sync_error_bits)
0513 {
0514     int r;
0515     struct dpages dp;
0516 
0517     r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
0518     if (r)
0519         return r;
0520 
0521     if (!io_req->notify.fn)
0522         return sync_io(io_req->client, num_regions, where,
0523                    io_req->bi_opf, &dp, sync_error_bits);
0524 
0525     return async_io(io_req->client, num_regions, where,
0526             io_req->bi_opf, &dp, io_req->notify.fn,
0527             io_req->notify.context);
0528 }
0529 EXPORT_SYMBOL(dm_io);
0530 
0531 int __init dm_io_init(void)
0532 {
0533     _dm_io_cache = KMEM_CACHE(io, 0);
0534     if (!_dm_io_cache)
0535         return -ENOMEM;
0536 
0537     return 0;
0538 }
0539 
0540 void dm_io_exit(void)
0541 {
0542     kmem_cache_destroy(_dm_io_cache);
0543     _dm_io_cache = NULL;
0544 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!