exofs: address_space_operations
Boaz Harrosh [Mon, 27 Oct 2008 17:31:34 +0000 (19:31 +0200)]
OK Now we start to read and write from osd-objects. We try to
collect at most contiguous pages as possible in a single write/read.
The first page index is the object's offset.

TODO:
   In 64-bit a single bio can carry at most 128 pages.
   Add support of chaining multiple bios

Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>

fs/exofs/exofs.h
fs/exofs/inode.c

index 825454d..0276242 100644 (file)
@@ -130,6 +130,9 @@ static inline struct exofs_i_info *exofs_i(struct inode *inode)
 /* inode.c               */
 void exofs_truncate(struct inode *inode);
 int exofs_setattr(struct dentry *, struct iattr *);
+int exofs_write_begin(struct file *file, struct address_space *mapping,
+               loff_t pos, unsigned len, unsigned flags,
+               struct page **pagep, void **fsdata);
 
 /*********************
  * operation vectors *
@@ -138,6 +141,9 @@ int exofs_setattr(struct dentry *, struct iattr *);
 extern const struct inode_operations exofs_file_inode_operations;
 extern const struct file_operations exofs_file_operations;
 
+/* inode.c           */
+extern const struct address_space_operations exofs_aops;
+
 /* symlink.c         */
 extern const struct inode_operations exofs_symlink_inode_operations;
 extern const struct inode_operations exofs_fast_symlink_inode_operations;
index b0bda1e..a3691d8 100644 (file)
@@ -35,6 +35,7 @@
 
 #include <linux/writeback.h>
 #include <linux/buffer_head.h>
+#include <scsi/scsi_device.h>
 
 #include "exofs.h"
 
 #  define EXOFS_DEBUG_OBJ_ISIZE 1
 #endif
 
+struct page_collect {
+       struct exofs_sb_info *sbi;
+       struct request_queue *req_q;
+       struct inode *inode;
+       unsigned expected_pages;
+
+       struct bio *bio;
+       unsigned nr_pages;
+       unsigned long length;
+       loff_t pg_first; /* keep 64bit also in 32-arches */
+};
+
+static void _pcol_init(struct page_collect *pcol, unsigned expected_pages,
+               struct inode *inode)
+{
+       struct exofs_sb_info *sbi = inode->i_sb->s_fs_info;
+       struct request_queue *req_q = sbi->s_dev->scsi_device->request_queue;
+
+       pcol->sbi = sbi;
+       pcol->req_q = req_q;
+       pcol->inode = inode;
+       pcol->expected_pages = expected_pages;
+
+       pcol->bio = NULL;
+       pcol->nr_pages = 0;
+       pcol->length = 0;
+       pcol->pg_first = -1;
+
+       EXOFS_DBGMSG("_pcol_init ino=0x%lx expected_pages=%u\n", inode->i_ino,
+                    expected_pages);
+}
+
+static void _pcol_reset(struct page_collect *pcol)
+{
+       pcol->expected_pages -= min(pcol->nr_pages, pcol->expected_pages);
+
+       pcol->bio = NULL;
+       pcol->nr_pages = 0;
+       pcol->length = 0;
+       pcol->pg_first = -1;
+       EXOFS_DBGMSG("_pcol_reset ino=0x%lx expected_pages=%u\n",
+                    pcol->inode->i_ino, pcol->expected_pages);
+
+       /* this is probably the end of the loop but in writes
+        * it might not end here. don't be left with nothing
+        */
+       if (!pcol->expected_pages)
+               pcol->expected_pages = 128;
+}
+
+static int pcol_try_alloc(struct page_collect *pcol)
+{
+       int pages = min_t(unsigned, pcol->expected_pages, BIO_MAX_PAGES);
+
+       for (; pages; pages >>= 1) {
+               pcol->bio = bio_alloc(GFP_KERNEL, pages);
+               if (likely(pcol->bio))
+                       return 0;
+       }
+
+       EXOFS_ERR("Failed to kcalloc expected_pages=%u\n",
+                 pcol->expected_pages);
+       return -ENOMEM;
+}
+
+static void pcol_free(struct page_collect *pcol)
+{
+       bio_put(pcol->bio);
+       pcol->bio = NULL;
+}
+
+static int pcol_add_page(struct page_collect *pcol, struct page *page,
+                        unsigned len)
+{
+       int added_len = bio_add_pc_page(pcol->req_q, pcol->bio, page, len, 0);
+       if (unlikely(len != added_len))
+               return -ENOMEM;
+
+       ++pcol->nr_pages;
+       pcol->length += len;
+       return 0;
+}
+
+static int update_read_page(struct page *page, int ret)
+{
+       if (ret == 0) {
+               /* Everything is OK */
+               SetPageUptodate(page);
+               if (PageError(page))
+                       ClearPageError(page);
+       } else if (ret == -EFAULT) {
+               /* In this case we were trying to read something that wasn't on
+                * disk yet - return a page full of zeroes.  This should be OK,
+                * because the object should be empty (if there was a write
+                * before this read, the read would be waiting with the page
+                * locked */
+               clear_highpage(page);
+
+               SetPageUptodate(page);
+               if (PageError(page))
+                       ClearPageError(page);
+               ret = 0; /* recovered error */
+               EXOFS_DBGMSG("recovered read error\n");
+       } else /* Error */
+               SetPageError(page);
+
+       return ret;
+}
+
+static void update_write_page(struct page *page, int ret)
+{
+       if (ret) {
+               mapping_set_error(page->mapping, ret);
+               SetPageError(page);
+       }
+       end_page_writeback(page);
+}
+
+/* Called at the end of reads, to optionally unlock pages and update their
+ * status.
+ */
+static int __readpages_done(struct osd_request *or, struct page_collect *pcol,
+                           bool do_unlock)
+{
+       struct bio_vec *bvec;
+       int i;
+       u64 resid;
+       u64 good_bytes;
+       u64 length = 0;
+       int ret = exofs_check_ok_resid(or, &resid, NULL);
+
+       osd_end_request(or);
+
+       if (likely(!ret))
+               good_bytes = pcol->length;
+       else if (!resid)
+               good_bytes = 0;
+       else
+               good_bytes = pcol->length - resid;
+
+       EXOFS_DBGMSG("readpages_done(0x%lx) good_bytes=0x%llx"
+                    " length=0x%lx nr_pages=%u\n",
+                    pcol->inode->i_ino, _LLU(good_bytes), pcol->length,
+                    pcol->nr_pages);
+
+       __bio_for_each_segment(bvec, pcol->bio, i, 0) {
+               struct page *page = bvec->bv_page;
+               struct inode *inode = page->mapping->host;
+               int page_stat;
+
+               if (inode != pcol->inode)
+                       continue; /* osd might add more pages at end */
+
+               if (likely(length < good_bytes))
+                       page_stat = 0;
+               else
+                       page_stat = ret;
+
+               EXOFS_DBGMSG("    readpages_done(0x%lx, 0x%lx) %s\n",
+                         inode->i_ino, page->index,
+                         page_stat ? "bad_bytes" : "good_bytes");
+
+               ret = update_read_page(page, page_stat);
+               if (do_unlock)
+                       unlock_page(page);
+               length += bvec->bv_len;
+       }
+
+       pcol_free(pcol);
+       EXOFS_DBGMSG("readpages_done END\n");
+       return ret;
+}
+
+/* callback of async reads */
+static void readpages_done(struct osd_request *or, void *p)
+{
+       struct page_collect *pcol = p;
+
+       __readpages_done(or, pcol, true);
+       atomic_dec(&pcol->sbi->s_curr_pending);
+       kfree(p);
+}
+
+static void _unlock_pcol_pages(struct page_collect *pcol, int ret, int rw)
+{
+       struct bio_vec *bvec;
+       int i;
+
+       __bio_for_each_segment(bvec, pcol->bio, i, 0) {
+               struct page *page = bvec->bv_page;
+
+               if (rw == READ)
+                       update_read_page(page, ret);
+               else
+                       update_write_page(page, ret);
+
+               unlock_page(page);
+       }
+       pcol_free(pcol);
+}
+
+static int read_exec(struct page_collect *pcol, bool is_sync)
+{
+       struct exofs_i_info *oi = exofs_i(pcol->inode);
+       struct osd_obj_id obj = {pcol->sbi->s_pid,
+                                       pcol->inode->i_ino + EXOFS_OBJ_OFF};
+       struct osd_request *or = NULL;
+       struct page_collect *pcol_copy = NULL;
+       loff_t i_start = pcol->pg_first << PAGE_CACHE_SHIFT;
+       int ret;
+
+       if (!pcol->bio)
+               return 0;
+
+       /* see comment in _readpage() about sync reads */
+       WARN_ON(is_sync && (pcol->nr_pages != 1));
+
+       or = osd_start_request(pcol->sbi->s_dev, GFP_KERNEL);
+       if (unlikely(!or)) {
+               ret = -ENOMEM;
+               goto err;
+       }
+
+       osd_req_read(or, &obj, pcol->bio, i_start);
+
+       if (is_sync) {
+               exofs_sync_op(or, pcol->sbi->s_timeout, oi->i_cred);
+               return __readpages_done(or, pcol, false);
+       }
+
+       pcol_copy = kmalloc(sizeof(*pcol_copy), GFP_KERNEL);
+       if (!pcol_copy) {
+               ret = -ENOMEM;
+               goto err;
+       }
+
+       *pcol_copy = *pcol;
+       ret = exofs_async_op(or, readpages_done, pcol_copy, oi->i_cred);
+       if (unlikely(ret))
+               goto err;
+
+       atomic_inc(&pcol->sbi->s_curr_pending);
+
+       EXOFS_DBGMSG("read_exec obj=0x%llx start=0x%llx length=0x%lx\n",
+                 obj.id, _LLU(i_start), pcol->length);
+
+       /* pages ownership was passed to pcol_copy */
+       _pcol_reset(pcol);
+       return 0;
+
+err:
+       if (!is_sync)
+               _unlock_pcol_pages(pcol, ret, READ);
+       kfree(pcol_copy);
+       if (or)
+               osd_end_request(or);
+       return ret;
+}
+
+/* readpage_strip is called either directly from readpage() or by the VFS from
+ * within read_cache_pages(), to add one more page to be read. It will try to
+ * collect as many contiguous pages as posible. If a discontinuity is
+ * encountered, or it runs out of resources, it will submit the previous segment
+ * and will start a new collection. Eventually caller must submit the last
+ * segment if present.
+ */
+static int readpage_strip(void *data, struct page *page)
+{
+       struct page_collect *pcol = data;
+       struct inode *inode = pcol->inode;
+       struct exofs_i_info *oi = exofs_i(inode);
+       loff_t i_size = i_size_read(inode);
+       pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
+       size_t len;
+       int ret;
+
+       /* FIXME: Just for debugging, will be removed */
+       if (PageUptodate(page))
+               EXOFS_ERR("PageUptodate(0x%lx, 0x%lx)\n", pcol->inode->i_ino,
+                         page->index);
+
+       if (page->index < end_index)
+               len = PAGE_CACHE_SIZE;
+       else if (page->index == end_index)
+               len = i_size & ~PAGE_CACHE_MASK;
+       else
+               len = 0;
+
+       if (!len || !obj_created(oi)) {
+               /* this will be out of bounds, or doesn't exist yet.
+                * Current page is cleared and the request is split
+                */
+               clear_highpage(page);
+
+               SetPageUptodate(page);
+               if (PageError(page))
+                       ClearPageError(page);
+
+               unlock_page(page);
+               EXOFS_DBGMSG("readpage_strip(0x%lx, 0x%lx) empty page,"
+                            " splitting\n", inode->i_ino, page->index);
+
+               return read_exec(pcol, false);
+       }
+
+try_again:
+
+       if (unlikely(pcol->pg_first == -1)) {
+               pcol->pg_first = page->index;
+       } else if (unlikely((pcol->pg_first + pcol->nr_pages) !=
+                  page->index)) {
+               /* Discontinuity detected, split the request */
+               ret = read_exec(pcol, false);
+               if (unlikely(ret))
+                       goto fail;
+               goto try_again;
+       }
+
+       if (!pcol->bio) {
+               ret = pcol_try_alloc(pcol);
+               if (unlikely(ret))
+                       goto fail;
+       }
+
+       if (len != PAGE_CACHE_SIZE)
+               zero_user(page, len, PAGE_CACHE_SIZE - len);
+
+       EXOFS_DBGMSG("    readpage_strip(0x%lx, 0x%lx) len=0x%zx\n",
+                    inode->i_ino, page->index, len);
+
+       ret = pcol_add_page(pcol, page, len);
+       if (ret) {
+               EXOFS_DBGMSG("Failed pcol_add_page pages[i]=%p "
+                         "this_len=0x%zx nr_pages=%u length=0x%lx\n",
+                         page, len, pcol->nr_pages, pcol->length);
+
+               /* split the request, and start again with current page */
+               ret = read_exec(pcol, false);
+               if (unlikely(ret))
+                       goto fail;
+
+               goto try_again;
+       }
+
+       return 0;
+
+fail:
+       /* SetPageError(page); ??? */
+       unlock_page(page);
+       return ret;
+}
+
+static int exofs_readpages(struct file *file, struct address_space *mapping,
+                          struct list_head *pages, unsigned nr_pages)
+{
+       struct page_collect pcol;
+       int ret;
+
+       _pcol_init(&pcol, nr_pages, mapping->host);
+
+       ret = read_cache_pages(mapping, pages, readpage_strip, &pcol);
+       if (ret) {
+               EXOFS_ERR("read_cache_pages => %d\n", ret);
+               return ret;
+       }
+
+       return read_exec(&pcol, false);
+}
+
+static int _readpage(struct page *page, bool is_sync)
+{
+       struct page_collect pcol;
+       int ret;
+
+       _pcol_init(&pcol, 1, page->mapping->host);
+
+       /* readpage_strip might call read_exec(,async) inside at several places
+        * but this is safe for is_async=0 since read_exec will not do anything
+        * when we have a single page.
+        */
+       ret = readpage_strip(&pcol, page);
+       if (ret) {
+               EXOFS_ERR("_readpage => %d\n", ret);
+               return ret;
+       }
+
+       return read_exec(&pcol, is_sync);
+}
+
+/*
+ * We don't need the file
+ */
+static int exofs_readpage(struct file *file, struct page *page)
+{
+       return _readpage(page, false);
+}
+
+/* Callback for osd_write. All writes are asynchronouse */
+static void writepages_done(struct osd_request *or, void *p)
+{
+       struct page_collect *pcol = p;
+       struct bio_vec *bvec;
+       int i;
+       u64 resid;
+       u64  good_bytes;
+       u64  length = 0;
+
+       int ret = exofs_check_ok_resid(or, NULL, &resid);
+
+       osd_end_request(or);
+       atomic_dec(&pcol->sbi->s_curr_pending);
+
+       if (likely(!ret))
+               good_bytes = pcol->length;
+       else if (!resid)
+               good_bytes = 0;
+       else
+               good_bytes = pcol->length - resid;
+
+       EXOFS_DBGMSG("writepages_done(0x%lx) good_bytes=0x%llx"
+                    " length=0x%lx nr_pages=%u\n",
+                    pcol->inode->i_ino, _LLU(good_bytes), pcol->length,
+                    pcol->nr_pages);
+
+       __bio_for_each_segment(bvec, pcol->bio, i, 0) {
+               struct page *page = bvec->bv_page;
+               struct inode *inode = page->mapping->host;
+               int page_stat;
+
+               if (inode != pcol->inode)
+                       continue; /* osd might add more pages to a bio */
+
+               if (likely(length < good_bytes))
+                       page_stat = 0;
+               else
+                       page_stat = ret;
+
+               update_write_page(page, page_stat);
+               unlock_page(page);
+               EXOFS_DBGMSG("    writepages_done(0x%lx, 0x%lx) status=%d\n",
+                            inode->i_ino, page->index, page_stat);
+
+               length += bvec->bv_len;
+       }
+
+       pcol_free(pcol);
+       kfree(pcol);
+       EXOFS_DBGMSG("writepages_done END\n");
+}
+
+static int write_exec(struct page_collect *pcol)
+{
+       struct exofs_i_info *oi = exofs_i(pcol->inode);
+       struct osd_obj_id obj = {pcol->sbi->s_pid,
+                                       pcol->inode->i_ino + EXOFS_OBJ_OFF};
+       struct osd_request *or = NULL;
+       struct page_collect *pcol_copy = NULL;
+       loff_t i_start = pcol->pg_first << PAGE_CACHE_SHIFT;
+       int ret;
+
+       if (!pcol->bio)
+               return 0;
+
+       or = osd_start_request(pcol->sbi->s_dev, GFP_KERNEL);
+       if (unlikely(!or)) {
+               EXOFS_ERR("write_exec: Faild to osd_start_request()\n");
+               ret = -ENOMEM;
+               goto err;
+       }
+
+       pcol_copy = kmalloc(sizeof(*pcol_copy), GFP_KERNEL);
+       if (!pcol_copy) {
+               EXOFS_ERR("write_exec: Faild to kmalloc(pcol)\n");
+               ret = -ENOMEM;
+               goto err;
+       }
+
+       *pcol_copy = *pcol;
+
+       osd_req_write(or, &obj, pcol_copy->bio, i_start);
+       ret = exofs_async_op(or, writepages_done, pcol_copy, oi->i_cred);
+       if (unlikely(ret)) {
+               EXOFS_ERR("write_exec: exofs_async_op() Faild\n");
+               goto err;
+       }
+
+       atomic_inc(&pcol->sbi->s_curr_pending);
+       EXOFS_DBGMSG("write_exec(0x%lx, 0x%llx) start=0x%llx length=0x%lx\n",
+                 pcol->inode->i_ino, pcol->pg_first, _LLU(i_start),
+                 pcol->length);
+       /* pages ownership was passed to pcol_copy */
+       _pcol_reset(pcol);
+       return 0;
+
+err:
+       _unlock_pcol_pages(pcol, ret, WRITE);
+       kfree(pcol_copy);
+       if (or)
+               osd_end_request(or);
+       return ret;
+}
+
+/* writepage_strip is called either directly from writepage() or by the VFS from
+ * within write_cache_pages(), to add one more page to be written to storage.
+ * It will try to collect as many contiguous pages as possible. If a
+ * discontinuity is encountered or it runs out of resources it will submit the
+ * previous segment and will start a new collection.
+ * Eventually caller must submit the last segment if present.
+ */
+static int writepage_strip(struct page *page,
+                          struct writeback_control *wbc_unused, void *data)
+{
+       struct page_collect *pcol = data;
+       struct inode *inode = pcol->inode;
+       struct exofs_i_info *oi = exofs_i(inode);
+       loff_t i_size = i_size_read(inode);
+       pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
+       size_t len;
+       int ret;
+
+       BUG_ON(!PageLocked(page));
+
+       ret = wait_obj_created(oi);
+       if (unlikely(ret))
+               goto fail;
+
+       if (page->index < end_index)
+               /* in this case, the page is within the limits of the file */
+               len = PAGE_CACHE_SIZE;
+       else {
+               len = i_size & ~PAGE_CACHE_MASK;
+
+               if (page->index > end_index || !len) {
+                       /* in this case, the page is outside the limits
+                        * (truncate in progress)
+                        */
+                       ret = write_exec(pcol);
+                       if (unlikely(ret))
+                               goto fail;
+                       if (PageError(page))
+                               ClearPageError(page);
+                       unlock_page(page);
+                       return 0;
+               }
+       }
+
+try_again:
+
+       if (unlikely(pcol->pg_first == -1)) {
+               pcol->pg_first = page->index;
+       } else if (unlikely((pcol->pg_first + pcol->nr_pages) !=
+                  page->index)) {
+               /* Discontinuity detected, split the request */
+               ret = write_exec(pcol);
+               if (unlikely(ret))
+                       goto fail;
+               goto try_again;
+       }
+
+       if (!pcol->bio) {
+               ret = pcol_try_alloc(pcol);
+               if (unlikely(ret))
+                       goto fail;
+       }
+
+       EXOFS_DBGMSG("    writepage_strip(0x%lx, 0x%lx) len=0x%zx\n",
+                    inode->i_ino, page->index, len);
+
+       ret = pcol_add_page(pcol, page, len);
+       if (unlikely(ret)) {
+               EXOFS_DBGMSG("Failed pcol_add_page "
+                            "nr_pages=%u total_length=0x%lx\n",
+                            pcol->nr_pages, pcol->length);
+
+               /* split the request, next loop will start again */
+               ret = write_exec(pcol);
+               if (unlikely(ret)) {
+                       EXOFS_DBGMSG("write_exec faild => %d", ret);
+                       goto fail;
+               }
+
+               goto try_again;
+       }
+
+       BUG_ON(PageWriteback(page));
+       set_page_writeback(page);
+
+       return 0;
+
+fail:
+       set_bit(AS_EIO, &page->mapping->flags);
+       unlock_page(page);
+       return ret;
+}
+
+static int exofs_writepages(struct address_space *mapping,
+                      struct writeback_control *wbc)
+{
+       struct page_collect pcol;
+       long start, end, expected_pages;
+       int ret;
+
+       start = wbc->range_start >> PAGE_CACHE_SHIFT;
+       end = (wbc->range_end == LLONG_MAX) ?
+                       start + mapping->nrpages :
+                       wbc->range_end >> PAGE_CACHE_SHIFT;
+
+       if (start || end)
+               expected_pages = min(end - start + 1, 32L);
+       else
+               expected_pages = mapping->nrpages;
+
+       EXOFS_DBGMSG("inode(0x%lx) wbc->start=0x%llx wbc->end=0x%llx"
+                    " m->nrpages=%lu start=0x%lx end=0x%lx\n",
+                    mapping->host->i_ino, wbc->range_start, wbc->range_end,
+                    mapping->nrpages, start, end);
+
+       _pcol_init(&pcol, expected_pages, mapping->host);
+
+       ret = write_cache_pages(mapping, wbc, writepage_strip, &pcol);
+       if (ret) {
+               EXOFS_ERR("write_cache_pages => %d\n", ret);
+               return ret;
+       }
+
+       return write_exec(&pcol);
+}
+
+static int exofs_writepage(struct page *page, struct writeback_control *wbc)
+{
+       struct page_collect pcol;
+       int ret;
+
+       _pcol_init(&pcol, 1, page->mapping->host);
+
+       ret = writepage_strip(page, NULL, &pcol);
+       if (ret) {
+               EXOFS_ERR("exofs_writepage => %d\n", ret);
+               return ret;
+       }
+
+       return write_exec(&pcol);
+}
+
+int exofs_write_begin(struct file *file, struct address_space *mapping,
+               loff_t pos, unsigned len, unsigned flags,
+               struct page **pagep, void **fsdata)
+{
+       int ret = 0;
+       struct page *page;
+
+       page = *pagep;
+       if (page == NULL) {
+               ret = simple_write_begin(file, mapping, pos, len, flags, pagep,
+                                        fsdata);
+               if (ret) {
+                       EXOFS_DBGMSG("simple_write_begin faild\n");
+                       return ret;
+               }
+
+               page = *pagep;
+       }
+
+        /* read modify write */
+       if (!PageUptodate(page) && (len != PAGE_CACHE_SIZE)) {
+               ret = _readpage(page, true);
+               if (ret) {
+                       /*SetPageError was done by _readpage. Is it ok?*/
+                       unlock_page(page);
+                       EXOFS_DBGMSG("__readpage_filler faild\n");
+               }
+       }
+
+       return ret;
+}
+
+static int exofs_write_begin_export(struct file *file,
+               struct address_space *mapping,
+               loff_t pos, unsigned len, unsigned flags,
+               struct page **pagep, void **fsdata)
+{
+       *pagep = NULL;
+
+       return exofs_write_begin(file, mapping, pos, len, flags, pagep,
+                                       fsdata);
+}
+
+const struct address_space_operations exofs_aops = {
+       .readpage       = exofs_readpage,
+       .readpages      = exofs_readpages,
+       .writepage      = exofs_writepage,
+       .writepages     = exofs_writepages,
+       .write_begin    = exofs_write_begin_export,
+       .write_end      = simple_write_end,
+};
+
 /******************************************************************************
  * INODE OPERATIONS
  *****************************************************************************/