* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*/
+#ifndef AUTOCONF_INCLUDED
#include <linux/config.h>
+#endif
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/string.h>
#define DEBUG_SUBSYSTEM S_LLITE
-#include <linux/lustre_mds.h>
-#include <linux/lustre_lite.h>
+//#include <lustre_mdc.h>
+#include <lustre_lite.h>
#include "llite_internal.h"
#include <linux/lustre_compat25.h>
static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
{
- struct inode *inode = page->mapping->host;
- struct obd_export *exp;
+ return ll_writepage(page);
+}
+
+/* It is safe to not check anything in invalidatepage/releasepage below
+ because they are run with page locked and all our io is happening with
+ locked page too */
+#ifdef HAVE_INVALIDATEPAGE_RETURN_INT
+static int ll_invalidatepage(struct page *page, unsigned long offset)
+{
+ if (offset)
+ return 0;
+ if (PagePrivate(page))
+ ll_removepage(page);
+ return 1;
+}
+#else
+static void ll_invalidatepage(struct page *page, unsigned long offset)
+{
+ if (offset == 0 && PagePrivate(page))
+ ll_removepage(page);
+}
+#endif
+static int ll_releasepage(struct page *page, int gfp_mask)
+{
+ if (PagePrivate(page))
+ ll_removepage(page);
+ return 1;
+}
+
+static int ll_set_page_dirty(struct page *page)
+{
struct ll_async_page *llap;
- int rc;
ENTRY;
+
+ llap = llap_from_page(page, LLAP_ORIGIN_UNKNOWN);
+ if (IS_ERR(llap))
+ RETURN(PTR_ERR(llap));
+
+ llap_write_pending(page->mapping->host, llap);
+ RETURN(__set_page_dirty_nobuffers(page));
+}
- LASSERT(!PageDirty(page));
- LASSERT(PageLocked(page));
+#define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
- exp = ll_i2obdexp(inode);
- if (exp == NULL)
- GOTO(out, rc = -EINVAL);
+static inline int ll_get_user_pages(int rw, unsigned long user_addr,
+ size_t size, struct page ***pages)
+{
+ int result = -ENOMEM;
+ int page_count;
- llap = llap_from_page(page);
- if (IS_ERR(llap))
- GOTO(out, rc = PTR_ERR(llap));
-
- page_cache_get(page);
- if (llap->llap_write_queued) {
- LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
- rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
- llap->llap_cookie,
- ASYNC_READY | ASYNC_URGENT);
+ /* set an arbitrary limit to prevent arithmetic overflow */
+ if (size > MAX_DIRECTIO_SIZE) {
+ *pages = NULL;
+ return -EFBIG;
+ }
+
+ page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
+ page_count -= user_addr >> CFS_PAGE_SHIFT;
+
+ OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
+ if (*pages) {
+ down_read(¤t->mm->mmap_sem);
+ result = get_user_pages(current, current->mm, user_addr,
+ page_count, (rw == READ), 0, *pages,
+ NULL);
+ up_read(¤t->mm->mmap_sem);
+ if (result < 0)
+ OBD_FREE(*pages, page_count * sizeof(**pages));
+ }
+
+ return result;
+}
+
+/* ll_free_user_pages - tear down page struct array
+ * @pages: array of page struct pointers underlying target buffer */
+static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
+{
+ int i;
+
+ for (i = 0; i < npages; i++) {
+ if (do_dirty)
+ set_page_dirty_lock(pages[i]);
+ page_cache_release(pages[i]);
+ }
+
+ OBD_FREE(pages, npages * sizeof(*pages));
+}
+
+static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
+ struct address_space *mapping,
+ struct lov_stripe_md *lsm,
+ size_t size, loff_t file_offset,
+ struct page **pages, int page_count)
+{
+ struct brw_page *pga;
+ struct obdo oa;
+ int opc, i, rc = 0;
+ size_t length;
+ struct obd_capa *ocapa;
+ loff_t file_offset_orig = file_offset;
+ ENTRY;
+
+ OBD_ALLOC(pga, sizeof(*pga) * page_count);
+ if (!pga) {
+ CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
+ (int)sizeof(*pga), page_count);
+ RETURN(-ENOMEM);
+ }
+
+ for (i = 0, length = size; length > 0;
+ length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
+ pga[i].pg = pages[i];
+ pga[i].off = file_offset;
+ /* To the end of the page, or the length, whatever is less */
+ pga[i].count = min_t(int, CFS_PAGE_SIZE -
+ (file_offset & ~CFS_PAGE_MASK),
+ length);
+ pga[i].flag = 0;
+ if (rw == READ)
+ POISON_PAGE(pages[i], 0x0d);
+ }
+
+ ll_inode_fill_obdo(inode, rw, &oa);
+
+ if (rw == WRITE) {
+ lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
+ LPROC_LL_DIRECT_WRITE, size);
+ opc = CAPA_OPC_OSS_WRITE;
+ llap_write_pending(inode, NULL);
} else {
- llap->llap_write_queued = 1;
- rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
- llap->llap_cookie, OBD_BRW_WRITE, 0, 0,
- 0, ASYNC_READY | ASYNC_URGENT);
- if (rc == 0)
- LL_CDEBUG_PAGE(D_PAGE, page, "mmap write queued\n");
- else
- llap->llap_write_queued = 0;
+ lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
+ LPROC_LL_DIRECT_READ, size);
+ opc = CAPA_OPC_OSS_RW;
+ }
+ ocapa = ll_osscapa_get(inode, current->fsuid, opc);
+ rc = obd_brw_rqset(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
+ ll_i2dtexp(inode), &oa, lsm, page_count, pga, NULL,
+ ocapa);
+ capa_put(ocapa);
+ if ((rc > 0) && (rw == WRITE)) {
+ lov_stripe_lock(lsm);
+ obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset_orig + rc, 0);
+ lov_stripe_unlock(lsm);
}
- if (rc)
- page_cache_release(page);
-out:
- if (rc)
- unlock_page(page);
+
+ OBD_FREE(pga, sizeof(*pga) * page_count);
RETURN(rc);
}
+/* This is the maximum size of a single O_DIRECT request, based on a 128kB
+ * kmalloc limit. We need to fit all of the brw_page structs, each one
+ * representing PAGE_SIZE worth of user data, into a single buffer, and
+ * then truncate this to be a full-sized RPC. This is 22MB for 4kB pages. */
+#define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
+ ~(PTLRPC_MAX_BRW_SIZE - 1))
+static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
+ const struct iovec *iov, loff_t file_offset,
+ unsigned long nr_segs)
+{
+ struct file *file = iocb->ki_filp;
+ struct inode *inode = file->f_mapping->host;
+ ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
+ struct ll_inode_info *lli = ll_i2info(inode);
+ unsigned long seg = 0;
+ size_t size = MAX_DIO_SIZE;
+ ENTRY;
+
+ if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
+ RETURN(-EBADF);
+
+ /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
+ if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
+ RETURN(-EINVAL);
+
+ CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
+ "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
+ inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
+ file_offset, file_offset, count >> CFS_PAGE_SHIFT,
+ MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
+
+ if (rw == WRITE)
+ ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
+ else
+ ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
+
+ /* Check that all user buffers are aligned as well */
+ for (seg = 0; seg < nr_segs; seg++) {
+ if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
+ (iov[seg].iov_len & ~CFS_PAGE_MASK))
+ RETURN(-EINVAL);
+ }
+
+ for (seg = 0; seg < nr_segs; seg++) {
+ size_t iov_left = iov[seg].iov_len;
+ unsigned long user_addr = (unsigned long)iov[seg].iov_base;
+
+ while (iov_left > 0) {
+ struct page **pages;
+ int page_count;
+ ssize_t result;
+
+ page_count = ll_get_user_pages(rw, user_addr,
+ min(size, iov_left),
+ &pages);
+ LASSERT(page_count != 0);
+ if (page_count > 0) {
+ result = ll_direct_IO_26_seg(rw, inode,
+ file->f_mapping,
+ lli->lli_smd,
+ min(size,iov_left),
+ file_offset, pages,
+ page_count);
+ ll_free_user_pages(pages, page_count, rw==READ);
+ } else {
+ result = 0;
+ }
+ if (page_count < 0 || result <= 0) {
+ /* If we can't allocate a large enough buffer
+ * for the request, shrink it to a smaller
+ * PAGE_SIZE multiple and try again.
+ * We should always be able to kmalloc for a
+ * page worth of page pointers = 4MB on i386. */
+ if ((page_count == -ENOMEM||result == -ENOMEM)&&
+ size > (CFS_PAGE_SIZE / sizeof(*pages)) *
+ CFS_PAGE_SIZE) {
+ size = ((((size / 2) - 1) |
+ ~CFS_PAGE_MASK) + 1) &
+ CFS_PAGE_MASK;
+ CDEBUG(D_VFSTRACE, "DIO size now %u\n",
+ (int)size);
+ continue;
+ }
+
+ if (tot_bytes > 0)
+ RETURN(tot_bytes);
+ RETURN(page_count < 0 ? page_count : result);
+ }
+
+ tot_bytes += result;
+ file_offset += result;
+ iov_left -= result;
+ user_addr += result;
+ }
+ }
+ RETURN(tot_bytes);
+}
+
struct address_space_operations ll_aops = {
- readpage: ll_readpage,
-// readpages: ll_readpages,
-// direct_IO: ll_direct_IO_26,
- writepage: ll_writepage_26,
- writepages: generic_writepages,
- set_page_dirty: __set_page_dirty_nobuffers,
- sync_page: NULL,
- prepare_write: ll_prepare_write,
- commit_write: ll_commit_write,
- removepage: ll_removepage,
- bmap: NULL
+ .readpage = ll_readpage,
+// .readpages = ll_readpages,
+ .direct_IO = ll_direct_IO_26,
+ .writepage = ll_writepage_26,
+ .writepages = generic_writepages,
+ .set_page_dirty = ll_set_page_dirty,
+ .sync_page = NULL,
+ .prepare_write = ll_prepare_write,
+ .commit_write = ll_commit_write,
+ .invalidatepage = ll_invalidatepage,
+ .releasepage = ll_releasepage,
+ .bmap = NULL
};