+ }
+ /*
+ * If dirtyend exceeds file size, chop it down. This should
+ * not occur unless there is a race.
+ */
+ if (NBOFF(bp) + bp->nb_dirtyend > (off_t)np->n_size) {
+ bp->nb_dirtyend = np->n_size - NBOFF(bp);
+ if (bp->nb_dirtyoff >= bp->nb_dirtyend)
+ bp->nb_dirtyoff = bp->nb_dirtyend = 0;
+ }
+ /*
+ * UBC doesn't handle partial pages, so we need to make sure
+ * that any pages left in the page cache are completely valid.
+ *
+ * Writes that are smaller than a block are delayed if they
+ * don't extend to the end of the block.
+ *
+ * If the block isn't (completely) cached, we may need to read
+ * in some parts of pages that aren't covered by the write.
+ * If the write offset (on) isn't page aligned, we'll need to
+ * read the start of the first page being written to. Likewise,
+ * if the offset of the end of the write (on+n) isn't page aligned,
+ * we'll need to read the end of the last page being written to.
+ *
+ * Notes:
+ * We don't want to read anything we're just going to write over.
+ * We don't want to issue multiple I/Os if we don't have to
+ * (because they're synchronous rpcs).
+ * We don't want to read anything we already have modified in the
+ * page cache.
+ */
+ if (!ISSET(bp->nb_flags, NB_NOCACHE) && !ISSET(bp->nb_flags, NB_CACHE) && (n < biosize)) {
+ int firstpg, lastpg, dirtypg;
+ int firstpgoff, lastpgoff;
+ start = end = -1;
+ firstpg = on/PAGE_SIZE;
+ firstpgoff = on & PAGE_MASK;
+ lastpg = (on+n-1)/PAGE_SIZE;
+ lastpgoff = (on+n) & PAGE_MASK;
+ if (firstpgoff && !NBPGVALID(bp,firstpg)) {
+ /* need to read start of first page */
+ start = firstpg * PAGE_SIZE;
+ end = start + firstpgoff;
+ }
+ if (lastpgoff && !NBPGVALID(bp,lastpg)) {
+ /* need to read end of last page */
+ if (start < 0)
+ start = (lastpg * PAGE_SIZE) + lastpgoff;
+ end = (lastpg + 1) * PAGE_SIZE;
+ }
+ if (end > start) {
+ /* need to read the data in range: start...end-1 */
+
+ /* first, check for dirty pages in between */
+ /* if there are, we'll have to do two reads because */
+ /* we don't want to overwrite the dirty pages. */
+ for (dirtypg=start/PAGE_SIZE; dirtypg <= (end-1)/PAGE_SIZE; dirtypg++)
+ if (NBPGDIRTY(bp,dirtypg))
+ break;
+
+ /* if start is at beginning of page, try */
+ /* to get any preceeding pages as well. */
+ if (!(start & PAGE_MASK)) {
+ /* stop at next dirty/valid page or start of block */
+ for (; start > 0; start-=PAGE_SIZE)
+ if (NBPGVALID(bp,((start-1)/PAGE_SIZE)))
+ break;
+ }
+
+ NFS_BUF_MAP(bp);
+ /* setup uio for read(s) */
+ boff = NBOFF(bp);
+ auio = uio_createwithbuffer(1, 0, UIO_SYSSPACE, UIO_READ,
+ &auio_buf, sizeof(auio_buf));
+
+ if (dirtypg <= (end-1)/PAGE_SIZE) {
+ /* there's a dirty page in the way, so just do two reads */
+ /* we'll read the preceding data here */
+ uio_reset(auio, boff + start, UIO_SYSSPACE, UIO_READ);
+ uio_addiov(auio, CAST_USER_ADDR_T(bp->nb_data + start), on - start);
+ error = nfs_read_rpc(np, auio, ctx);
+ if (error) /* couldn't read the data, so treat buffer as NOCACHE */
+ SET(bp->nb_flags, (NB_NOCACHE|NB_STABLE));
+ if (uio_resid(auio) > 0) {
+ FSDBG(516, bp, (caddr_t)uio_curriovbase(auio) - bp->nb_data, uio_resid(auio), 0xd00dee01);
+ bzero(CAST_DOWN(caddr_t, uio_curriovbase(auio)), uio_resid(auio));
+ }
+ if (!error) {
+ /* update validoff/validend if necessary */
+ if ((bp->nb_validoff < 0) || (bp->nb_validoff > start))
+ bp->nb_validoff = start;
+ if ((bp->nb_validend < 0) || (bp->nb_validend < on))
+ bp->nb_validend = on;
+ if ((off_t)np->n_size > boff + bp->nb_validend)
+ bp->nb_validend = min(np->n_size - (boff + start), biosize);
+ /* validate any pages before the write offset */
+ for (; start < on/PAGE_SIZE; start+=PAGE_SIZE)
+ NBPGVALID_SET(bp, start/PAGE_SIZE);
+ }
+ /* adjust start to read any trailing data */
+ start = on+n;
+ }
+
+ /* if end is at end of page, try to */
+ /* get any following pages as well. */
+ if (!(end & PAGE_MASK)) {
+ /* stop at next valid page or end of block */
+ for (; end < biosize; end+=PAGE_SIZE)
+ if (NBPGVALID(bp,end/PAGE_SIZE))
+ break;
+ }
+
+ if (((boff+start) >= (off_t)np->n_size) ||
+ ((start >= on) && ((boff + on + n) >= (off_t)np->n_size))) {
+ /*
+ * Either this entire read is beyond the current EOF
+ * or the range that we won't be modifying (on+n...end)
+ * is all beyond the current EOF.
+ * No need to make a trip across the network to
+ * read nothing. So, just zero the buffer instead.
+ */
+ FSDBG(516, bp, start, end - start, 0xd00dee00);
+ bzero(bp->nb_data + start, end - start);
+ error = 0;
+ } else if (!ISSET(bp->nb_flags, NB_NOCACHE)) {
+ /* now we'll read the (rest of the) data */
+ uio_reset(auio, boff + start, UIO_SYSSPACE, UIO_READ);
+ uio_addiov(auio, CAST_USER_ADDR_T(bp->nb_data + start), end - start);
+ error = nfs_read_rpc(np, auio, ctx);
+ if (error) /* couldn't read the data, so treat buffer as NOCACHE */
+ SET(bp->nb_flags, (NB_NOCACHE|NB_STABLE));
+ if (uio_resid(auio) > 0) {
+ FSDBG(516, bp, (caddr_t)uio_curriovbase(auio) - bp->nb_data, uio_resid(auio), 0xd00dee02);
+ bzero(CAST_DOWN(caddr_t, uio_curriovbase(auio)), uio_resid(auio));
+ }
+ }
+ if (!error) {
+ /* update validoff/validend if necessary */
+ if ((bp->nb_validoff < 0) || (bp->nb_validoff > start))
+ bp->nb_validoff = start;
+ if ((bp->nb_validend < 0) || (bp->nb_validend < end))
+ bp->nb_validend = end;
+ if ((off_t)np->n_size > boff + bp->nb_validend)
+ bp->nb_validend = min(np->n_size - (boff + start), biosize);
+ /* validate any pages before the write offset's page */
+ for (; start < (off_t)trunc_page_32(on); start+=PAGE_SIZE)
+ NBPGVALID_SET(bp, start/PAGE_SIZE);
+ /* validate any pages after the range of pages being written to */
+ for (; (end - 1) > (off_t)round_page_32(on+n-1); end-=PAGE_SIZE)
+ NBPGVALID_SET(bp, (end-1)/PAGE_SIZE);
+ }
+ /* Note: pages being written to will be validated when written */
+ }
+ }
+
+ if (ISSET(bp->nb_flags, NB_ERROR)) {
+ error = bp->nb_error;
+ nfs_buf_release(bp, 1);
+ goto out;
+ }
+
+ nfs_node_lock_force(np);
+ np->n_flag |= NMODIFIED;
+ nfs_node_unlock(np);
+
+ NFS_BUF_MAP(bp);
+ error = uiomove((char *)bp->nb_data + on, n, uio);
+ if (error) {
+ SET(bp->nb_flags, NB_ERROR);
+ nfs_buf_release(bp, 1);
+ goto out;
+ }
+
+ /* validate any pages written to */
+ start = on & ~PAGE_MASK;
+ for (; start < on+n; start += PAGE_SIZE) {
+ NBPGVALID_SET(bp, start/PAGE_SIZE);
+ /*
+ * This may seem a little weird, but we don't actually set the
+ * dirty bits for writes. This is because we keep the dirty range
+ * in the nb_dirtyoff/nb_dirtyend fields. Also, particularly for
+ * delayed writes, when we give the pages back to the VM we don't
+ * want to keep them marked dirty, because when we later write the
+ * buffer we won't be able to tell which pages were written dirty
+ * and which pages were mmapped and dirtied.
+ */
+ }
+ if (bp->nb_dirtyend > 0) {
+ bp->nb_dirtyoff = min(on, bp->nb_dirtyoff);
+ bp->nb_dirtyend = max((on + n), bp->nb_dirtyend);
+ } else {
+ bp->nb_dirtyoff = on;
+ bp->nb_dirtyend = on + n;
+ }
+ if (bp->nb_validend <= 0 || bp->nb_validend < bp->nb_dirtyoff ||
+ bp->nb_validoff > bp->nb_dirtyend) {
+ bp->nb_validoff = bp->nb_dirtyoff;
+ bp->nb_validend = bp->nb_dirtyend;
+ } else {
+ bp->nb_validoff = min(bp->nb_validoff, bp->nb_dirtyoff);
+ bp->nb_validend = max(bp->nb_validend, bp->nb_dirtyend);
+ }
+ if (!ISSET(bp->nb_flags, NB_CACHE))
+ nfs_buf_normalize_valid_range(np, bp);
+
+ /*
+ * Since this block is being modified, it must be written
+ * again and not just committed.
+ */
+ if (ISSET(bp->nb_flags, NB_NEEDCOMMIT)) {
+ nfs_node_lock_force(np);
+ if (ISSET(bp->nb_flags, NB_NEEDCOMMIT)) {
+ np->n_needcommitcnt--;
+ CHECK_NEEDCOMMITCNT(np);
+ }
+ CLR(bp->nb_flags, NB_NEEDCOMMIT);
+ nfs_node_unlock(np);
+ }
+
+ if (ioflag & IO_SYNC) {
+ error = nfs_buf_write(bp);
+ if (error)
+ goto out;
+ } else if (((n + on) == biosize) || (ioflag & IO_APPEND) ||
+ (ioflag & IO_NOCACHE) || ISSET(bp->nb_flags, NB_NOCACHE)) {
+ SET(bp->nb_flags, NB_ASYNC);
+ error = nfs_buf_write(bp);
+ if (error)
+ goto out;
+ } else {
+ /* If the block wasn't already delayed: charge for the write */
+ if (!ISSET(bp->nb_flags, NB_DELWRI)) {
+ proc_t p = vfs_context_proc(ctx);
+ if (p && p->p_stats)
+ OSIncrementAtomicLong(&p->p_stats->p_ru.ru_oublock);
+ }
+ nfs_buf_write_delayed(bp);
+ }
+ if (np->n_needcommitcnt >= NFS_A_LOT_OF_NEEDCOMMITS)
+ nfs_flushcommits(np, 1);
+
+ } while (uio_resid(uio) > 0 && n > 0);
+
+out:
+ nfs_node_lock_force(np);
+ np->n_wrbusy--;
+ nfs_node_unlock(np);
+ nfs_data_unlock(np);
+ FSDBG_BOT(515, np, uio_offset(uio), uio_resid(uio), error);
projects / apple / xnu.git / blobdiff