Merge tag '6.3-rc-smb3-client-fixes' of git://git.samba.org/sfrench/cifs-2.6
authorLinus Torvalds <torvalds@linux-foundation.org>
Thu, 23 Feb 2023 01:12:44 +0000 (17:12 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 23 Feb 2023 01:12:44 +0000 (17:12 -0800)
Pull cifs client updates from Steve French:
 "The largest subset of this is from David Howells et al: making the
  cifs/smb3 driver pass iov_iters down to the lowest layers, directly to
  the network transport rather than passing lists of pages around,
  helping multiple areas:

   - Pin user pages, thereby fixing the race between concurrent DIO read
     and fork, where the pages containing the DIO read buffer may end up
     belonging to the child process and not the parent - with the result
     that the parent might not see the retrieved data.

   - cifs shouldn't take refs on pages extracted from non-user-backed
     iterators (eg. KVEC). With these changes, cifs will apply the
     appropriate cleanup.

   - Making it easier to transition to using folios in cifs rather than
     pages by dealing with them through BVEC and XARRAY iterators.

   - Allowing cifs to use the new splice function

  The remainder are:

   - fixes for stable, including various fixes for uninitialized memory,
     wrong length field causing mount issue to very old servers,
     important directory lease fixes and reconnect fixes

   - cleanups (unused code removal, change one element array usage, and
     a change form strtobool to kstrtobool, and Kconfig cleanups)

   - SMBDIRECT (RDMA) fixes including iov_iter integration and UAF fixes

   - reconnect fixes

   - multichannel fixes, including improving channel allocation (to
     least used channel)

   - remove the last use of lock_page_killable by moving to
     folio_lock_killable"

* tag '6.3-rc-smb3-client-fixes' of git://git.samba.org/sfrench/cifs-2.6: (46 commits)
  update internal module version number for cifs.ko
  cifs: update ip_addr for ses only for primary chan setup
  cifs: use tcon allocation functions even for dummy tcon
  cifs: use the least loaded channel for sending requests
  cifs: DIO to/from KVEC-type iterators should now work
  cifs: Remove unused code
  cifs: Build the RDMA SGE list directly from an iterator
  cifs: Change the I/O paths to use an iterator rather than a page list
  cifs: Add a function to read into an iter from a socket
  cifs: Add some helper functions
  cifs: Add a function to Hash the contents of an iterator
  cifs: Add a function to build an RDMA SGE list from an iterator
  netfs: Add a function to extract an iterator into a scatterlist
  netfs: Add a function to extract a UBUF or IOVEC into a BVEC iterator
  cifs: Implement splice_read to pass down ITER_BVEC not ITER_PIPE
  splice: Export filemap/direct_splice_read()
  iov_iter: Add a function to extract a page list from an iterator
  iov_iter: Define flags to qualify page extraction.
  splice: Add a func to do a splice from an O_DIRECT file without ITER_PIPE
  splice: Add a func to do a splice from a buffered file without ITER_PIPE
  ...

21 files changed:
1  2 
block/bio.c
block/blk-map.c
fs/cifs/cifsacl.c
fs/cifs/cifsfs.c
fs/cifs/cifsfs.h
fs/cifs/cifsglob.h
fs/cifs/cifsproto.h
fs/cifs/cifssmb.c
fs/cifs/connect.c
fs/cifs/dir.c
fs/cifs/file.c
fs/cifs/fscache.c
fs/cifs/inode.c
fs/cifs/link.c
fs/cifs/smb2file.c
fs/ksmbd/smb2pdu.c
fs/netfs/iterator.c
fs/splice.c
include/linux/fs.h
include/linux/uio.h
lib/iov_iter.c

diff --cc block/bio.c
Simple merge
diff --cc block/blk-map.c
Simple merge
Simple merge
Simple merge
Simple merge
Simple merge
Simple merge
Simple merge
Simple merge
diff --cc fs/cifs/dir.c
Simple merge
diff --cc fs/cifs/file.c
Simple merge
Simple merge
diff --cc fs/cifs/inode.c
Simple merge
diff --cc fs/cifs/link.c
Simple merge
Simple merge
Simple merge
index 0000000000000000000000000000000000000000,80d7ff440cacd6fbaa9bbeecaad0b5c81356e2e5..f00d43b8ac0a8c6c51ebfadebb359aa1152fdfbf
mode 000000,100644..100644
--- /dev/null
@@@ -1,0 -1,371 +1,369 @@@
 -                      bv[npages + i].bv_page   = *pages++;
 -                      bv[npages + i].bv_offset = offset;
 -                      bv[npages + i].bv_len    = len - offset;
+ // SPDX-License-Identifier: GPL-2.0-or-later
+ /* Iterator helpers.
+  *
+  * Copyright (C) 2022 Red Hat, Inc. All Rights Reserved.
+  * Written by David Howells (dhowells@redhat.com)
+  */
+ #include <linux/export.h>
+ #include <linux/slab.h>
+ #include <linux/mm.h>
+ #include <linux/uio.h>
+ #include <linux/scatterlist.h>
+ #include <linux/netfs.h>
+ #include "internal.h"
+ /**
+  * netfs_extract_user_iter - Extract the pages from a user iterator into a bvec
+  * @orig: The original iterator
+  * @orig_len: The amount of iterator to copy
+  * @new: The iterator to be set up
+  * @extraction_flags: Flags to qualify the request
+  *
+  * Extract the page fragments from the given amount of the source iterator and
+  * build up a second iterator that refers to all of those bits.  This allows
+  * the original iterator to disposed of.
+  *
+  * @extraction_flags can have ITER_ALLOW_P2PDMA set to request peer-to-peer DMA be
+  * allowed on the pages extracted.
+  *
+  * On success, the number of elements in the bvec is returned, the original
+  * iterator will have been advanced by the amount extracted.
+  *
+  * The iov_iter_extract_mode() function should be used to query how cleanup
+  * should be performed.
+  */
+ ssize_t netfs_extract_user_iter(struct iov_iter *orig, size_t orig_len,
+                               struct iov_iter *new,
+                               iov_iter_extraction_t extraction_flags)
+ {
+       struct bio_vec *bv = NULL;
+       struct page **pages;
+       unsigned int cur_npages;
+       unsigned int max_pages;
+       unsigned int npages = 0;
+       unsigned int i;
+       ssize_t ret;
+       size_t count = orig_len, offset, len;
+       size_t bv_size, pg_size;
+       if (WARN_ON_ONCE(!iter_is_ubuf(orig) && !iter_is_iovec(orig)))
+               return -EIO;
+       max_pages = iov_iter_npages(orig, INT_MAX);
+       bv_size = array_size(max_pages, sizeof(*bv));
+       bv = kvmalloc(bv_size, GFP_KERNEL);
+       if (!bv)
+               return -ENOMEM;
+       /* Put the page list at the end of the bvec list storage.  bvec
+        * elements are larger than page pointers, so as long as we work
+        * 0->last, we should be fine.
+        */
+       pg_size = array_size(max_pages, sizeof(*pages));
+       pages = (void *)bv + bv_size - pg_size;
+       while (count && npages < max_pages) {
+               ret = iov_iter_extract_pages(orig, &pages, count,
+                                            max_pages - npages, extraction_flags,
+                                            &offset);
+               if (ret < 0) {
+                       pr_err("Couldn't get user pages (rc=%zd)\n", ret);
+                       break;
+               }
+               if (ret > count) {
+                       pr_err("get_pages rc=%zd more than %zu\n", ret, count);
+                       break;
+               }
+               count -= ret;
+               ret += offset;
+               cur_npages = DIV_ROUND_UP(ret, PAGE_SIZE);
+               if (npages + cur_npages > max_pages) {
+                       pr_err("Out of bvec array capacity (%u vs %u)\n",
+                              npages + cur_npages, max_pages);
+                       break;
+               }
+               for (i = 0; i < cur_npages; i++) {
+                       len = ret > PAGE_SIZE ? PAGE_SIZE : ret;
++                      bvec_set_page(bv + npages + i, *pages++, len - offset, offset);
+                       ret -= len;
+                       offset = 0;
+               }
+               npages += cur_npages;
+       }
+       iov_iter_bvec(new, orig->data_source, bv, npages, orig_len - count);
+       return npages;
+ }
+ EXPORT_SYMBOL_GPL(netfs_extract_user_iter);
+ /*
+  * Extract and pin a list of up to sg_max pages from UBUF- or IOVEC-class
+  * iterators, and add them to the scatterlist.
+  */
+ static ssize_t netfs_extract_user_to_sg(struct iov_iter *iter,
+                                       ssize_t maxsize,
+                                       struct sg_table *sgtable,
+                                       unsigned int sg_max,
+                                       iov_iter_extraction_t extraction_flags)
+ {
+       struct scatterlist *sg = sgtable->sgl + sgtable->nents;
+       struct page **pages;
+       unsigned int npages;
+       ssize_t ret = 0, res;
+       size_t len, off;
+       /* We decant the page list into the tail of the scatterlist */
+       pages = (void *)sgtable->sgl + array_size(sg_max, sizeof(struct scatterlist));
+       pages -= sg_max;
+       do {
+               res = iov_iter_extract_pages(iter, &pages, maxsize, sg_max,
+                                            extraction_flags, &off);
+               if (res < 0)
+                       goto failed;
+               len = res;
+               maxsize -= len;
+               ret += len;
+               npages = DIV_ROUND_UP(off + len, PAGE_SIZE);
+               sg_max -= npages;
+               for (; npages < 0; npages--) {
+                       struct page *page = *pages;
+                       size_t seg = min_t(size_t, PAGE_SIZE - off, len);
+                       *pages++ = NULL;
+                       sg_set_page(sg, page, len, off);
+                       sgtable->nents++;
+                       sg++;
+                       len -= seg;
+                       off = 0;
+               }
+       } while (maxsize > 0 && sg_max > 0);
+       return ret;
+ failed:
+       while (sgtable->nents > sgtable->orig_nents)
+               put_page(sg_page(&sgtable->sgl[--sgtable->nents]));
+       return res;
+ }
+ /*
+  * Extract up to sg_max pages from a BVEC-type iterator and add them to the
+  * scatterlist.  The pages are not pinned.
+  */
+ static ssize_t netfs_extract_bvec_to_sg(struct iov_iter *iter,
+                                       ssize_t maxsize,
+                                       struct sg_table *sgtable,
+                                       unsigned int sg_max,
+                                       iov_iter_extraction_t extraction_flags)
+ {
+       const struct bio_vec *bv = iter->bvec;
+       struct scatterlist *sg = sgtable->sgl + sgtable->nents;
+       unsigned long start = iter->iov_offset;
+       unsigned int i;
+       ssize_t ret = 0;
+       for (i = 0; i < iter->nr_segs; i++) {
+               size_t off, len;
+               len = bv[i].bv_len;
+               if (start >= len) {
+                       start -= len;
+                       continue;
+               }
+               len = min_t(size_t, maxsize, len - start);
+               off = bv[i].bv_offset + start;
+               sg_set_page(sg, bv[i].bv_page, len, off);
+               sgtable->nents++;
+               sg++;
+               sg_max--;
+               ret += len;
+               maxsize -= len;
+               if (maxsize <= 0 || sg_max == 0)
+                       break;
+               start = 0;
+       }
+       if (ret > 0)
+               iov_iter_advance(iter, ret);
+       return ret;
+ }
+ /*
+  * Extract up to sg_max pages from a KVEC-type iterator and add them to the
+  * scatterlist.  This can deal with vmalloc'd buffers as well as kmalloc'd or
+  * static buffers.  The pages are not pinned.
+  */
+ static ssize_t netfs_extract_kvec_to_sg(struct iov_iter *iter,
+                                       ssize_t maxsize,
+                                       struct sg_table *sgtable,
+                                       unsigned int sg_max,
+                                       iov_iter_extraction_t extraction_flags)
+ {
+       const struct kvec *kv = iter->kvec;
+       struct scatterlist *sg = sgtable->sgl + sgtable->nents;
+       unsigned long start = iter->iov_offset;
+       unsigned int i;
+       ssize_t ret = 0;
+       for (i = 0; i < iter->nr_segs; i++) {
+               struct page *page;
+               unsigned long kaddr;
+               size_t off, len, seg;
+               len = kv[i].iov_len;
+               if (start >= len) {
+                       start -= len;
+                       continue;
+               }
+               kaddr = (unsigned long)kv[i].iov_base + start;
+               off = kaddr & ~PAGE_MASK;
+               len = min_t(size_t, maxsize, len - start);
+               kaddr &= PAGE_MASK;
+               maxsize -= len;
+               ret += len;
+               do {
+                       seg = min_t(size_t, len, PAGE_SIZE - off);
+                       if (is_vmalloc_or_module_addr((void *)kaddr))
+                               page = vmalloc_to_page((void *)kaddr);
+                       else
+                               page = virt_to_page(kaddr);
+                       sg_set_page(sg, page, len, off);
+                       sgtable->nents++;
+                       sg++;
+                       sg_max--;
+                       len -= seg;
+                       kaddr += PAGE_SIZE;
+                       off = 0;
+               } while (len > 0 && sg_max > 0);
+               if (maxsize <= 0 || sg_max == 0)
+                       break;
+               start = 0;
+       }
+       if (ret > 0)
+               iov_iter_advance(iter, ret);
+       return ret;
+ }
+ /*
+  * Extract up to sg_max folios from an XARRAY-type iterator and add them to
+  * the scatterlist.  The pages are not pinned.
+  */
+ static ssize_t netfs_extract_xarray_to_sg(struct iov_iter *iter,
+                                         ssize_t maxsize,
+                                         struct sg_table *sgtable,
+                                         unsigned int sg_max,
+                                         iov_iter_extraction_t extraction_flags)
+ {
+       struct scatterlist *sg = sgtable->sgl + sgtable->nents;
+       struct xarray *xa = iter->xarray;
+       struct folio *folio;
+       loff_t start = iter->xarray_start + iter->iov_offset;
+       pgoff_t index = start / PAGE_SIZE;
+       ssize_t ret = 0;
+       size_t offset, len;
+       XA_STATE(xas, xa, index);
+       rcu_read_lock();
+       xas_for_each(&xas, folio, ULONG_MAX) {
+               if (xas_retry(&xas, folio))
+                       continue;
+               if (WARN_ON(xa_is_value(folio)))
+                       break;
+               if (WARN_ON(folio_test_hugetlb(folio)))
+                       break;
+               offset = offset_in_folio(folio, start);
+               len = min_t(size_t, maxsize, folio_size(folio) - offset);
+               sg_set_page(sg, folio_page(folio, 0), len, offset);
+               sgtable->nents++;
+               sg++;
+               sg_max--;
+               maxsize -= len;
+               ret += len;
+               if (maxsize <= 0 || sg_max == 0)
+                       break;
+       }
+       rcu_read_unlock();
+       if (ret > 0)
+               iov_iter_advance(iter, ret);
+       return ret;
+ }
+ /**
+  * netfs_extract_iter_to_sg - Extract pages from an iterator and add ot an sglist
+  * @iter: The iterator to extract from
+  * @maxsize: The amount of iterator to copy
+  * @sgtable: The scatterlist table to fill in
+  * @sg_max: Maximum number of elements in @sgtable that may be filled
+  * @extraction_flags: Flags to qualify the request
+  *
+  * Extract the page fragments from the given amount of the source iterator and
+  * add them to a scatterlist that refers to all of those bits, to a maximum
+  * addition of @sg_max elements.
+  *
+  * The pages referred to by UBUF- and IOVEC-type iterators are extracted and
+  * pinned; BVEC-, KVEC- and XARRAY-type are extracted but aren't pinned; PIPE-
+  * and DISCARD-type are not supported.
+  *
+  * No end mark is placed on the scatterlist; that's left to the caller.
+  *
+  * @extraction_flags can have ITER_ALLOW_P2PDMA set to request peer-to-peer DMA
+  * be allowed on the pages extracted.
+  *
+  * If successul, @sgtable->nents is updated to include the number of elements
+  * added and the number of bytes added is returned.  @sgtable->orig_nents is
+  * left unaltered.
+  *
+  * The iov_iter_extract_mode() function should be used to query how cleanup
+  * should be performed.
+  */
+ ssize_t netfs_extract_iter_to_sg(struct iov_iter *iter, size_t maxsize,
+                                struct sg_table *sgtable, unsigned int sg_max,
+                                iov_iter_extraction_t extraction_flags)
+ {
+       if (maxsize == 0)
+               return 0;
+       switch (iov_iter_type(iter)) {
+       case ITER_UBUF:
+       case ITER_IOVEC:
+               return netfs_extract_user_to_sg(iter, maxsize, sgtable, sg_max,
+                                               extraction_flags);
+       case ITER_BVEC:
+               return netfs_extract_bvec_to_sg(iter, maxsize, sgtable, sg_max,
+                                               extraction_flags);
+       case ITER_KVEC:
+               return netfs_extract_kvec_to_sg(iter, maxsize, sgtable, sg_max,
+                                               extraction_flags);
+       case ITER_XARRAY:
+               return netfs_extract_xarray_to_sg(iter, maxsize, sgtable, sg_max,
+                                                 extraction_flags);
+       default:
+               pr_err("%s(%u) unsupported\n", __func__, iov_iter_type(iter));
+               WARN_ON_ONCE(1);
+               return -EIO;
+       }
+ }
+ EXPORT_SYMBOL_GPL(netfs_extract_iter_to_sg);
diff --cc fs/splice.c
Simple merge
Simple merge
Simple merge
diff --cc lib/iov_iter.c
Simple merge