[Drbd-dev] [PATCH RFC PKS/PMEM 00/58] PMEM: Introduce stray write protection for PMEM

[Drbd-dev] [PATCH RFC PKS/PMEM 00/58] PMEM: Introduce stray write protection for PMEM

[ira.weiny]

From: Ira Weiny <ira.weiny@intel.com>

Should a stray write in the kernel occur persistent memory is affected more
than regular memory.  A write to the wrong area of memory could result in
latent data corruption which will will persist after a reboot.  PKS provides a
nice way to restrict access to persistent memory kernel mappings, while
providing fast access when needed.

Since the last RFC[1] this patch set has grown quite a bit.  It now depends on
the core patches submitted separately.

	https://lore.kernel.org/lkml/20201009194258.3207172-1-ira.weiny@intel.com/

And contained in the git tree here:

	https://github.com/weiny2/linux-kernel/tree/pks-rfc-v3

However, functionally there is only 1 major change from the last RFC.
Specifically, kmap() is most often used within a single thread in a 'map/do
something/unmap' pattern.  In fact this is the pattern used in ~90% of the
callers of kmap().  This pattern works very well for the pmem use case and the
testing which was done.  However, there were another ~20-30 kmap users which do
not follow this pattern.  Some of them seem to expect the mapping to be
'global' while others require a detailed audit to be sure.[2][3]

While we don't anticipate global mappings to pmem there is a danger in
changing the semantics of kmap().  Effectively, this would cause an unresolved
page fault with little to no information about why.

There were a number of options considered.

1) Attempt to change all the thread local kmap() calls to kmap_atomic()
2) Introduce a flags parameter to kmap() to indicate if the mapping should be
   global or not
3) Change ~20-30 call sites to 'kmap_global()' to indicate that they require a
   global mapping of the pages
4) Change ~209 call sites to 'kmap_thread()' to indicate that the mapping is to
   be used within that thread of execution only

Option 1 is simply not feasible kmap_atomic() is not the same semantic as
kmap() within a single tread.  Option 2 would require all of the call sites of
kmap() to change.  Option 3 seems like a good minimal change but there is a
danger that new code may miss the semantic change of kmap() and not get the
behavior intended for future users.  Therefore, option #4 was chosen.

To handle the global PKRS state in the most efficient manner possible.  We
lazily override the thread specific PKRS key value only when needed because we
anticipate PKS to not be needed will not be needed most of the time.  And even
when it is used 90% of the time it is a thread local call.


[1] https://lore.kernel.org/lkml/20200717072056.73134-1-ira.weiny@intel.com/

[2] The following list of callers continue calling kmap() (utilizing the global
PKRS).  It would be nice if more of them could be converted to kmap_thread()

	drivers/firewire/net.c:         ptr = kmap(dev->broadcast_rcv_buffer.pages[u]);
	drivers/gpu/drm/i915/gem/i915_gem_pages.c:              return kmap(sg_page(sgt->sgl));
	drivers/gpu/drm/ttm/ttm_bo_util.c:              map->virtual = kmap(map->page);
	drivers/infiniband/hw/qib/qib_user_sdma.c:      mpage = kmap(page);
	drivers/misc/vmw_vmci/vmci_host.c:      context->notify = kmap(context->notify_page) + (uva & (PAGE_SIZE - 1));
	drivers/misc/xilinx_sdfec.c:            addr = kmap(pages[i]);
	drivers/mmc/host/usdhi6rol0.c:  host->pg.mapped         = kmap(host->pg.page);
	drivers/mmc/host/usdhi6rol0.c:  host->pg.mapped = kmap(host->pg.page);
	drivers/mmc/host/usdhi6rol0.c:  host->pg.mapped = kmap(host->pg.page);
	drivers/nvme/target/tcp.c:              iov->iov_base = kmap(sg_page(sg)) + sg->offset + sg_offset;
	drivers/scsi/libiscsi_tcp.c:            segment->sg_mapped = kmap(sg_page(sg));
	drivers/target/iscsi/iscsi_target.c:            iov[i].iov_base = kmap(sg_page(sg)) + sg->offset + page_off;
	drivers/target/target_core_transport.c:         return kmap(sg_page(sg)) + sg->offset;
	fs/btrfs/check-integrity.c:             block_ctx->datav[i] = kmap(block_ctx->pagev[i]);
	fs/ceph/dir.c:          cache_ctl->dentries = kmap(cache_ctl->page);
	fs/ceph/inode.c:                ctl->dentries = kmap(ctl->page);
	fs/erofs/zpvec.h:               kmap_atomic(ctor->curr) : kmap(ctor->curr);
	lib/scatterlist.c:              miter->addr = kmap(miter->page) + miter->__offset;
	net/ceph/pagelist.c:    pl->mapped_tail = kmap(page);
	net/ceph/pagelist.c:            pl->mapped_tail = kmap(page);
	virt/kvm/kvm_main.c:                    hva = kmap(page);

[3] The following appear to follow the same pattern as ext2 which was converted
after some code audit.  So I _think_ they too could be converted to
k[un]map_thread().

	fs/freevxfs/vxfs_subr.c|75| kmap(pp);
	fs/jfs/jfs_metapage.c|102| kmap(page);
	fs/jfs/jfs_metapage.c|156| kmap(page);
	fs/minix/dir.c|72| kmap(page);
	fs/nilfs2/dir.c|195| kmap(page);
	fs/nilfs2/ifile.h|24| void *kaddr = kmap(ibh->b_page);
	fs/ntfs/aops.h|78| kmap(page);
	fs/ntfs/compress.c|574| kmap(page);
	fs/qnx6/dir.c|32| kmap(page);
	fs/qnx6/dir.c|58| kmap(*p = page);
	fs/qnx6/inode.c|190| kmap(page);
	fs/qnx6/inode.c|557| kmap(page);
	fs/reiserfs/inode.c|2397| kmap(bh_result->b_page);
	fs/reiserfs/xattr.c|444| kmap(page);
	fs/sysv/dir.c|60| kmap(page);
	fs/sysv/dir.c|262| kmap(page);
	fs/ufs/dir.c|194| kmap(page);
	fs/ufs/dir.c|562| kmap(page);


Ira Weiny (58):
  x86/pks: Add a global pkrs option
  x86/pks/test: Add testing for global option
  memremap: Add zone device access protection
  kmap: Add stray access protection for device pages
  kmap: Introduce k[un]map_thread
  kmap: Introduce k[un]map_thread debugging
  drivers/drbd: Utilize new kmap_thread()
  drivers/firmware_loader: Utilize new kmap_thread()
  drivers/gpu: Utilize new kmap_thread()
  drivers/rdma: Utilize new kmap_thread()
  drivers/net: Utilize new kmap_thread()
  fs/afs: Utilize new kmap_thread()
  fs/btrfs: Utilize new kmap_thread()
  fs/cifs: Utilize new kmap_thread()
  fs/ecryptfs: Utilize new kmap_thread()
  fs/gfs2: Utilize new kmap_thread()
  fs/nilfs2: Utilize new kmap_thread()
  fs/hfs: Utilize new kmap_thread()
  fs/hfsplus: Utilize new kmap_thread()
  fs/jffs2: Utilize new kmap_thread()
  fs/nfs: Utilize new kmap_thread()
  fs/f2fs: Utilize new kmap_thread()
  fs/fuse: Utilize new kmap_thread()
  fs/freevxfs: Utilize new kmap_thread()
  fs/reiserfs: Utilize new kmap_thread()
  fs/zonefs: Utilize new kmap_thread()
  fs/ubifs: Utilize new kmap_thread()
  fs/cachefiles: Utilize new kmap_thread()
  fs/ntfs: Utilize new kmap_thread()
  fs/romfs: Utilize new kmap_thread()
  fs/vboxsf: Utilize new kmap_thread()
  fs/hostfs: Utilize new kmap_thread()
  fs/cramfs: Utilize new kmap_thread()
  fs/erofs: Utilize new kmap_thread()
  fs: Utilize new kmap_thread()
  fs/ext2: Use ext2_put_page
  fs/ext2: Utilize new kmap_thread()
  fs/isofs: Utilize new kmap_thread()
  fs/jffs2: Utilize new kmap_thread()
  net: Utilize new kmap_thread()
  drivers/target: Utilize new kmap_thread()
  drivers/scsi: Utilize new kmap_thread()
  drivers/mmc: Utilize new kmap_thread()
  drivers/xen: Utilize new kmap_thread()
  drivers/firmware: Utilize new kmap_thread()
  drives/staging: Utilize new kmap_thread()
  drivers/mtd: Utilize new kmap_thread()
  drivers/md: Utilize new kmap_thread()
  drivers/misc: Utilize new kmap_thread()
  drivers/android: Utilize new kmap_thread()
  kernel: Utilize new kmap_thread()
  mm: Utilize new kmap_thread()
  lib: Utilize new kmap_thread()
  powerpc: Utilize new kmap_thread()
  samples: Utilize new kmap_thread()
  dax: Stray access protection for dax_direct_access()
  nvdimm/pmem: Stray access protection for pmem->virt_addr
  [dax|pmem]: Enable stray access protection

 Documentation/core-api/protection-keys.rst    |  11 +-
 arch/powerpc/mm/mem.c                         |   4 +-
 arch/x86/entry/common.c                       |  28 +++
 arch/x86/include/asm/pkeys.h                  |   6 +-
 arch/x86/include/asm/pkeys_common.h           |   8 +-
 arch/x86/kernel/process.c                     |  74 ++++++-
 arch/x86/mm/fault.c                           | 193 ++++++++++++++----
 arch/x86/mm/pkeys.c                           |  88 ++++++--
 drivers/android/binder_alloc.c                |   4 +-
 drivers/base/firmware_loader/fallback.c       |   4 +-
 drivers/base/firmware_loader/main.c           |   4 +-
 drivers/block/drbd/drbd_main.c                |   4 +-
 drivers/block/drbd/drbd_receiver.c            |  12 +-
 drivers/dax/device.c                          |   2 +
 drivers/dax/super.c                           |   2 +
 drivers/firmware/efi/capsule-loader.c         |   6 +-
 drivers/firmware/efi/capsule.c                |   4 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c       |  12 +-
 drivers/gpu/drm/gma500/gma_display.c          |   4 +-
 drivers/gpu/drm/gma500/mmu.c                  |  10 +-
 drivers/gpu/drm/i915/gem/i915_gem_shmem.c     |   4 +-
 .../drm/i915/gem/selftests/i915_gem_context.c |   4 +-
 .../drm/i915/gem/selftests/i915_gem_mman.c    |   8 +-
 drivers/gpu/drm/i915/gt/intel_ggtt_fencing.c  |   4 +-
 drivers/gpu/drm/i915/gt/intel_gtt.c           |   4 +-
 drivers/gpu/drm/i915/gt/shmem_utils.c         |   4 +-
 drivers/gpu/drm/i915/i915_gem.c               |   8 +-
 drivers/gpu/drm/i915/i915_gpu_error.c         |   4 +-
 drivers/gpu/drm/i915/selftests/i915_perf.c    |   4 +-
 drivers/gpu/drm/radeon/radeon_ttm.c           |   4 +-
 drivers/infiniband/hw/hfi1/sdma.c             |   4 +-
 drivers/infiniband/hw/i40iw/i40iw_cm.c        |  10 +-
 drivers/infiniband/sw/siw/siw_qp_tx.c         |  14 +-
 drivers/md/bcache/request.c                   |   4 +-
 drivers/misc/vmw_vmci/vmci_queue_pair.c       |  12 +-
 drivers/mmc/host/mmc_spi.c                    |   4 +-
 drivers/mmc/host/sdricoh_cs.c                 |   4 +-
 drivers/mtd/mtd_blkdevs.c                     |  12 +-
 drivers/net/ethernet/intel/igb/igb_ethtool.c  |   4 +-
 .../net/ethernet/intel/ixgbe/ixgbe_ethtool.c  |   4 +-
 drivers/nvdimm/pmem.c                         |   6 +
 drivers/scsi/ipr.c                            |   8 +-
 drivers/scsi/pmcraid.c                        |   8 +-
 drivers/staging/rts5208/rtsx_transport.c      |   4 +-
 drivers/target/target_core_iblock.c           |   4 +-
 drivers/target/target_core_rd.c               |   4 +-
 drivers/target/target_core_transport.c        |   4 +-
 drivers/xen/gntalloc.c                        |   4 +-
 fs/afs/dir.c                                  |  16 +-
 fs/afs/dir_edit.c                             |  16 +-
 fs/afs/mntpt.c                                |   4 +-
 fs/afs/write.c                                |   4 +-
 fs/aio.c                                      |   4 +-
 fs/binfmt_elf.c                               |   4 +-
 fs/binfmt_elf_fdpic.c                         |   4 +-
 fs/btrfs/check-integrity.c                    |   4 +-
 fs/btrfs/compression.c                        |   4 +-
 fs/btrfs/inode.c                              |  16 +-
 fs/btrfs/lzo.c                                |  24 +--
 fs/btrfs/raid56.c                             |  34 +--
 fs/btrfs/reflink.c                            |   8 +-
 fs/btrfs/send.c                               |   4 +-
 fs/btrfs/zlib.c                               |  32 +--
 fs/btrfs/zstd.c                               |  20 +-
 fs/cachefiles/rdwr.c                          |   4 +-
 fs/cifs/cifsencrypt.c                         |   6 +-
 fs/cifs/file.c                                |  16 +-
 fs/cifs/smb2ops.c                             |   8 +-
 fs/cramfs/inode.c                             |  10 +-
 fs/ecryptfs/crypto.c                          |   8 +-
 fs/ecryptfs/read_write.c                      |   8 +-
 fs/erofs/super.c                              |   4 +-
 fs/erofs/xattr.c                              |   4 +-
 fs/exec.c                                     |  10 +-
 fs/ext2/dir.c                                 |   8 +-
 fs/ext2/ext2.h                                |   8 +
 fs/ext2/namei.c                               |  15 +-
 fs/f2fs/f2fs.h                                |   8 +-
 fs/freevxfs/vxfs_immed.c                      |   4 +-
 fs/fuse/readdir.c                             |   4 +-
 fs/gfs2/bmap.c                                |   4 +-
 fs/gfs2/ops_fstype.c                          |   4 +-
 fs/hfs/bnode.c                                |  14 +-
 fs/hfs/btree.c                                |  20 +-
 fs/hfsplus/bitmap.c                           |  20 +-
 fs/hfsplus/bnode.c                            | 102 ++++-----
 fs/hfsplus/btree.c                            |  18 +-
 fs/hostfs/hostfs_kern.c                       |  12 +-
 fs/io_uring.c                                 |   4 +-
 fs/isofs/compress.c                           |   4 +-
 fs/jffs2/file.c                               |   8 +-
 fs/jffs2/gc.c                                 |   4 +-
 fs/nfs/dir.c                                  |  20 +-
 fs/nilfs2/alloc.c                             |  34 +--
 fs/nilfs2/cpfile.c                            |   4 +-
 fs/ntfs/aops.c                                |   4 +-
 fs/reiserfs/journal.c                         |   4 +-
 fs/romfs/super.c                              |   4 +-
 fs/splice.c                                   |   4 +-
 fs/ubifs/file.c                               |  16 +-
 fs/vboxsf/file.c                              |  12 +-
 fs/zonefs/super.c                             |   4 +-
 include/linux/entry-common.h                  |   3 +
 include/linux/highmem.h                       |  63 +++++-
 include/linux/memremap.h                      |   1 +
 include/linux/mm.h                            |  43 ++++
 include/linux/pkeys.h                         |   6 +-
 include/linux/sched.h                         |   8 +
 include/trace/events/kmap_thread.h            |  56 +++++
 init/init_task.c                              |   6 +
 kernel/fork.c                                 |  18 ++
 kernel/kexec_core.c                           |   8 +-
 lib/Kconfig.debug                             |   8 +
 lib/iov_iter.c                                |  12 +-
 lib/pks/pks_test.c                            | 138 +++++++++++--
 lib/test_bpf.c                                |   4 +-
 lib/test_hmm.c                                |   8 +-
 mm/Kconfig                                    |  13 ++
 mm/debug.c                                    |  23 +++
 mm/memory.c                                   |   8 +-
 mm/memremap.c                                 |  90 ++++++++
 mm/swapfile.c                                 |   4 +-
 mm/userfaultfd.c                              |   4 +-
 net/ceph/messenger.c                          |   4 +-
 net/core/datagram.c                           |   4 +-
 net/core/sock.c                               |   8 +-
 net/ipv4/ip_output.c                          |   4 +-
 net/sunrpc/cache.c                            |   4 +-
 net/sunrpc/xdr.c                              |   8 +-
 net/tls/tls_device.c                          |   4 +-
 samples/vfio-mdev/mbochs.c                    |   4 +-
 131 files changed, 1284 insertions(+), 565 deletions(-)
 create mode 100644 include/trace/events/kmap_thread.h

-- 
2.28.0.rc0.12.gb6a658bd00c9

[Drbd-dev] [PATCH RFC PKS/PMEM 04/58] kmap: Add stray access protection for device pages

[ira.weiny]

From: Ira Weiny <ira.weiny@intel.com>

Device managed pages may have additional protections.  These protections
need to be removed prior to valid use by kernel users.

Check for special treatment of device managed pages in kmap and take
action if needed.  We use kmap as an interface for generic kernel code
because under normal circumstances it would be a bug for general kernel
code to not use kmap prior to accessing kernel memory.  Therefore, this
should allow any valid kernel users to seamlessly use these pages
without issues.

Because of the critical nature of kmap it must be pointed out that the
over head on regular DRAM is carefully implemented to be as fast as
possible.  Furthermore the underlying MSR write required on device pages
when protected is better than a normal MSR write.

Specifically, WRMSR(MSR_IA32_PKRS) is not serializing but still
maintains ordering properties similar to WRPKRU.  The current SDM
section on PKRS needs updating but should be the same as that of WRPKRU.
So to quote from the WRPKRU text:

	WRPKRU will never execute speculatively. Memory accesses
	affected by PKRU register will not execute (even speculatively)
	until all prior executions of WRPKRU have completed execution
	and updated the PKRU register.

Still this will make accessing pmem more expensive from the kernel but
the overhead is minimized and many pmem users access this memory through
user page mappings which are not affected at all.

Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
---
 include/linux/highmem.h | 32 +++++++++++++++++++++++++++++++-
 1 file changed, 31 insertions(+), 1 deletion(-)

diff --git a/include/linux/highmem.h b/include/linux/highmem.h
index 14e6202ce47f..2a9806e3b8d2 100644
--- a/include/linux/highmem.h
+++ b/include/linux/highmem.h
@@ -8,6 +8,7 @@
 #include <linux/mm.h>
 #include <linux/uaccess.h>
 #include <linux/hardirq.h>
+#include <linux/memremap.h>
 
 #include <asm/cacheflush.h>
 
@@ -31,6 +32,20 @@ static inline void invalidate_kernel_vmap_range(void *vaddr, int size)
 
 #include <asm/kmap_types.h>
 
+static inline void dev_page_enable_access(struct page *page, bool global)
+{
+	if (!page_is_access_protected(page))
+		return;
+	dev_access_enable(global);
+}
+
+static inline void dev_page_disable_access(struct page *page, bool global)
+{
+	if (!page_is_access_protected(page))
+		return;
+	dev_access_disable(global);
+}
+
 #ifdef CONFIG_HIGHMEM
 extern void *kmap_atomic_high_prot(struct page *page, pgprot_t prot);
 extern void kunmap_atomic_high(void *kvaddr);
@@ -55,6 +70,11 @@ static inline void *kmap(struct page *page)
 	else
 		addr = kmap_high(page);
 	kmap_flush_tlb((unsigned long)addr);
+	/*
+	 * Even non-highmem pages may have additional access protections which
+	 * need to be checked and potentially enabled.
+	 */
+	dev_page_enable_access(page, true);
 	return addr;
 }
 
@@ -63,6 +83,11 @@ void kunmap_high(struct page *page);
 static inline void kunmap(struct page *page)
 {
 	might_sleep();
+	/*
+	 * Even non-highmem pages may have additional access protections which
+	 * need to be checked and potentially disabled.
+	 */
+	dev_page_disable_access(page, true);
 	if (!PageHighMem(page))
 		return;
 	kunmap_high(page);
@@ -85,6 +110,7 @@ static inline void *kmap_atomic_prot(struct page *page, pgprot_t prot)
 {
 	preempt_disable();
 	pagefault_disable();
+	dev_page_enable_access(page, false);
 	if (!PageHighMem(page))
 		return page_address(page);
 	return kmap_atomic_high_prot(page, prot);
@@ -137,6 +163,7 @@ static inline unsigned long totalhigh_pages(void) { return 0UL; }
 static inline void *kmap(struct page *page)
 {
 	might_sleep();
+	dev_page_enable_access(page, true);
 	return page_address(page);
 }
 
@@ -146,6 +173,7 @@ static inline void kunmap_high(struct page *page)
 
 static inline void kunmap(struct page *page)
 {
+	dev_page_disable_access(page, true);
 #ifdef ARCH_HAS_FLUSH_ON_KUNMAP
 	kunmap_flush_on_unmap(page_address(page));
 #endif
@@ -155,6 +183,7 @@ static inline void *kmap_atomic(struct page *page)
 {
 	preempt_disable();
 	pagefault_disable();
+	dev_page_enable_access(page, false);
 	return page_address(page);
 }
 #define kmap_atomic_prot(page, prot)	kmap_atomic(page)
@@ -216,7 +245,8 @@ static inline void kmap_atomic_idx_pop(void)
 #define kunmap_atomic(addr)                                     \
 do {                                                            \
 	BUILD_BUG_ON(__same_type((addr), struct page *));       \
-	kunmap_atomic_high(addr);                                  \
+	dev_page_disable_access(kmap_to_page(addr), false);     \
+	kunmap_atomic_high(addr);                               \
 	pagefault_enable();                                     \
 	preempt_enable();                                       \
 } while (0)
-- 
2.28.0.rc0.12.gb6a658bd00c9

[Drbd-dev] [PATCH RFC PKS/PMEM 05/58] kmap: Introduce k[un]map_thread

[ira.weiny]

From: Ira Weiny <ira.weiny@intel.com>

To correctly support the semantics of kmap() with Kernel protection keys
(PKS), kmap() may be required to set the protections on multiple
processors (globally).  Enabling PKS globally can be very expensive
depending on the requested operation.  Furthermore, enabling a domain
globally reduces the protection afforded by PKS.

Most kmap() (Aprox 209 of 229) callers use the map within a single thread and
have no need for the protection domain to be enabled globally.  However, the
remaining callers do not follow this pattern and, as best I can tell, expect
the mapping to be 'global' and available to any thread who may access the
mapping.[1]

We don't anticipate global mappings to pmem, however in general there is a
danger in changing the semantics of kmap().  Effectively, this would cause an
unresolved page fault with little to no information about why the failure
occurred.

To resolve this a number of options were considered.

1) Attempt to change all the thread local kmap() calls to kmap_atomic()[2]
2) Introduce a flags parameter to kmap() to indicate if the mapping should be
   global or not
3) Change ~20 call sites to 'kmap_global()' to indicate that they require a
   global enablement of the pages.
4) Change ~209 call sites to 'kmap_thread()' to indicate that the mapping is to
   be used within that thread of execution only

Option 1 is simply not feasible.  Option 2 would require all of the call sites
of kmap() to change.  Option 3 seems like a good minimal change but there is a
danger that new code may miss the semantic change of kmap() and not get the
behavior the developer intended.  Therefore, #4 was chosen.

Subsequent patches will convert most ~90% of the kmap callers to this new call
leaving about 10% of the existing kmap callers to enable PKS globally.

Cc: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
---
 include/linux/highmem.h | 34 ++++++++++++++++++++++++++--------
 1 file changed, 26 insertions(+), 8 deletions(-)

diff --git a/include/linux/highmem.h b/include/linux/highmem.h
index 2a9806e3b8d2..ef7813544719 100644
--- a/include/linux/highmem.h
+++ b/include/linux/highmem.h
@@ -60,7 +60,7 @@ static inline void kmap_flush_tlb(unsigned long addr) { }
 #endif
 
 void *kmap_high(struct page *page);
-static inline void *kmap(struct page *page)
+static inline void *__kmap(struct page *page, bool global)
 {
 	void *addr;
 
@@ -74,20 +74,20 @@ static inline void *kmap(struct page *page)
 	 * Even non-highmem pages may have additional access protections which
 	 * need to be checked and potentially enabled.
 	 */
-	dev_page_enable_access(page, true);
+	dev_page_enable_access(page, global);
 	return addr;
 }
 
 void kunmap_high(struct page *page);
 
-static inline void kunmap(struct page *page)
+static inline void __kunmap(struct page *page, bool global)
 {
 	might_sleep();
 	/*
 	 * Even non-highmem pages may have additional access protections which
 	 * need to be checked and potentially disabled.
 	 */
-	dev_page_disable_access(page, true);
+	dev_page_disable_access(page, global);
 	if (!PageHighMem(page))
 		return;
 	kunmap_high(page);
@@ -160,10 +160,10 @@ static inline struct page *kmap_to_page(void *addr)
 
 static inline unsigned long totalhigh_pages(void) { return 0UL; }
 
-static inline void *kmap(struct page *page)
+static inline void *__kmap(struct page *page, bool global)
 {
 	might_sleep();
-	dev_page_enable_access(page, true);
+	dev_page_enable_access(page, global);
 	return page_address(page);
 }
 
@@ -171,9 +171,9 @@ static inline void kunmap_high(struct page *page)
 {
 }
 
-static inline void kunmap(struct page *page)
+static inline void __kunmap(struct page *page, bool global)
 {
-	dev_page_disable_access(page, true);
+	dev_page_disable_access(page, global);
 #ifdef ARCH_HAS_FLUSH_ON_KUNMAP
 	kunmap_flush_on_unmap(page_address(page));
 #endif
@@ -238,6 +238,24 @@ static inline void kmap_atomic_idx_pop(void)
 
 #endif
 
+static inline void *kmap(struct page *page)
+{
+	return __kmap(page, true);
+}
+static inline void kunmap(struct page *page)
+{
+	__kunmap(page, true);
+}
+
+static inline void *kmap_thread(struct page *page)
+{
+	return __kmap(page, false);
+}
+static inline void kunmap_thread(struct page *page)
+{
+	__kunmap(page, false);
+}
+
 /*
  * Prevent people trying to call kunmap_atomic() as if it were kunmap()
  * kunmap_atomic() should get the return value of kmap_atomic, not the page.
-- 
2.28.0.rc0.12.gb6a658bd00c9

[Drbd-dev] [PATCH RFC PKS/PMEM 07/58] drivers/drbd: Utilize new kmap_thread()

[ira.weiny]

From: Ira Weiny <ira.weiny@intel.com>

The kmap() calls in this driver are localized to a single thread.  To
avoid the over head of global PKRS updates use the new kmap_thread()
call.

Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
---
 drivers/block/drbd/drbd_main.c     |  4 ++--
 drivers/block/drbd/drbd_receiver.c | 12 ++++++------
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/drivers/block/drbd/drbd_main.c b/drivers/block/drbd/drbd_main.c
index 573dbf6f0c31..f0d0c6b0745e 100644
--- a/drivers/block/drbd/drbd_main.c
+++ b/drivers/block/drbd/drbd_main.c
@@ -1532,9 +1532,9 @@ static int _drbd_no_send_page(struct drbd_peer_device *peer_device, struct page
 	int err;
 
 	socket = peer_device->connection->data.socket;
-	addr = kmap(page) + offset;
+	addr = kmap_thread(page) + offset;
 	err = drbd_send_all(peer_device->connection, socket, addr, size, msg_flags);
-	kunmap(page);
+	kunmap_thread(page);
 	if (!err)
 		peer_device->device->send_cnt += size >> 9;
 	return err;
diff --git a/drivers/block/drbd/drbd_receiver.c b/drivers/block/drbd/drbd_receiver.c
index 422363daa618..4704bc0564e2 100644
--- a/drivers/block/drbd/drbd_receiver.c
+++ b/drivers/block/drbd/drbd_receiver.c
@@ -1951,13 +1951,13 @@ read_in_block(struct drbd_peer_device *peer_device, u64 id, sector_t sector,
 	page = peer_req->pages;
 	page_chain_for_each(page) {
 		unsigned len = min_t(int, ds, PAGE_SIZE);
-		data = kmap(page);
+		data = kmap_thread(page);
 		err = drbd_recv_all_warn(peer_device->connection, data, len);
 		if (drbd_insert_fault(device, DRBD_FAULT_RECEIVE)) {
 			drbd_err(device, "Fault injection: Corrupting data on receive\n");
 			data[0] = data[0] ^ (unsigned long)-1;
 		}
-		kunmap(page);
+		kunmap_thread(page);
 		if (err) {
 			drbd_free_peer_req(device, peer_req);
 			return NULL;
@@ -1992,7 +1992,7 @@ static int drbd_drain_block(struct drbd_peer_device *peer_device, int data_size)
 
 	page = drbd_alloc_pages(peer_device, 1, 1);
 
-	data = kmap(page);
+	data = kmap_thread(page);
 	while (data_size) {
 		unsigned int len = min_t(int, data_size, PAGE_SIZE);
 
@@ -2001,7 +2001,7 @@ static int drbd_drain_block(struct drbd_peer_device *peer_device, int data_size)
 			break;
 		data_size -= len;
 	}
-	kunmap(page);
+	kunmap_thread(page);
 	drbd_free_pages(peer_device->device, page, 0);
 	return err;
 }
@@ -2033,10 +2033,10 @@ static int recv_dless_read(struct drbd_peer_device *peer_device, struct drbd_req
 	D_ASSERT(peer_device->device, sector == bio->bi_iter.bi_sector);
 
 	bio_for_each_segment(bvec, bio, iter) {
-		void *mapped = kmap(bvec.bv_page) + bvec.bv_offset;
+		void *mapped = kmap_thread(bvec.bv_page) + bvec.bv_offset;
 		expect = min_t(int, data_size, bvec.bv_len);
 		err = drbd_recv_all_warn(peer_device->connection, mapped, expect);
-		kunmap(bvec.bv_page);
+		kunmap_thread(bvec.bv_page);
 		if (err)
 			return err;
 		data_size -= expect;
-- 
2.28.0.rc0.12.gb6a658bd00c9

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 24/58] fs/freevxfs: Utilize new kmap_thread()

[Christoph Hellwig]

> -	kaddr = kmap(pp);
> +	kaddr = kmap_thread(pp);
>  	memcpy(kaddr, vip->vii_immed.vi_immed + offset, PAGE_SIZE);
> -	kunmap(pp);
> +	kunmap_thread(pp);

You only Cced me on this particular patch, which means I have absolutely
no idea what kmap_thread and kunmap_thread actually do, and thus can't
provide an informed review.

That being said I think your life would be a lot easier if you add
helpers for the above code sequence and its counterpart that copies
to a potential hughmem page first, as that hides the implementation
details from most users.

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 33/58] fs/cramfs: Utilize new kmap_thread()

[Nicolas Pitre]

On Fri, 9 Oct 2020, ira.weiny@intel.com wrote:

> From: Ira Weiny <ira.weiny@intel.com>
> 
> The kmap() calls in this FS are localized to a single thread.  To avoid
> the over head of global PKRS updates use the new kmap_thread() call.
> 
> Cc: Nicolas Pitre <nico@fluxnic.net>
> Signed-off-by: Ira Weiny <ira.weiny@intel.com>

Acked-by: Nicolas Pitre <nico@fluxnic.net>

>  fs/cramfs/inode.c | 10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
> 
> diff --git a/fs/cramfs/inode.c b/fs/cramfs/inode.c
> index 912308600d39..003c014a42ed 100644
> --- a/fs/cramfs/inode.c
> +++ b/fs/cramfs/inode.c
> @@ -247,8 +247,8 @@ static void *cramfs_blkdev_read(struct super_block *sb, unsigned int offset,
>  		struct page *page = pages[i];
>  
>  		if (page) {
> -			memcpy(data, kmap(page), PAGE_SIZE);
> -			kunmap(page);
> +			memcpy(data, kmap_thread(page), PAGE_SIZE);
> +			kunmap_thread(page);
>  			put_page(page);
>  		} else
>  			memset(data, 0, PAGE_SIZE);
> @@ -826,7 +826,7 @@ static int cramfs_readpage(struct file *file, struct page *page)
>  
>  	maxblock = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
>  	bytes_filled = 0;
> -	pgdata = kmap(page);
> +	pgdata = kmap_thread(page);
>  
>  	if (page->index < maxblock) {
>  		struct super_block *sb = inode->i_sb;
> @@ -914,13 +914,13 @@ static int cramfs_readpage(struct file *file, struct page *page)
>  
>  	memset(pgdata + bytes_filled, 0, PAGE_SIZE - bytes_filled);
>  	flush_dcache_page(page);
> -	kunmap(page);
> +	kunmap_thread(page);
>  	SetPageUptodate(page);
>  	unlock_page(page);
>  	return 0;
>  
>  err:
> -	kunmap(page);
> +	kunmap_thread(page);
>  	ClearPageUptodate(page);
>  	SetPageError(page);
>  	unlock_page(page);
> -- 
> 2.28.0.rc0.12.gb6a658bd00c9
> 
> 

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 33/58] fs/cramfs: Utilize new kmap_thread()

[Dan Williams]

On Fri, Oct 9, 2020 at 12:52 PM <ira.weiny@intel.com> wrote:
>
> From: Ira Weiny <ira.weiny@intel.com>
>
> The kmap() calls in this FS are localized to a single thread.  To avoid
> the over head of global PKRS updates use the new kmap_thread() call.
>
> Cc: Nicolas Pitre <nico@fluxnic.net>
> Signed-off-by: Ira Weiny <ira.weiny@intel.com>
> ---
>  fs/cramfs/inode.c | 10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/fs/cramfs/inode.c b/fs/cramfs/inode.c
> index 912308600d39..003c014a42ed 100644
> --- a/fs/cramfs/inode.c
> +++ b/fs/cramfs/inode.c
> @@ -247,8 +247,8 @@ static void *cramfs_blkdev_read(struct super_block *sb, unsigned int offset,
>                 struct page *page = pages[i];
>
>                 if (page) {
> -                       memcpy(data, kmap(page), PAGE_SIZE);
> -                       kunmap(page);
> +                       memcpy(data, kmap_thread(page), PAGE_SIZE);
> +                       kunmap_thread(page);

Why does this need a sleepable kmap? This looks like a textbook
kmap_atomic() use case.

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 33/58] fs/cramfs: Utilize new kmap_thread()

[Matthew Wilcox]

On Tue, Oct 13, 2020 at 11:44:29AM -0700, Dan Williams wrote:
> On Fri, Oct 9, 2020 at 12:52 PM <ira.weiny@intel.com> wrote:
> >
> > From: Ira Weiny <ira.weiny@intel.com>
> >
> > The kmap() calls in this FS are localized to a single thread.  To avoid
> > the over head of global PKRS updates use the new kmap_thread() call.
> >
> > Cc: Nicolas Pitre <nico@fluxnic.net>
> > Signed-off-by: Ira Weiny <ira.weiny@intel.com>
> > ---
> >  fs/cramfs/inode.c | 10 +++++-----
> >  1 file changed, 5 insertions(+), 5 deletions(-)
> >
> > diff --git a/fs/cramfs/inode.c b/fs/cramfs/inode.c
> > index 912308600d39..003c014a42ed 100644
> > --- a/fs/cramfs/inode.c
> > +++ b/fs/cramfs/inode.c
> > @@ -247,8 +247,8 @@ static void *cramfs_blkdev_read(struct super_block *sb, unsigned int offset,
> >                 struct page *page = pages[i];
> >
> >                 if (page) {
> > -                       memcpy(data, kmap(page), PAGE_SIZE);
> > -                       kunmap(page);
> > +                       memcpy(data, kmap_thread(page), PAGE_SIZE);
> > +                       kunmap_thread(page);
> 
> Why does this need a sleepable kmap? This looks like a textbook
> kmap_atomic() use case.

There's a lot of code of this form.  Could we perhaps have:

static inline void copy_to_highpage(struct page *to, void *vfrom, unsigned int size)
{
	char *vto = kmap_atomic(to);

	memcpy(vto, vfrom, size);
	kunmap_atomic(vto);
}

in linux/highmem.h ?

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 33/58] fs/cramfs: Utilize new kmap_thread()

[Dan Williams]

On Tue, Oct 13, 2020 at 12:37 PM Matthew Wilcox <willy@infradead.org> wrote:
>
> On Tue, Oct 13, 2020 at 11:44:29AM -0700, Dan Williams wrote:
> > On Fri, Oct 9, 2020 at 12:52 PM <ira.weiny@intel.com> wrote:
> > >
> > > From: Ira Weiny <ira.weiny@intel.com>
> > >
> > > The kmap() calls in this FS are localized to a single thread.  To avoid
> > > the over head of global PKRS updates use the new kmap_thread() call.
> > >
> > > Cc: Nicolas Pitre <nico@fluxnic.net>
> > > Signed-off-by: Ira Weiny <ira.weiny@intel.com>
> > > ---
> > >  fs/cramfs/inode.c | 10 +++++-----
> > >  1 file changed, 5 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/fs/cramfs/inode.c b/fs/cramfs/inode.c
> > > index 912308600d39..003c014a42ed 100644
> > > --- a/fs/cramfs/inode.c
> > > +++ b/fs/cramfs/inode.c
> > > @@ -247,8 +247,8 @@ static void *cramfs_blkdev_read(struct super_block *sb, unsigned int offset,
> > >                 struct page *page = pages[i];
> > >
> > >                 if (page) {
> > > -                       memcpy(data, kmap(page), PAGE_SIZE);
> > > -                       kunmap(page);
> > > +                       memcpy(data, kmap_thread(page), PAGE_SIZE);
> > > +                       kunmap_thread(page);
> >
> > Why does this need a sleepable kmap? This looks like a textbook
> > kmap_atomic() use case.
>
> There's a lot of code of this form.  Could we perhaps have:
>
> static inline void copy_to_highpage(struct page *to, void *vfrom, unsigned int size)
> {
>         char *vto = kmap_atomic(to);
>
>         memcpy(vto, vfrom, size);
>         kunmap_atomic(vto);
> }
>
> in linux/highmem.h ?

Nice, yes, that could also replace the local ones in lib/iov_iter.c
(memcpy_{to,from}_page())

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 33/58] fs/cramfs: Utilize new kmap_thread()

[Al Viro]

On Tue, Oct 13, 2020 at 08:36:43PM +0100, Matthew Wilcox wrote:

> static inline void copy_to_highpage(struct page *to, void *vfrom, unsigned int size)
> {
> 	char *vto = kmap_atomic(to);
> 
> 	memcpy(vto, vfrom, size);
> 	kunmap_atomic(vto);
> }
> 
> in linux/highmem.h ?

You mean, like
static void memcpy_from_page(char *to, struct page *page, size_t offset, size_t len)
{
        char *from = kmap_atomic(page);
        memcpy(to, from + offset, len);
        kunmap_atomic(from);
}

static void memcpy_to_page(struct page *page, size_t offset, const char *from, size_t len)
{
        char *to = kmap_atomic(page);
        memcpy(to + offset, from, len);
        kunmap_atomic(to);
}

static void memzero_page(struct page *page, size_t offset, size_t len)
{
        char *addr = kmap_atomic(page);
        memset(addr + offset, 0, len);
        kunmap_atomic(addr);
}

in lib/iov_iter.c?  FWIW, I don't like that "highpage" in the name and
highmem.h as location - these make perfect sense regardless of highmem;
they are normal memory operations with page + offset used instead of
a pointer...

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 33/58] fs/cramfs: Utilize new kmap_thread()

[Ira Weiny]

On Tue, Oct 13, 2020 at 08:36:43PM +0100, Matthew Wilcox wrote:
> On Tue, Oct 13, 2020 at 11:44:29AM -0700, Dan Williams wrote:
> > On Fri, Oct 9, 2020 at 12:52 PM <ira.weiny@intel.com> wrote:
> > >
> > > From: Ira Weiny <ira.weiny@intel.com>
> > >
> > > The kmap() calls in this FS are localized to a single thread.  To avoid
> > > the over head of global PKRS updates use the new kmap_thread() call.
> > >
> > > Cc: Nicolas Pitre <nico@fluxnic.net>
> > > Signed-off-by: Ira Weiny <ira.weiny@intel.com>
> > > ---
> > >  fs/cramfs/inode.c | 10 +++++-----
> > >  1 file changed, 5 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/fs/cramfs/inode.c b/fs/cramfs/inode.c
> > > index 912308600d39..003c014a42ed 100644
> > > --- a/fs/cramfs/inode.c
> > > +++ b/fs/cramfs/inode.c
> > > @@ -247,8 +247,8 @@ static void *cramfs_blkdev_read(struct super_block *sb, unsigned int offset,
> > >                 struct page *page = pages[i];
> > >
> > >                 if (page) {
> > > -                       memcpy(data, kmap(page), PAGE_SIZE);
> > > -                       kunmap(page);
> > > +                       memcpy(data, kmap_thread(page), PAGE_SIZE);
> > > +                       kunmap_thread(page);
> > 
> > Why does this need a sleepable kmap? This looks like a textbook
> > kmap_atomic() use case.
> 
> There's a lot of code of this form.  Could we perhaps have:
> 
> static inline void copy_to_highpage(struct page *to, void *vfrom, unsigned int size)
> {
> 	char *vto = kmap_atomic(to);
> 
> 	memcpy(vto, vfrom, size);
> 	kunmap_atomic(vto);
> }
> 
> in linux/highmem.h ?

Christoph had the same idea.  I'll work on it.

Ira

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 33/58] fs/cramfs: Utilize new kmap_thread()

[Ira Weiny]

On Tue, Oct 13, 2020 at 09:01:49PM +0100, Al Viro wrote:
> On Tue, Oct 13, 2020 at 08:36:43PM +0100, Matthew Wilcox wrote:
> 
> > static inline void copy_to_highpage(struct page *to, void *vfrom, unsigned int size)
> > {
> > 	char *vto = kmap_atomic(to);
> > 
> > 	memcpy(vto, vfrom, size);
> > 	kunmap_atomic(vto);
> > }
> > 
> > in linux/highmem.h ?
> 
> You mean, like
> static void memcpy_from_page(char *to, struct page *page, size_t offset, size_t len)
> {
>         char *from = kmap_atomic(page);
>         memcpy(to, from + offset, len);
>         kunmap_atomic(from);
> }
> 
> static void memcpy_to_page(struct page *page, size_t offset, const char *from, size_t len)
> {
>         char *to = kmap_atomic(page);
>         memcpy(to + offset, from, len);
>         kunmap_atomic(to);
> }
> 
> static void memzero_page(struct page *page, size_t offset, size_t len)
> {
>         char *addr = kmap_atomic(page);
>         memset(addr + offset, 0, len);
>         kunmap_atomic(addr);
> }
> 
> in lib/iov_iter.c?  FWIW, I don't like that "highpage" in the name and
> highmem.h as location - these make perfect sense regardless of highmem;
> they are normal memory operations with page + offset used instead of
> a pointer...

I was thinking along those lines as well especially because of the direction
this patch set takes kmap().

Thanks for pointing these out to me.  How about I lift them to a common header?
But if not highmem.h where?

Ira

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 24/58] fs/freevxfs: Utilize new kmap_thread()

[Ira Weiny]

On Tue, Oct 13, 2020 at 12:25:44PM +0100, Christoph Hellwig wrote:
> > -	kaddr = kmap(pp);
> > +	kaddr = kmap_thread(pp);
> >  	memcpy(kaddr, vip->vii_immed.vi_immed + offset, PAGE_SIZE);
> > -	kunmap(pp);
> > +	kunmap_thread(pp);
> 
> You only Cced me on this particular patch, which means I have absolutely
> no idea what kmap_thread and kunmap_thread actually do, and thus can't
> provide an informed review.

Sorry the list was so big I struggled with who to CC and on which patches.

> 
> That being said I think your life would be a lot easier if you add
> helpers for the above code sequence and its counterpart that copies
> to a potential hughmem page first, as that hides the implementation
> details from most users.

Matthew Wilcox and Al Viro have suggested similar ideas.

https://lore.kernel.org/lkml/20201013205012.GI2046448@iweiny-DESK2.sc.intel.com/

Ira

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 05/58] kmap: Introduce k[un]map_thread

[Thomas Gleixner]

Ira,

On Fri, Oct 09 2020 at 12:49, ira weiny wrote:
> From: Ira Weiny <ira.weiny@intel.com>
>
> To correctly support the semantics of kmap() with Kernel protection keys
> (PKS), kmap() may be required to set the protections on multiple
> processors (globally).  Enabling PKS globally can be very expensive
> depending on the requested operation.  Furthermore, enabling a domain
> globally reduces the protection afforded by PKS.
>
> Most kmap() (Aprox 209 of 229) callers use the map within a single thread and
> have no need for the protection domain to be enabled globally.  However, the
> remaining callers do not follow this pattern and, as best I can tell, expect
> the mapping to be 'global' and available to any thread who may access the
> mapping.[1]
>
> We don't anticipate global mappings to pmem, however in general there is a
> danger in changing the semantics of kmap().  Effectively, this would cause an
> unresolved page fault with little to no information about why the failure
> occurred.
>
> To resolve this a number of options were considered.
>
> 1) Attempt to change all the thread local kmap() calls to kmap_atomic()[2]
> 2) Introduce a flags parameter to kmap() to indicate if the mapping should be
>    global or not
> 3) Change ~20 call sites to 'kmap_global()' to indicate that they require a
>    global enablement of the pages.
> 4) Change ~209 call sites to 'kmap_thread()' to indicate that the mapping is to
>    be used within that thread of execution only
>
> Option 1 is simply not feasible.  Option 2 would require all of the call sites
> of kmap() to change.  Option 3 seems like a good minimal change but there is a
> danger that new code may miss the semantic change of kmap() and not get the
> behavior the developer intended.  Therefore, #4 was chosen.

There is Option #5:

Convert the thread local kmap() invocations to the proposed kmap_local()
interface which is coming along [1].

That solves a couple of issues:

 1) It relieves the current kmap_atomic() usage sites from the implict
    pagefault/preempt disable semantics which apply even when
    CONFIG_HIGHMEM is disabled. kmap_local() still can be invoked from
    atomic context.

 2) Due to #1 it allows to replace the conditional usage of kmap() and
    kmap_atomic() for purely thread local mappings.

 3) It puts the burden on the HIGHMEM inflicted systems

 4) It is actually more efficient for most of the pure thread local use
    cases on HIGHMEM inflicted systems because it avoids the overhead of
    the global lock and the potential kmap slot exhaustion. A potential
    preemption will be more expensive, but that's not really the case we
    want to optimize for.

 5) It solves the RT issue vs. kmap_atomic()

So instead of creating yet another variety of kmap() which is just
scratching the particular PKRS itch, can we please consolidate all of
that on the wider reaching kmap_local() approach?

Thanks,

        tglx
     
[1] https://lore.kernel.org/lkml/20201103092712.714480842@linutronix.de/

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 05/58] kmap: Introduce k[un]map_thread

[Ira Weiny]

On Tue, Nov 10, 2020 at 02:13:56AM +0100, Thomas Gleixner wrote:
> Ira,
> 
> On Fri, Oct 09 2020 at 12:49, ira weiny wrote:
> > From: Ira Weiny <ira.weiny@intel.com>
> >
> > To correctly support the semantics of kmap() with Kernel protection keys
> > (PKS), kmap() may be required to set the protections on multiple
> > processors (globally).  Enabling PKS globally can be very expensive
> > depending on the requested operation.  Furthermore, enabling a domain
> > globally reduces the protection afforded by PKS.
> >
> > Most kmap() (Aprox 209 of 229) callers use the map within a single thread and
> > have no need for the protection domain to be enabled globally.  However, the
> > remaining callers do not follow this pattern and, as best I can tell, expect
> > the mapping to be 'global' and available to any thread who may access the
> > mapping.[1]
> >
> > We don't anticipate global mappings to pmem, however in general there is a
> > danger in changing the semantics of kmap().  Effectively, this would cause an
> > unresolved page fault with little to no information about why the failure
> > occurred.
> >
> > To resolve this a number of options were considered.
> >
> > 1) Attempt to change all the thread local kmap() calls to kmap_atomic()[2]
> > 2) Introduce a flags parameter to kmap() to indicate if the mapping should be
> >    global or not
> > 3) Change ~20 call sites to 'kmap_global()' to indicate that they require a
> >    global enablement of the pages.
> > 4) Change ~209 call sites to 'kmap_thread()' to indicate that the mapping is to
> >    be used within that thread of execution only
> >
> > Option 1 is simply not feasible.  Option 2 would require all of the call sites
> > of kmap() to change.  Option 3 seems like a good minimal change but there is a
> > danger that new code may miss the semantic change of kmap() and not get the
> > behavior the developer intended.  Therefore, #4 was chosen.
> 
> There is Option #5:

There is now yes.  :-D

> 
> Convert the thread local kmap() invocations to the proposed kmap_local()
> interface which is coming along [1].

I've been trying to follow that thread.

> 
> That solves a couple of issues:
> 
>  1) It relieves the current kmap_atomic() usage sites from the implict
>     pagefault/preempt disable semantics which apply even when
>     CONFIG_HIGHMEM is disabled. kmap_local() still can be invoked from
>     atomic context.
> 
>  2) Due to #1 it allows to replace the conditional usage of kmap() and
>     kmap_atomic() for purely thread local mappings.
> 
>  3) It puts the burden on the HIGHMEM inflicted systems
> 
>  4) It is actually more efficient for most of the pure thread local use
>     cases on HIGHMEM inflicted systems because it avoids the overhead of
>     the global lock and the potential kmap slot exhaustion. A potential
>     preemption will be more expensive, but that's not really the case we
>     want to optimize for.
> 
>  5) It solves the RT issue vs. kmap_atomic()
> 
> So instead of creating yet another variety of kmap() which is just
> scratching the particular PKRS itch, can we please consolidate all of
> that on the wider reaching kmap_local() approach?

Yes I agree.  We absolutely don't want more kmap*() calls and I was hoping to
dovetail into your kmap_local() work.[2]

I've pivoted away from this work a bit to clean up all the
kmap()/memcpy*()/kunmaps() as discussed elsewhere in the thread first.[3]  I
was hoping your work would land and then I could s/kmap_thread()/kmap_local()/
on all of these patches.

Also, we can convert the new memcpy_*_page() calls to kmap_local() as well.
[For now my patch just uses kmap_atomic().]

I've not looked at all of the patches in your latest version.  Have you
included converting any of the kmap() call sites?  I thought you were more
focused on converting the kmap_atomic() to kmap_local()?

Ira

> 
> Thanks,
> 
>         tglx
>      
> [1] https://lore.kernel.org/lkml/20201103092712.714480842@linutronix.de/

[2] https://lore.kernel.org/lkml/20201012195354.GC2046448@iweiny-DESK2.sc.intel.com/
[3] https://lore.kernel.org/lkml/20201009213434.GA839@sol.localdomain/
    https://lore.kernel.org/lkml/20201013200149.GI3576660@ZenIV.linux.org.uk/

Re: [Drbd-dev] [PATCH RFC PKS/PMEM 05/58] kmap: Introduce k[un]map_thread

[Thomas Gleixner]

On Mon, Nov 09 2020 at 20:59, Ira Weiny wrote:
> On Tue, Nov 10, 2020 at 02:13:56AM +0100, Thomas Gleixner wrote:
> Also, we can convert the new memcpy_*_page() calls to kmap_local() as well.
> [For now my patch just uses kmap_atomic().]
>
> I've not looked at all of the patches in your latest version.  Have you
> included converting any of the kmap() call sites?  I thought you were more
> focused on converting the kmap_atomic() to kmap_local()?

I did not touch any of those yet, but it's a logical consequence to
convert all kmap() instances which are _not_ creating a global mapping
over to it.

Thanks,

        tglx