Skip to contentSkip to conversation

xen/x86: Always strip xen.efi

PATCH In review Andrew Cooper 6 participants 19 messages Jun 8 – Jun 17

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xen/x86: Always strip xen.efi

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Teddy Astie <[email protected]>
CC: Frediano Ziglio <[email protected]>
CC: Oleksii Kurochko <[email protected]>
CC: Marek Marczykowski-Górecki <[email protected]>
CC: Daniel P. Smith <[email protected]>

For 4.22. This was posted previously as

https://lore.kernel.org/xen-devel/[email protected]/T/#u

but merged the two patches and rewritten the commit message to make it clear that failing to strip xen.efi is causing boot failures.

Previously xen.efi.elf was produced but it's unclear why, and unnecessaerily different, so I've dropped it.

While this does want backporting, it can't be. Xen 4.21 and older still build test with Ubuntu 16.04 and choke

---
 .gitignore            |  1 +
 CHANGELOG.md          |  3 +++
 docs/misc/efi.pandoc  |  8 +-------
 xen/Kconfig.debug     |  9 ++-------
 xen/Makefile          | 19 -------------------
 xen/arch/x86/Makefile | 11 ++++-------
 xen/arch/x86/arch.mk  |  7 -------
 7 files changed, 11 insertions(+), 47 deletions(-)
.gitignore +1 -0
index bfc7bdf043c3..49e2c6961768 100644
@@ -224,6 +224,7 @@ tools/flask/policy/xenpolicy-*
 xen/xen
 xen/suppression-list.txt
 xen/xen-syms
+xen/xen-syms.efi
 xen/xen-syms.map
 xen/xen.*
 
CHANGELOG.md +3 -0
index 5cf19372a361..71d1e9ab8c69 100644
@@ -14,6 +14,9 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
  - On x86:
    - Enable pf-fixup option by default for PVH dom0.
    - The libxenguest bzImage loader now uses the system liblz4 library.
+   - The install-time environment variable INSTALL_EFI_STRIP no longer exists.
+     xen.efi is always stripped, while the symbols remain available in
+     xen-syms.efi.
 
 ### Added
  - Support for per-domain Xenstore quota in C xenstored (includes
docs/misc/efi.pandoc +1 -7
index 8198a7f063cf..0a3fd67076fc 100644
@@ -20,13 +20,7 @@ Xen to load the configuration file even if multiboot modules are found.
 Once built, `make install-xen` will place the resulting binary directly into
 the EFI boot partition, provided `EFI_VENDOR` is set in the environment (and
 `EFI_MOUNTPOINT` is overridden as needed, should the default of `/boot/efi` not
-match your system). When built with debug info, the binary can be quite large.
-Setting `INSTALL_EFI_STRIP=1` in the environment will cause it to be stripped
-of debug info in the process of installing. `INSTALL_EFI_STRIP` can also be set
-to any combination of options suitable to pass to `strip`, in case the default
-ones don't do. The xen.efi binary will also be installed in `/usr/lib64/efi/`,
-unless `EFI_DIR` is set in the environment to override this default. This
-binary will not be stripped in the process.
+match your system).
 
 The binary itself will require a configuration file (names with the `.efi`
 extension of the binary's name replaced by `.cfg`, and - until an existing
xen/Kconfig.debug +2 -7
index d900d926c555..fcd3fc3d36cf 100644
@@ -147,12 +147,7 @@ config DEBUG_INFO
 	  Say Y here if you want to build Xen with debug information. This
 	  information is needed e.g. for doing crash dump analysis of the
 	  hypervisor via the "crash" tool.
-	  Saying Y will increase the size of the xen-syms and xen.efi
-	  binaries. In case the space on the EFI boot partition is rather
-	  limited, you may want to install a stripped variant of xen.efi in
-	  the EFI boot partition (look for "INSTALL_EFI_STRIP" in
-	  docs/misc/efi.pandoc for more information - when not using
-	  "make install-xen" for installing xen.efi, stripping needs to be
-	  done outside the Xen build environment).
+	  Saying Y will increase the size of the xen-syms and xen-syms.efi
+	  binaries.
 
 endmenu
xen/Makefile +0 -19
index 1f11610b5f68..0f9b56fc399d 100644
@@ -493,22 +493,6 @@ endif
 .PHONY: _build
 _build: $(TARGET)$(CONFIG_XEN_INSTALL_SUFFIX)
 
-# Strip
-#
-# INSTALL_EFI_STRIP, if defined, will cause xen.efi to be stripped before it
-# is installed. If INSTALL_EFI_STRIP is '1', then the default option(s) below
-# will be used. Otherwise, INSTALL_EFI_STRIP value will be used as the
-# option(s) to the strip command.
-ifdef INSTALL_EFI_STRIP
-
-ifeq ($(INSTALL_EFI_STRIP),1)
-efi-strip-opt := --strip-debug --keep-file-symbols
-else
-efi-strip-opt := $(INSTALL_EFI_STRIP)
-endif
-
-endif
-
 .PHONY: _install
 _install: D=$(DESTDIR)
 _install: T=$(notdir $(TARGET))
@@ -535,9 +519,6 @@ _install: $(TARGET)$(CONFIG_XEN_INSTALL_SUFFIX)
 		ln -sf $(T)-$(XEN_FULLVERSION).efi $(D)$(EFI_DIR)/$(T)-$(XEN_VERSION).efi; \
 		ln -sf $(T)-$(XEN_FULLVERSION).efi $(D)$(EFI_DIR)/$(T).efi; \
 		if [ -n '$(EFI_MOUNTPOINT)' -a -n '$(EFI_VENDOR)' ]; then \
-			$(if $(efi-strip-opt), \
-			     $(STRIP) $(efi-strip-opt) -p -o $(TARGET).efi.stripped $(TARGET).efi && \
-			     $(INSTALL_DATA) $(TARGET).efi.stripped $(D)$(EFI_MOUNTPOINT)/efi/$(EFI_VENDOR)/$(T)-$(XEN_FULLVERSION).efi ||) \
 			$(INSTALL_DATA) $(TARGET).efi $(D)$(EFI_MOUNTPOINT)/efi/$(EFI_VENDOR)/$(T)-$(XEN_FULLVERSION).efi; \
 		elif [ "$(D)" = "$(patsubst $(shell cd $(XEN_ROOT) && pwd)/%,%,$(D))" ]; then \
 			echo 'EFI installation only partially done (EFI_VENDOR not set)' >&2; \
xen/arch/x86/Makefile +4 -7
index 47dd6c50fe88..01ed7302202e 100644
@@ -196,10 +196,7 @@ note_file_option ?= $(note_file)
 
 extra-$(XEN_BUILD_PE) += efi.lds
 ifeq ($(XEN_BUILD_PE),y)
-$(TARGET).efi: $(objtree)/prelink.o $(note_file) $(obj)/efi.lds $(obj)/efi/relocs-dummy.o $(obj)/efi/mkreloc
-ifeq ($(CONFIG_DEBUG_INFO),y)
-	$(if $(filter --strip-debug,$(EFI_LDFLAGS)),echo,:) "Will strip debug info from $(@F)"
-endif
+$(TARGET)-syms.efi: $(objtree)/prelink.o $(note_file) $(obj)/efi.lds $(obj)/efi/relocs-dummy.o $(obj)/efi/mkreloc
 	$(objtree)/tools/symbols $(all_symbols) --source-name=$(@F).S --empty \
 		> $(dot-target).0s.S
 	$(MAKE) $(build)=$(@D) .$(@F).0s.o
@@ -233,10 +230,10 @@ endif
 	$(NM) -pa --format=sysv $@ \
 		| $(objtree)/tools/symbols --all-symbols --xensyms --sysv --sort \
 		> [email protected]
-ifeq ($(CONFIG_DEBUG_INFO),y)
-	$(if $(filter --strip-debug,$(EFI_LDFLAGS)),:$(space))$(OBJCOPY) -O elf64-x86-64 $@ [email protected]
-endif
 	rm -f $(dot-target).[0-9]* $(@D)/..$(@F).[0-9]*
+
+$(TARGET).efi: $(TARGET)-syms.efi
+	$(STRIP) $< -o $@
 ifeq ($(CONFIG_XEN_IBT),y)
 	$(SHELL) $(srctree)/tools/check-endbr.sh $@
 endif
xen/arch/x86/arch.mk +0 -7
index cd0602a79aaf..199adc1a0967 100644
@@ -95,13 +95,6 @@ EFI_LDFLAGS := $(patsubst -m%,-mi386pep,$(LDFLAGS)) --subsystem=10 --enable-long
 LD_PE_check_cmd = $(call ld-option,$(EFI_LDFLAGS) --image-base=0x100000000 -o $(efi-check).efi $(efi-check).o)
 XEN_BUILD_PE := $(LD_PE_check_cmd)
 
-# If the above failed, it may be merely because of the linker not dealing well
-# with debug info. Try again with stripping it.
-ifeq ($(CONFIG_DEBUG_INFO)-$(XEN_BUILD_PE),y-n)
-EFI_LDFLAGS += --strip-debug
-XEN_BUILD_PE := $(LD_PE_check_cmd)
-endif
-
 ifeq ($(XEN_BUILD_PE),y)
 
 # Check if the linker produces fixups in PE by default

base-commit: 37df17d2f903503c619713eb01e488f2cb1a257f

Signature · 1 line
2.39.5
#1 lore raw
Quoted 14 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Quoted 4 lines

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

While I see Ubuntu 16.04 dropped, how is the "require a less broken toolchain" addressed? By implicitly disabling xen.efi build on broken toolchain? Maybe README should have a note about needing newer Binutils for xen.efi? Currently it says just Binutils 2.25. There is a section about optional build deps, maybe add there something like "GNU Binutils X.Y (required for building xen.efi)", if the version is known, or at least "GNU Binutils capable of producing non-broken PE files (required for building xen.efi)" if the version is not known.

Quoted 187 lines

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Teddy Astie <[email protected]>
CC: Frediano Ziglio <[email protected]>
CC: Oleksii Kurochko <[email protected]>
CC: Marek Marczykowski-Górecki <[email protected]>
CC: Daniel P. Smith <[email protected]>

For 4.22. This was posted previously as

https://lore.kernel.org/xen-devel/[email protected]/T/#u

but merged the two patches and rewritten the commit message to make it clear that failing to strip xen.efi is causing boot failures.

Previously xen.efi.elf was produced but it's unclear why, and unnecessaerily different, so I've dropped it.

While this does want backporting, it can't be. Xen 4.21 and older still build test with Ubuntu 16.04 and choke

---
 .gitignore            |  1 +
 CHANGELOG.md          |  3 +++
 docs/misc/efi.pandoc  |  8 +-------
 xen/Kconfig.debug     |  9 ++-------
 xen/Makefile          | 19 -------------------
 xen/arch/x86/Makefile | 11 ++++-------
 xen/arch/x86/arch.mk  |  7 -------
 7 files changed, 11 insertions(+), 47 deletions(-)
.gitignore +1 -0
index bfc7bdf043c3..49e2c6961768 100644
@@ -224,6 +224,7 @@ tools/flask/policy/xenpolicy-*
 xen/xen
 xen/suppression-list.txt
 xen/xen-syms
+xen/xen-syms.efi
 xen/xen-syms.map
 xen/xen.*
 
CHANGELOG.md +3 -0
index 5cf19372a361..71d1e9ab8c69 100644
@@ -14,6 +14,9 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
  - On x86:
    - Enable pf-fixup option by default for PVH dom0.
    - The libxenguest bzImage loader now uses the system liblz4 library.
+   - The install-time environment variable INSTALL_EFI_STRIP no longer exists.
+     xen.efi is always stripped, while the symbols remain available in
+     xen-syms.efi.
 
 ### Added
  - Support for per-domain Xenstore quota in C xenstored (includes
docs/misc/efi.pandoc +1 -7
index 8198a7f063cf..0a3fd67076fc 100644
@@ -20,13 +20,7 @@ Xen to load the configuration file even if multiboot modules are found.
 Once built, `make install-xen` will place the resulting binary directly into
 the EFI boot partition, provided `EFI_VENDOR` is set in the environment (and
 `EFI_MOUNTPOINT` is overridden as needed, should the default of `/boot/efi` not
-match your system). When built with debug info, the binary can be quite large.
-Setting `INSTALL_EFI_STRIP=1` in the environment will cause it to be stripped
-of debug info in the process of installing. `INSTALL_EFI_STRIP` can also be set
-to any combination of options suitable to pass to `strip`, in case the default
-ones don't do. The xen.efi binary will also be installed in `/usr/lib64/efi/`,
-unless `EFI_DIR` is set in the environment to override this default. This
-binary will not be stripped in the process.
+match your system).
 
 The binary itself will require a configuration file (names with the `.efi`
 extension of the binary's name replaced by `.cfg`, and - until an existing
xen/Kconfig.debug +2 -7
index d900d926c555..fcd3fc3d36cf 100644
@@ -147,12 +147,7 @@ config DEBUG_INFO
 	  Say Y here if you want to build Xen with debug information. This
 	  information is needed e.g. for doing crash dump analysis of the
 	  hypervisor via the "crash" tool.
-	  Saying Y will increase the size of the xen-syms and xen.efi
-	  binaries. In case the space on the EFI boot partition is rather
-	  limited, you may want to install a stripped variant of xen.efi in
-	  the EFI boot partition (look for "INSTALL_EFI_STRIP" in
-	  docs/misc/efi.pandoc for more information - when not using
-	  "make install-xen" for installing xen.efi, stripping needs to be
-	  done outside the Xen build environment).
+	  Saying Y will increase the size of the xen-syms and xen-syms.efi
+	  binaries.
 
 endmenu
xen/Makefile +0 -19
index 1f11610b5f68..0f9b56fc399d 100644
@@ -493,22 +493,6 @@ endif
 .PHONY: _build
 _build: $(TARGET)$(CONFIG_XEN_INSTALL_SUFFIX)
 
-# Strip
-#
-# INSTALL_EFI_STRIP, if defined, will cause xen.efi to be stripped before it
-# is installed. If INSTALL_EFI_STRIP is '1', then the default option(s) below
-# will be used. Otherwise, INSTALL_EFI_STRIP value will be used as the
-# option(s) to the strip command.
-ifdef INSTALL_EFI_STRIP
-
-ifeq ($(INSTALL_EFI_STRIP),1)
-efi-strip-opt := --strip-debug --keep-file-symbols
-else
-efi-strip-opt := $(INSTALL_EFI_STRIP)
-endif
-
-endif
-
 .PHONY: _install
 _install: D=$(DESTDIR)
 _install: T=$(notdir $(TARGET))
@@ -535,9 +519,6 @@ _install: $(TARGET)$(CONFIG_XEN_INSTALL_SUFFIX)
 		ln -sf $(T)-$(XEN_FULLVERSION).efi $(D)$(EFI_DIR)/$(T)-$(XEN_VERSION).efi; \
 		ln -sf $(T)-$(XEN_FULLVERSION).efi $(D)$(EFI_DIR)/$(T).efi; \
 		if [ -n '$(EFI_MOUNTPOINT)' -a -n '$(EFI_VENDOR)' ]; then \
-			$(if $(efi-strip-opt), \
-			     $(STRIP) $(efi-strip-opt) -p -o $(TARGET).efi.stripped $(TARGET).efi && \
-			     $(INSTALL_DATA) $(TARGET).efi.stripped $(D)$(EFI_MOUNTPOINT)/efi/$(EFI_VENDOR)/$(T)-$(XEN_FULLVERSION).efi ||) \
 			$(INSTALL_DATA) $(TARGET).efi $(D)$(EFI_MOUNTPOINT)/efi/$(EFI_VENDOR)/$(T)-$(XEN_FULLVERSION).efi; \
 		elif [ "$(D)" = "$(patsubst $(shell cd $(XEN_ROOT) && pwd)/%,%,$(D))" ]; then \
 			echo 'EFI installation only partially done (EFI_VENDOR not set)' >&2; \
xen/arch/x86/Makefile +4 -7
index 47dd6c50fe88..01ed7302202e 100644
@@ -196,10 +196,7 @@ note_file_option ?= $(note_file)
 
 extra-$(XEN_BUILD_PE) += efi.lds
 ifeq ($(XEN_BUILD_PE),y)
-$(TARGET).efi: $(objtree)/prelink.o $(note_file) $(obj)/efi.lds $(obj)/efi/relocs-dummy.o $(obj)/efi/mkreloc
-ifeq ($(CONFIG_DEBUG_INFO),y)
-	$(if $(filter --strip-debug,$(EFI_LDFLAGS)),echo,:) "Will strip debug info from $(@F)"
-endif
+$(TARGET)-syms.efi: $(objtree)/prelink.o $(note_file) $(obj)/efi.lds $(obj)/efi/relocs-dummy.o $(obj)/efi/mkreloc
 	$(objtree)/tools/symbols $(all_symbols) --source-name=$(@F).S --empty \
 		> $(dot-target).0s.S
 	$(MAKE) $(build)=$(@D) .$(@F).0s.o
@@ -233,10 +230,10 @@ endif
 	$(NM) -pa --format=sysv $@ \
 		| $(objtree)/tools/symbols --all-symbols --xensyms --sysv --sort \
 		> [email protected]
-ifeq ($(CONFIG_DEBUG_INFO),y)
-	$(if $(filter --strip-debug,$(EFI_LDFLAGS)),:$(space))$(OBJCOPY) -O elf64-x86-64 $@ [email protected]
-endif
 	rm -f $(dot-target).[0-9]* $(@D)/..$(@F).[0-9]*
+
+$(TARGET).efi: $(TARGET)-syms.efi
+	$(STRIP) $< -o $@
 ifeq ($(CONFIG_XEN_IBT),y)
 	$(SHELL) $(srctree)/tools/check-endbr.sh $@
 endif
xen/arch/x86/arch.mk +0 -7
index cd0602a79aaf..199adc1a0967 100644
@@ -95,13 +95,6 @@ EFI_LDFLAGS := $(patsubst -m%,-mi386pep,$(LDFLAGS)) --subsystem=10 --enable-long
 LD_PE_check_cmd = $(call ld-option,$(EFI_LDFLAGS) --image-base=0x100000000 -o $(efi-check).efi $(efi-check).o)
 XEN_BUILD_PE := $(LD_PE_check_cmd)
 
-# If the above failed, it may be merely because of the linker not dealing well
-# with debug info. Try again with stripping it.
-ifeq ($(CONFIG_DEBUG_INFO)-$(XEN_BUILD_PE),y-n)
-EFI_LDFLAGS += --strip-debug
-XEN_BUILD_PE := $(LD_PE_check_cmd)
-endif
-
 ifeq ($(XEN_BUILD_PE),y)
 
 # Check if the linker produces fixups in PE by default

base-commit: 37df17d2f903503c619713eb01e488f2cb1a257f

Signature · 1 line
2.39.5
Signature · 3 lines
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab
#2 lore raw
Quoted 19 lines · Marek Marczykowski-Górecki
Quoted 14 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Quoted 4 lines

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

Quoted 8 lines

While I see Ubuntu 16.04 dropped, how is the "require a less broken toolchain" addressed? By implicitly disabling xen.efi build on broken toolchain? Maybe README should have a note about needing newer Binutils for xen.efi? Currently it says just Binutils 2.25. There is a section about optional build deps, maybe add there something like "GNU Binutils X.Y (required for building xen.efi)", if the version is known, or at least "GNU Binutils capable of producing non-broken PE files (required for building xen.efi)" if the version is not known.

xen.efi has never had any relation to the README minimum toolchain version.

It has always probed the toolchain, and silently turned itself off it doesn't like the result.  In this case, we drop one of the "lets work around this bug different" checks which ends up excluding the problem revision.

If you prefer, I could re-split the patch, and state on the first patch that it's a prerequisite to be able to use $(STRIP) in the second patch ?

binutils' PE+ support is horribly buggy and Xen is the only user in this area.  At some point, 2.46 (practically bleeding edge) is going to be required, seeing as it's the first version of bintuils where we don't need to hexedit the PE+ header in order to satisfy the signing process.

~Andrew

#3 lore raw
Quoted 29 lines · Andrew Cooper
Quoted 19 lines · Marek Marczykowski-Górecki
Quoted 14 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Quoted 4 lines

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

Quoted 8 lines

While I see Ubuntu 16.04 dropped, how is the "require a less broken toolchain" addressed? By implicitly disabling xen.efi build on broken toolchain? Maybe README should have a note about needing newer Binutils for xen.efi? Currently it says just Binutils 2.25. There is a section about optional build deps, maybe add there something like "GNU Binutils X.Y (required for building xen.efi)", if the version is known, or at least "GNU Binutils capable of producing non-broken PE files (required for building xen.efi)" if the version is not known.

Quoted 6 lines

xen.efi has never had any relation to the README minimum toolchain version.

It has always probed the toolchain, and silently turned itself off it doesn't like the result.  In this case, we drop one of the "lets work around this bug different" checks which ends up excluding the problem revision.

Ok, in that case

Quoted 9 lines

If you prefer, I could re-split the patch, and state on the first patch that it's a prerequisite to be able to use $(STRIP) in the second patch ?

binutils' PE+ support is horribly buggy and Xen is the only user in this area.  At some point, 2.46 (practically bleeding edge) is going to be required, seeing as it's the first version of bintuils where we don't need to hexedit the PE+ header in order to satisfy the signing process.

~Andrew

Signature · 3 lines
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab
Acked-by on v1 1/1
#4 lore raw
In reply to Andrew Cooper
Quoted 37 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Teddy Astie <[email protected]>
CC: Frediano Ziglio <[email protected]>
CC: Oleksii Kurochko <[email protected]>
CC: Marek Marczykowski-Górecki <[email protected]>
CC: Daniel P. Smith <[email protected]>

For 4.22. This was posted previously as

   https://lore.kernel.org/xen-devel/[email protected]/T/#u

but merged the two patches and rewritten the commit message to make it clear that failing to strip xen.efi is causing boot failures.

Quoted 6 lines

Previously xen.efi.elf was produced but it's unclear why, and unnecessaerily different, so I've dropped it.

While this does want backporting, it can't be. Xen 4.21 and older still build test with Ubuntu 16.04 and choke
---

LGTM to be in 4.22:

and ...

Quoted 31 lines
  .gitignore            |  1 +
  CHANGELOG.md          |  3 +++>   docs/misc/efi.pandoc  |  8 +-------
  xen/Kconfig.debug     |  9 ++-------
  xen/Makefile          | 19 -------------------
  xen/arch/x86/Makefile | 11 ++++-------
  xen/arch/x86/arch.mk  |  7 -------
  7 files changed, 11 insertions(+), 47 deletions(-)
.gitignore +1 -0
index bfc7bdf043c3..49e2c6961768 100644
@@ -224,6 +224,7 @@ tools/flask/policy/xenpolicy-*
  xen/xen
  xen/suppression-list.txt
  xen/xen-syms
+xen/xen-syms.efi
  xen/xen-syms.map
  xen/xen.*
 
CHANGELOG.md +3 -0
index 5cf19372a361..71d1e9ab8c69 100644
@@ -14,6 +14,9 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
   - On x86:
     - Enable pf-fixup option by default for PVH dom0.
     - The libxenguest bzImage loader now uses the system liblz4 library.
+   - The install-time environment variable INSTALL_EFI_STRIP no longer exists.
+     xen.efi is always stripped, while the symbols remain available in
+     xen-syms.efi.

...

Thanks.

~ Oleksii

#5 lore raw
In reply to Andrew Cooper
Quoted 14 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Quoted 4 lines

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

We should then bump the minimum required GNU binutils version in the README, as strip is also part of the binutils suite itself?

Quoted 51 lines

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Teddy Astie <[email protected]>
CC: Frediano Ziglio <[email protected]>
CC: Oleksii Kurochko <[email protected]>
CC: Marek Marczykowski-Górecki <[email protected]>
CC: Daniel P. Smith <[email protected]>

For 4.22. This was posted previously as

https://lore.kernel.org/xen-devel/[email protected]/T/#u

but merged the two patches and rewritten the commit message to make it clear that failing to strip xen.efi is causing boot failures.

Previously xen.efi.elf was produced but it's unclear why, and unnecessaerily different, so I've dropped it.

While this does want backporting, it can't be. Xen 4.21 and older still build test with Ubuntu 16.04 and choke

---
 .gitignore            |  1 +
 CHANGELOG.md          |  3 +++
 docs/misc/efi.pandoc  |  8 +-------
 xen/Kconfig.debug     |  9 ++-------
 xen/Makefile          | 19 -------------------
 xen/arch/x86/Makefile | 11 ++++-------
 xen/arch/x86/arch.mk  |  7 -------
 7 files changed, 11 insertions(+), 47 deletions(-)
.gitignore +1 -0
index bfc7bdf043c3..49e2c6961768 100644
@@ -224,6 +224,7 @@ tools/flask/policy/xenpolicy-*
 xen/xen
 xen/suppression-list.txt
 xen/xen-syms
+xen/xen-syms.efi
 xen/xen-syms.map
 xen/xen.*
 
CHANGELOG.md +0 -0
index 5cf19372a361..71d1e9ab8c69 100644

CHANGELOG.md The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/) · line 19

Quoted 6 lines
CHANGELOG.md +3 -0
@@ -14,6 +14,9 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
  - On x86:
    - Enable pf-fixup option by default for PVH dom0.
    - The libxenguest bzImage loader now uses the system liblz4 library.
+   - The install-time environment variable INSTALL_EFI_STRIP no longer exists.
+     xen.efi is always stripped, while the symbols remain available in
+     xen-syms.efi.

This is not x86-only, AFAICT ARM also seems to have a rune to generate a xen.efi image, which will be affected by the removal of INSTALL_EFI_STRIP?

Quoted 95 lines
CHANGELOG.md +0 -0
 ### Added
  - Support for per-domain Xenstore quota in C xenstored (includes
docs/misc/efi.pandoc +1 -7
index 8198a7f063cf..0a3fd67076fc 100644
@@ -20,13 +20,7 @@ Xen to load the configuration file even if multiboot modules are found.
 Once built, `make install-xen` will place the resulting binary directly into
 the EFI boot partition, provided `EFI_VENDOR` is set in the environment (and
 `EFI_MOUNTPOINT` is overridden as needed, should the default of `/boot/efi` not
-match your system). When built with debug info, the binary can be quite large.
-Setting `INSTALL_EFI_STRIP=1` in the environment will cause it to be stripped
-of debug info in the process of installing. `INSTALL_EFI_STRIP` can also be set
-to any combination of options suitable to pass to `strip`, in case the default
-ones don't do. The xen.efi binary will also be installed in `/usr/lib64/efi/`,
-unless `EFI_DIR` is set in the environment to override this default. This
-binary will not be stripped in the process.
+match your system).
 
 The binary itself will require a configuration file (names with the `.efi`
 extension of the binary's name replaced by `.cfg`, and - until an existing
xen/Kconfig.debug +2 -7
index d900d926c555..fcd3fc3d36cf 100644
@@ -147,12 +147,7 @@ config DEBUG_INFO
 	  Say Y here if you want to build Xen with debug information. This
 	  information is needed e.g. for doing crash dump analysis of the
 	  hypervisor via the "crash" tool.
-	  Saying Y will increase the size of the xen-syms and xen.efi
-	  binaries. In case the space on the EFI boot partition is rather
-	  limited, you may want to install a stripped variant of xen.efi in
-	  the EFI boot partition (look for "INSTALL_EFI_STRIP" in
-	  docs/misc/efi.pandoc for more information - when not using
-	  "make install-xen" for installing xen.efi, stripping needs to be
-	  done outside the Xen build environment).
+	  Saying Y will increase the size of the xen-syms and xen-syms.efi
+	  binaries.
 
 endmenu
xen/Makefile +0 -19
index 1f11610b5f68..0f9b56fc399d 100644
@@ -493,22 +493,6 @@ endif
 .PHONY: _build
 _build: $(TARGET)$(CONFIG_XEN_INSTALL_SUFFIX)
 
-# Strip
-#
-# INSTALL_EFI_STRIP, if defined, will cause xen.efi to be stripped before it
-# is installed. If INSTALL_EFI_STRIP is '1', then the default option(s) below
-# will be used. Otherwise, INSTALL_EFI_STRIP value will be used as the
-# option(s) to the strip command.
-ifdef INSTALL_EFI_STRIP
-
-ifeq ($(INSTALL_EFI_STRIP),1)
-efi-strip-opt := --strip-debug --keep-file-symbols
-else
-efi-strip-opt := $(INSTALL_EFI_STRIP)
-endif
-
-endif
-
 .PHONY: _install
 _install: D=$(DESTDIR)
 _install: T=$(notdir $(TARGET))
@@ -535,9 +519,6 @@ _install: $(TARGET)$(CONFIG_XEN_INSTALL_SUFFIX)
 		ln -sf $(T)-$(XEN_FULLVERSION).efi $(D)$(EFI_DIR)/$(T)-$(XEN_VERSION).efi; \
 		ln -sf $(T)-$(XEN_FULLVERSION).efi $(D)$(EFI_DIR)/$(T).efi; \
 		if [ -n '$(EFI_MOUNTPOINT)' -a -n '$(EFI_VENDOR)' ]; then \
-			$(if $(efi-strip-opt), \
-			     $(STRIP) $(efi-strip-opt) -p -o $(TARGET).efi.stripped $(TARGET).efi && \
-			     $(INSTALL_DATA) $(TARGET).efi.stripped $(D)$(EFI_MOUNTPOINT)/efi/$(EFI_VENDOR)/$(T)-$(XEN_FULLVERSION).efi ||) \
 			$(INSTALL_DATA) $(TARGET).efi $(D)$(EFI_MOUNTPOINT)/efi/$(EFI_VENDOR)/$(T)-$(XEN_FULLVERSION).efi; \
 		elif [ "$(D)" = "$(patsubst $(shell cd $(XEN_ROOT) && pwd)/%,%,$(D))" ]; then \
 			echo 'EFI installation only partially done (EFI_VENDOR not set)' >&2; \
xen/arch/x86/Makefile +1 -4
index 47dd6c50fe88..01ed7302202e 100644
@@ -196,10 +196,7 @@ note_file_option ?= $(note_file)
 
 extra-$(XEN_BUILD_PE) += efi.lds
 ifeq ($(XEN_BUILD_PE),y)
-$(TARGET).efi: $(objtree)/prelink.o $(note_file) $(obj)/efi.lds $(obj)/efi/relocs-dummy.o $(obj)/efi/mkreloc
-ifeq ($(CONFIG_DEBUG_INFO),y)
-	$(if $(filter --strip-debug,$(EFI_LDFLAGS)),echo,:) "Will strip debug info from $(@F)"
-endif
+$(TARGET)-syms.efi: $(objtree)/prelink.o $(note_file) $(obj)/efi.lds $(obj)/efi/relocs-dummy.o $(obj)/efi/mkreloc
 	$(objtree)/tools/symbols $(all_symbols) --source-name=$(@F).S --empty \
 		> $(dot-target).0s.S
 	$(MAKE) $(build)=$(@D) .$(@F).0s.o

xen/arch/x86/Makefile endif · line 236

Quoted 10 lines
xen/arch/x86/Makefile +3 -3
@@ -233,10 +230,10 @@ endif
 	$(NM) -pa --format=sysv $@ \
 		| $(objtree)/tools/symbols --all-symbols --xensyms --sysv --sort \
 		> [email protected]
-ifeq ($(CONFIG_DEBUG_INFO),y)
-	$(if $(filter --strip-debug,$(EFI_LDFLAGS)),:$(space))$(OBJCOPY) -O elf64-x86-64 $@ [email protected]
-endif
 	rm -f $(dot-target).[0-9]* $(@D)/..$(@F).[0-9]*
+
+$(TARGET).efi: $(TARGET)-syms.efi
+	$(STRIP) $< -o $@

I'm not that good with Makefiles, but don't we need a similar adjustment to strip the .efi generated for ARM?

Thanks, Roger.

#6 lore raw
Quoted 67 lines · Roger Pau Monné
Quoted 14 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Quoted 4 lines

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

We should then bump the minimum required GNU binutils version in the README, as strip is also part of the binutils suite itself?

Quoted 45 lines

---
CC: Jan Beulich <[email protected]>
CC: Roger Pau Monné <[email protected]>
CC: Teddy Astie <[email protected]>
CC: Frediano Ziglio <[email protected]>
CC: Oleksii Kurochko <[email protected]>
CC: Marek Marczykowski-Górecki <[email protected]>
CC: Daniel P. Smith <[email protected]>

For 4.22. This was posted previously as

https://lore.kernel.org/xen-devel/[email protected]/T/#u

but merged the two patches and rewritten the commit message to make it clear that failing to strip xen.efi is causing boot failures.

Previously xen.efi.elf was produced but it's unclear why, and unnecessaerily different, so I've dropped it.

While this does want backporting, it can't be. Xen 4.21 and older still build test with Ubuntu 16.04 and choke

---
 .gitignore            |  1 +
 CHANGELOG.md          |  3 +++
 docs/misc/efi.pandoc  |  8 +-------
 xen/Kconfig.debug     |  9 ++-------
 xen/Makefile          | 19 -------------------
 xen/arch/x86/Makefile | 11 ++++-------
 xen/arch/x86/arch.mk  |  7 -------
 7 files changed, 11 insertions(+), 47 deletions(-)
.gitignore +1 -0
index bfc7bdf043c3..49e2c6961768 100644
@@ -224,6 +224,7 @@ tools/flask/policy/xenpolicy-*
 xen/xen
 xen/suppression-list.txt
 xen/xen-syms
+xen/xen-syms.efi
 xen/xen-syms.map
 xen/xen.*
 
Quoted 14 lines

CHANGELOG.md The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/) · line 19

Quoted 11 lines
CHANGELOG.md +3 -0
index 5cf19372a361..71d1e9ab8c69 100644
@@ -14,6 +14,9 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
  - On x86:
    - Enable pf-fixup option by default for PVH dom0.
    - The libxenguest bzImage loader now uses the system liblz4 library.
+   - The install-time environment variable INSTALL_EFI_STRIP no longer exists.
+     xen.efi is always stripped, while the symbols remain available in
+     xen-syms.efi.

This is not x86-only, AFAICT ARM also seems to have a rune to generate a xen.efi image, which will be affected by the removal of INSTALL_EFI_STRIP?

x86 and ARM are entirely different, with x86 being the weird one.

ARM, like every other systems codebase trying to do EFI, is a raw binary with hand-coded MZ/PE+ headers.  In our case, it's deliberately a polygot which is both EFI and Linux zImage.

Notice how the rule is this:

$(TARGET): $(TARGET)-syms
        $(OBJCOPY) -O binary -S $< $@
ifeq ($(CONFIG_ARM_64),y)
        ln -sf $(@F) [email protected]
endif

~Andrew

Acked-by on v1 1/1
#7 lore raw
In reply to Andrew Cooper
Quoted 14 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Quoted 7 lines

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

Thanks, Roger.

#8 lore raw
In reply to Andrew Cooper
Quoted 8 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

I'm not happy with how things are put here. There's nothing abnormal about including about anything. What is abnormal is the manufacturing of a 32-bit ELF binary from a 64-bit one by mkelf32, to please bootloaders. An EFI binary should be permitted to include whatever data it wants, and firmware should be able to load it as long as memory permits. I don't expect you mean to indicate that problematic systems don't have 45Mb available at boot.

Including debug info can be a waste of I/O bandwidth and memory, when the loader doesn't skip loading those .debug_* sections (for valid or bogus reasons).

Quoted 9 lines

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

Jan

#9 lore raw
Quoted 9 lines · Jan Beulich
Quoted 4 lines · Andrew Cooper

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

I should have added: An option may be to link twice: Once with debug info included, and once with it stripped. Personally I trust the linker creating the various headers, including the section ones, more than strip's (or objcopy's). Yet then I can only repeat my observation that linking PE+ from ELF inputs looks to be significantly slower than linking ELF -> ELF.

Jan

#10 lore raw
Quoted 12 lines · Jan Beulich
Quoted 9 lines · Jan Beulich
Quoted 4 lines · Andrew Cooper

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

Quoted 8 lines

I should have added: An option may be to link twice: Once with debug info included, and once with it stripped. Personally I trust the linker creating the various headers, including the section ones, more than strip's (or objcopy's). Yet then I can only repeat my observation that linking PE+ from ELF inputs looks to be significantly slower than linking ELF -> ELF.

Jan

That was also attempted. See previous versions. And no, it does not work.

Frediano

#11 lore raw
Quoted 13 lines · Frediano Ziglio
Quoted 12 lines · Jan Beulich
Quoted 9 lines · Jan Beulich
Quoted 4 lines · Andrew Cooper

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

Quoted 7 lines
Quoted 5 lines

I should have added: An option may be to link twice: Once with debug info included, and once with it stripped. Personally I trust the linker creating the various headers, including the section ones, more than strip's (or objcopy's). Yet then I can only repeat my observation that linking PE+ from ELF inputs looks to be significantly slower than linking ELF -> ELF.

That was also attempted. See previous versions. And no, it does not work.

How exactly does it not work? When stripping debug info while linking (as we now do for the first two passes), the resulting image should be both small enough and correct. What am I missing? The only caveat I'm aware of is the Eclair scan, where we should avoid doing any work for the "auxiliary" linking step (the one not producing the binary that's actually going to be used for running Xen).

Jan

#12 lore raw
Quoted 23 lines · Jan Beulich
Quoted 13 lines · Frediano Ziglio
Quoted 12 lines · Jan Beulich
Quoted 9 lines · Jan Beulich
Quoted 4 lines · Andrew Cooper

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

Quoted 7 lines
Quoted 5 lines

I should have added: An option may be to link twice: Once with debug info included, and once with it stripped. Personally I trust the linker creating the various headers, including the section ones, more than strip's (or objcopy's). Yet then I can only repeat my observation that linking PE+ from ELF inputs looks to be significantly slower than linking ELF -> ELF.

That was also attempted. See previous versions. And no, it does not work.

Quoted 8 lines

How exactly does it not work? When stripping debug info while linking (as we now do for the first two passes), the resulting image should be both small enough and correct. What am I missing? The only caveat I'm aware of is the Eclair scan, where we should avoid doing any work for the "auxiliary" linking step (the one not producing the binary that's actually going to be used for running Xen).

Jan

One thing I remember was the build-id was not the same and debugging tools could not work.

Frediano

#13 lore raw
Quoted 32 lines · Frediano Ziglio
Quoted 30 lines · Jan Beulich
Quoted 13 lines · Frediano Ziglio
Quoted 12 lines · Jan Beulich
Quoted 9 lines · Jan Beulich
Quoted 4 lines · Andrew Cooper

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

Quoted 7 lines
Quoted 5 lines

I should have added: An option may be to link twice: Once with debug info included, and once with it stripped. Personally I trust the linker creating the various headers, including the section ones, more than strip's (or objcopy's). Yet then I can only repeat my observation that linking PE+ from ELF inputs looks to be significantly slower than linking ELF -> ELF.

That was also attempted. See previous versions. And no, it does not work.

How exactly does it not work? When stripping debug info while linking (as we now do for the first two passes), the resulting image should be both small enough and correct. What am I missing? The only caveat I'm aware of is the Eclair scan, where we should avoid doing any work for the "auxiliary" linking step (the one not producing the binary that's actually going to be used for running Xen).

Quoted 5 lines
Quoted 1 line

Jan

One thing I remember was the build-id was not the same and debugging tools could not work.

Another reason was that the generated binary with debug information was not working with various tools anyway (gdb complained about it and also binutils again for converting to ELF to use other tools like I think live patching).
About this I'm less sure about.

Frediano

#14 lore raw
In reply to Frediano Ziglio
Quoted 31 lines · Frediano Ziglio
Quoted 21 lines · Jan Beulich
Quoted 19 lines · Frediano Ziglio
Quoted 17 lines · Jan Beulich
Quoted 9 lines · Jan Beulich
Quoted 4 lines · Andrew Cooper

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

I should have added: An option may be to link twice: Once with debug info included, and once with it stripped. Personally I trust the linker creating the various headers, including the section ones, more than strip's (or objcopy's). Yet then I can only repeat my observation that linking PE+ from ELF inputs looks to be significantly slower than linking ELF -> ELF.

Quoted 8 lines
Quoted 1 line

That was also attempted. See previous versions. And no, it does not work.

How exactly does it not work? When stripping debug info while linking (as we now do for the first two passes), the resulting image should be both small enough and correct. What am I missing? The only caveat I'm aware of is the Eclair scan, where we should avoid doing any work for the "auxiliary" linking step (the one not producing the binary that's actually going to be used for running Xen).

Quoted 2 lines

One thing I remember was the build-id was not the same and debugging tools could not work.

Hmm, yes, that's a little ugly, but can likely be dealt with by using --build-id=0x<hexdigits> to replicate the build-id that was generated for the main binary.

Jan

#15 lore raw
Quoted 32 lines · Jan Beulich
Quoted 31 lines · Frediano Ziglio
Quoted 29 lines · Jan Beulich
Quoted 13 lines · Frediano Ziglio
Quoted 12 lines · Jan Beulich
Quoted 9 lines · Jan Beulich
Quoted 4 lines · Andrew Cooper

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

Quoted 7 lines
Quoted 5 lines

I should have added: An option may be to link twice: Once with debug info included, and once with it stripped. Personally I trust the linker creating the various headers, including the section ones, more than strip's (or objcopy's). Yet then I can only repeat my observation that linking PE+ from ELF inputs looks to be significantly slower than linking ELF -> ELF.

That was also attempted. See previous versions. And no, it does not work.

How exactly does it not work? When stripping debug info while linking (as we now do for the first two passes), the resulting image should be both small enough and correct. What am I missing? The only caveat I'm aware of is the Eclair scan, where we should avoid doing any work for the "auxiliary" linking step (the one not producing the binary that's actually going to be used for running Xen).

Quoted 6 lines
Quoted 2 lines

One thing I remember was the build-id was not the same and debugging tools could not work.

Hmm, yes, that's a little ugly, but can likely be dealt with by using --build-id=0x<hexdigits> to replicate the build-id that was generated for the main binary.

IMO linking twice (with and without debug symbols) has a great risk potential of producing different layout of the binary. While arguably it would be a bug in the build scripts, it doesn't matter with strip approach (used by virtually every other project I've seen).

Signature · 3 lines
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab
#16 lore raw
Quoted 36 lines · Marek Marczykowski-Górecki
Quoted 35 lines · Jan Beulich
Quoted 33 lines · Frediano Ziglio
Quoted 21 lines · Jan Beulich
Quoted 19 lines · Frediano Ziglio
Quoted 17 lines · Jan Beulich
Quoted 9 lines · Jan Beulich
Quoted 4 lines · Andrew Cooper

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

I should have added: An option may be to link twice: Once with debug info included, and once with it stripped. Personally I trust the linker creating the various headers, including the section ones, more than strip's (or objcopy's). Yet then I can only repeat my observation that linking PE+ from ELF inputs looks to be significantly slower than linking ELF -> ELF.

Quoted 8 lines
Quoted 1 line

That was also attempted. See previous versions. And no, it does not work.

How exactly does it not work? When stripping debug info while linking (as we now do for the first two passes), the resulting image should be both small enough and correct. What am I missing? The only caveat I'm aware of is the Eclair scan, where we should avoid doing any work for the "auxiliary" linking step (the one not producing the binary that's actually going to be used for running Xen).

One thing I remember was the build-id was not the same and debugging tools could not work.

Quoted 7 lines
Quoted 3 lines

Hmm, yes, that's a little ugly, but can likely be dealt with by using --build-id=0x<hexdigits> to replicate the build-id that was generated for the main binary.

IMO linking twice (with and without debug symbols) has a great risk potential of producing different layout of the binary. While arguably it would be a bug in the build scripts,

Indeed, as it would further undermine reproducible builds. That said, we could certainly put some (a lot of?) sanity checking in place.

Quoted 2 lines

it doesn't matter with strip
approach (used by virtually every other project I've seen).

Yet I expect that's "strip" on ELF binaries, not PE ones?

Jan

#17 lore raw
In reply to Jan Beulich
Quoted 10 lines · Jan Beulich
Quoted 8 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

Quoted 6 lines

I'm not happy with how things are put here. There's nothing abnormal about including about anything. What is abnormal is the manufacturing of a 32-bit ELF binary from a 64-bit one by mkelf32, to please bootloaders. An EFI binary should be permitted to include whatever data it wants, and firmware should be able to load it as long as memory permits. I don't expect you mean to indicate that problematic systems don't have 45Mb available at boot.

Meanwhile it occurred to me to further ask: How would unified images work on such systems imposing size restrictions? They're pretty likely to be yet larger, especially if the initrd is rather big.

Jan

#18 lore raw
In reply to Jan Beulich
Quoted 18 lines · Jan Beulich
Quoted 8 lines · Andrew Cooper

From: Frediano Ziglio <[email protected]>

xen.efi with debugging symbols is ~45MB, down to ~9.3MB when stripped. Multiple firmwares (as seen by QubesOS, Trenchboot, and XenServer) are unable to boot xen.efi when debugging symbols are included.

Either way, having debug symbols by default is abnormal and contrary to how the non-EFI path works.

I'm not happy with how things are put here. There's nothing abnormal about including about anything. What is abnormal is the manufacturing of a 32-bit ELF binary from a 64-bit one by mkelf32, to please bootloaders. An EFI binary should be permitted to include whatever data it wants, and firmware should be able to load it as long as memory permits. I don't expect you mean to indicate that problematic systems don't have 45Mb available at boot.

Quoted 4 lines

Including debug info can be a waste of I/O bandwidth and memory, when the loader doesn't skip loading those .debug_* sections (for valid or bogus reasons).

One reason, at least for secure boot, is to compute the hash of the file. The hash includes almost everything, excluding the header checksum, the signature section header and the signature section (all that must be read too).

Quoted 5 lines
Quoted 5 lines

Produce xen-syms.efi unconditionally, just like xen-syms. If CONFIG_DEBUG_INFO is enabled, these will contain debug symbols, and if not, then not. When xen-syms is processed by mkelf32, the debug symbols are simply discarded. For xen-syms.efi, call $(STRIP) to produce xen.efi.

Quoted 9 lines
Quoted 4 lines

Some old versions of binutils ld managed to produce efi files which the matching version of strip couldn't process. This includes Binutils 2.26 included in Ubuntu 16.04. Delete the workaround for this bug, and require a less broken toolchain.

And we're certain newer versions of strip don't do any harm to the binaries? Already towards Frediano's posting I said that having looked at how things work there, I'm far from certain.

Yes, software contains bugs and in this area binutils has quite a history. What we know for sure is that a specific problem has been fixed. Are all the bugs fixed? Probably not. I don't see a valid reason to wait to have some kind of "bug free" version.

Quoted 2 lines

Jan

Frediano