Jake reports that the IS_ARM_MEMORY_REGION_ATTRIBUTES_SECURE() macro is
no longer accurate since commit 852227a9d5 ("ArmPkg/Mmu: Remove
handling of NONSECURE memory regions").
Fortunately, it only affects the NS bit in level 1 short descriptors,
which is ignored when executing in non-secure mode. And given that
running UEFI in the secure world is not a use case we aim to support,
let's just drop this logic altogether.
Reported-by: Jake Garver <jake@nvidia.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Now that we have a sane API to set and clear memory permissions that
works the same on ARM and AArch64, we no longer have a need for the
individual set/clear no-access/read-only/no-exec helpers so let's drop
them.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Currently, ArmSetMemoryAttributes () takes a combination of
EFI_MEMORY_xx constants describing the memory type and permission
attributes that should be set on a region of memory. In cases where the
memory type is omitted, we assume that the memory permissions being set
are final, and that existing memory permissions can be discarded.
This is problematic, because we aim to map memory non-executable
(EFI_MEMORY_XP) by default, and only relax this requirement for code
regions that are mapped read-only (EFI_MEMORY_RO). Currently, setting
one permission clears the other, and so code managing these permissions
has to be aware of the existing permissions in order to be able to
preserve them, and this is not always tractable (e.g., the UEFI memory
attribute protocol implements an abstraction that promises to preserve
memory permissions that it is not operating on explicitly).
So let's add an AttributeMask parameter to ArmSetMemoryAttributes(),
which is permitted to be non-zero if no memory type is being provided,
in which case only memory permission attributes covered in the mask will
be affected by the update.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Oliver Smith-Denny <osde@linux.microsoft.com>
Reviewed-by: Michael Kubacki <michael.kubacki@microsoft.com>
The helper that updates live page table entries writes a zero entry,
invalidates the covered address range from the TLBs, and finally writes
the actual entry. This ensures that no TLB conflicts can occur.
Writing the final entry needs to complete before any translations can be
performed, as otherwise, the zero entry, which describes an invalid
translation, may be observed by the page table walker, resulting in a
translation fault. For this reason, the final write is followed by a DSB
barrier instruction.
However, this barrier will not stall the pipeline, and instruction
fetches may still hit this invalid translation, as has been observed and
reported by Oliver. To ensure that the new translation is fully active
before returning from this helper, we have to insert an ISB barrier as
well.
Reported-by: Oliver Steffen <osteffen@redhat.com>
Tested-by: Oliver Steffen <osteffen@redhat.com>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Acked-by: Michael D Kinney <michael.d.kinney@intel.com>
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
As the ASM_FUNC() macro performs a section switch, the preceding
.balign directive applies the alignment constraint to the current
location in the previous section. As the linker may not merge the
sections in-order, ArmReplaceLiveTranslationEntry() may be left
unaligned.
Replace the explicit invocation of .balign with the ASM_FUNC_ALIGN()
macro, which guarantees the alignment constraint is applied correctly.
To make sure related issues are reliably caught in the future, align the
end of the function before checking the total occupied size. This
ensures crossing a 0x200 boundary will cause a compilation error.
Signed-off-by: Marvin Häuser <mhaeuser@posteo.de>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
__FUNCTION__ is a pre-standard extension that gcc and Visual C++ among
others support, while __func__ was standardized in C99.
Since it's more standard, replace __FUNCTION__ with __func__ throughout
ArmPkg.
Signed-off-by: Rebecca Cran <rebecca@bsdio.com>
Reviewed-by: Michael D Kinney <michael.d.kinney@intel.com>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
To prepare for the enablement of booting EFI with the SCTLR.WXN control
enabled, which makes all writeable memory regions non-executable by
default, introduce a memory type that we will use to describe the flash
region that carries the SEC and PEIM modules that execute in place. Even
if these are implicitly read-only due to the ROM nature, they need to be
mapped with read-only attributes in the page tables to be able to
execute from them.
Also add the XP counterpart which will be used for all normal DRAM right
at the outset.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Non-secure memory is a distinction that only matters when executing code
in the secure world that reasons about the secure vs non-secure address
spaces. EDK2 was not designed for that, and the AArch64 version of the
MMU handling library already treats them as identical, so let's just
drop the ARM memory region types that mark memory as 'non-secure'
explicitly.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Currently, the ARM MMU page table logic will break down any block entry
that overlaps with the region being mapped, even if the block entry in
question is using the same attributes as the new region.
This means that creating a non-executable mapping inside a region that
is already mapped non-executable at a coarser granularity may trigger a
call to AllocatePages (), which may recurse back into the page table
code to update the attributes on the newly allocated page tables.
Let's avoid this, by preserving the block entry if it already covers the
region being mapped with the correct attributes.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Implement support for read-protected memory by wiring it up to the
access flag in the page table descriptor. The resulting mapping is
implicitly non-writable and non-executable as well, but this is good
enough for implementing this attribute, as we never rely on write or
execute permissions without read permissions.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Currently, the MMU code that is supposed to clear the RO or XP
attributes from a region just clears both unconditionally. This
approximates the desired behavior to some extent, but it does mean that
setting the RO bit first on a code region, and then clearing the XP bit
results both RO and XP being cleared, and we end up with writable code,
and avoiding that is the point of all these protections.
Once we introduce RP support, this will only get worse, so let's fix
this up, by reshuffling the attribute update code to take the entry mask
from the caller, and use the mask to preserve other attributes when
clearing RO or XP.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Split the ARM permission fields in the short descriptors into an access
flag and AP[2:1] as per the recommendation in the ARM ARM. This makes
the access flag available separately, which allows us to implement
EFI_MEMORY_RP memory analogous to how it will be implemented for
AArch64.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
With large page support out of the picture, we can treat bits 1 and 0 of
the page descriptor as individual valid and XN bits, instead of treating
XN as a page type. Doing so aligns the handling of the attribute with
the section descriptor layout, as well as the XN handling on AArch64,
and this is beneficial for maintainability.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Large page support on 32-bit ARM is essentially a glorified contiguous
bit where 16 consecutive entries describing a contiguous range with the
same attributes are presented in a way that permits the TLB to cache its
translation with a single entry.
This was never wired up completely, and does not add a lot of value in
EFI, where the page granularity is 4k and we expect to be able to set RO
and XP permissions on individual pages.
Given that large page support complicates the handling of the XN bit at
the page level (which is in a different place depending on whether the
page is small or large), let's just rip it out.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
In order to reduce the likelihood that we will need to rely on the logic
that disables and re-enables the MMU for updating a page table entry
safely, expose the XIP version of the helper routine via a HOB and use
it instead of the one that is copied into DRAM. Since the XIP copy is
already clean to the PoC, and will never end up getting unmapped during
a block entry split, we can use it safely without any cache maintenance,
and without running the risk of pulling the rug from under our feet when
updating an entry by going through an invalid mapping.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Acked-by: Leif Lindholm <quic_llindhol@quicinc.com>
Permit the use of this library with the MMU and caches already enabled.
This removes the need for any cache maintenance for coherency, and is
generally better for robustness and performance, especially when running
under virtualization.
Note that this means we have to defer assignment of TTBR0 until the
page tables are ready to be used, and so UpdateRegionMapping() can no
longer read back TTBR0 directly to discover the root table address.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Acked-by: Leif Lindholm <quic_llindhol@quicinc.com>
When updating a page table descriptor in a way that requires break
before make, we temporarily disable the MMU to ensure that we don't
unmap the memory region that the code itself is executing from.
However, this is a condition we can check in a straight-forward manner,
and if the regions are disjoint, we don't have to bother with the MMU
controls, and we can just perform an ordinary break before make.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
Drop the optimization that replaces table entries with block entries and
frees the page tables in the subhierarchy that is being replaced. This
rarely occurs in practice anyway, and will require more elaborate TLB
maintenance once we switch to a different approach where we no longer
disable the MMU and nuke the TLB entirely every time we update a
descriptor in a way that requires break-before-make (BBM).
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
In an effort to clean the documentation of the above
package, remove duplicated words, and fix a typo while at it.
Signed-off-by: Pierre Gondois <pierre.gondois@arm.com>
Reviewed-by: Sami Mujawar <sami.muajwar@arm.com>
Reviewed-by: Leif Lindholm <quic_llindhol@quicinc.com>
RVCT is obsolete and no longer used.
Remove support for it.
Signed-off-by: Rebecca Cran <quic_rcran@quicinc.com>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
We never run any code at EL0, and so it would seem that any access
permissions set for EL0 (via the AP[1] attribute in the page tables) are
irrelevant. We currently set EL0 and EL1 permissions to the same value
arbitrarily.
However, this causes problems on hardware like the Apple M1 running the
MacOS hypervisor framework, which enters EL1 with SCTLR_EL1.SPAN
enabled, causing the Privileged Access Never (PAN) feature to be enabled
on any exception taken to EL1, including the IRQ exceptions that handle
our timer interrupt. When PAN is enabled, EL1 has no access to any
mappings that are also accessible to EL0, causing the firmware to crash
if it attempts to access such a mapping.
Even though it is debatable whether or not SCTLR_EL1.SPAN should be
disabled at entry or whether the firmware should put all UNKNOWN bits in
all system registers in a consistent state (which it should), using EL0
permissions serves no purpose whatsoever so let's fix that regardless.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Tested-by: Alexander Graf <agraf@csgraf.de>
Acked-by: Leif Lindholm <leif@nuviainc.com>
REF: https://bugzilla.tianocore.org/show_bug.cgi?id=3737
Apply uncrustify changes to .c/.h files in the ArmPkg package
Cc: Andrew Fish <afish@apple.com>
Cc: Leif Lindholm <leif@nuviainc.com>
Cc: Michael D Kinney <michael.d.kinney@intel.com>
Signed-off-by: Michael Kubacki <michael.kubacki@microsoft.com>
Reviewed-by: Andrew Fish <afish@apple.com>
REF: https://bugzilla.tianocore.org/show_bug.cgi?id=3739
Update all use of EFI_D_* defines in DEBUG() macros to DEBUG_* defines.
Cc: Andrew Fish <afish@apple.com>
Cc: Leif Lindholm <leif@nuviainc.com>
Cc: Michael Kubacki <michael.kubacki@microsoft.com>
Signed-off-by: Michael D Kinney <michael.d.kinney@intel.com>
Reviewed-by: Sami Mujawar <sami.mujawar@arm.com>
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
The 'cspell' CI test detected some small typos in ArmPkg.
Correct them.
Cc: Bret Barkelew <bret.barkelew@microsoft.com>
Cc: Sean Brogan <sean.brogan@microsoft.com>
Cc: Leif Lindholm <leif@nuviainc.com>
Cc: Ard Biesheuvel <ardb+tianocore@kernel.org>
Cc: Sami Mujawar <sami.mujawar@arm.com>
Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
Reviewed-by: Sami Mujawar <sami.mujawar@arm.com>
This patch fixes the following Ecc reported error:
There should be no initialization of a variable as
part of its declaration
Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
This patch fixes the following Ecc reported error:
Non-Boolean comparisons should use a compare operator
(==, !=, >, < >=, <=)
Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
The header of the file is not formatted properly, making
the Ecc tool crash when running on the ArmPkg.
The following command was run:
./BaseTools/BinWrappers/PosixLike/Ecc
-c BaseTools/Source/Python/Ecc/config.ini
-e BaseTools/Source/Python/Ecc/exception.xml
-t ArmPkg -r ArmPkgEcc.xls
Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
The function ArmReplaceLiveTranslationEntry () is passed as a VOID
pointer to WriteBackDataCacheRange (). This produces the following
warning on VS2019:
warning C4152: nonstandard extension, function/data pointer
conversion in expression
This change explicitly casts the argument to the formal parameter
type VOID*.
This can be reproduced with the following build command:
build -b DEBUG -a AARCH64 -t VS2019 -p ArmPkg/ArmPkg.dsc
-m ArmPkg/Library/ArmMmuLib/ArmMmuPeiLib.inf
Signed-off-by: Michael Kubacki <michael.kubacki@microsoft.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
REF:https://bugzilla.tianocore.org/show_bug.cgi?id=2835
There's several occurrences of a UINT64 or an EFI_PHYSICAL_ADDRESS
being assigned to a UINT32 value in ArmMmuLib. These result in
warning C4244 in VS2019:
warning C4244: '=': conversion from 'UINT64' to 'UINT32', possible
loss of data
warning C4244: '=': conversion from 'EFI_PHYSICAL_ADDRESS' to
'UINT32', possible loss of data
This change explicitly casts the values to UINT32.
These can be reproduced with the following build command:
build -b DEBUG -a ARM -t VS2019 -p ArmPkg/ArmPkg.dsc
-m ArmPkg/Library/ArmMmuLib/ArmMmuBaseLib.inf
Signed-off-by: Michael Kubacki <michael.kubacki@microsoft.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
While building with the following command line:
build -b DEBUG -a AARCH64 -t VS2017 -p MdeModulePkg\MdeModulePkg.dsc
A missing cast triggers the following warning, then triggering an error:
ArmPkg/Library/ArmMmuLib/AArch64/ArmMmuLibCore.c(652):
warning C4152: nonstandard extension, function/data pointer
conversion in expression
This patch first casts the function pointer to (UINTN), then to (VOID *),
followowing the C99 standard s6.3.2.3 "Pointer", paragraphs 5 and 6.
This suppresses the warning.
Signed-off-by: Pierre Gondois <pierre.gondois@arm.com>
Suggested-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
TT_ATTR_INDX_INVALID is #define'd but never used so drop it. Note
that this leaves a CPP macro of the same name in CpuDxe, but there,
it is actually being used, and although the name suggests that this
value is somehow defined by the architecture, this is really not the
case and it only has meaning within the scope of CpuDxe's implementation.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Only a single call to GetRootTranslationTableInfo() remains, which
only provides the root table level. So let's create a new static
helper function that returns just this value, and use it instead.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
LookupAddresstoRootTable() uses a loop to go over its MaxAddress
argument, essentially to do a log2() and determine how many bits are
needed to represent it. Since the argument is the result of a shift-left
expression, there is some room for improvement here, and we can simply
use the bit count directly to calculate the value of T0SZ. At the same
time, we can omit calling GetRootTranslationTableInfo() to determine the
number of root table entries, and add a new helper that applies the
trivial calculation directly.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
The routine PageAttributeToGcdAttribute() is exported by ArmMmuLib
but only ever used in the implementation of CpuDxe. So let's move
the function there and make it STATIC.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Currently, depending on the size of the region being (re)mapped, the
page table manipulation code may replace a table entry with a block entry,
even if the existing table entry uses different mapping attributes to
describe different parts of the region it covers. This is undesirable, and
instead, we should avoid doing so unless we are disregarding the original
attributes anyway. And if we make such a replacement, we should free all
the page tables that have become orphaned in the process.
So let's implement this, by taking the table entry path through the code
for block sized regions if a table entry already exists, and the clear
mask is set (which means we are preserving attributes from the existing
mapping). And when we do replace a table entry with a block entry, free
all the pages that are no longer referenced.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Reviewed-by: Ashish Singhal <ashishsingha@nvidia.com>
Tested-by: Ashish Singhal <ashishsingha@nvidia.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
Given how the meaning of the attribute bits for page table entry types
is slightly awkward, and changes between levels, add some helpers to
abstract from this.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Reviewed-by: Ashish Singhal <ashishsingha@nvidia.com>
Tested-by: Ashish Singhal <ashishsingha@nvidia.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
FreePageTablesRecursive () traverses the page table tree depth first
to free all pages that it finds, without taking into account the
level at which it is operating.
Since TT_TYPE_TABLE_ENTRY aliases TT_TYPE_BLOCK_ENTRY_LEVEL3, we cannot
distinguish table entries from block entries unless we take the level
into account, and so we may be dereferencing garbage if we happen to
try and free a hierarchy of page tables that has level 3 pages in it.
Let's fix this by passing the level into FreePageTablesRecursive (),
and limit the recursion to levels < 3.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Reviewed-by: Ashish Singhal <ashishsingha@nvidia.com>
Tested-by: Ashish Singhal <ashishsingha@nvidia.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
Some cosmetic fixups to the AArch64 MMU code:
- reflow overly long lines unless it hurts legibility
- add/remove whitespace according to the [de facto] coding style
- use camel case for goto labels
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Message-Id: <20200307091008.14918-3-ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
This is the AARCH64 counterpart of commit 1f3b1eb308, to remove
a pointless check against the memory type of the allocations that the
page tables happened to land in. On ArmV8, we use writeback cacheable
exclusively for all memory.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Message-Id: <20200307091008.14918-2-ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
As it turns out, ARMv8 also permits accesses made with the MMU and
caches off to hit in the caches, so to ensure that any modifications
we make before enabling the MMU are visible afterwards as well, we
should invalidate page tables right after allocation like we do now on
ARM, if the MMU is still disabled at that point.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Message-Id: <20200307083849.8940-3-ard.biesheuvel@linaro.org>
Replace the slightly overcomplicated page table management code with
a simplified, recursive implementation that should be far easier to
reason about.
Note that, as a side effect, this extends the per-entry cache invalidation
that we do on page table entries to block and page entries, whereas the
previous change inadvertently only affected the creation of table entries.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Message-Id: <20200307083849.8940-2-ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
We already expect normal memory to be mapped writeback cacheable if
EDK2 itself is to make use of it, so doing an early sanity check on
the memory type of the allocation that the page tables happened to
land in isn't very useful. So let's drop it.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
The expression passed into ArmSetTTBR0 () in ArmConfigureMmu() is
sub-optimal at several levels:
- TranslationTable is already aligned, and if it wasn't, doing it
here wouldn't help
- TTBRAttributes is guaranteed not to have any bits set outside of
the 0x7f mask, so the mask operation is pointless as well,
- an additional (UINTN) cast for good measure is also not needed.
So simplify the expression.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
In the AARCH64 version of ArmMmuLib, we are currently relying on
set/way invalidation to ensure that the caches are in a consistent
state with respect to main memory once we turn the MMU on. Even if
set/way operations were the appropriate method to achieve this, doing
an invalidate-all first and then populating the page table entries
creates a window where page table entries could be loaded speculatively
into the caches before we modify them, and shadow the new values that
we write there.
So let's get rid of the blanket clean/invalidate operations, and
instead, update ArmUpdateTranslationTableEntry () to invalidate each
page table entry *after* it is written if the MMU is still disabled
at this point.
On ARMv8, it is guaranteed that memory accesses done by the page table
walker are cache coherent, and so we can ignore the case where the
MMU is on.
Since the MMU and D-cache are already off when we reach this point, we
can drop the MMU and D-cache disables as well. Maintenance of the I-cache
is unnecessary, since we are not modifying any code, and the installed
mapping is guaranteed to be 1:1. This means we can also leave it enabled
while the page table population code is running.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
In the ARM version of ArmMmuLib, we are currently relying on set/way
invalidation to ensure that the caches are in a consistent state with
respect to main memory once we turn the MMU on. Even if set/way
operations were the appropriate method to achieve this, doing an
invalidate-all first and then populating the page table entries creates
a window where page table entries could be loaded speculatively into
the caches before we modify them, and shadow the new values that we
write there.
So let's get rid of the blanket clean/invalidate operations, and instead,
invalidate each page table right after allocating it, and each section
entry after it is updated (to address all the little corner cases that the
ARMv7 spec permits), and invalidate sets of level 2 entries in blocks,
using the generic invalidation routine from CacheMaintenanceLib
On ARMv7, cache maintenance may be required also when the MMU is
enabled, in case the page table walker is not cache coherent. However,
the code being updated here is guaranteed to run only when the MMU is
still off, and so we can disregard the case when the MMU and caches
are on.
Since the MMU and D-cache are already off when we reach this point, we
can drop the MMU and D-cache disables as well. Maintenance of the I-cache
is unnecessary, since we are not modifying any code, and the installed
mapping is guaranteed to be 1:1. This means we can also leave it enabled
while the page table population code is running.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Instead of overallocating memory and align the resulting base address
manually, use the AllocateAlignedPages () helper, which achieves the
same, and might even manage that without leaking a chunk of memory of
the same size as the allocation itself.
While at it, fix up a variable declaration in the same hunk, and drop
a comment whose contents add nothing to the following line of code.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Unlike the AArch64 implementation of ArmMmuLib, which combines the
initial page table population code with the code that runs at later
stages to manage permission attributes in the page tables, ARM uses
two completely separate sets of routines for this.
Since ArmMmuLib is a static library, we can prevent duplication of
this code between different users, which usually only need one or
the other. (Note that LTO should also achieve the same.)
This also makes it easier to reason about modifying the cache
maintenance handling, and replace the set/way ops with by-VA
ops, since the code that performs the set/way ops only executes
when the MMU is still off.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Make the CONSTRUCTOR define in the .INF AARCH64 only, so we can drop
the empty stub that exists for ARM.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Leif Lindholm <leif@nuviainc.com>
Ard Biesheuvel
Marvin Häuser
Rebecca Cran
Pierre Gondois
Rebecca Cran
Michael Kubacki
Michael D Kinney
Ard Biesheuvel
Ard Biesheuvel
Leif Lindholm
Antoine Cœur