From mboxrd@z Thu Jan 1 00:00:00 1970 Authentication-Results: mx.groups.io; dkim=missing; spf=pass (domain: redhat.com, ip: 209.132.183.28, mailfrom: lersek@redhat.com) Received: from mx1.redhat.com (mx1.redhat.com [209.132.183.28]) by groups.io with SMTP; Thu, 26 Sep 2019 11:23:58 -0700 Received: from smtp.corp.redhat.com (int-mx05.intmail.prod.int.phx2.redhat.com [10.5.11.15]) (using TLSv1.2 with cipher AECDH-AES256-SHA (256/256 bits)) (No client certificate requested) by mx1.redhat.com (Postfix) with ESMTPS id 833647FDCA; Thu, 26 Sep 2019 18:23:57 +0000 (UTC) Received: from lacos-laptop-7.usersys.redhat.com (ovpn-120-49.rdu2.redhat.com [10.10.120.49]) by smtp.corp.redhat.com (Postfix) with ESMTP id B6BA55D6A7; Thu, 26 Sep 2019 18:23:54 +0000 (UTC) Subject: Re: [PATCH V1 0/5] UEFI Variable SMI Reduction To: Michael Kubacki , devel@edk2.groups.io Cc: Dandan Bi , Ard Biesheuvel , Eric Dong , Liming Gao , Michael D Kinney , Ray Ni , Jian J Wang , Hao A Wu , Jiewen Yao References: <20190926045046.34592-1-michael.a.kubacki@intel.com> From: "Laszlo Ersek" Message-ID: Date: Thu, 26 Sep 2019 20:23:53 +0200 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.9.1 MIME-Version: 1.0 In-Reply-To: <20190926045046.34592-1-michael.a.kubacki@intel.com> X-Scanned-By: MIMEDefang 2.79 on 10.5.11.15 X-Greylist: Sender IP whitelisted, not delayed by milter-greylist-4.5.16 (mx1.redhat.com [10.5.110.27]); Thu, 26 Sep 2019 18:23:57 +0000 (UTC) Content-Type: text/plain; charset=utf-8 Content-Language: en-US Content-Transfer-Encoding: 7bit Hello Michael, On 09/26/19 06:50, Michael Kubacki wrote: > REF:https://bugzilla.tianocore.org/show_bug.cgi?id=2220 > > Overview > --------- > This patch series reduces SMM usage when using VariableSmmRuntimeDxe with > VariableSmm. It does so by eliminating SMM usage for runtime service > GetVariable () and GetNextVariableName () invocations. Most UEFI variable > usage in typical systems after the variable store is initialized > (e.g. manufacturing boots) is due to GetVariable ( ) and > GetNextVariableName () not SetVariable (). GetVariable () calls can regularly > exceed 100 per boot while SetVariable () calls typically remain less than 10 > per boot. By focusing on the common case, the majority of overhead associated > with SMM can be avoided while still using existing and proven code for > operations such as variable authentication that require an isolated execution > environment. > > * Advantage: Reduces overall system SMM usage > * Disadvantage: Requires more Runtime data memory usage In a perfect world, I would carefully scrutinize this patch set, and respond with comments. In the real world, I have hardly enough time to read the blurb :/ So I'll have to defer to the other reviewers on this patch. I'd like to spell out another "disadvantage" however. Admittedly it's quite a corner case. The disadvantage in my case is that, by eliminating SMM from variable *reads*, OVMF will lose its simplest method to exercise the SMM driver stack. Namely, right now, if you boot OVMF (built with -D SMM_REQUIRE), and at root prompt in the Linux guest, you run: # taskset -c 0 efibootmgr # taskset -c 1 efibootmgr then you very easily test the SMM machinery (through Boot####, BootOrder, BootNext variable *reads*). In addition, the "taskset" commands above force the guest Linux kernel to initiate the GetVariable runtime service call -- and therefore entering SMM -- on CPU#0 (BSP) vs. CPU#1 (AP). This difference (that is, BSP vs. AP being used for the runtime service call) used to expose *extreme* timing and stability differences in the edk2 SMM stack, dependent on the SMI delivery method used. To this day, the above two commands remain part of our stock regression-tests for the health of the SMM stack -- the commands are executed in the guest OS both after normal boot, and after S3 resume. https://github.com/tianocore/tianocore.github.io/wiki/Testing-SMM-with-QEMU,-KVM-and-libvirt#uefi-variable-access-test Therefore, would it be possible to make the feature dependent on a new FeaturePCD? Or, if that would complicate the code too much, perhaps new module INF files could be introduced (library instances, or even drivers) that a platform could choose to select in DSC files, perhaps dependent on a -D build flag. There could be a set of INF files for the current behavior, and another set of INF files for the "read cache" behavior. And the related C source files would not have to be littered with if (FeaturePcdGet (...)) { // // read cache // } else { // // traditional behavior // } Of course, with the read cache feature, SMM entry could still be forced in OVMF through non-volatile variable *writes* -- but writes are not without side-effects on the varstore, and they depend on extra conditions relative to reads. You mention "SMM cache" below, and I'm not entirely clear when exactly that would be used, in favor of the "runtime cache". It seems that the "SMM cache" would primarily serve SMM callers. If the FeaturePCD could be used for forcing the use of "SMM cache" for the normal GetVariable runtime service too, I think that might cover my use case. I don't need GetVariable to access flash, i.e. caching per se is fine; I'd just need GetVariable to continue exercising the SMM stack *in OVMF* -- even if that path is slower than desirable, for physical platforms. Thanks! Laszlo > Initial Performance Observations > --------------------------------- > * With these proposed changes, an Intel Atom based SoC saw GetVariable ( ) > time for an existing variable reduce from ~220us to ~5us. > > Major Changes > -------------- > 1. Two UEFI variable caches will be maintained. > a. "Runtime Cache" - Maintained in VariableSmmRuntimeDxe. Used to serve > runtime service GetVariable () and GetNextVariableName () callers. > b. "SMM cache" - Maintained in VariableSmm to service SMM GetVariable () > and GetNextVariableName () callers. > i. A cache in SMRAM is retained so SMM modules do not operate on data > outside SMRAM. > 2. A new UEFI variable read and write flow will be used as described below. > > At any given time, the two caches would be coherent. On a variable write, the > runtime cache is only updated after validation in SMM and, in the case of a > non-volatile UEFI variable, the variable must also be successfully written to > non-volatile storage. > > Prior RFC Feedback Addressed > ----------------------------- > RFC sent Sept. 5, 2019: https://edk2.groups.io/g/devel/message/46939 > > 1. UEFI variable data retrieval from a ring 0 buffer > > A common concern with this proposed set of changes is the potential security > threat presented by serving runtime services callers from a ring 0 memory > buffer of EfiRuntimeServicesData type. This conclusion was that this change > does not fundamentally alter the attack surface. The UEFI variable Runtime > Services are invoked from ring 0 and the data already travels through ring > 0 buffers (such as the SMM communicate buffer) to reach the caller. Even > today if ring 0 is assumed to be malicious, the malicious code may keep one > AP in a loop to monitor the communication data, when the BSP gets an > (authenticated) variable. When the communication buffer is updated and the > status is set to EFI_SUCCESS, the AP may modify the communication buffer > contents such the tampered data is returned to the BSP caller. Or an > interrupt handler on the BSP may alter the communication buffer contents > before the data is returned to the caller. In summary, this was not found to > introduce any attack not possible today. > > 2. VarCheckLib impact > > VarCheckLib plays a role in SetVariable () calls. This patch series only > changes GetVariable () behavior. Therefore, VarCheckLib is expected to > have no impact due to these changes. > > Testing Performed > ------------------ > This code was tested with the master branch of edk2 on an Intel Kaby Lake U > Intel Whiskey Lake U reference validation platform. The set of tests performed > included: > > 1. Boot from S5 to Windows 10 OS with SMM variables enabled. > 2. Boot from S5 to Ubuntu 18.04.1 LTS with SMM variable enabled. > 3. Boot from S5 to EFI shell with DXE variables enabled. > 4. Dump UEFI variable store at shell with dmpstore to verify contents. > 5. Dump NvStorage FV from SPI flash after boot to verify contents written. > 6. Dump UEFI variable statistics with VariableInfo at shell. > 7. Boot with emulated variables enabled. > 8. Cycles of adding and deleting a UEFI variable to verify cache results. > 9. Set OsIndications to stop at FW UI to verify cache load of non-volatile > contents. > > Why Keep SMM on Variable Writes > -------------------------------- > * SMM provides a ubiquitous isolated execution environment in x86 for > authenticated UEFI variables. > * BIOS region SPI flash write restrictions to SMM in platforms today can > be retained. > > Today's UEFI Variable Cache (for reference) > -------------------------------------------- > * Maintained in SMRAM via VariableSmm. > * A "write-through" cache of variable data in the form of a UEFI variable > store. > * Non-volatile and volatile variables are maintained in separate buffers > (variable stores). > > Runtime & SMM Cache Coherency > ------------------------------ > The non-volatile cache should always accurately reflect non-volatile storage > contents (done today) and the "SMM cache" and "Runtime cache" should always be > coherent on access. The runtime cache is updated by VariableSmm. > > Updating both caches from within a SMM SetVariable () operation is fairly > straightforward but a race condition can occur if an SMI occurs during the > execution of runtime code reading from the runtime cache. To handle this case, > a runtime cache read lock is introduced that explicitly moves pending updates > from SMM to the runtime cache if an SMM update occurs while the runtime cache > is locked. Note that it is not expected a Runtime services call will interrupt > SMM processing since all CPU cores rendezvous in SMM. > > New Key Elements for Coherence > ------------------------------- > Runtime DXE (VariableSmmRuntimeDxe) > 1. RuntimeCacheReadLock - A global lock used to lock read access to the > runtime cache. > 2. RuntimeCachePendingUpdate - A global flag used to notify runtime code of a > pending cache update in SMM. > > SMM (VariableSmm) > 1. FlushRuntimeCachePendingUpdate SMI - A SW SMI handler that synchronizes > the runtime cache buffer with the SMM > cache buffer. > > Proposed Runtime DXE Read Flow > ------------------------------- > 1. Wait for RuntimeCacheReadLock to be free > 2. Acquire RuntimeCacheReadLock > 3. If RuntimeCachePendingUpdate flag (rare) is set then: > 3.a. Trigger FlushRuntimeCachePendingUpdate SMI > 3.b. Verify RuntimeCachePendingUpdate flag is cleared > 4. Perform read from RuntimeCache > 5. Release RuntimeCacheReadLock > > Proposed FlushRuntimeCachePendingUpdate SMI > -------------------------------------------- > 1. If RuntimeCachePendingUpdate flag is not set: > 1.a. Return > 2. Copy the data at RuntimeCachePendingOffset of RuntimeCachePendingLength to > RuntimeCache > 3. Clear the RuntimeCachePendingUpdate flag > > Proposed SMM Write Flow > ------------------------ > 1. Perform variable authentication and non-volatile write. If either fail, > return an error to the caller. > 2. If RuntimeCacheReadLock is set then: > 2.a. Set RuntimeCachePendingUpdate flag > 2.b. Update RuntimeCachePendingOffset and RuntimeCachePendingLength to > cover the a superset of the pending chunk (for simplicity, the > entire variable store is currently synchronized). > 3. Else: > 3.a. Update RuntimeCache > 4. Update SmmCache > - Note: RT read cannot occur during SMI processing since all cores are > locked in SMM. > > Cc: Dandan Bi > Cc: Ard Biesheuvel > Cc: Eric Dong > Cc: Laszlo Ersek > Cc: Liming Gao > Cc: Michael D Kinney > Cc: Ray Ni > Cc: Jian J Wang > Cc: Hao A Wu > Cc: Jiewen Yao > Signed-off-by: Michael Kubacki > > Michael Kubacki (5): > MdeModulePkg/Variable: Consolidate common parsing functions > MdeModulePkg VariableInfo: Always consider RT DXE and SMM stats > MdeModulePkg/Variable: Add RT GetVariable() cache support > MdeModulePkg/Variable: Add RT GetNextVariableName() cache support > MdeModulePkg/VariableSmm: Remove unused SMI handler functions > > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableRuntimeDxe.inf | 6 + > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableSmm.inf | 6 + > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableSmmRuntimeDxe.inf | 32 +- > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableStandaloneMm.inf | 11 + > MdeModulePkg/Include/Guid/SmmVariableCommon.h | 33 +- > MdeModulePkg/Universal/Variable/RuntimeDxe/Variable.h | 158 +--- > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableNonVolatile.h | 25 + > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableParsing.h | 342 ++++++++ > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableRuntimeCache.h | 47 ++ > MdeModulePkg/Application/VariableInfo/VariableInfo.c | 37 +- > MdeModulePkg/Universal/Variable/RuntimeDxe/Variable.c | 828 ++------------------ > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableExLib.c | 11 +- > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableNonVolatile.c | 28 + > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableParsing.c | 816 +++++++++++++++++++ > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableRuntimeCache.c | 153 ++++ > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableSmm.c | 213 +++-- > MdeModulePkg/Universal/Variable/RuntimeDxe/VariableSmmRuntimeDxe.c | 726 +++++++++++++---- > 17 files changed, 2298 insertions(+), 1174 deletions(-) > create mode 100644 MdeModulePkg/Universal/Variable/RuntimeDxe/VariableNonVolatile.h > create mode 100644 MdeModulePkg/Universal/Variable/RuntimeDxe/VariableParsing.h > create mode 100644 MdeModulePkg/Universal/Variable/RuntimeDxe/VariableRuntimeCache.h > create mode 100644 MdeModulePkg/Universal/Variable/RuntimeDxe/VariableNonVolatile.c > create mode 100644 MdeModulePkg/Universal/Variable/RuntimeDxe/VariableParsing.c > create mode 100644 MdeModulePkg/Universal/Variable/RuntimeDxe/VariableRuntimeCache.c >