From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received-SPF: Pass (sender SPF authorized) identity=mailfrom; client-ip=2607:f8b0:4003:c06::244; helo=mail-oi0-x244.google.com; envelope-from=vladimir.olovyannikov@broadcom.com; receiver=edk2-devel@lists.01.org Received: from mail-oi0-x244.google.com (mail-oi0-x244.google.com [IPv6:2607:f8b0:4003:c06::244]) (using TLSv1.2 with cipher ECDHE-RSA-AES128-GCM-SHA256 (128/128 bits)) (No client certificate requested) by ml01.01.org (Postfix) with ESMTPS id 792072194D3AE for ; Mon, 24 Sep 2018 10:44:57 -0700 (PDT) Received: by mail-oi0-x244.google.com with SMTP id k12-v6so17605501oiw.8 for ; Mon, 24 Sep 2018 10:44:57 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=broadcom.com; s=google; h=mime-version:references:in-reply-to:from:date:message-id:subject:to :cc:content-transfer-encoding; bh=tQeYJaQF2fyu1STlQN409pfg3is1+esfpFxxhcRiVvA=; b=L44G4fKHEArVgSzHXs3ozKVrAfBGmMeBbMq5Xrpppr108ZKtFX00pI0tNzdKV0MK7f qRXoR4lufnHNskAtlT3Ni5oW059GuZEo4KTO1Cr6yqLNfRbc0jDet9WVK9YC4kM0uaCN XLOVb3FOA2oqXYr706iKAUoDBryuGA8xmXm2U= X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:mime-version:references:in-reply-to:from:date :message-id:subject:to:cc:content-transfer-encoding; bh=tQeYJaQF2fyu1STlQN409pfg3is1+esfpFxxhcRiVvA=; b=j4Ygl+3U2ITMWkTHqo4xE4ujk1iWT55tKGZRYoZJgP5IAdRa1AUtXRaN8zzPEjseFe XefdmRk3v7LbyxCgMyi0Ecjt9vhtnHUTnvETfugBmLfY9QbTcvjcwm79GheIaoEHFUCB GpprEapjxNzQZieRkA+JFcoK/NGmII0D/t2JWTymn7hl8MU6VNIAj3znh5uOBTx3Z0Uh I0nviasa7i90h81h7HZvxSKYfFzwCLV/7gsfeFF8qHX9beZO3/Rot4ULG9qrLJtJGlx+ xxCzjYxWCNHjTQC3JKCkA1XnUezIA++jcvdeONlfv8jUFjDB1MApWEzxJ0BJDkm4uNDd 5S7A== X-Gm-Message-State: APzg51AMRgAfCWX3e8uuOHPAm8OufFQ7O0DoXdmf91RF4U1sB3weBBXr 4M+W2cff/WklbtT6O6v4Vt47UaaFr5RklMKZg05hFQ== X-Google-Smtp-Source: ANB0VdYT0WwDUiGlNcqKwge8JzZdQTRdi4ULq/S5kdEzFF363qo5fgefT35cTc1hF9ukJFt7nvXQ9zaDx/3AdKLkd/I= X-Received: by 2002:aca:f244:: with SMTP id q65-v6mr6418067oih.268.1537811096105; Mon, 24 Sep 2018 10:44:56 -0700 (PDT) MIME-Version: 1.0 References: <4F31CED5-DCB4-4834-A0F2-52D410C2F5BF@apple.com> In-Reply-To: From: Vladimir Olovyannikov Date: Mon, 24 Sep 2018 10:44:44 -0700 Message-ID: To: Ard Biesheuvel Cc: Andrew Fish , Bill Paul , "edk2-devel@lists.01.org" Subject: Re: Stack issue after warm UEFI reset and MMU enabling on an Armv8 platform X-BeenThere: edk2-devel@lists.01.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: EDK II Development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Mon, 24 Sep 2018 17:44:57 -0000 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: quoted-printable On Fri, Sep 21, 2018 at 5:14 PM Ard Biesheuvel wrote: > > > > On Fri 21 Sep 2018 at 16:57, Andrew Fish wrote: >> >> >> >> > On Sep 21, 2018, at 4:24 PM, Vladimir Olovyannikov wrote: >> > >> > On Thu, Sep 20, 2018 at 2:52 PM Vladimir Olovyannikov >> > wrote: >> >> >> >> On Wed, Sep 19, 2018 at 5:21 PM Bill Paul wrote= : >> >>> >> >>> Of all the gin joints in all the towns in all the world, Vladimir >> >>> Olovyannikov >> >>> had to walk into mine at 16:58 on Wednesday 19 September 2018 and sa= y: >> >>> >> >>>>> From: Ard Biesheuvel [mailto:ard.biesheuvel@linaro.org] >> >>>>> Sent: Wednesday, September 19, 2018 4:38 PM >> >>>>> To: Vladimir Olovyannikov >> >>>>> Cc: edk2-devel@lists.01.org >> >>>>> Subject: Re: Stack issue after warm UEFI reset and MMU enabling on= an >> >>>>> Armv8 platform >> >>>>> >> >>>>> >> >>>>> On 19 September 2018 at 15:55, Vladimir Olovyannikov >> >>>>> >> >>>>> wrote: >> >>>>>> Hi All, >> >>>>>> >> >>>>>> I need UEFI experts help on the problem with Armv8 board on warm = UEFI >> >>>>>> reset. >> >>>>>> Cold reset works fine. >> >>>>>> >> >>>>>> Here is how I set up a warm reset: >> >>>>>> >> >>>>>> STATIC >> >>>>>> EFI_STATUS >> >>>>>> ShutdownUefiBootServices ( >> >>>>>> >> >>>>>> VOID >> >>>>>> ) >> >>>>>> >> >>>>>> { >> >>>>>> >> >>>>>> EFI_STATUS Status; >> >>>>>> UINTN MemoryMapSize; >> >>>>>> EFI_MEMORY_DESCRIPTOR *MemoryMap; >> >>>>>> UINTN MapKey; >> >>>>>> UINTN DescriptorSize; >> >>>>>> UINT32 DescriptorVersion; >> >>>>>> UINTN Pages; >> >>>>>> >> >>>>>> MemoryMap =3D NULL; >> >>>>>> MemoryMapSize =3D 0; >> >>>>>> Pages =3D 0; >> >>>>>> >> >>>>>> do { >> >>>>>> >> >>>>>> Status =3D gBS->GetMemoryMap ( >> >>>>>> >> >>>>>> &MemoryMapSize, >> >>>>>> MemoryMap, >> >>>>>> &MapKey, >> >>>>>> &DescriptorSize, >> >>>>>> &DescriptorVersion >> >>>>>> ); >> >>>>>> >> >>>>>> if (Status =3D=3D EFI_BUFFER_TOO_SMALL) { >> >>>>>> >> >>>>>> Pages =3D EFI_SIZE_TO_PAGES (MemoryMapSize) + 1; >> >>>>>> MemoryMap =3D AllocatePages (Pages); >> >>>>>> >> >>>>>> // >> >>>>>> // Get System MemoryMap >> >>>>>> // >> >>>>>> Status =3D gBS->GetMemoryMap ( >> >>>>>> >> >>>>>> &MemoryMapSize, >> >>>>>> MemoryMap, >> >>>>>> &MapKey, >> >>>>>> &DescriptorSize, >> >>>>>> &DescriptorVersion >> >>>>>> ); >> >>>>>> >> >>>>>> } >> >>>>>> >> >>>>>> // Don't do anything between the GetMemoryMap() and >> >>>>>> ExitBootServices() if (!EFI_ERROR(Status)) { >> >>>>>> >> >>>>>> Status =3D gBS->ExitBootServices (gImageHandle, MapKey); >> >>>>>> if (EFI_ERROR(Status)) { >> >>>>>> >> >>>>>> FreePages (MemoryMap, Pages); >> >>>>>> MemoryMap =3D NULL; >> >>>>>> MemoryMapSize =3D 0; >> >>>>>> >> >>>>>> } >> >>>>>> >> >>>>>> } >> >>>>>> >> >>>>>> } while (EFI_ERROR(Status)); >> >>>>>> >> >>>>>> return Status; >> >>>>>> >> >>>>>> } >> >>>>>> >> >>>>>> Then perform >> >>>>>> ArmCleanDataCache (); >> >>>>>> ArmInvalidateDataCache (); >> >>>>>> ArmDisableInstructionCache (); >> >>>>>> ArmInvalidateInstructionCache (); >> >>>>> >> >>>>> These don't do anything useful on ARM. You can only reliably perfo= rm >> >>>>> cache >> >>>>> maintenance by virtual address. >> >>>> >> >>>> So, should I just remove them altogether? >> >>>> >> >>>>>> ArmDisableMmu (); >> >>>>> >> >>>>> ... so after this call returns, all bets are off with regards to >> >>>>> whether >> >>>>> what is popped from the stack is actually what we pushed when we >> >>>>> entered >> >>>>> the function. >> >>>> >> >>>> OK, thank you for explanation. >> >>>> But this call returns back into ResetLib implementation as it shoul= d, >> >>>> and >> >>>> then there is a direct jump to the start of FV. >> >>>> Am I doing anything wrong here? >> >>>> Then, up to the point of enabling of MMU the stack is OK. But right >> >>>> after >> >>>> enabling MMU it points at _ModuleEntryPoint end of function in >> >>>> DxeCoreEntryPoint.c >> >>>> Am I missing anything? Maybe some stack cleanup before jumping to t= he >> >>>> start >> >>>> of FV? >> >>> >> >>> When the MMU is enabled, does the mapping for the stack pages change= ? That >> >>> is, >> >>> could the stack now be mapped to different physical page now? >> >> Thanks for ideas Bill, >> >> No, the mapping stays the same. >> >> The issue is only with warm reset, and only on an A72 board. >> >> There is another platform on A53 sharing the same code, which has no = issues >> >> with warm reset. >> >> I cannot explain why. >> >>> >> >>> Instead of showing a stack trace, can you dump the stack pages and c= ompare >> >>> the >> >>> before and after contents? >> >> I can clearly see that before and after contents are different. >> >>> >> >>> Assuming the same physical memory pages are still being used, then t= here >> >>> could >> >>> be a cache flushing problem. What could happen is: >> >>> >> >>> - some stack memory has been touched recently and is now in the data= cache >> >>> - changes are made, which are written to the cache, but not yet flus= hed >> >>> out to >> >>> RAM >> >>> - enabling the MMU causes a full invalidate of the cache >> >>> >> >>> Now when you look at the stack, you see the earlier contents that we= re in >> >>> RAM >> >>> -- the changes previously only written to the cache have been lost. >> >>> >> >>> Enabling/disabling caches and MMU is always tricky. I can't say for = sure, >> >>> but >> >>> I wouldn't be surprised if there's some subtle bug that causes a flu= sh >> >>> operation to be missed and things may just work by coincidence in th= e cold >> >>> start case. >> >> I might be missing something preparing for warm reset. >> >> Disabling interrupts does not help though. >> >> Ard, I switched off all DMA-capable devices, so am just booting into = UEFI >> >> with no disks or network, >> >> disabled interrupts. The issue is here. Any ideas on how to debug it = and >> >> fix? >> > Update: when I add this as an experiment: >> > UINTN StackBottom; >> > >> > __asm__ volatile ("mov %0,SP" : "=3Dr"(StackBottom)); >> > WriteBackInvalidateDataCacheRange((VOID *)StackBottom, >> > PcdGet64(PcdSystemMemorySize) + >> > PcdGet64(PcdSystemMemoryBase) - StackBottom); >> > then the stack data were not corrupted anymore on the next ArmEnableMm= u() call, >> > and warm reset worked (though unreliably, can throw exception on >> > memcpy down the road on UEFI boot; probably because I invalidated only >> > the stack area in the experiment). >> > Considering that A53 board does not have issues, does this means that >> > ArmInvalidateDataCache() implementation is useless for A72? >> > Based on this, should the approach be "find all data regions and >> > invalidate them using InvalidateDataCacheRange()"? >> >> Vladimir, >> >> We hit an issue like this a while back on x86 and it turned out our sequ= ence was dependent on the C compiler code generation not touching a specifi= c register. It might be worth while to disassemble this code and take a loo= k at the assembler. I'd also point out the __asm__ volatile is a serializin= g event to the C memory model so it could change how the C code was optimiz= ed. You could try __asm__ volatile with a no-op instruction to see if it r= eally is the SP read at this point that fixes the issue. But it is likely t= he bug will be more obvious if you look at the assembly. Andrew, Thanks for points. I used that __asm__ directive just as an experiment to see if InvalidateDataCacheRange would make any difference and it did. I could replace it with InvalidateDataCacheRange(DEFAULT_STACK_BOTTOM, 0x260) and it would work the same way. I verified assembly code and do not see any issues. The issue is fixed not by SP read, but by using InvalidateDataCacheRange() API instead of ArmInvalidateDataCache(). Now I can do say 4-10 successful warm resets in the row which eventually crashes on the next warm reset. Need to find the reason for that crash. > > > Is your primary FV hosted in DRAM? Are you sure it has not been corrupted= by the time you attempt your warm reboot? Yes, it is hosted at the start of DRAM from 0x80000000. SystemMemorySize Pcd is declared as 240M from 0x80000000. Another memory is declared by MemoryPeiLib. (1.76G). As I can see all modules are loaded into lower 240M space. DxeCore is at 0x8ceaa000. Other modules are loaded below it. As I told, if I use InvalidateDataCacheRange API for the stack area, I am able to get warm reset to work (for several consecutive times), then eventually there is a crash on warm reset. I will compare contents nevertheless, just to make sure. > > > > >> >> >> >>> >> >>> -Bill >> >>> >> >>>>>> Then jump to start of FV: >> >>>>>> >> >>>>>> typedef >> >>>>>> VOID >> >>>>>> >> >>>>>> (EFIAPI *START_FV)( >> >>>>>> >> >>>>>> VOID >> >>>>>> >> >>>>>> ); >> >>>>>> StartOfFv =3D (START_FV)(UINTN)PcdGet64(PcdFvBaseAddress); >> >>>>>> StartOfFv (); >> >>>>>> >> >>>>>> Now this is what happens on warm reset: >> >>>>>> reset -c warm >> >>>>>> 1. Until ArmEnableMmu() gets called, everything works as expected= . >> >>>>>> >> >>>>>> Here is the stack right before ArmEnableMmu() is called: >> >>>>>> ArmConfigureMmu+0x4f8 >> >>>>>> InitMmu+0x24 >> >>>>>> MemoryPeim+0x440 >> >>>>>> PrePiMain+0x114 >> >>>>>> PrimaryMain+0x68 >> >>>>>> CEntryPoint+0xC4 >> >>>>>> EL2:0x00000000800008BC >> >>>>>> ----- End of stack info ----- >> >>>>>> >> >>>>>> 2. Here is the stack as soon as Mmu is enabled with ArmEnableMmu(= ) : >> >>>>>> ArmConfigureMmu+0x4fc <-- This one is correct, at line 745 in >> >>>>>> >> >>>>>> ArmConfigureMmu() in ArmPkg/Library/ArmMmuLib/AArch64/ArmMmuLibCo= re.c >> >>>>>> (return EFI_SUCCESS) >> >>>>>> >> >>>>>> _ModuleEntryPoint+0x24 <-- Wrong. This points directly to >> >>>>>> >> >>>>>> ASSERT(FALSE); and to CpuDeadLoop() in DxeCoreEntryPoint.c, lines >> >>>>>> 59-60. >> >>>>>> >> >>>>>> El2:0x000000008E5E8300 <-- Absolutely bogus >> >>>>>> >> >>>>>> --- End of stack info --- >> >>>>>> >> >>>>>> So, as soon as ArmEnableMmu() exits, execution jumps directly to >> >>>>>> CpuDeadLoop() in DxeCoreEntryPoint of _ModuleEntryPoint(). >> >>>>>> >> >>>>>> Would be grateful for any advice. >> >>>>>> >> >>>>>> Thank you, >> >>>>>> Vladimir >> >>>> >> >>>> _______________________________________________ >> >>>> edk2-devel mailing list >> >>>> edk2-devel@lists.01.org >> >>>> https://lists.01.org/mailman/listinfo/edk2-devel >> >>> -- >> >>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D >> >>> -Bill Paul (510) 749-2329 | Senior Member of Technical St= aff, >> >>> wpaul@windriver.com | Master of Unix-Fu - Wind River >> >>> Systems >> >>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D >> >>> "I put a dollar in a change machine. Nothing changed." - George Ca= rlin >> >>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D >> > _______________________________________________ >> > edk2-devel mailing list >> > edk2-devel@lists.01.org >> > https://lists.01.org/mailman/listinfo/edk2-devel >> >> _______________________________________________ >> edk2-devel mailing list >> edk2-devel@lists.01.org >> https://lists.01.org/mailman/listinfo/edk2-devel