On 09/08/2022 17:53, Alexandru Elisei wrote:
Hi, On Tue, Aug 09, 2022 at 04:53:18PM +0100, Nikos Nikoleris wrote:On 09/08/2022 15:22, Alexandru Elisei wrote:On Tue, Aug 09, 2022 at 02:53:34PM +0100, Nikos Nikoleris wrote:Hi Alex, On 09/08/2022 10:15, Alexandru Elisei wrote:asm_mmu_disable is overly ambitious and provably incorrect: 1. It tries to clean and invalidate the data caches for the *entire* memory, which is highly unnecessary, as it's very unlikely that a test will write to the entire memory, and even more unlikely that a test will modify the text section of the test image.While it appears that we don't modify the text section, there is some loading happening before we start executing a test. Are you sure that the loader doesn't leave the memory dirty?Yes, it's in the boot protocol for Linux [1]. I also mentioned this in the commit message for the previous patch. [1] https://elixir.bootlin.com/linux/v5.19/source/Documentation/arm64/booting.rst#L180I see, thanks! Right now {asm_,}mmu_disable() is not used anywhere. So this patch will introduce the assumption that mmu_disable() can be safely called only if we didn't perform any writes, outside the test's stack, doesn't it?This patch introduces the assumption that the code that disables the MMU will do the necessary cache maintenance. I should reword the commit message to make it clearer.When we add support for EFI, there is a lot happening from efi_main() until we get to the point where we can mmu_disable(), cleaning just the (new) stack of the test seems risky.Well, that's an understatement, the code disabling the MMU definitly needs to do the necessary cache maintenance! asm_mmu_disable() is not a silver bullet that removes the need to do any cache maintenace, the previous patch explains what needs to be done and why. If you're looking for inspiration about what maintenance to be done for UEFI, I suggest you look there. Or even better, you can reuse that code, which I think is the better approach for the UEFI series going forward, but that's a discussion for the UEFI thread.
I would argue that asm_mmu_disable() was "a silver bullet" and your patch is changing it. How do you choose what is reasonable for asm_mmu_disable to clean? Why should it clean the stack?
2. There is no corresponding dcache invalidate command for the entire memory in asm_mmu_enable, leaving it up to the test that disabled the MMU to do the cache maintenance in an asymmetrical fashion: only for re-enabling the MMU, but not for disabling it. 3. It's missing the DMB SY memory barrier to ensure that the dcache maintenance is performed after the last store executed in program order before calling asm_mmu_disable.I am not sure why this is needed. In general, iiuc, a store to location x followed by a DC CVAC to x in program order don't need an barrier (see Arm ARM ARM DDI 0487G.b "Data cache maintenance instructions" at K11.5.1 andJust a note, the latest public version is H.a. K11.5.1 looks to me like it deals with ordering of the cache maintenance operations with regards to memory accesses that are *after* the CMO in program order, this patch is about memory accesses that are *before* the CMO in program order.The AArch64 example in K11.5.1 has a memory instruction before and after the CMO: STR W5, [X1] DC CVAC, X1 DMB ISH STR W0, [X4] The first store and the DC CVAC access the same cache line and there is no need for a memory barrier in between. The second store is assumed to be to a different location and that's why we need a barrier to order it with respect to the DC CVAC.It's explained why the DMB is not necessary in the section that you've referenced. I'll reproduce the paragraph: "All data cache instructions, other than DC ZVA, that specify an address: Execute in program order relative to loads or stores that have all of the following properties: —Access an address in Normal memory with either Inner Write Through or Inner Write Back attributes within the same cache line of minimum size, as indicated by CTR_EL0.DMinLine. —Use an address with the same cacheability attributes as the address passed to the data cache instruction." Both the store and the dcache clean access the same cache line, indexed by the adress in register X1. Does that make sense to you?"Ordering and completion of data and instruction cache instructions" at D4-2656). It doesn't hurt to have it but I think it's unnecessary.D4-2656 is about PAC, I assume you meant D4-2636 judging from the section name (please correct me if I'm wrong): > "All data cache instructions, other than DC ZVA, that specify an address: [..] Can execute in any order relative to loads or stores that access any address with the Device memory attribute, or with Normal memory with Inner Non-cacheable attribute unless a DMB or DSB is executed between the instructions." Since the maintenance is performed with the MMU off, I think the DMB SY is required as per the architecture. I prefer to keep the maintenance after the MMU is disabled, to allow for any kind of translation table setups that a test might conjure up (a test in theory can create and install its own translation tables).Right, so between the stores and the DC CVAC, we've switched the MMU off, in which case the DMB SY might be necessary. I was missing this part.^^^^^^^^^^^^^^^^^^^ might be necessary or might be **unnecessary**? I would say that it's definitely unecessary according to the architecture, not "might be".
Well you had successfully convinced me that since we're switching the MMU off, there needs to be a barrier to ensure that the dc cvac is ordered with respect to prior stores. Switching the MMU off means that stores could be executed with different memory attributes (e.g., Normal, Inner-Shareable, Writeback) than the DC CVAC (Device-nGnRnE). Some type of barrier might be needed. This is what your patch is doing.
The benefits of this design choice (switch the MMU off then clean data) are still unclear to me. This patch is modifying the CMO operation to perform only a clean. Why can't we clean the data cache before we switch off the MMU and use the same translation we used to write to it.What do you mean by "translation"? Same VA to PA mapping? Or same address attributes? If it's the latter, the architecture is pretty clear that this is correct and expected.
Both.
If it's the VA to PA mapping, asm_mmu_disable is called with an identify mapped stack, otherwise following the ret at the end of the function, asm_mmu_disable would not return to the calling function (when the MMU is disabled, all addresses are flat mapped, and x30/lr will point to some bogus address). mmu_disable() even has an assert to check that the stack is identify mapped. So I really do think that the order of the operations is correct. Unless you can prove otherwise. Why is it so important to you that the dcache is cleaned with the MMU on? It's correct either way, so I'm interested to know why you are so keen on doing it with the MMU enabled. I've already told you my reason for doing it with the MMU disabled, I'm waiting to hear yours.
Sorry, maybe I am missing something. As far as I remember, your argument was that invalidating the cache before switching the MMU off was pointless, for Normal Memory any kind of speculation might result in fetching data to the cache. I agree. But this patch changes the CMO we use and it doesn't invalidate any more. What was the argument for cleaning the cache after switching the MMU off?
I am happy either way, I am just trying to understand :) Thanks, Nikos
Thanks, AlexThanks, NikosThanks, AlexThanks, NikosFix all of the issues in one go, by doing the cache maintenance only for the stack, as that is out of the control of the C code, and add the missing memory barrier. The code used to test that mmu_disable works correctly is similar to the code used to test commit 410b3bf09e76 ("arm/arm64: Perform dcache clean + invalidate after turning MMU off"), with extra cache maintenance added: +#include <alloc_page.h> +#include <asm/cacheflush.h> +#include <asm/mmu.h> int main(int argc, char **argv) { + int *x = alloc_page(); + bool pass = true; + int i; + + for (i = 0; i < 1000000; i++) { + *x = 0x42; + dcache_clean_addr_poc((unsigned long)x); + mmu_disable(); + if (*x != 0x42) { + pass = false; + break; + } + *x = 0x50; + /* Needed for the invalidation only. */ + dcache_clean_inval_addr_poc((unsigned long)x); + mmu_enable(current_thread_info()->pgtable); + if (*x != 0x50) { + pass = false; + break; + } + } + report(pass, "MMU disable cache maintenance"); Signed-off-by: Alexandru Elisei <alexandru.elisei@xxxxxxx> --- arm/cstart.S | 11 ++++++----- arm/cstart64.S | 11 +++++------ 2 files changed, 11 insertions(+), 11 deletions(-) diff --git a/arm/cstart.S b/arm/cstart.S index fc7c558802f1..b27de44f30a6 100644 --- a/arm/cstart.S +++ b/arm/cstart.S @@ -242,11 +242,12 @@ asm_mmu_disable: mcr p15, 0, r0, c1, c0, 0 isb - ldr r0, =__phys_offset - ldr r0, [r0] - ldr r1, =__phys_end - ldr r1, [r1] - dcache_by_line_op dccimvac, sy, r0, r1, r2, r3 + dmb sy + mov r0, sp + lsr r0, #THREAD_SHIFT + lsl r0, #THREAD_SHIFT + add r1, r0, #THREAD_SIZE + dcache_by_line_op dccmvac, sy, r0, r1, r3, r4 mov pc, lr diff --git a/arm/cstart64.S b/arm/cstart64.S index 1ce6b9e14d23..af4970775298 100644 --- a/arm/cstart64.S +++ b/arm/cstart64.S @@ -283,12 +283,11 @@ asm_mmu_disable: msr sctlr_el1, x0 isb - /* Clean + invalidate the entire memory */ - adrp x0, __phys_offset - ldr x0, [x0, :lo12:__phys_offset] - adrp x1, __phys_end - ldr x1, [x1, :lo12:__phys_end] - dcache_by_line_op civac, sy, x0, x1, x2, x3 + dmb sy + mov x9, sp + and x9, x9, #THREAD_MASK + add x10, x9, #THREAD_SIZE + dcache_by_line_op cvac, sy, x9, x10, x11, x12 ret
