Le 20/09/2022 à 04:32, Benjamin Gray a écrit : > On Mon, 2022-09-19 at 07:16 +0000, Christophe Leroy wrote: >> Why would it be unpredictable ? Only one page is mapped. If it >> crosses >> the boundary, __put_kernel_nofault() will fail in a controled manner. >> I see no point in doing the check before every write. > > Oh I didn't see that get_vm_area automatically adds a guard. You're > right then, it's redundant. I had assumed there could be a writeable > mapping directly after. > >> And while you are thinking about alignment, don't forget that dcbst >> and >> icbi apply on a give cacheline. If your memory crosses a cacheline >> you >> may have a problem. > > Yeah, though this applies to the existing patch_instruction too (in > theory; prefixed instructions cannot cross a 64 byte boundary, but the > ISA does not specify minimum cacheline size). I don't have a nice > solution right now though. The flush needs to be done on the effective > address (i.e. text poke address) according to the ISA, but the text > poke address is only valid within the IRQ save region. So non-prefixed > instruction patching would either pay for some kind of check, or need > special casing.
As far as I know, cachelines are minimum 64 bytes on PPC64 aren't they ? > > Maybe an "is aligned" flag in a generic __patch_text to make the extra > flush conditional is good enough? Well, if the cacheline is already flushed, the operation will be a nop anyway so it shouldn't cost much to flush both addresses, after all it is the sync that costs and you'll still have only one. > > > Related to EA based flushing, data patching ought to run 'dcbst' on the > 'exec_addr' too. So given the icache flush only needs to be applied to > instruction patching, and data flush only to data patching, I plan to > move those exec_addr syncs outside of __patch_text to the relevant > public instruction/data specific entry points. Why should it run 'dcbst' on the 'exec_addr' at all ? We have not written anything there. Anyway, powerpc handles cachelines by physical address, so no matter which EA you use as far as it is done. And dcbst is handled as a write by the MMU, so you just can't apply it on the read-only exec address. > >>>>> + case 8: >>>>> + __put_kernel_nofault(dest, src, u64, >>>>> failed); >>>>> + break; >>>> >>>> Is case 8 needed for PPC32 ? >>> >>> I don't have a particular need for it, but the underlying >>> __put_kernel_nofault is capable of it so I included it. >> >> Well, not including it will allow you to pass the source as a 'long' >> as >> mentionned above. > > I checked the machine code of a 32 bit build, but it still passes the > pointer in a register. I also checked all 3 ABI docs and they say a > pointer is the same size as a long. Could you clarify when a pointer is > passed on the stack but not a long? > > Or do you mean that we could pass the pointed-to data in a register and > skip the pointer altogether? That seems like a good choice, but > __put_kernel_nofault takes a pointer to the source and the > implementation is very complex. I don't think I can safely write the > modified version we'd need at this point. Yes I meant to pass the value instead of passing a pointer to the value. When you pass a pointer to a value, it forces gcc to put that value in memory, namely in the stack. While when you pass the value directly , then it gets passed in a register. So I think you have to pass the value and only change it to a pointer to that value at the time you are calling __put_kernel_nofault(). That way gcc is able to handle it efficiently and most of the time voids going via memory. Today raw_patch_instruction() is : c0017ebc <raw_patch_instruction>: c0017ebc: 90 83 00 00 stw r4,0(r3) c0017ec0: 7c 00 18 6c dcbst 0,r3 c0017ec4: 7c 00 04 ac hwsync c0017ec8: 7c 00 1f ac icbi 0,r3 c0017ecc: 7c 00 04 ac hwsync c0017ed0: 4c 00 01 2c isync c0017ed4: 38 60 00 00 li r3,0 c0017ed8: 4e 80 00 20 blr c0017edc: 38 60 ff ff li r3,-1 c0017ee0: 4e 80 00 20 blr Here r4 is the value to be written. With your patch, extract from __patch_text() is: c0017fdc: 28 05 00 04 cmplwi r5,4 c0017fe0: 41 82 00 74 beq c0018054 <__patch_text+0x98> c0017fe4: 41 81 00 40 bgt c0018024 <__patch_text+0x68> c0017fe8: 28 05 00 01 cmplwi r5,1 c0017fec: 41 82 00 74 beq c0018060 <__patch_text+0xa4> c0017ff0: 28 05 00 02 cmplwi r5,2 c0017ff4: 40 82 00 0c bne c0018000 <__patch_text+0x44> c0017ff8: a1 24 00 00 lhz r9,0(r4) c0017ffc: b1 23 00 00 sth r9,0(r3) c0018000: 7c 00 18 6c dcbst 0,r3 c0018004: 7c 00 04 ac hwsync c0018008: 2c 06 00 00 cmpwi r6,0 c001800c: 38 60 00 00 li r3,0 c0018010: 4d 82 00 20 beqlr c0018014: 7c 00 3f ac icbi 0,r7 c0018018: 7c 00 04 ac hwsync c001801c: 4c 00 01 2c isync c0018020: 4e 80 00 20 blr c0018024: 28 05 00 08 cmplwi r5,8 c0018028: 40 a2 ff d8 bne c0018000 <__patch_text+0x44> c001802c: 81 44 00 00 lwz r10,0(r4) c0018030: 81 64 00 04 lwz r11,4(r4) c0018034: 91 43 00 00 stw r10,0(r3) c0018038: 91 63 00 04 stw r11,4(r3) c001803c: 7c 00 18 6c dcbst 0,r3 c0018040: 7c 00 04 ac hwsync c0018044: 2c 06 00 00 cmpwi r6,0 c0018048: 38 60 00 00 li r3,0 c001804c: 4d 82 00 20 beqlr c0018050: 4b ff ff c4 b c0018014 <__patch_text+0x58> c0018054: 81 24 00 00 lwz r9,0(r4) c0018058: 91 23 00 00 stw r9,0(r3) c001805c: 4b ff ff a4 b c0018000 <__patch_text+0x44> c0018060: 89 24 00 00 lbz r9,0(r4) c0018064: 99 23 00 00 stb r9,0(r3) c0018068: 4b ff ff 98 b c0018000 <__patch_text+0x44> c001806c: 38 60 ff ff li r3,-1 c0018070: 4e 80 00 20 blr c0018074: 38 60 ff f2 li r3,-14 c0018078: 4e 80 00 20 blr So as you can see, r4 now is a memory pointer and the data has to be loaded from there. Christophe
