summary refs log tree commit diff stats
path: root/target (follow)
Commit message (Collapse)AuthorAgeFilesLines
...
| * | target/arm: Add FEAT_NV to max, neoverse-n2, neoverse-v1 CPUsPeter Maydell2024-01-092-3/+6
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Enable FEAT_NV on the 'max' CPU, and stop filtering it out for the Neoverse N2 and Neoverse V1 CPUs. We continue to downgrade FEAT_NV2 support to FEAT_NV for the latter two CPU types. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Handle FEAT_NV page table attribute changesPeter Maydell2024-01-091-0/+21
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | FEAT_NV requires that when HCR_EL2.{NV,NV1} == {1,1} the handling of some of the page table attribute bits changes for the EL1&0 translation regime: * for block and page descriptors: - bit [54] holds PXN, not UXN - bit [53] is RES0, and the effective value of UXN is 0 - bit [6], AP[1], is treated as 0 * for table descriptors, when hierarchical permissions are enabled: - bit [60] holds PXNTable, not UXNTable - bit [59] is RES0 - bit [61], APTable[0] is treated as 0 Implement these changes to the page table attribute handling. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Treat LDTR* and STTR* as LDR/STR when NV, NV1 is 1, 1Peter Maydell2024-01-091-2/+4
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | FEAT_NV requires (per I_JKLJK) that when HCR_EL2.{NV,NV1} is {1,1} the unprivileged-access instructions LDTR, STTR etc behave as normal loads and stores. Implement the check that handles this. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Don't honour PSTATE.PAN when HCR_EL2.{NV, NV1} == {1, 1}Peter Maydell2024-01-091-0/+3
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | For FEAT_NV, when HCR_EL2.{NV,NV1} is {1,1} PAN is always disabled even when the PSTATE.PAN bit is set. Implement this by having arm_pan_enabled() return false in this situation. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Always use arm_pan_enabled() when checking if PAN is enabledPeter Maydell2024-01-091-11/+11
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Currently the code in target/arm/helper.c mostly checks the PAN bits in env->pstate or env->uncached_cpsr directly when it wants to know if PAN is enabled, because in most callsites we know whether we are in AArch64 or AArch32. We do have an arm_pan_enabled() function, but we only use it in a few places where the code might run in either an AArch32 or AArch64 context. For FEAT_NV, when HCR_EL2.{NV,NV1} is {1,1} PAN is always disabled even when the PSTATE.PAN bit is set, the "is PAN enabled" test becomes more complicated. Make all places that check for PAN use arm_pan_enabled(), so we have a place to put the FEAT_NV test. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Trap registers when HCR_EL2.{NV, NV1} == {1, 1}Peter Maydell2024-01-091-4/+41
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | When HCR_EL2.{NV,NV1} is {1,1} we must trap five extra registers to EL2: VBAR_EL1, ELR_EL1, SPSR_EL1, SCXTNUM_EL1 and TFSR_EL1. Implement these traps. This trap does not apply when FEAT_NV2 is implemented and enabled; include the check that HCR_EL2.NV2 is 0 here, to save us having to come back and add it later. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Set SPSR_EL1.M correctly when nested virt is enabledPeter Maydell2024-01-091-0/+6
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | FEAT_NV requires that when HCR_EL2.{NV,NV1} == {1,0} and an exception is taken from EL1 to EL1 then the reported EL in SPSR_EL1.M should be EL2, not EL1. Implement this behaviour. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Make NV reads of CurrentEL return EL2Peter Maydell2024-01-091-2/+7
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | FEAT_NV requires that when HCR_EL2.NV is set reads of the CurrentEL register from EL1 always report EL2 rather than the real EL. Implement this. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Trap sysreg accesses for FEAT_NVPeter Maydell2024-01-095-10/+77
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | For FEAT_NV, accesses to system registers and instructions from EL1 which would normally UNDEF there but which work in EL2 need to instead be trapped to EL2. Detect this both for "we know this will UNDEF at translate time" and "we found this UNDEFs at runtime", and make the affected registers trap to EL2 instead. The Arm ARM defines the set of registers that should trap in terms of their names; for our implementation this would be both awkward and inefficent as a test, so we instead trap based on the opc1 field of the sysreg. The regularity of the architectural choice of encodings for sysregs means that in practice this captures exactly the correct set of registers. Regardless of how we try to define the registers this trapping applies to, there's going to be a certain possibility of breakage if new architectural features introduce new registers that don't follow the current rules (FEAT_MEC is one example already visible in the released sysreg XML, though not yet in the Arm ARM). This approach seems to me to be straightforward and likely to require a minimum of manual overrides. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Move FPU/SVE/SME access checks up above ARM_CP_SPECIAL_MASK checkPeter Maydell2024-01-091-7/+8
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | In handle_sys() we don't do the check for whether the register is marked as needing an FPU/SVE/SME access check until after we've handled the special cases covered by ARM_CP_SPECIAL_MASK. This is conceptually the wrong way around, because if for example we happen to implement an FPU-access-checked register as ARM_CP_NOP, we should do the access check first. Move the access checks up so they are with all the other access checks, not sandwiched between the special-case read/write handling and the normal-case read/write handling. This doesn't change behaviour at the moment, because we happen not to define any cpregs with both ARM_CPU_{FPU,SVE,SME} and one of the cases dealt with by ARM_CP_SPECIAL_MASK. Moving this code also means we have the correct place to put the FEAT_NV/FEAT_NV2 access handling, which should come after the access checks and before we try to do any read/write action. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Make EL2 cpreg accessfns safe for FEAT_NV EL1 accessesPeter Maydell2024-01-092-8/+69
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | FEAT_NV and FEAT_NV2 will allow EL1 to attempt to access cpregs that only exist at EL2. This means we're going to want to run their accessfns when the CPU is at EL1. In almost all cases, the behaviour we want is "the accessfn returns OK if at EL1". Mostly the accessfn already does the right thing; in a few cases we need to explicitly check that the EL is not 1 before applying various trap controls, or split out an accessfn used both for an _EL1 and an _EL2 register into two so we can handle the FEAT_NV case correctly for the _EL2 register. There are two registers where we want the accessfn to trap for a FEAT_NV EL1 access: VSTTBR_EL2 and VSTCR_EL2 should UNDEF an access from NonSecure EL1, not trap to EL2 under FEAT_NV. The way we have written sel2_access() already results in this behaviour. We can identify the registers we care about here because they all have opc1 == 4 or 5. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: *_EL12 registers should UNDEF when HCR_EL2.E2H is 0Peter Maydell2024-01-092-1/+18
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The alias registers like SCTLR_EL12 only exist when HCR_EL2.E2H is 1; they should UNDEF otherwise. We weren't implementing this. Add an intercept of the accessfn for these aliases, and implement the UNDEF check. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Record correct opcode fields in cpreg for E2H aliasesPeter Maydell2024-01-091-0/+35
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | For FEAT_VHE, we define a set of register aliases, so that for instance: * the SCTLR_EL1 either accesses the real SCTLR_EL1, or (if E2H is 1) SCTLR_EL2 * a new SCTLR_EL12 register accesses SCTLR_EL1 if E2H is 1 However when we create the 'new_reg' cpreg struct for the SCTLR_EL12 register, we duplicate the information in the SCTLR_EL1 cpreg, which means the opcode fields are those of SCTLR_EL1, not SCTLR_EL12. This is a problem for code which looks at the cpreg opcode fields to determine behaviour (e.g. in access_check_cp_reg()). In practice the current checks we do there don't intersect with the *_EL12 registers, but for FEAT_NV this will become a problem. Write the correct values from the encoding into the new_reg struct. This restores the invariant that the cpreg that you get back from the hashtable has opcode fields that match the key you used to retrieve it. When we call the readfn or writefn for the target register, we pass it the cpreg struct for that target register, not the one for the alias, in case the readfn/writefn want to look at the opcode fields to determine behaviour. This means we need to interpose custom read/writefns for the e12 aliases. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Allow use of upper 32 bits of TBFLAG_A64Peter Maydell2024-01-091-3/+5
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The TBFLAG_A64 TB flag bits go in flags2, which for AArch64 guests we know is 64 bits. However at the moment we use FIELD_EX32() and FIELD_DP32() to read and write these bits, which only works for bits 0 to 31. Since we're about to add a flag that uses bit 32, switch to FIELD_EX64() and FIELD_DP64() so that this will work. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Always honour HCR_EL2.TSC when HCR_EL2.NV is setPeter Maydell2024-01-091-3/+13
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The HCR_EL2.TSC trap for trapping EL1 execution of SMC instructions has a behaviour change for FEAT_NV when EL3 is not implemented: * in older architecture versions TSC was required to have no effect (i.e. the SMC insn UNDEFs) * with FEAT_NV, when HCR_EL2.NV == 1 the trap must apply (i.e. SMC traps to EL2, as it already does in all cases when EL3 is implemented) * in newer architecture versions, the behaviour either without FEAT_NV or with FEAT_NV and HCR_EL2.NV == 0 is relaxed to an IMPDEF choice between UNDEF and trap-to-EL2 (i.e. it is permitted to always honour HCR_EL2.TSC) for AArch64 only Add the condition to honour the trap bit when HCR_EL2.NV == 1. We leave the HCR_EL2.NV == 0 case with the existing (UNDEF) behaviour, as our IMPDEF choice (both because it avoids a behaviour change for older CPU models and because we'd have to distinguish AArch32 from AArch64 if we opted to trap to EL2). Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Enable trapping of ERET for FEAT_NVPeter Maydell2024-01-094-7/+16
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | When FEAT_NV is turned on via the HCR_EL2.NV bit, ERET instructions are trapped, with the same syndrome information as for the existing FEAT_FGT fine-grained trap (in the pseudocode this is handled in AArch64.CheckForEretTrap()). Rename the DisasContext and tbflag bits to reflect that they are no longer exclusively for FGT traps, and set the tbflag bit when FEAT_NV is enabled as well as when the FGT is enabled. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Implement HCR_EL2.AT handlingPeter Maydell2024-01-091-6/+15
| | | | | | | | | | | | | | | | | | | | | | | | | | | The FEAT_NV HCR_EL2.AT bit enables trapping of some address translation instructions from EL1 to EL2. Implement this behaviour. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Handle HCR_EL2 accesses for bits introduced with FEAT_NVPeter Maydell2024-01-092-1/+10
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | FEAT_NV defines three new bits in HCR_EL2: NV, NV1 and AT. When the feature is enabled, allow these bits to be written, and flush the TLBs for the bits which affect page table interpretation. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
| * | target/arm: Set CTR_EL0.{IDC,DIC} for the 'max' CPUPeter Maydell2024-01-091-0/+10
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The CTR_EL0 register has some bits which allow the implementation to tell the guest that it does not need to do cache maintenance for data-to-instruction coherence and instruction-to-data coherence. QEMU doesn't emulate caches and so our cache maintenance insns are all NOPs. We already have some models of specific CPUs where we set these bits (e.g. the Neoverse V1), but the 'max' CPU still uses the settings it inherits from Cortex-A57. Set the bits for 'max' as well, so the guest doesn't need to do unnecessary work. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Tested-by: Miguel Luis <miguel.luis@oracle.com>
* | | target/riscv: Ensure mideleg is set correctly on resetAlistair Francis2024-01-101-0/+8
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Bits 10, 6, 2 and 12 of mideleg are read only 1 when the Hypervisor is enabled. We currently only set them on accesses to mideleg, but they aren't correctly set on reset. Let's ensure they are always the correct value. Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1617 Signed-off-by: Alistair Francis <alistair.francis@wdc.com> Reviewed-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Message-ID: <20240108001328.280222-4-alistair.francis@wdc.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: Don't adjust vscause for exceptionsAlistair Francis2024-01-101-2/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We have been incorrectly adjusting both the interrupt and exception cause when using the hypervisor extension and trapping to VS-mode. This patch changes the conditional to ensure we only adjust the cause for interrupts and not exceptions. Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1708 Signed-off-by: Alistair Francis <alistair.francis@wdc.com> Reviewed-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Message-ID: <20240108001328.280222-3-alistair.francis@wdc.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: Assert that the CSR numbers will be correctAlistair Francis2024-01-101-1/+4
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The CSRs will always be between either CSR_MHPMCOUNTER3 and CSR_MHPMCOUNTER31 or CSR_MHPMCOUNTER3H and CSR_MHPMCOUNTER31H. So although ctr_index can't be negative, Coverity doesn't know this and it isn't obvious to human readers either. Let's add an assert to ensure that Coverity knows the values will be within range. To simplify the code let's also change the RV32 adjustment. Fixes: Coverity CID 1523910 Signed-off-by: Alistair Francis <alistair.francis@wdc.com> Reviewed-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Message-ID: <20240108001328.280222-2-alistair.francis@wdc.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: pmp: Ignore writes when RW=01 and MML=0Ivan Klokov2024-01-101-1/+1
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This patch changes behavior on writing RW=01 to pmpcfg with MML=0. RWX filed is form of collective WARL with the combination of pmpcfg.RW=01 remains reserved for future standard use. According to definition of WARL writing the CSR has no other side effect. But current implementation change architectural state and change system behavior. After writing we will get unreadable-unwriteble region regardless on the previous state. On the other side WARL said that we should read legal value and nothing says about what we should write. Current behavior change system state regardless of whether we read this register or not. Fixes: ac66f2f0 ("target/riscv: pmp: Ignore writes when RW=01") Signed-off-by: Ivan Klokov <ivan.klokov@syntacore.com> Reviewed-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231220153205.11072-1-ivan.klokov@syntacore.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/kvm: add RVV and Vector CSR regsDaniel Henrique Barboza2024-01-101-0/+74
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Add support for RVV and Vector CSR KVM regs vstart, vl and vtype. Support for vregs[] requires KVM side changes and an extra reg (vlenb) and will be added later. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218204321.75757-5-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/kvm: do PR_RISCV_V_SET_CONTROL during realize()Daniel Henrique Barboza2024-01-101-0/+29
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Linux RISC-V vector documentation (Document/arch/riscv/vector.rst) mandates a prctl() in order to allow an userspace thread to use the Vector extension from the host. This is something to be done in realize() time, after init(), when we already decided whether we're using RVV or not. We don't have a realize() callback for KVM yet, so add kvm_cpu_realize() and enable RVV for the thread via PR_RISCV_V_SET_CONTROL. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218204321.75757-4-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/kvm.c: remove group setting of KVM AIA if the machine only has ↵Yong-Xuan Wang2024-01-101-14/+17
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | 1 socket The emulated AIA within the Linux kernel restores the HART index of the IMSICs according to the configured AIA settings. During this process, the group setting is used only when the machine partitions harts into groups. It's unnecessary to set the group configuration if the machine has only one socket, as its address space might not contain the group shift. Signed-off-by: Yong-Xuan Wang <yongxuan.wang@sifive.com> Reviewed-by: Jim Shu <jim.shu@sifive.com> Reviewed-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Message-ID: <20231218090543.22353-2-yongxuan.wang@sifive.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add rva22s64 cpuDaniel Henrique Barboza2024-01-102-0/+9
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Add a new profile CPU 'rva22s64' to work as an alias of -cpu rv64i,rva22s64 Like the existing rva22u64 CPU already does with the RVA22U64 profile. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-27-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add RVA22S64 profileDaniel Henrique Barboza2024-01-101-0/+32
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The RVA22S64 profile consists of the following: - all mandatory extensions of RVA22U64; - priv spec v1.12.0; - satp mode sv39; - Ssccptr, a cache related named feature that we're assuming always enable since we don't implement a cache; - Other named features already implemented: Sstvecd, Sstvala, Sscounterenw; - the new Svade named feature that was recently added. Most of the work is already done, so this patch is enough to implement the profile. After this patch, the 'rva22s64' user flag alone can be used with the rva64i CPU to boot Linux: -cpu rv64i,rva22s64=true This is the /proc/cpuinfo with this profile enabled: # cat /proc/cpuinfo processor : 0 hart : 0 isa : rv64imafdc_zicbom_zicbop_zicboz_zicntr_zicsr_zifencei_zihintpause_zihpm_zfhmin_zca_zcd_zba_zbb_zbs_zkt_svinval_svpbmt mmu : sv39 Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-26-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add 'parent' in profile descriptionDaniel Henrique Barboza2024-01-103-1/+15
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Certain S-mode profiles, like RVA22S64 and RVA23S64, mandate all the mandatory extensions of their respective U-mode profiles. RVA22S64 includes all mandatory extensions of RVA22U64, and the same happens with RVA23 profiles. Add a 'parent' field to allow profiles to enable other profiles. This will allow us to describe S-mode profiles by specifying their parent U-mode profile, then adding just the S-mode specific extensions. We're naming the field 'parent' to consider the possibility of other uses (e.g. a s-mode profile including a previous s-mode profile) in the future. Suggested-by: Andrew Jones <ajones@ventanamicro.com> Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-25-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add satp_mode profile supportDaniel Henrique Barboza2024-01-103-0/+42
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | 'satp_mode' is a requirement for supervisor profiles like RVA22S64. User-mode/application profiles like RVA22U64 doesn't care. Add 'satp_mode' to the profile description. If a profile requires it, set it during cpu_set_profile(). We'll also check it during finalize() to validate if the running config implements the profile. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Acked-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-24-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/cpu.c: add riscv_cpu_is_32bit()Daniel Henrique Barboza2024-01-102-1/+7
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Next patch will need to retrieve if a given RISCVCPU is 32 or 64 bit. The existing helper riscv_is_32bit() (hw/riscv/boot.c) will always check the first CPU of a given hart array, not any given CPU. Create a helper to retrieve the info for any given CPU, not the first CPU of the hart array. The helper is using the same 32 bit check that riscv_cpu_satp_mode_finalize() was doing. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-23-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/cpu.c: finalize satp_mode earlierDaniel Henrique Barboza2024-01-101-8/+8
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Profiles will need to validate satp_mode during their own finalize methods. This will occur inside riscv_tcg_cpu_finalize_features() for TCG. Given that satp_mode does not have any pre-req from the accelerator finalize() method, it's safe to finalize it earlier. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-22-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add priv ver restriction to profilesDaniel Henrique Barboza2024-01-103-0/+34
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Some profiles, like RVA22S64, has a priv_spec requirement. Make this requirement explicit for all profiles. We'll validate this requirement finalize() time and, in case the user chooses an incompatible priv_spec while activating a profile, a warning will be shown. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-21-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: implement svadeDaniel Henrique Barboza2024-01-103-0/+7
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | 'svade' is a RVA22S64 profile requirement, a profile we're going to add shortly. It is a named feature (i.e. not a formal extension, not defined in riscv,isa DT at this moment) defined in [1] as: "Page-fault exceptions are raised when a page is accessed when A bit is clear, or written when D bit is clear.". As far as the spec goes, 'svade' is one of the two distinct modes of handling PTE_A and PTE_D. The other way, i.e. update PTE_A/PTE_D when they're cleared, is defined by the 'svadu' extension. Checking cpu_helper.c, get_physical_address(), we can verify that QEMU is compliant with that: we will update PTE_A/PTE_D if 'svadu' is enabled, or throw a page-fault exception if 'svadu' isn't enabled. So, as far as we're concerned, 'svade' translates to 'svadu must be disabled'. We'll implement it like 'zic64b': an internal flag that profiles can enable. The flag will not be exposed to users. [1] https://github.com/riscv/riscv-profiles/blob/main/profiles.adoc Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-20-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add 'rva22u64' CPUDaniel Henrique Barboza2024-01-103-0/+27
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This CPU was suggested by Alistair [1] and others during the profile design discussions. It consists of the bare 'rv64i' CPU with rva22u64 enabled by default, like an alias of '-cpu rv64i,rva22u64=true'. Users now have an even easier way of consuming this user-mode profile by doing '-cpu rva22u64'. Extensions can be enabled/disabled at will on top of it. We can boot Linux with this "user-mode" CPU by doing: -cpu rva22u64,sv39=true,s=true,zifencei=true [1] https://lore.kernel.org/qemu-riscv/CAKmqyKP7xzZ9Sx=-Lbx2Ob0qCfB7Z+JO944FQ2TQ+49mqo0q_Q@mail.gmail.com/ Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-19-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | riscv-qmp-cmds.c: add profile flags in cpu-model-expansionDaniel Henrique Barboza2024-01-101-0/+14
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Expose all profile flags for all CPUs when executing query-cpu-model-expansion. This will allow callers to quickly determine if a certain profile is implemented by a given CPU. This includes vendor CPUs - the fact that they don't have profile user flags doesn't mean that they don't implement the profile. After this change it's possible to quickly determine if our stock CPUs implement the existing rva22u64 profile. Here's a few examples: $ ./build/qemu-system-riscv64 -S -M virt -display none -qmp tcp:localhost:1234,server,wait=off $ ./scripts/qmp/qmp-shell localhost:1234 Welcome to the QMP low-level shell! Connected to QEMU 8.1.50 - As expected, the 'max' CPU implements the rva22u64 profile. (QEMU) query-cpu-model-expansion type=full model={"name":"max"} {"return": {"model": {"name": "rv64", "props": {... "rva22u64": true, ...}}}} - rv64 is missing "zba", "zbb", "zbs", "zkt" and "zfhmin": query-cpu-model-expansion type=full model={"name":"rv64"} {"return": {"model": {"name": "rv64", "props": {... "rva22u64": false, ...}}}} query-cpu-model-expansion type=full model={"name":"rv64", "props":{"zba":true,"zbb":true,"zbs":true,"zkt":true,"zfhmin":true}} {"return": {"model": {"name": "rv64", "props": {... "rva22u64": true, ...}}}} We have no vendor CPUs that supports rva22u64 (veyron-v1 is the closest - it is missing just 'zkt'). In short, aside from the 'max' CPU, we have no CPUs that supports rva22u64 by default. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-18-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: validate profiles during finalizeDaniel Henrique Barboza2024-01-101-0/+69
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Enabling a profile and then disabling some of its mandatory extensions is a valid use. It can be useful for debugging and testing. But the common expected use of enabling a profile is to enable all its mandatory extensions. Add an user warning when mandatory extensions from an enabled profile are disabled in the command line. We're also going to disable the profile flag in this case since the profile must include all the mandatory extensions. This flag can be exposed by QMP to indicate the actual profile state after the CPU is realized. After this patch, this will throw warnings: -cpu rv64,rva22u64=true,zihintpause=false,zicbom=false,zicboz=false qemu-system-riscv64: warning: Profile rva22u64 mandates disabled extension zihintpause qemu-system-riscv64: warning: Profile rva22u64 mandates disabled extension zicbom qemu-system-riscv64: warning: Profile rva22u64 mandates disabled extension zicboz Note that the following will NOT throw warnings because the profile is being enabled last, hence all its mandatory extensions will be enabled: -cpu rv64,zihintpause=false,zicbom=false,zicboz=false,rva22u64=true Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-17-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: honor user choice for G MISA bitsDaniel Henrique Barboza2024-01-101-25/+48
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | RVG behaves like a profile: a single flag enables a set of bits. Right now we're considering user choice when handling RVG and zicsr/zifencei and ignoring user choice on MISA bits. We'll add user warnings for profiles when the user disables its mandatory extensions in the next patch. We'll do the same thing with RVG now to keep consistency between RVG and profile handling. First and foremost, create a new RVG only helper to avoid clogging riscv_cpu_validate_set_extensions(). We do not want to annoy users with RVG warnings like we did in the past (see 9b9741c38f), thus we'll only warn if RVG was user set and the user disabled a RVG extension in the command line. For every RVG MISA bit (IMAFD), zicsr and zifencei, the logic then becomes: - if enabled, do nothing; - if disabled and not user set, enable it; - if disabled and user set, throw a warning that it's a RVG mandatory extension. This same logic will be used for profiles in the next patch. Note that this is a behavior change, where we would error out if the user disabled either zicsr or zifencei. As long as users are explicitly disabling things in the command line we'll let them have a go at it, at least in this step. We'll error out later in the validation if needed. Other notable changes from the previous RVG code: - use riscv_cpu_write_misa_bit() instead of manually updating both env->misa_ext and env->misa_ext_mask; - set zicsr and zifencei directly. We're already checking if they were user set and priv version will never fail for these extensions, making cpu_cfg_ext_auto_update() redundant. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-16-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: add hash table insert helpersDaniel Henrique Barboza2024-01-101-12/+16
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Previous patches added several g_hash_table_insert() patterns. Add two helpers, one for each user hash, to make the code cleaner. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-15-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: handle profile MISA bitsDaniel Henrique Barboza2024-01-101-0/+21
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The profile support is handling multi-letter extensions only. Let's add support for MISA bits as well. We'll go through every known MISA bit. If the profile doesn't declare the bit as mandatory, ignore it. Otherwise, set the bit in env->misa_ext and env->misa_ext_mask. Now that we're setting profile MISA bits, one can use the rv64i CPU to boot Linux using the following options: -cpu rv64i,rva22u64=true,rv39=true,s=true,zifencei=true In the near future, when rva22s64 (where, 's', 'zifencei' and sv39 are mandatory), is implemented, rv64i will be able to boot Linux loading rva22s64 and no additional flags. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: LIU Zhiwei <zhiwei_liu@linux.alibaba.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-14-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: add riscv_cpu_write_misa_bit()Daniel Henrique Barboza2024-01-101-14/+18
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We have two instances of the setting/clearing a MISA bit from env->misa_ext and env->misa_ext_mask pattern. And the next patch will end up adding one more. Create a helper to avoid code repetition. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Reviewed-by: LIU Zhiwei <zhiwei_liu@linux.alibaba.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Message-ID: <20231218125334.37184-13-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: add MISA user options hashDaniel Henrique Barboza2024-01-101-1/+14
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We already track user choice for multi-letter extensions because we needed to honor user choice when enabling/disabling extensions during realize(). We refrained from adding the same mechanism for MISA extensions since we didn't need it. Profile support requires tne need to check for user choice for MISA extensions, so let's add the corresponding hash now. It works like the existing multi-letter hash (multi_ext_user_opts) but tracking MISA bits options in the cpu_set_misa_ext_cfg() callback. Note that we can't re-use the same hash from multi-letter extensions because that hash uses cpu->cfg offsets as keys, while for MISA extensions we're using MISA bits as keys. After adding the user hash in cpu_set_misa_ext_cfg(), setting default values with object_property_set_bool() in add_misa_properties() will end up marking the user choice hash with them. Set the default value manually to avoid it. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Reviewed-by: LIU Zhiwei <zhiwei_liu@linux.alibaba.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Message-ID: <20231218125334.37184-12-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: add user flag for profile supportDaniel Henrique Barboza2024-01-101-0/+80
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The TCG emulation implements all the extensions described in the RVA22U64 profile, both mandatory and optional. The mandatory extensions will be enabled via the profile flag. We'll leave the optional extensions to be enabled by hand. Given that this is the first profile we're implementing in TCG we'll need some ground work first: - all profiles declared in riscv_profiles[] will be exposed to users. TCG is the main accelerator we're considering when adding profile support in QEMU, so for now it's safe to assume that all profiles in riscv_profiles[] will be relevant to TCG; - we'll not support user profile settings for vendor CPUs. The flags will still be exposed but users won't be able to change them; - profile support, albeit available for all non-vendor CPUs, will be based on top of the new 'rv64i' CPU. Setting a profile to 'true' means enable all mandatory extensions of this profile, setting it to 'false' will disable all mandatory profile extensions of the CPU, which will obliterate preset defaults. This is not a problem for a bare CPU like rv64i but it can allow for silly scenarios when using other CPUs. E.g. an user can do "-cpu rv64,rva22u64=false" and have a bunch of default rv64 extensions disabled. The recommended way of using profiles is the rv64i CPU, but users are free to experiment. For now we'll handle multi-letter extensions only. MISA extensions need additional steps that we'll take care later. At this point we can boot a Linux buildroot using rva22u64 using the following options: -cpu rv64i,rva22u64=true,sv39=true,g=true,c=true,s=true Note that being an usermode/application profile we still need to explicitly set 's=true' to enable Supervisor mode to boot Linux. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-11-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/kvm: add 'rva22u64' flag as unavailableDaniel Henrique Barboza2024-01-101-1/+6
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | KVM does not have the means to support enabling the rva22u64 profile. The main reasons are: - we're missing support for some mandatory rva22u64 extensions in the KVM module; - we can't make promises about enabling a profile since it all depends on host support in the end. We'll revisit this decision in the future if needed. For now mark the 'rva22u64' profile as unavailable when running a KVM CPU: $ qemu-system-riscv64 -machine virt,accel=kvm -cpu rv64,rva22u64=true qemu-system-riscv64: can't apply global rv64-riscv-cpu.rva22u64=true: 'rva22u64' is not available with KVM Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Reviewed-by: LIU Zhiwei <zhiwei_liu@linux.alibaba.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Message-ID: <20231218125334.37184-10-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add rva22u64 profile definitionDaniel Henrique Barboza2024-01-102-0/+44
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The rva22U64 profile, described in: https://github.com/riscv/riscv-profiles/blob/main/profiles.adoc#rva22-profiles Contains a set of CPU extensions aimed for 64-bit userspace applications. Enabling this set to be enabled via a single user flag makes it convenient to enable a predictable set of features for the CPU, giving users more predicability when running/testing their workloads. QEMU implements all possible extensions of this profile. All the so called 'synthetic extensions' described in the profile that are cache related are ignored/assumed enabled (Za64rs, Zic64b, Ziccif, Ziccrse, Ziccamoa, Zicclsm) since we do not implement a cache model. An abstraction called RISCVCPUProfile is created to store the profile. 'ext_offsets' contains mandatory extensions that QEMU supports. Same thing with the 'misa_ext' mask. Optional extensions must be enabled manually in the command line if desired. The design here is to use the common target/riscv/cpu.c file to store the profile declaration and export it to the accelerator files. Each accelerator is then responsible to expose it (or not) to users and how to enable the extensions. Next patches will implement the profile for TCG and KVM. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Acked-by: Alistair Francis <alistair.francis@wdc.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Message-ID: <20231218125334.37184-9-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | riscv-qmp-cmds.c: expose named features in cpu_model_expansionDaniel Henrique Barboza2024-01-101-5/+25
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Named features (zic64b the sole example at this moment) aren't expose to users, thus we need another way to expose them. Go through each named feature, get its boolean value, do the needed conversions (bool to qbool, qbool to QObject) and add it to output dict. Another adjustment is needed: named features are evaluated during finalize(), so riscv_cpu_finalize_features() needs to be mandatory regardless of whether we have an input dict or not. Otherwise zic64b will always return 'false', which is incorrect: the default values of cache blocksizes ([cbom/cbop/cboz]_blocksize) are set to 64, satisfying the conditions for zic64b. Here's an API usage example after this patch: $ ./build/qemu-system-riscv64 -S -M virt -display none -qmp tcp:localhost:1234,server,wait=off $ ./scripts/qmp/qmp-shell localhost:1234 Welcome to the QMP low-level shell! Connected to QEMU 8.1.50 (QEMU) query-cpu-model-expansion type=full model={"name":"rv64"} {"return": {"model": {"name": "rv64", "props": {... "zic64b": true, ...}}}} zic64b is set to 'true', as expected, since all cache sizes are 64 bytes by default. If we change one of the cache blocksizes, zic64b is returned as 'false': (QEMU) query-cpu-model-expansion type=full model={"name":"rv64","props":{"cbom_blocksize":128}} {"return": {"model": {"name": "rv64", "props": {... "zic64b": false, ...}}}} Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-8-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: add 'zic64b' supportDaniel Henrique Barboza2024-01-104-0/+34
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | zic64b is defined in the RVA22U64 profile [1] as a named feature for "Cache blocks must be 64 bytes in size, naturally aligned in the address space". It's a fantasy name for 64 bytes cache blocks. The RVA22U64 profile mandates this feature, meaning that applications using this profile expects 64 bytes cache blocks. To make the upcoming RVA22U64 implementation complete, we'll zic64b as a 'named feature', not a regular extension. This means that: - it won't be exposed to users; - it won't be written in riscv,isa. This will be extended to other named extensions in the future, so we're creating some common boilerplate for them as well. zic64b is default to 'true' since we're already using 64 bytes blocks. If any cache block size (cbo{m,p,z}_blocksize) is changed to something different than 64, zic64b is set to 'false'. Our profile implementation will then be able to check the current state of zic64b and take the appropriate action (e.g. throw a warning). [1] https://github.com/riscv/riscv-profiles/releases/download/v1.0/profiles.pdf Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-7-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add zicbop extension flagDaniel Henrique Barboza2024-01-102-0/+5
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | QEMU already implements zicbom (Cache Block Management Operations) and zicboz (Cache Block Zero Operations). Commit 59cb29d6a5 ("target/riscv: add Zicbop cbo.prefetch{i, r, m} placeholder") added placeholders for what would be the instructions for zicbop (Cache Block Prefetch Operations), which are now no-ops. The RVA22U64 profile mandates zicbop, which means that applications that run with this profile might expect zicbop to be present in the riscv,isa DT and might behave badly if it's absent. Adding zicbop as an extension will make our future RVA22U64 implementation more in line with what userspace expects and, if/when cache block prefetch operations became relevant to QEMU, we already have the extension flag to turn then on/off as needed. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-6-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv: add rv64i CPUDaniel Henrique Barboza2024-01-102-0/+48
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We don't have any form of a 'bare bones' CPU. rv64, our default CPUs, comes with a lot of defaults. This is fine for most regular uses but it's not suitable when more control of what is actually loaded in the CPU is required. A bare-bones CPU would be annoying to deal with if not by profile support, a way to load a multitude of extensions with a single flag. Profile support is going to be implemented shortly, so let's add a CPU for it. The new 'rv64i' CPU will have only RVI loaded. It is inspired in the profile specification that dictates, for RVA22U64 [1]: "RVA22U64 Mandatory Base RV64I is the mandatory base ISA for RVA22U64" And so it seems that RV64I is the mandatory base ISA for all profiles listed in [1], making it an ideal CPU to use with profile support. rv64i is a CPU of type TYPE_RISCV_BARE_CPU. It has a mix of features from pre-existent CPUs: - it allows extensions to be enabled, like generic CPUs; - it will not inherit extension defaults, like vendor CPUs. This is the minimum extension set to boot OpenSBI and buildroot using rv64i: ./build/qemu-system-riscv64 -nographic -M virt \ -cpu rv64i,sv39=true,g=true,c=true,s=true,u=true Our minimal riscv,isa in this case will be: # cat /proc/device-tree/cpus/cpu@0/riscv,isa rv64imafdc_zicntr_zicsr_zifencei_zihpm_zca_zcd# [1] https://github.com/riscv/riscv-profiles/blob/main/profiles.adoc Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-5-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
* | | target/riscv/tcg: update priv_ver on user_set extensionsDaniel Henrique Barboza2024-01-101-0/+32
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We'll add a new bare CPU type that won't have any default priv_ver. This means that the CPU will default to priv_ver = 0, i.e. 1.10.0. At the same we'll allow these CPUs to enable extensions at will, but then, if the extension has a priv_ver newer than 1.10, we'll end up disabling it. Users will then need to manually set priv_ver to something other than 1.10 to enable the extensions they want, which is not ideal. Change the setter() of extensions to allow user enabled extensions to bump the priv_ver of the CPU. This will make it convenient for users to enable extensions for CPUs that doesn't set a default priv_ver. This change does not affect any existing CPU: vendor CPUs does not allow extensions to be enabled, and generic CPUs are already set to priv_ver LATEST. Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com> Reviewed-by: Andrew Jones <ajones@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.francis@wdc.com> Message-ID: <20231218125334.37184-4-dbarboza@ventanamicro.com> Signed-off-by: Alistair Francis <alistair.francis@wdc.com>