Skip to contentSkip to conversation

target/s390x: Check storage keys in the TPROT instruction

RFC PATCH In review Thomas Huth 2 participants 3 messages May 2, 2022

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 target/s390x: Check storage keys in the TPROT instruction

TPROT allows to specify an access key that should be used for checking with the storage key of the destination page, to see whether an access is allowed or not. Honor this storage key checking now in the emulated TPROT instruction, too.

Since we need the absolute address of the page (for getting the storage key), we are now also calling mmu_translate() directly instead of going via s390_cpu_virt_mem_check_write/read() - since we're only interested in one page, and since mmu_translate() does not try to inject excetions directly (it reports them via the return code instead), this is likely the better function to use in TPROT anyway.

---
This fixes the new TPROT-related storage key checks in this new kvm-unit-tests patch:
https://lore.kernel.org/kvm/[email protected]/

 target/s390x/cpu.h            |  1 +
 target/s390x/tcg/mem_helper.c | 61 ++++++++++++++++++++++++++++-------
 2 files changed, 50 insertions(+), 12 deletions(-)
target/s390x/cpu.h +1 -0
index 7d6d01325b..348950239f 100644
@@ -328,6 +328,7 @@ extern const VMStateDescription vmstate_s390_cpu;
 /* Control register 0 bits */
 #define CR0_LOWPROT             0x0000000010000000ULL
 #define CR0_SECONDARY           0x0000000004000000ULL
+#define CR0_STOR_PROT_OVERRIDE  0x0000000001000000ULL
 #define CR0_EDAT                0x0000000000800000ULL
 #define CR0_AFP                 0x0000000000040000ULL
 #define CR0_VECTOR              0x0000000000020000ULL
target/s390x/tcg/mem_helper.c +49 -12

92 lines not on this page · raw message

Signature · 1 line
2.27.0
#1 lore raw
Quoted 12 lines · Thomas Huth

TPROT allows to specify an access key that should be used for checking with the storage key of the destination page, to see whether an access is allowed or not. Honor this storage key checking now in the emulated TPROT instruction, too.

Since we need the absolute address of the page (for getting the storage key), we are now also calling mmu_translate() directly instead of going via s390_cpu_virt_mem_check_write/read() - since we're only interested in one page, and since mmu_translate() does not try to inject excetions directly (it reports them via the return code instead), this is likely the better function to use in TPROT anyway.

Quoted 5 lines

---
This fixes the new TPROT-related storage key checks in this new kvm-unit-tests patch:
https://lore.kernel.org/kvm/[email protected]/

Thanks for having a go at this.
The key checking logic looks good to me; the expressions get a bit unwieldy, but that is a style thing.
However, I'm wondering whether it would be better to mirror what the kernel is doing and address the

* TODO: key-controlled protection. Only CPU accesses make use of the
* PSW key. CSS accesses are different - we have to pass in the key.

in mmu_handle_skey, then tprot emulation would just relay the result of trying the translation with store or fetch, passing in the key.

Quoted 25 lines
 target/s390x/cpu.h            |  1 +
 target/s390x/tcg/mem_helper.c | 61 ++++++++++++++++++++++++++++-------
 2 files changed, 50 insertions(+), 12 deletions(-)
target/s390x/cpu.h +1 -0
index 7d6d01325b..348950239f 100644
@@ -328,6 +328,7 @@ extern const VMStateDescription vmstate_s390_cpu;
 /* Control register 0 bits */
 #define CR0_LOWPROT             0x0000000010000000ULL
 #define CR0_SECONDARY           0x0000000004000000ULL
+#define CR0_STOR_PROT_OVERRIDE  0x0000000001000000ULL
 #define CR0_EDAT                0x0000000000800000ULL
 #define CR0_AFP                 0x0000000000040000ULL
 #define CR0_VECTOR              0x0000000000020000ULL
target/s390x/tcg/mem_helper.c +0 -0
index fc52aa128b..1e0309bbe8 100644
@@ -2141,43 +2141,80 @@ uint32_t HELPER(testblock)(CPUS390XState *env, uint64_t real_addr)
     return 0;
 }
 

[...]

Quoted 10 lines
target/s390x/tcg/mem_helper.c +6 -1
+
 uint32_t HELPER(tprot)(CPUS390XState *env, uint64_t a1, uint64_t a2)
 {
     S390CPU *cpu = env_archcpu(env);
-    CPUState *cs = env_cpu(env);
+    const int tp_acc = (a2 & SK_ACC_MASK) >> 4;
+    uint8_t skey;
+    int acc, pgm_code, pflags;
+    target_ulong abs_addr;
+    uint64_t asc = cpu->env.psw.mask & PSW_MASK_ASC;

target/s390x/tcg/mem_helper.c uint32_t HELPER(testblock)(CPUS390XState *env, uint64_t real_addr) · line 2176

Quoted 10 lines
target/s390x/tcg/mem_helper.c +3 -3
+    uint64_t tec;
 
     /*
      * TODO: we currently don't handle all access protection types
-     * (including access-list and key-controlled) as well as AR mode.
+     * (including access-list) as well as AR mode.
      */
-    if (!s390_cpu_virt_mem_check_write(cpu, a1, 0, 1)) {
-        /* Fetching permitted; storing permitted */
+    pgm_code = mmu_translate(env, a1, true, asc, &abs_addr, &pflags, &tec);

I don't like the use of true to indicate a store here, when values other than 0 and 1 are possible. Any reason not to use MMU_DATA_STORE?

A comment about fetch protection override might be nice here:
/*
* Since fetch protection override may apply to half of page 0 only,
* it need not be considered in the following.
*/

Quoted 52 lines
target/s390x/tcg/mem_helper.c +23 -8
+    if (!pgm_code) {
+        /* Fetching permitted; storing permitted - but still check skeys */
+        skey = get_skey(abs_addr);
+        acc = (skey & SK_ACC_MASK) >> 4;
+        if (tp_acc != 0 && tp_acc != acc &&
+            !((env->cregs[0] & CR0_STOR_PROT_OVERRIDE) && acc == 9)) {
+            if (skey & SK_F) {
+                return 2;
+            } else {
+                return 1;
+            }
+        }
         return 0;
     }
 
-    if (env->int_pgm_code == PGM_PROTECTION) {
+    if (pgm_code == PGM_PROTECTION) {
         /* retry if reading is possible */
-        cs->exception_index = -1;
-        if (!s390_cpu_virt_mem_check_read(cpu, a1, 0, 1)) {
+        pgm_code = mmu_translate(env, a1, false, asc, &abs_addr, &pflags, &tec);
+        if (!pgm_code) {
             /* Fetching permitted; storing not permitted */
+            skey = get_skey(abs_addr);
+            acc = (skey & SK_ACC_MASK) >> 4;
+            if ((skey & SK_F) && tp_acc != 0 && tp_acc != acc &&
+                !((env->cregs[0] & CR0_STOR_PROT_OVERRIDE) && acc == 9)) {
+                return 2;
+            }
             return 1;
         }
     }
 
-    switch (env->int_pgm_code) {
+    switch (pgm_code) {
     case PGM_PROTECTION:
         /* Fetching not permitted; storing not permitted */
-        cs->exception_index = -1;
         return 2;
     case PGM_ADDRESSING:
     case PGM_TRANS_SPEC:
         /* exceptions forwarded to the guest */
-        s390_cpu_virt_mem_handle_exc(cpu, GETPC());
-        return 0;
+        tcg_s390_program_interrupt(env, pgm_code, GETPC());
     }
 
     /* Translation not available */
-    cs->exception_index = -1;
     return 3;
 }
#2 lore raw
Quoted 49 lines · Janis Schoetterl-Glausch
Quoted 12 lines · Thomas Huth

TPROT allows to specify an access key that should be used for checking with the storage key of the destination page, to see whether an access is allowed or not. Honor this storage key checking now in the emulated TPROT instruction, too.

Since we need the absolute address of the page (for getting the storage key), we are now also calling mmu_translate() directly instead of going via s390_cpu_virt_mem_check_write/read() - since we're only interested in one page, and since mmu_translate() does not try to inject excetions directly (it reports them via the return code instead), this is likely the better function to use in TPROT anyway.

Quoted 5 lines

---
This fixes the new TPROT-related storage key checks in this new kvm-unit-tests patch:
https://lore.kernel.org/kvm/[email protected]/

Thanks for having a go at this.
The key checking logic looks good to me; the expressions get a bit unwieldy, but that is a style thing.
However, I'm wondering whether it would be better to mirror what the kernel is doing and address the

* TODO: key-controlled protection. Only CPU accesses make use of the
* PSW key. CSS accesses are different - we have to pass in the key.

in mmu_handle_skey, then tprot emulation would just relay the result of trying the translation with store or fetch, passing in the key.

Quoted 19 lines
 target/s390x/cpu.h            |  1 +
 target/s390x/tcg/mem_helper.c | 61 ++++++++++++++++++++++++++++-------
 2 files changed, 50 insertions(+), 12 deletions(-)
target/s390x/cpu.h +1 -0
index 7d6d01325b..348950239f 100644
@@ -328,6 +328,7 @@ extern const VMStateDescription vmstate_s390_cpu;
 /* Control register 0 bits */
 #define CR0_LOWPROT             0x0000000010000000ULL
 #define CR0_SECONDARY           0x0000000004000000ULL
+#define CR0_STOR_PROT_OVERRIDE  0x0000000001000000ULL
 #define CR0_EDAT                0x0000000000800000ULL
 #define CR0_AFP                 0x0000000000040000ULL
 #define CR0_VECTOR              0x0000000000020000ULL
target/s390x/tcg/mem_helper.c +0 -0
index fc52aa128b..1e0309bbe8 100644
Quoted 29 lines
Quoted 6 lines
target/s390x/tcg/mem_helper.c +0 -0
@@ -2141,43 +2141,80 @@ uint32_t HELPER(testblock)(CPUS390XState *env, uint64_t real_addr)
     return 0;
 }
 

[...]

Quoted 20 lines
target/s390x/tcg/mem_helper.c +9 -4
+
 uint32_t HELPER(tprot)(CPUS390XState *env, uint64_t a1, uint64_t a2)
 {
     S390CPU *cpu = env_archcpu(env);
-    CPUState *cs = env_cpu(env);
+    const int tp_acc = (a2 & SK_ACC_MASK) >> 4;
+    uint8_t skey;
+    int acc, pgm_code, pflags;
+    target_ulong abs_addr;
+    uint64_t asc = cpu->env.psw.mask & PSW_MASK_ASC;
+    uint64_t tec;
 
     /*
      * TODO: we currently don't handle all access protection types
-     * (including access-list and key-controlled) as well as AR mode.
+     * (including access-list) as well as AR mode.
      */
-    if (!s390_cpu_virt_mem_check_write(cpu, a1, 0, 1)) {
-        /* Fetching permitted; storing permitted */
+    pgm_code = mmu_translate(env, a1, true, asc, &abs_addr, &pflags, &tec);

mmu_translate/mmu_handle_skey sets the change bit for stores, whereas TPROT specifies that it doesn't.
Not sure what the best way to handle this is.
Additional pretend fetch/store access modes?

Quoted 9 lines

I don't like the use of true to indicate a store here, when values other than 0 and 1 are possible. Any reason not to use MMU_DATA_STORE?

A comment about fetch protection override might be nice here:
/*
* Since fetch protection override may apply to half of page 0 only,
* it need not be considered in the following.
*/

Disregard that, it's not true, TPROT does honor fetch-protection override, I just made a mistake while adding a test for it to the kvm-unit-test.

[...]