On 3/11/21 12:16 PM, David Hildenbrand wrote:
On 11.03.21 18:52, Richard Henderson wrote:
On 3/11/21 10:17 AM, David Hildenbrand wrote:
+#if !defined(CONFIG_USER_ONLY)
+ /*
+ * For !CONFIG_USER_ONLY, we cannot rely on TLB_INVALID_MASK or
haddr==NULL
+ * to detect if there was an exception during tlb_fill().
+ */
+ env->tlb_fill_exc = 0;
+#endif
+ flags = probe_access_flags(env, vaddr1, access_type, mmu_idx,
+ nofault, &haddr1, ra);
+#if !defined(CONFIG_USER_ONLY)
+ if (env->tlb_fill_exc) {
+ return env->tlb_fill_exc;
+ }
+#else
+ if (!haddr1) {
+ env->__excp_addr = vaddr1;
+ return PGM_ADDRESSING;
+ }
+#endif
The assumption of PGM_ADDRESSING is incorrect here -- it could still be
PGM_PROTECTION, depending on how the page is mapped.
Interesting, I was only looking at the s390x tlb_fill() implementation. But I
assume these checks are performed in common code.
Actually, no. It's a common bug in our linux-user targets, where we don't fill
in the SIGSEGV si_code correctly.
See e.g. 8db94ab4e5d.
Thanks, maybe I can factor that out in a nice way. I guess we could do the
access via probe_access_flags() and only on error do the page_get_flags()?
Yes, we could do that. It's certainly better for !nofault, which is the
common-case user of this function.
r~
