bsd-user/main.c | 37 +++++++++++++++++++------------------ 1 file changed, 19 insertions(+), 18 deletions(-)
Do not hard-code guest_base at 32GB.
Do not override mmap_next_start for reserved_va.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
Hi Warner,
With the blitz-trial branch you provided, the host libc allocates
thread-local storage within the [32GB, 36GB) region that you currently
assume is free.
The armv7-hello program happens to map on top of this thread-local
storage, and then we crash later accessing some host TLS variable.
While the linux-user probe_guest_base is significantly more complex,
we are also trying to handle 32-bit hosts. I think freebsd is always
assuming 64-bit hosts, which makes this simpler.
r~
---
bsd-user/main.c | 37 +++++++++++++++++++------------------
1 file changed, 19 insertions(+), 18 deletions(-)
diff --git a/bsd-user/main.c b/bsd-user/main.c
index f500ec292b..9760aad9f6 100644
--- a/bsd-user/main.c
+++ b/bsd-user/main.c
@@ -50,20 +50,8 @@
int do_strace;
-/*
- * Going hand in hand with the va space needed (see below), we need
- * to find a host address to map the guest to. Assume that qemu
- * itself doesn't need memory above 32GB (or that we don't collide
- * with anything interesting). This is selected rather arbitrarily,
- * but seems to produce good results in tests to date.
- */
-# if HOST_LONG_BITS >= 64
-uintptr_t guest_base = 0x800000000ul; /* at 32GB */
-bool have_guest_base = true;
-#else
-uintptr_t guest_base; /* TODO: use sysctl to find big enough hole */
+uintptr_t guest_base;
bool have_guest_base;
-#endif
static bool opt_one_insn_per_tb;
static const char *cpu_model;
static const char *cpu_type;
@@ -522,10 +510,6 @@ int main(int argc, char **argv)
target_environ = envlist_to_environ(envlist, NULL);
envlist_free(envlist);
- if (reserved_va) {
- mmap_next_start = reserved_va + 1;
- }
-
{
Error *err = NULL;
if (seed_optarg != NULL) {
@@ -543,7 +527,24 @@ int main(int argc, char **argv)
* Now that page sizes are configured we can do
* proper page alignment for guest_base.
*/
- guest_base = HOST_PAGE_ALIGN(guest_base);
+ if (have_guest_base) {
+ if (guest_base & ~qemu_host_page_mask) {
+ error_report("Selected guest base not host page aligned");
+ exit(1);
+ }
+ } else if (reserved_va) {
+ void *p = mmap(NULL, reserved_va + 1, PROT_NONE, MAP_GUARD, -1, 0);
+ if (p == MAP_FAILED) {
+ const char *err = strerror(errno);
+ char *sz = size_to_str(reserved_va + 1);
+
+ error_report("Cannot allocate %s bytes for guest address space: %s",
+ sz, err);
+ exit(1);
+ }
+ guest_base = (uintptr_t)p;
+ have_guest_base = true;
+ }
if (loader_exec(filename, argv + optind, target_environ, regs, info,
&bprm) != 0) {
--
2.41.0On Thu, Jul 27, 2023 at 10:11 AM Richard Henderson <
richard.henderson@linaro.org> wrote:
> Do not hard-code guest_base at 32GB.
> Do not override mmap_next_start for reserved_va.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>
> Hi Warner,
>
> With the blitz-trial branch you provided, the host libc allocates
> thread-local storage within the [32GB, 36GB) region that you currently
> assume is free.
>
> The armv7-hello program happens to map on top of this thread-local
> storage, and then we crash later accessing some host TLS variable.
>
> While the linux-user probe_guest_base is significantly more complex,
> we are also trying to handle 32-bit hosts. I think freebsd is always
> assuming 64-bit hosts, which makes this simpler.
>
Double mapping makes sense for problems to arise. I'd not have thought
this would be the problem, but it does eliminate a bunch of code that I'd
thought suspect (made worse by my trying to 'fix' old kludges with new
kludges of my own).
Yes. FreeBSD's bsd-user binary will only run on 64-bit hosts. The project
has started phasing out support for 32-bit hosts, and the role of bsd-user
(package builder tool) is such that 32-bit hosts don't make sense.
I've tested this out, and it works for me. Any chance we can get this into
8.1 as a bug fix for the last minute breakage of bsd-user (without this and
another patch, the static hello world that used to work broke). I can send
that second patch out for review. I can queue this fix in the mean time for
whenever the tree opens up.
Reviewed by: Warner Losh <imp@bsdimp.com>
>
> r~
> ---
> bsd-user/main.c | 37 +++++++++++++++++++------------------
> 1 file changed, 19 insertions(+), 18 deletions(-)
>
> diff --git a/bsd-user/main.c b/bsd-user/main.c
> index f500ec292b..9760aad9f6 100644
> --- a/bsd-user/main.c
> +++ b/bsd-user/main.c
> @@ -50,20 +50,8 @@
>
> int do_strace;
>
> -/*
> - * Going hand in hand with the va space needed (see below), we need
> - * to find a host address to map the guest to. Assume that qemu
> - * itself doesn't need memory above 32GB (or that we don't collide
> - * with anything interesting). This is selected rather arbitrarily,
> - * but seems to produce good results in tests to date.
> - */
> -# if HOST_LONG_BITS >= 64
> -uintptr_t guest_base = 0x800000000ul; /* at 32GB */
> -bool have_guest_base = true;
> -#else
> -uintptr_t guest_base; /* TODO: use sysctl to find big enough hole */
> +uintptr_t guest_base;
> bool have_guest_base;
> -#endif
> static bool opt_one_insn_per_tb;
> static const char *cpu_model;
> static const char *cpu_type;
> @@ -522,10 +510,6 @@ int main(int argc, char **argv)
> target_environ = envlist_to_environ(envlist, NULL);
> envlist_free(envlist);
>
> - if (reserved_va) {
> - mmap_next_start = reserved_va + 1;
> - }
> -
> {
> Error *err = NULL;
> if (seed_optarg != NULL) {
> @@ -543,7 +527,24 @@ int main(int argc, char **argv)
> * Now that page sizes are configured we can do
> * proper page alignment for guest_base.
> */
> - guest_base = HOST_PAGE_ALIGN(guest_base);
> + if (have_guest_base) {
> + if (guest_base & ~qemu_host_page_mask) {
> + error_report("Selected guest base not host page aligned");
> + exit(1);
> + }
> + } else if (reserved_va) {
> + void *p = mmap(NULL, reserved_va + 1, PROT_NONE, MAP_GUARD, -1,
> 0);
> + if (p == MAP_FAILED) {
> + const char *err = strerror(errno);
> + char *sz = size_to_str(reserved_va + 1);
> +
> + error_report("Cannot allocate %s bytes for guest address
> space: %s",
> + sz, err);
> + exit(1);
> + }
> + guest_base = (uintptr_t)p;
> + have_guest_base = true;
> + }
>
> if (loader_exec(filename, argv + optind, target_environ, regs, info,
> &bprm) != 0) {
> --
> 2.41.0
>
>
On 7/27/23 17:09, Warner Losh wrote: > Yes. FreeBSD's bsd-user binary will only run on 64-bit hosts. The project > has started phasing out support for 32-bit hosts, and the role of bsd-user > (package builder tool) is such that 32-bit hosts don't make sense. Ok, fine. > I've tested this out, and it works for me. Any chance we can get this into > 8.1 as a bug fix for the last minute breakage of bsd-user (without this and > another patch, the static hello world that used to work broke). I can send > that second patch out for review. I can queue this fix in the mean time for > whenever the tree opens up. Yes. I can queue this to tcg-next, which will have another round before rc2. Send the alignment patch tomorrow and I can include that too. r~
On Thu, Jul 27, 2023 at 6:31 PM Richard Henderson < richard.henderson@linaro.org> wrote: > On 7/27/23 17:09, Warner Losh wrote: > > Yes. FreeBSD's bsd-user binary will only run on 64-bit hosts. The project > > has started phasing out support for 32-bit hosts, and the role of > bsd-user > > (package builder tool) is such that 32-bit hosts don't make sense. > > Ok, fine. > > > I've tested this out, and it works for me. Any chance we can get this > into > > 8.1 as a bug fix for the last minute breakage of bsd-user (without this > and > > another patch, the static hello world that used to work broke). I can > send > > that second patch out for review. I can queue this fix in the mean time > for > > whenever the tree opens up. > > Yes. I can queue this to tcg-next, which will have another round before > rc2. > Send the alignment patch tomorrow and I can include that too. > Great! I've just sent the patch. Thanks! With it, I can now get 'hello world' working again... This should also give my GSoC student a good place to merge against to finish off his mmap and other patches. Warner
© 2016 - 2024 Red Hat, Inc.