]> www.infradead.org Git - users/hch/block.git/commitdiff
um: fix time-travel syscall scheduling hack
authorJohannes Berg <johannes.berg@intel.com>
Tue, 27 Aug 2024 14:05:01 +0000 (16:05 +0200)
committerRichard Weinberger <richard@nod.at>
Thu, 12 Sep 2024 18:46:23 +0000 (20:46 +0200)
The schedule() call there really never did anything at
least since the introduction of the EEVDF scheduler,
but now I found a case where we permanently hang in a
loop of -ERESTARTNOINTR (due to locking.) Work around
it by making any syscalls with error return take time
(and then schedule after) so we cannot hang in such a
loop forever.

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: Richard Weinberger <richard@nod.at>
arch/um/kernel/skas/syscall.c

index 9ee19e566da3bdcf073c45edfb004906c42f6cc4..b09e85279d2b8c84487a1d7ea35bfdbf16943024 100644 (file)
 #include <sysdep/syscalls.h>
 #include <linux/time-internal.h>
 #include <asm/unistd.h>
+#include <asm/delay.h>
 
 void handle_syscall(struct uml_pt_regs *r)
 {
        struct pt_regs *regs = container_of(r, struct pt_regs, regs);
        int syscall;
 
-       /*
-        * If we have infinite CPU resources, then make every syscall also a
-        * preemption point, since we don't have any other preemption in this
-        * case, and kernel threads would basically never run until userspace
-        * went to sleep, even if said userspace interacts with the kernel in
-        * various ways.
-        */
-       if (time_travel_mode == TT_MODE_INFCPU ||
-           time_travel_mode == TT_MODE_EXTERNAL)
-               schedule();
-
        /* Initialize the syscall number and default return value. */
        UPT_SYSCALL_NR(r) = PT_SYSCALL_NR(r->gp);
        PT_REGS_SET_SYSCALL_RETURN(regs, -ENOSYS);
@@ -41,9 +31,25 @@ void handle_syscall(struct uml_pt_regs *r)
                goto out;
 
        syscall = UPT_SYSCALL_NR(r);
-       if (syscall >= 0 && syscall < __NR_syscalls)
-               PT_REGS_SET_SYSCALL_RETURN(regs,
-                               EXECUTE_SYSCALL(syscall, regs));
+       if (syscall >= 0 && syscall < __NR_syscalls) {
+               unsigned long ret = EXECUTE_SYSCALL(syscall, regs);
+
+               PT_REGS_SET_SYSCALL_RETURN(regs, ret);
+
+               /*
+                * An error value here can be some form of -ERESTARTSYS
+                * and then we'd just loop. Make any error syscalls take
+                * some time, so that it won't just loop if something is
+                * not ready, and hopefully other things will make some
+                * progress.
+                */
+               if (IS_ERR_VALUE(ret) &&
+                   (time_travel_mode == TT_MODE_INFCPU ||
+                    time_travel_mode == TT_MODE_EXTERNAL)) {
+                       um_udelay(1);
+                       schedule();
+               }
+       }
 
 out:
        syscall_trace_leave(regs);