Boot Linux faster!

Check our new training course

Boot Linux faster!

Check our new training course
and Creative Commons CC-BY-SA
lecture and lab materials

Bootlin logo

Elixir Cross Referencer

 * Copyright (c) 2010-2015 Wind River Systems, Inc.
 * SPDX-License-Identifier: Apache-2.0

 * @file
 * @brief Kernel swapper code for IA-32
 * This module implements the _Swap() routine for the IA-32 architecture.
 * Note that the file arch/x86/include/swapstk.h defines
 * a representation of the save stack frame generated by _Swap() in order
 * to generate offsets (in the form of absolute symbols) for consumption by
 * host tools.  Please update swapstk.h if changing the structure of the
 * save frame on the stack.

#include <kernel_structs.h>
#include <arch/x86/asm.h>
#include <offsets_short.h>

	/* exports (internal APIs) */


	/* externs */

 * @brief Initiate a cooperative context switch
 * The _Swap() routine is invoked by various kernel services to effect
 * a cooperative context switch.  Prior to invoking _Swap(), the
 * caller disables interrupts (via irq_lock) and the return 'key'
 * is passed as a parameter to _Swap().  The 'key' actually represents
 * the EFLAGS register prior to disabling interrupts via a 'cli' instruction.
 * Given that _Swap() is called to effect a cooperative context switch, only
 * the non-volatile integer registers need to be saved in the TCS of the
 * outgoing thread.  The restoration of the integer registers of the incoming
 * thread depends on whether that thread was preemptively context switched out.
 * The _INT_ACTIVE and _EXC_ACTIVE bits in the k_thread->thread_state field
 * will signify that the thread was preemptively context switched out, and thus
 * both the volatile and non-volatile integer registers need to be restored.
 * The non-volatile registers need to be scrubbed to ensure they contain no
 * sensitive information that could compromise system security.  This is to
 * make sure that information will not be leaked from one application to
 * another via these volatile registers.
 * Here, the integer registers (EAX, ECX, EDX) have been scrubbed.  Any changes
 * to this routine that alter the values of these registers MUST be reviewed
 * for potential security impacts.
 * Floating point registers are handled using a lazy save/restore
 * mechanism since it's expected relatively few threads will be created
 * with the K_FP_REGS or K_SSE_REGS option bits.  The kernel data structure
 * maintains a 'current_fp' field to keep track of the thread that "owns"
 * the floating point registers.  Floating point registers consist of
 * ST0->ST7 (x87 FPU and MMX registers) and XMM0 -> XMM7.
 * All floating point registers are considered 'volatile' thus they will
 * only be saved/restored when a preemptive context switch occurs.
 * Floating point registers are currently NOT scrubbed, and are subject to
 * potential security leaks.
 * @return -EAGAIN, or a return value set by a call to
 * _set_thread_return_value()
 * C function prototype:
 * unsigned int _Swap (unsigned int eflags);

	/* save EFLAGS on stack right before return address, just as SYSV would
	 * have done
	pushl	0(%esp)
	movl	%eax, 4(%esp)
	 * Push all non-volatile registers onto the stack; do not copy
	 * any of these registers into the k_thread.  Only the 'esp' register
	 * after all the pushes have been performed) will be stored in the
	 * k_thread.

	pushl	%edi

	movl	$_kernel, %edi

	pushl	%esi
	pushl	%ebx
	pushl	%ebp

	 * Carve space for the return value. Setting it to a defafult of
	 * -EAGAIN eliminates the need for the timeout code to set it.
	 * If another value is ever needed, it can be modified with
	 * _set_thread_return_value().

	pushl   _k_neg_eagain

	/* save esp into k_thread structure */

	movl	_kernel_offset_to_current(%edi), %ecx
	movl	%esp, _thread_offset_to_esp(%ecx)

	/* Register the context switch */
	call	_sys_k_event_logger_context_switch
	movl	_kernel_offset_to_ready_q_cache(%edi), %eax

	 * At this point, the %eax register contains the 'k_thread *' of the
	 * thread to be swapped in, and %edi still contains &_kernel.

	 * Clear the CR0[TS] bit (in the event the current thread
	 * doesn't have floating point enabled) to prevent the "device not
	 * available" exception when executing the subsequent fxsave/fnsave
	 * and/or fxrstor/frstor instructions.
	 * Indeed, it's possible that none of the aforementioned instructions
	 * need to be executed, for example, the incoming thread doesn't
	 * utilize floating point operations.  However, the code responsible
	 * for setting the CR0[TS] bit appropriately for the incoming thread
	 * (just after the 'restoreContext_NoFloatSwap' label) will leverage
	 * the fact that the following 'clts' was performed already.


	 * Determine whether the incoming thread utilizes floating point registers
	 * _and_ whether the thread was context switched out preemptively.

	testb	$_FP_USER_MASK, _thread_offset_to_user_options(%eax)
	je 	restoreContext_NoFloatSwap

	 * The incoming thread uses floating point registers:
	 * Was it the last thread to use floating point registers?
	 * If so, there there is no need to restore the floating point context.

	movl	_kernel_offset_to_current_fp(%edi), %ebx
	cmpl	%ebx, %eax
	je	restoreContext_NoFloatSwap

	 * The incoming thread uses floating point registers and it was _not_
	 * the last thread to use those registers:
	 * Check whether the current FP context actually needs to be saved
	 * before swapping in the context of the incoming thread.

	testl	%ebx, %ebx
	jz	restoreContext_NoFloatSave

	 * The incoming thread uses floating point registers and it was _not_
	 * the last thread to use those registers _and_ the current FP context
	 * needs to be saved.
	 * Given that the ST[0] -> ST[7] and XMM0 -> XMM7 registers are all
	 * 'volatile', only save the registers if the "current FP context"
	 * was preemptively context switched.

	testb	$_INT_OR_EXC_MASK, _thread_offset_to_thread_state(%ebx)
	je	restoreContext_NoFloatSave

	testb	$K_SSE_REGS, _thread_offset_to_user_options(%ebx)
	je	x87FloatSave

	 * 'fxsave' does NOT perform an implicit 'fninit', therefore issue an
	 * 'fninit' to ensure a "clean" FPU state for the incoming thread
	 * (for the case when the fxrstor is not executed).

	fxsave	_thread_offset_to_preempFloatReg(%ebx)
	jmp	floatSaveDone

#endif /* CONFIG_SSE */

	/* 'fnsave' performs an implicit 'fninit' after saving state! */

	fnsave	 _thread_offset_to_preempFloatReg(%ebx)

	/* fall through to 'floatSaveDone' */


	 * Restore floating point context of the incoming thread.

	 * Again, given that the ST[0] -> ST[7] and XMM0 -> XMM7 registers are
	 * all 'volatile', only restore the registers if the incoming thread
	 * was previously preemptively context switched out.

	testb   $_INT_OR_EXC_MASK, _thread_offset_to_thread_state(%eax)
	je 	restoreContext_NoFloatRestore

	testb	$K_SSE_REGS, _thread_offset_to_user_options(%eax)
	je	x87FloatRestore

	fxrstor	_thread_offset_to_preempFloatReg(%eax)
	jmp	floatRestoreDone


#endif /* CONFIG_SSE */

	frstor	_thread_offset_to_preempFloatReg(%eax)

	/* fall through to 'floatRestoreDone' */


	/* record that the incoming thread "owns" the floating point registers */

	movl	%eax, _kernel_offset_to_current_fp(%edi)

	 * Branch point when none of the floating point registers need to be
	 * swapped because: a) the incoming thread does not use them OR
	 * b) the incoming thread is the last thread that used those registers.


	 * Leave CR0[TS] clear if incoming thread utilizes the floating point
	 * registers

	testb	$_FP_USER_MASK, _thread_offset_to_user_options(%eax)
	jne	CROHandlingDone

	 * The incoming thread does NOT currently utilize the floating point
	 * registers, so set CR0[TS] to ensure the "device not available"
	 * exception occurs on the first attempt to access a x87 FPU, MMX,
	 * or XMM register.

	movl %cr0, %edx
	orl $0x8, %edx
	movl %edx, %cr0


#endif /* CONFIG_FP_SHARING */

	/* update _kernel.current to reflect incoming thread */

	movl    %eax, _kernel_offset_to_current(%edi)

	/* recover task/fiber stack pointer from k_thread */

	movl	_thread_offset_to_esp(%eax), %esp

	/* load return value from a possible _set_thread_return_value() */

	popl	%eax

	/* pop the non-volatile registers from the stack */

	popl	%ebp
	popl	%ebx
	popl	%esi
	popl	%edi

	 * %eax may contain one of these values:
	 * - the return value for _Swap() that was set up by a call to
	 * _set_thread_return_value()
	 * - in legacy kernel, same value as %ebx, which is non-volatile
	 * - in unified kernel, -EINVAL

	/* Utilize the 'eflags' parameter to _Swap() */

	pushl	4(%esp)
	testl	$0x200, (%esp)
	jz	skipIntLatencyStop

	/* save %eax since it used as the return value for _Swap */
	pushl	%eax
	/* interrupts are being reenabled, stop accumulating time */
	call	_int_latency_stop
	/* restore _Swap's %eax */
	popl	%eax

	/* Remember that eflags we stuck into the stack before the return
	 * address? need to get it out of there since the calling convention
	 * will not do that for us.
	popl	%edx
	movl	%edx, (%esp)