- 19 May, 2015 40 commits
-
-
Ingo Molnar authored
So 6 years ago we made the FPU fpstate dynamically allocated: aa283f49 ("x86, fpu: lazy allocation of FPU area - v5") 61c4628b ("x86, fpu: split FPU state from task struct - v5") In hindsight this was a mistake: - it complicated context allocation failure handling, such as: /* kthread execs. TODO: cleanup this horror. */ if (WARN_ON(fpstate_alloc_init(fpu))) force_sig(SIGKILL, tsk); - it caused us to enable irqs in fpu__restore(): local_irq_enable(); /* * does a slab alloc which can sleep */ if (fpstate_alloc_init(fpu)) { /* * ran out of memory! */ do_group_exit(SIGKILL); return; } local_irq_disable(); - it (slightly) slowed down task creation/destruction by adding slab allocation/free pattens. - it made access to context contents (slightly) slower by adding one more pointer dereference. The motivation for the dynamic allocation was two-fold: - reduce memory consumption by non-FPU tasks - allocate and handle only the necessary amount of context for various XSAVE processors that have varying hardware frame sizes. These days, with glibc using SSE memcpy by default and GCC optimizing for SSE/AVX by default, the scope of FPU using apps on an x86 system is much larger than it was 6 years ago. For example on a freshly installed Fedora 21 desktop system, with a recent kernel, all non-kthread tasks have used the FPU shortly after bootup. Also, even modern embedded x86 CPUs try to support the latest vector instruction set - so they'll too often use the larger xstate frame sizes. So remove the dynamic allocation complication by embedding the FPU fpstate in task_struct again. This should make the FPU a lot more accessible to all sorts of atomic contexts. We could still optimize for the xstate frame size in the future, by moving the state structure to the last element of task_struct, and allocating only a part of that. This change is kept minimal by still keeping the ctx_alloc()/free() routines (that now do nothing substantial) - we'll remove them in the following patches. Reviewed-by:
Borislav Petkov <bp@alien8.de> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by:
Ingo Molnar <mingo@kernel.org>
-
Ingo Molnar authored
x86/fpu: Optimize copy_fpregs_to_fpstate() by removing the FNCLEX synchronization with FP exceptions So we have the following ancient code in copy_fpregs_to_fpstate(): if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) { asm volatile("fnclex"); goto drop_fpregs; } which clears pending FPU exceptions and then drops registers, which causes the next FP instruction of the saved context to re-load the saved FPU state, with all pending exceptions marked properly, and will re-start the exception handling mechanism in the hardware. Since FPU exceptions are always issued on instruction boundaries, in particular on the next FP instruction following the exception generating instruction, there's no fear of getting an FP exception asynchronously. They were truly asynchronous back in the IRQ13 days, when the FPU was a weird and expensive co-processor that did its own processing, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. With the introduction of optimized XSAVE support there's a new complication: if the xstate features bit indicates that a particular state component is unused (in 'init state'), then the hardware does not guarantee that the XSAVE (et al) instruction keeps the underlying FPU state image in memory valid and current. In practice this means that the hardware won't write it, and the exceptions flag in the state might be an older version, with it still being set. This meant that we had to check the xfeatures flag as well, adding another memory load and branch to a critical hot path of the scheduler. So optimize all this by removing both the old quirk and the new check, and straight-line optimizing the most common cases with likely() hints. Quite a bit of code gets removed this way: arch/x86/kernel/process_64.o: text data bss dec filename 5484 8 0 5492 process_64.o.before 5416 8 0 5424 process_64.o.after Now there's also a chance that some weird behavior or erratum was masked by our IRQ13 handling quirk (or that I misunderstood the nature of the quirk), and that this change triggers some badness. There's no real good way to protect against that possibility other than keeping this change well isolated, well commented and well bisectable. If you bisect a weird (or not so weird) breakage to this commit then please let us know! Reviewed-by: -
Ingo Molnar authored
So fpu_save_init() is a historic name that got its name when the only way the FPU state was FNSAVE, which cleared (well, destroyed) the FPU state after saving it. Nowadays the name is misleading, because ever since the introduction of FXSAVE (and more modern FPU saving instructions) the 'we need to reload the FPU state' part is only true if there's a pending FPU exception [*], which is almost never the case. So rename it to copy_fpregs_to_fpstate() to make it clear what's happening. Also add a few comments about why we cannot keep registers in certain cases. Also clean up the control flow a bit, to make it more apparent when we are dropping/keeping FP registers, and to optimize the common case (of keeping fpregs) some more. [*] Probably not true anymore, modern instructions always leave the FPU state intact, even if exceptions are pending: because pending FP exceptions are posted on the next FP instruction, not asynchronously. They were truly asynchronous back in the IRQ13 case, and we had to synchronize with them, but that code is not working anymore: we don't have IRQ13 mapped in the IDT anymore. But a cleanup patch is obviously not the place to change subtle behavior. Reviewed-by: -
Ingo Molnar authored
Especially the irq_ts_save() function is pretty bloaty, generating over a dozen instructions, so uninline them. Even though the API is used rarely, the space savings are measurable: text data bss dec hex filename 13331995 2572920 1634304 17539219 10ba093 vmlinux.before 13331739 2572920 1634304 17538963 10b9f93 vmlinux.after ( This also allows the removal of an include file inclusion from fpu/api.h, speeding up the kernel build slightly. ) Reviewed-by:
-
Ingo Molnar authored
There are a number of FPU internal function prototypes and an inline function in fpu/api.h, mostly placed so historically as the code grew over the years. Move them over into fpu/internal.h where they belong. (Add sched.h include to stackprotector.h which incorrectly relied on getting it from fpu/api.h.) fpu/api.h is now a pure file that only contains FPU APIs intended for driver use. Reviewed-by:
-
Ingo Molnar authored
Both inline functions call an inline function unconditionally, so we already pay the function call based clobbering cost. Uninline them. This saves quite a bit of code in various performance sensitive code paths: text data bss dec hex filename 13321334 2569888 1634304 17525526 10b6b16 vmlinux.before 13320246 2569888 1634304 17524438 10b66d6 vmlinux.after Reviewed-by:
-
Ingo Molnar authored
It's an internal method, not a driver API, so move it from fpu/api.h to fpu/internal.h. Reviewed-by:
-
Ingo Molnar authored
Extend the comments of the FPU init code, and fix old ones. Reviewed-by:
-
Ingo Molnar authored
Reorder init methods in order of their relationship and usage, to form coherent blocks throughout the whole file. Reviewed-by:
-
Ingo Molnar authored
To bring it in line with the other init_system*() methods. Reviewed-by:
-
Ingo Molnar authored
Now that fpu__detect() has become an empty layer around fpu__init_system(), eliminate it and make fpu__init_system() the main system initialization routine. Reviewed-by:
-
Ingo Molnar authored
Move the fpu__init_system_early_generic() call into fpu__init_system(), which hosts all the system init calls. Expose fpu__init_system() to other modules - this will be our main and only system init function. Reviewed-by:
-
Ingo Molnar authored
check_fpu() currently relies on being called early in the init sequence, when CR0::TS has not been set up yet. Save/restore CR0::TS across this function, to make it invariant to init ordering. This way we'll be able to move the generic FPU setup routines earlier in the init sequence. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: -
Ingo Molnar authored
Create separate fpu/bugs.c code so that if we read generic FPU code we don't have to wade through all the bugcheck related code first. Reviewed-by:
-
Ingo Molnar authored
There's a !FPU related sanity check in fpu__init_cpu_generic(), which is executed on every CPU onlining - even though we should do this only once, and during system init. Move this check to fpu__init_system_early_generic(). Reviewed-by:
-
Ingo Molnar authored
Move the generic bits of fpu__detect() into fpu__init_system_early_generic(). We'll move some other code here too in a followup patch. Reviewed-by:
-
Ingo Molnar authored
Factor out the generic bits from fpu__init_system(). Rename mxcsr_feature_mask_init() to fpu__init_system_mxcsr() to bring it in line with the rest of the nomenclature. Reviewed-by:
-
Ingo Molnar authored
Factor out the generic bits from fpu__init_cpu(), to create a flat sequence of per CPU initialization function calls: fpu__init_cpu_generic(); fpu__init_cpu_xstate(); fpu__init_cpu_ctx_switch(); Reviewed-by:
-
Ingo Molnar authored
After the latest round of cleanups, fpu__cpu_init() has become a simple call to fpu__init_cpu(). Rename fpu__init_cpu() to fpu__cpu_init() and remove the extra layer. Reviewed-by:
-
Ingo Molnar authored
We are now doing the fpu__init_cpu_ctx_switch() call from fpu__init_cpu(), so there's no need to call it from fpu__init_system() anymore. Reviewed-by:
-
Ingo Molnar authored
fpu__cpu_init() is called on every CPU, so it is the wrong place to call fpu__init_system() from. Call it from fpu__detect(): this is early CPU init code, but we already have CPU features detected, so we can call the system-wide FPU init code from here. Reviewed-by:
-
Ingo Molnar authored
fpu__init_cpu() is currently called from fpu__init_system(), which is the wrong place for it: call it from the proper high level per CPU init function, fpu__init_cpu(). Note, we still keep the old call site as well, because it depends on having proper CR0::TS setup. We'll fix this in the next patch. Reviewed-by:
-
Ingo Molnar authored
The fpstate_xstate_init_size() function sets up a basic xstate_size, called during fpu__detect() currently. Its real dependency is to be called before fpu__init_system_xstate(). So move the function call site into fpu__init_system(), to right before the fpu__init_system_xstate() call. Also add a once-per-boot flag to fpstate_xstate_init_size(), we'll remove this quirk later once we've cleaned up the init dependencies. This moves the two related functions closer to each other and makes them both part of the _init_system() functionality. Currently we do the fpstate_xstate_init_size() Reviewed-by:
-
Ingo Molnar authored
mxcsr_feature_mask_init() depends on TS being cleared, as it executes an FXSAVE instruction. After later changes we will move the TS setup into fpu__init_cpu(), which will interact with this - so clear the TS flag explicitly. Reviewed-by:
-
Ingo Molnar authored
So fpu__ctx_switch_init() has two aspects: a once per bootup functionality that sets up a capability flag, and a per CPU functionality that sets CR0::TS. Split the function. Note that at this stage we still have duplicate calls into these methods, as both the _system() and the _cpu() methods are run on all CPUs, with lower level on_boot_cpu flags filtering out the duplicates where needed. So add TS flag clearing as well, to handle the aftermath of early CPU init sequences that might call in without having eager-fpu set - don't assume the TS flag is cleared. Calling each from its respective init level will happen later on. Reviewed-by:
-
Ingo Molnar authored
It's not an xsave specific function anymore, so rename it accordingly and also clean it up a bit: - remove the obsolete __init_refok, as the code paths are not mixed anymore - rename it from eager_fpu_init() to fpu__ctx_switch_init() - remove stray 'return;' - make it static to its only user Reviewed-by:
-
Ingo Molnar authored
Move eager_fpu_init() and the 'eagerfpu' boot parameter handling function to the generic FPU init file: it's generic FPU functionality. Reviewed-by:
-
Ingo Molnar authored
The FPU context switch type (lazy or eager) setup code is split into two places currently - move it all to eager_fpu_init(). Note that the code we move will now be executed on non-xstate CPUs as well, but this should be safe: both xfeatures_mask and cpu_has_xsaveopt is 0 there. Reviewed-by:
-
Ingo Molnar authored
It's a pure xstate method now, no need for this duplicate call. Reviewed-by:
-
Ingo Molnar authored
The legacy FPU init image is used on older CPUs who don't run xstate init. But the init code is called within setup_init_fpu_buf(), an xstate method. Move this legacy init out of the xstate code and put it into fpu/init.c. Reviewed-by:
-
Ingo Molnar authored
Only call xstate system setup routines from fpu__init_system(). Likewise, don't call fpu__init_cpu_xstate() from fpu__init_system(). Reviewed-by:
-
Ingo Molnar authored
Expand fpu__init_system_xstate() and fpu__init_cpu_xstate() calls into xsave_init() calls. (This will allow us to call the proper versions in higher level FPU init code later on.) No change in functionality. Reviewed-by:
-
Ingo Molnar authored
Linearize the call sequence in xsave_init(): fpu__init_system_xstate(); fpu__init_cpu_xstate(); We do this by propagating the boot-once quirk into fpu__init_system_xstate(). fpu__init_cpu_xstate() is safe to be called multiple time. Reviewed-by:
-
Ingo Molnar authored
Now that legacy code can execute fpu__init_cpu_xstate() in xsave_init(), we can move the once per boot legacy check into fpu__init_system_xstate(), where it belongs. Reviewed-by:
-
Ingo Molnar authored
fpu__init_system_xstate() does an FPU capability check that is better done in fpu__init_cpu_xstate(). This will allow us to call fpu__init_cpu_xstate() directly on legacy CPUs as well. Reviewed-by:
-
Ingo Molnar authored
Rename existing xstate init functions along the system/cpu init principles: fpu__init_system_xstate(): called once per system bootup fpu__init_cpu_xstate(): called per CPU onlining Also make the fpu__init_cpu_xstate() early code invariant: if xfeatures_mask is not set yet then don't crash but return. Reviewed-by:
-
Ingo Molnar authored
There are two kinds of FPU initialization sequences necessary to bring FPU functionality up: once per system bootup activities, such as detection, feature initialization, etc. of attributes that are shared by all CPUs in the system - and per cpu initialization sequences run when a CPU is brought online (either during bootup or during CPU hotplug onlining), such as CR0/CR4 register setting, etc. The FPU code is mixing these roles together, with no clear distinction. Start sorting this out by splitting the main FPU detection routine (fpu__cpu_init()) into two parts: fpu__init_system() for one per system init activities, and fpu__init_cpu() for the per CPU onlining init activities. Note that xstate_init() is called from both variants for the time being, because it has a dual nature as well. We'll fix that in upcoming patches. Just do the split and call it as we used to before, don't introduce any change in initialization behavior yet, beyond duplicate (and harmless) fpu__init_cpu() and xstate_init() calls - which we'll fix in later patches. Reviewed-by:
-
Ingo Molnar authored
Make init_xstate_buf allocated statically at build time. This structure's maximum size is around 1KB - and it's allocated even on most modern embedded x86 CPUs which strive for FPU instruction set parity with desktop and server CPUs, so it's not like we can save much on smaller systems. This removes the last bootmem allocation from the FPU init path, allowing it to be called earlier in the boot sequence. Reviewed-by:
-
Ingo Molnar authored
Remove the dependency on the init_xstate_buf == NULL check to implement once-per-bootup logic in eager_fpu_init(), by making setup_init_fpu_buf() run once per bootup explicitly. This is in preparation to make init_xstate_buf statically allocated. The various boot-once quirks in the FPU init code will be removed in a later cleanup stage. Reviewed-by:
-
Ingo Molnar authored
There's only 8 xstate bits at the moment, and it's not like we can support unknown bits - so put xstate_offsets[] and xstate_sizes[] into static allocation. This is in preparation to be able to call the FPU init code earlier, when there's no bootmem available yet. Reviewed-by:
-
