xref: /linux/drivers/gpu/nova-core/gpu.rs (revision 82b78182eacf82c1847c6f1fd93d91c15efb69cf)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 use kernel::{
4     device,
5     devres::Devres,
6     fmt,
7     io::Io,
8     num::Bounded,
9     pci,
10     prelude::*,
11     sync::Arc, //
12 };
13 
14 use crate::{
15     bounded_enum,
16     driver::Bar0,
17     falcon::{
18         gsp::Gsp as GspFalcon,
19         sec2::Sec2 as Sec2Falcon,
20         Falcon, //
21     },
22     fb::SysmemFlush,
23     gsp::Gsp,
24     regs,
25 };
26 
27 mod hal;
28 
29 macro_rules! define_chipset {
30     ({ $($variant:ident = $value:expr),* $(,)* }) =>
31     {
32         /// Enum representation of the GPU chipset.
33         #[derive(fmt::Debug, Copy, Clone, PartialOrd, Ord, PartialEq, Eq)]
34         pub(crate) enum Chipset {
35             $($variant = $value),*,
36         }
37 
38         impl Chipset {
39             pub(crate) const ALL: &'static [Chipset] = &[
40                 $( Chipset::$variant, )*
41             ];
42 
43             ::kernel::macros::paste!(
44             /// Returns the name of this chipset, in lowercase.
45             ///
46             /// # Examples
47             ///
48             /// ```
49             /// let chipset = Chipset::GA102;
50             /// assert_eq!(chipset.name(), "ga102");
51             /// ```
52             pub(crate) const fn name(&self) -> &'static str {
53                 match *self {
54                 $(
55                     Chipset::$variant => stringify!([<$variant:lower>]),
56                 )*
57                 }
58             }
59             );
60         }
61 
62         // TODO[FPRI]: replace with something like derive(FromPrimitive)
63         impl TryFrom<u32> for Chipset {
64             type Error = kernel::error::Error;
65 
66             fn try_from(value: u32) -> Result<Self, Self::Error> {
67                 match value {
68                     $( $value => Ok(Chipset::$variant), )*
69                     _ => Err(ENODEV),
70                 }
71             }
72         }
73     }
74 }
75 
76 define_chipset!({
77     // Turing
78     TU102 = 0x162,
79     TU104 = 0x164,
80     TU106 = 0x166,
81     TU117 = 0x167,
82     TU116 = 0x168,
83     // Ampere
84     GA100 = 0x170,
85     GA102 = 0x172,
86     GA103 = 0x173,
87     GA104 = 0x174,
88     GA106 = 0x176,
89     GA107 = 0x177,
90     // Hopper
91     GH100 = 0x180,
92     // Ada
93     AD102 = 0x192,
94     AD103 = 0x193,
95     AD104 = 0x194,
96     AD106 = 0x196,
97     AD107 = 0x197,
98     // Blackwell GB10x
99     GB100 = 0x1a0,
100     GB102 = 0x1a2,
101     // Blackwell GB20x
102     GB202 = 0x1b2,
103     GB203 = 0x1b3,
104     GB205 = 0x1b5,
105     GB206 = 0x1b6,
106     GB207 = 0x1b7,
107 });
108 
109 impl Chipset {
110     pub(crate) const fn arch(self) -> Architecture {
111         match self {
112             Self::TU102 | Self::TU104 | Self::TU106 | Self::TU117 | Self::TU116 => {
113                 Architecture::Turing
114             }
115             Self::GA100 | Self::GA102 | Self::GA103 | Self::GA104 | Self::GA106 | Self::GA107 => {
116                 Architecture::Ampere
117             }
118             Self::GH100 => Architecture::Hopper,
119             Self::AD102 | Self::AD103 | Self::AD104 | Self::AD106 | Self::AD107 => {
120                 Architecture::Ada
121             }
122             Self::GB100 | Self::GB102 => Architecture::BlackwellGB10x,
123             Self::GB202 | Self::GB203 | Self::GB205 | Self::GB206 | Self::GB207 => {
124                 Architecture::BlackwellGB20x
125             }
126         }
127     }
128 
129     /// Returns `true` if this chipset requires the PIO-loaded bootloader in order to boot FWSEC.
130     ///
131     /// This includes all chipsets < GA102.
132     pub(crate) const fn needs_fwsec_bootloader(self) -> bool {
133         matches!(self.arch(), Architecture::Turing) || matches!(self, Self::GA100)
134     }
135 }
136 
137 // TODO
138 //
139 // The resulting strings are used to generate firmware paths, hence the
140 // generated strings have to be stable.
141 //
142 // Hence, replace with something like strum_macros derive(Display).
143 //
144 // For now, redirect to fmt::Debug for convenience.
145 impl fmt::Display for Chipset {
146     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
147         write!(f, "{self:?}")
148     }
149 }
150 
151 bounded_enum! {
152     /// Enum representation of the GPU generation.
153     #[derive(fmt::Debug, Copy, Clone)]
154     pub(crate) enum Architecture with TryFrom<Bounded<u32, 6>> {
155         Turing = 0x16,
156         Ampere = 0x17,
157         Hopper = 0x18,
158         Ada = 0x19,
159         BlackwellGB10x = 0x1a,
160         BlackwellGB20x = 0x1b,
161     }
162 }
163 
164 #[derive(Clone, Copy)]
165 pub(crate) struct Revision {
166     major: Bounded<u8, 4>,
167     minor: Bounded<u8, 4>,
168 }
169 
170 impl From<regs::NV_PMC_BOOT_42> for Revision {
171     fn from(boot0: regs::NV_PMC_BOOT_42) -> Self {
172         Self {
173             major: boot0.major_revision().cast(),
174             minor: boot0.minor_revision().cast(),
175         }
176     }
177 }
178 
179 impl fmt::Display for Revision {
180     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
181         write!(f, "{:x}.{:x}", self.major, self.minor)
182     }
183 }
184 
185 /// Structure holding a basic description of the GPU: `Chipset` and `Revision`.
186 #[derive(Clone, Copy)]
187 pub(crate) struct Spec {
188     chipset: Chipset,
189     revision: Revision,
190 }
191 
192 impl Spec {
193     fn new(dev: &device::Device, bar: &Bar0) -> Result<Spec> {
194         // Some brief notes about boot0 and boot42, in chronological order:
195         //
196         // NV04 through NV50:
197         //
198         //    Not supported by Nova. boot0 is necessary and sufficient to identify these GPUs.
199         //    boot42 may not even exist on some of these GPUs.
200         //
201         // Fermi through Volta:
202         //
203         //     Not supported by Nova. boot0 is still sufficient to identify these GPUs, but boot42
204         //     is also guaranteed to be both present and accurate.
205         //
206         // Turing and later:
207         //
208         //     Supported by Nova. Identified by first checking boot0 to ensure that the GPU is not
209         //     from an earlier (pre-Fermi) era, and then using boot42 to precisely identify the GPU.
210         //     Somewhere in the Rubin timeframe, boot0 will no longer have space to add new GPU IDs.
211 
212         let boot0 = bar.read(regs::NV_PMC_BOOT_0);
213 
214         if boot0.is_older_than_fermi() {
215             return Err(ENODEV);
216         }
217 
218         let boot42 = bar.read(regs::NV_PMC_BOOT_42);
219         Spec::try_from(boot42).inspect_err(|_| {
220             dev_err!(dev, "Unsupported chipset: {}\n", boot42);
221         })
222     }
223 }
224 
225 impl TryFrom<regs::NV_PMC_BOOT_42> for Spec {
226     type Error = Error;
227 
228     fn try_from(boot42: regs::NV_PMC_BOOT_42) -> Result<Self> {
229         Ok(Self {
230             chipset: boot42.chipset()?,
231             revision: boot42.into(),
232         })
233     }
234 }
235 
236 impl fmt::Display for Spec {
237     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
238         f.write_fmt(fmt!(
239             "Chipset: {}, Architecture: {:?}, Revision: {}",
240             self.chipset,
241             self.chipset.arch(),
242             self.revision
243         ))
244     }
245 }
246 
247 /// Structure holding the resources required to operate the GPU.
248 #[pin_data]
249 pub(crate) struct Gpu {
250     spec: Spec,
251     /// MMIO mapping of PCI BAR 0
252     bar: Arc<Devres<Bar0>>,
253     /// System memory page required for flushing all pending GPU-side memory writes done through
254     /// PCIE into system memory, via sysmembar (A GPU-initiated HW memory-barrier operation).
255     sysmem_flush: SysmemFlush,
256     /// GSP falcon instance, used for GSP boot up and cleanup.
257     gsp_falcon: Falcon<GspFalcon>,
258     /// SEC2 falcon instance, used for GSP boot up and cleanup.
259     sec2_falcon: Falcon<Sec2Falcon>,
260     /// GSP runtime data. Temporarily an empty placeholder.
261     #[pin]
262     gsp: Gsp,
263 }
264 
265 impl Gpu {
266     pub(crate) fn new<'a>(
267         pdev: &'a pci::Device<device::Bound>,
268         devres_bar: Arc<Devres<Bar0>>,
269         bar: &'a Bar0,
270     ) -> impl PinInit<Self, Error> + 'a {
271         try_pin_init!(Self {
272             spec: Spec::new(pdev.as_ref(), bar).inspect(|spec| {
273                 dev_info!(pdev,"NVIDIA ({})\n", spec);
274             })?,
275 
276             // We must wait for GFW_BOOT completion before doing any significant setup on the GPU.
277             _: {
278                 hal::gpu_hal(spec.chipset).wait_gfw_boot_completion(bar)
279                     .inspect_err(|_| dev_err!(pdev, "GFW boot did not complete\n"))?;
280             },
281 
282             sysmem_flush: SysmemFlush::register(pdev.as_ref(), bar, spec.chipset)?,
283 
284             gsp_falcon: Falcon::new(
285                 pdev.as_ref(),
286                 spec.chipset,
287             )
288             .inspect(|falcon| falcon.clear_swgen0_intr(bar))?,
289 
290             sec2_falcon: Falcon::new(pdev.as_ref(), spec.chipset)?,
291 
292             gsp <- Gsp::new(pdev),
293 
294             _: { gsp.boot(pdev, bar, spec.chipset, gsp_falcon, sec2_falcon)? },
295 
296             bar: devres_bar,
297         })
298     }
299 
300     /// Called when the corresponding [`Device`](device::Device) is unbound.
301     ///
302     /// Note: This method must only be called from `Driver::unbind`.
303     pub(crate) fn unbind(&self, dev: &device::Device<device::Core>) {
304         kernel::warn_on!(self
305             .bar
306             .access(dev)
307             .inspect(|bar| self.sysmem_flush.unregister(bar))
308             .is_err());
309     }
310 }
311