#include <linux/seq_file.h> // For seq_* functions and types
#include <linux/version.h>  // Macros to detect kernel version

#include "nvdebug_linux.h"

// Uncomment to expand channel status information when printing the runlist
#define DETAILED_CHANNEL_INFO

#ifdef DETAILED_CHANNEL_INFO
/* Print channel details using PCCSR (Programmable Channel Control System RAM?)
  @param s      Pointer to state from seq_file subsystem to pass to seq_printf
  @param g      Pointer to our internal GPU state
  @param chid   ID of channel to print details on, range [0, 512)
  @param prefix Text string to prefix each line with, or empty string
*/
static int runlist_detail_seq_show_chan(struct seq_file *s, struct nvdebug_state *g, uint32_t chid, char *prefix) {
	channel_ctrl_t chan;
	uint64_t instance_ptr;

	if ((chan.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(chid))) == -1)
		return -EIO;
	instance_ptr = (uint64_t)chan.inst_ptr << 12;
	// Don't print write-only fields
	seq_printf(s, "%s|= Channel Info ======|\n", prefix);
	seq_printf(s, "%s| Enabled:           %d|\n", prefix, chan.enable);
	seq_printf(s, "%s| Next:              %d|\n", prefix, chan.next);
	seq_printf(s, "%s| PBDMA Faulted:     %d|\n", prefix, chan.pbdma_faulted);
	seq_printf(s, "%s| ENG Faulted:       %d|\n", prefix, chan.eng_faulted);
	seq_printf(s, "%s| Status:           %2d|\n", prefix, chan.status);
	seq_printf(s, "%s| Busy:              %d|\n", prefix, chan.busy);
	seq_printf(s, "%s| Instance PTR:       |\n", prefix);
	seq_printf(s, "%s|   %#018llx|\n", prefix, instance_ptr);
	seq_printf(s, "%s| %20s|\n", prefix, target_to_text(chan.inst_target));
	seq_printf(s, "%s| Instance bound:    %d|\n", prefix, chan.inst_bind);
	return 0;
}

/* `runlist_detail_seq_show_chan()`, but for Ampere+
  @param runlist_pri_base Base of the RLRAM region for this runlist

  `runlist_pri_base` is necessary, since Channel RAM is now per-runlist on
  Ampere+, and its location is configured in Runlist RAM.
*/
static int runlist_detail_seq_show_chan_ga100(struct seq_file *s, struct nvdebug_state *g, uint32_t chid, char *prefix, uint32_t runlist_pri_base) {
	runlist_channel_config_t channel_config;
	channel_ctrl_ga100_t chan;

	// Channel RAM is subsidiary to Runlist RAM (ie. per-runlist) on Ampere+
	if ((channel_config.raw = nvdebug_readl(g, runlist_pri_base + NV_RUNLIST_CHANNEL_CONFIG_GA100)) == -1)
		return -EIO;
	if ((chan.raw = nvdebug_readl(g, (((uint32_t)channel_config.bar0_offset << 4) + chid * 4))) == -1)
		return -EIO;
	seq_printf(s, "%s|= Channel Info ======|\n", prefix);
	seq_printf(s, "%s| Enabled:           %d|\n", prefix, chan.enable);
	seq_printf(s, "%s| Next:              %d|\n", prefix, chan.next);
	seq_printf(s, "%s| Busy:              %d|\n", prefix, chan.busy);
	seq_printf(s, "%s| PBDMA Faulted:     %d|\n", prefix, chan.pbdma_faulted);
	seq_printf(s, "%s| ENG Faulted:       %d|\n", prefix, chan.eng_faulted);
	seq_printf(s, "%s| On PBDMA:          %d|\n", prefix, chan.on_pbdma);
	seq_printf(s, "%s| On ENG:            %d|\n", prefix, chan.on_eng);
	seq_printf(s, "%s| Pending:           %d|\n", prefix, chan.pending);
	seq_printf(s, "%s| CTX Reload:        %d|\n", prefix, chan.ctx_reload);
	seq_printf(s, "%s| PBDMA Busy:        %d|\n", prefix, chan.pbdma_busy);
	seq_printf(s, "%s| ENG Busy:          %d|\n", prefix, chan.eng_busy);
	seq_printf(s, "%s| Acquire Fail:      %d|\n", prefix, chan.acquire_fail);
	return 0;
}
#endif

#if LINUX_VERSION_CODE < KERNEL_VERSION(4,19,0)
// Bug workaround. See comment in runlist_file_seq_start()
static loff_t pos_fixup;
#endif

static void *runlist_file_seq_start(struct seq_file *s, loff_t *pos) {
	static struct runlist_iter rl_iter;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(s->file)];
	// *pos == 0 for first call after read of file
	if (*pos == 0) {
		int err = get_runlist_iter(g, seq2gpuidx(s), &rl_iter);
		if (err)
			return ERR_PTR(err);
		// Don't try to print an empty runlist
		if (rl_iter.len <= 0)
			return NULL;
		return &rl_iter;
	}
	// If we're resuming an earlier print
	if (*pos < rl_iter.len) {
#if LINUX_VERSION_CODE < KERNEL_VERSION(4,19,0)
		// There's a nasty bug prior to 4.19-rc1 that if the buffer overflows, the
		// last update to `pos` is not saved. Work around that here by reloading a
		// saved copy of `pos`.
		if (!pos_fixup)
			return NULL;
		*pos = pos_fixup;
#endif
		return &rl_iter;
	}
	// When called with *pos != 0, we already traversed the runlist
	return NULL;
}

static void* runlist_file_seq_next(struct seq_file *s, void *raw_rl_iter,
				   loff_t *pos) {
	struct runlist_iter* rl_iter = raw_rl_iter;
	void *ret = NULL;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(s->file)];
	// Advance by one TSG or channel
	(*pos)++;
	rl_iter->curr_entry += NV_RL_ENTRY_SIZE(g);
	// Verify we haven't reached the end of the runlist
	// len is the num of tsg entries + total num of channel entries
	if (*pos < rl_iter->len) {
		ret = rl_iter;
	}
#if LINUX_VERSION_CODE < KERNEL_VERSION(4,19,0)
	// Bug workaround. See comment in runlist_file_seq_start()
	pos_fixup = ret ? *pos : 0;
#endif
	if (rl_iter->entries_left_in_tsg)
		rl_iter->entries_left_in_tsg--;
	return ret;
}

static void runlist_file_seq_stop(struct seq_file *s, void *raw_rl_iter) {
	// No cleanup needed
}

// _show() must be idempotent. This function will be rerun if the seq_printf
// buffer was too small.
static int runlist_file_seq_show(struct seq_file *s, void *raw_rl_iter) {
	struct runlist_iter *rl_iter = raw_rl_iter;
	void *entry = rl_iter->curr_entry;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(s->file)];
	if (entry_type(g, entry) == ENTRY_TYPE_TSG) {
		if (rl_iter->entries_left_in_tsg) {
			printk(KERN_WARNING "[nvdebug] Found TSG ID%d @ %px when %d channels were still expected under the previous TSG in the runlist!\n", tsgid(g, entry), entry, rl_iter->entries_left_in_tsg);
			while (rl_iter->entries_left_in_tsg--)
				seq_printf(s, "[missing channel]\n");
		}
		rl_iter->entries_left_in_tsg = tsg_length(g, entry) + 1;
		seq_printf(s, "+---- TSG Entry %-3d---+\n", tsgid(g, entry));
		seq_printf(s, "| Scale: %-13d|\n", timeslice_scale(g, entry));
		seq_printf(s, "| Timeout: %-11d|\n", timeslice_timeout(g, entry));
		seq_printf(s, "| Length: %-12d|\n", tsg_length(g, entry));
		seq_printf(s, "+---------------------+\n");
	} else {
		char *indt = "";
		u64 instance_ptr = 0;
		if (rl_iter->entries_left_in_tsg)
			indt = "  ";
		// Reconstruct pointer to channel instance block
		if (g->chip_id >= NV_CHIP_ID_VOLTA) {
			instance_ptr = ((struct gv100_runlist_chan*)entry)->inst_ptr_hi;
			instance_ptr <<= 32;
		}
		instance_ptr |= inst_ptr_lo(g, entry) << 12;
		// Print channel information from runlist
		seq_printf(s, "%s+- Channel Entry %-4d-+\n", indt, chid(g, entry));
		if (g->chip_id >= NV_CHIP_ID_VOLTA)
			seq_printf(s, "%s| Runqueue Selector: %d|\n", indt,
			           ((struct gv100_runlist_chan*)entry)->runqueue_selector);
		// Not populated on Kepler [ex: gk104 in Bonham (Quadro K5000)], and
		// populated but unused on Pascal [ex: gp104 in Bonham (GTX 1080 Ti)].
		// (The aperture field may be incorrectly populated as INVALID, but the
		// context still works on the aformentioned Pascal GPU.)
		seq_printf(s, "%s| Instance PTR:       |\n", indt);
		seq_printf(s, "%s|   %#018llx|\n", indt, instance_ptr);
		seq_printf(s, "%s| %20s|\n", indt, target_to_text(inst_target(g, entry)));
#ifdef DETAILED_CHANNEL_INFO
		// Print channel info from PCCSR/Channel RAM and the instance block
		if (g->chip_id < NV_CHIP_ID_AMPERE)
			runlist_detail_seq_show_chan(s, g, chid(g, entry), indt);
		else
			runlist_detail_seq_show_chan_ga100(s, g, chid(g, entry), indt, rl_iter->runlist_pri_base);
#endif
		seq_printf(s, "%s+---------------------+\n", indt);
	}
	return 0;
}

static const struct seq_operations runlist_file_seq_ops = {
	.start = runlist_file_seq_start,
	.next = runlist_file_seq_next,
	.stop = runlist_file_seq_stop,
	.show = runlist_file_seq_show,
};

static int runlist_file_open(struct inode *inode, struct file *f) {
	return seq_open(f, &runlist_file_seq_ops);
}

struct file_operations runlist_file_ops = {
	.open = runlist_file_open,
	.read = seq_read,
	.llseek = seq_lseek,
	.release = seq_release,
};

ssize_t preempt_tsg_file_write(struct file *f, const char __user *buffer,
                               size_t count, loff_t *off) {
	uint32_t target_tsgid, target_runlist_ram;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(f)];
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	int err = kstrtou32_from_user(buffer, count, 0, &target_tsgid);
	if (err)
		return err;

	// TSG IDs are a 12-bit field, so make sure the request is in-range
	if (target_tsgid > MAX_TSGID)
		return -ERANGE;

	// (Ab)use the PDE_DATA field for the index into which Runlist RAM this TSG
	// ID is scoped to (only applicable on Ampere+)
	if (g->chip_id >= NV_CHIP_ID_AMPERE)
		target_runlist_ram = file2gpuidx(f);
	else
		target_runlist_ram = 0;

	// Execute preemption
	if ((err = preempt_tsg(g, target_runlist_ram, target_tsgid)))
		return err;

	return count;
}

struct file_operations preempt_tsg_file_ops = {
	.write = preempt_tsg_file_write,
	.llseek = default_llseek,
};

ssize_t resubmit_runlist_file_write(struct file *f, const char __user *buffer,
                                    size_t count, loff_t *off) {
	uint32_t target_runlist;
	struct nvdebug_state *g = &g_nvdebug_state[file2gpuidx(f)];
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	int err = kstrtou32_from_user(buffer, count, 0, &target_runlist);
	if (err)
		return err;

	// resubmit_runlist() checks that target_runlist is valid
	if ((err = resubmit_runlist(g, target_runlist)))
		return err;

	return count;
}

struct file_operations resubmit_runlist_file_ops = {
	.write = resubmit_runlist_file_write,
	.llseek = default_llseek,
};


ssize_t disable_channel_file_write(struct file *f, const char __user *buffer,
                                   size_t count, loff_t *off) {
	uint32_t target_channel;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(f)];
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	int err = kstrtou32_from_user(buffer, count, 0, &target_channel);
	if (err)
		return err;

	if (g->chip_id < NV_CHIP_ID_AMPERE) {
		channel_ctrl_t chan;
		if (target_channel > MAX_CHID)
			return -ERANGE;
		// Read current configuration
		if ((chan.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(target_channel))) == -1)
			return -EIO;
		// Request disablement
		chan.enable_clear = true;
		nvdebug_writeq(g, NV_PCCSR_CHANNEL_INST(target_channel), chan.raw);
	} else {
		uint32_t runlist_reg_base, chram_base, channel_max;
		runlist_channel_config_t channel_config;
		channel_ctrl_ga100_t chan;
		// (Ab)use the PDE_DATA field for the runlist ID
		if ((err = get_runlist_ram(g, file2gpuidx(f), &runlist_reg_base)))
			return err;
		// Channel RAM is subsidiary to Runlist RAM (ie. per-runlist) on Ampere
		if ((channel_config.raw = nvdebug_readl(g, runlist_reg_base + NV_RUNLIST_CHANNEL_CONFIG_GA100)) == -1)
			return -EIO;
		channel_max = 1u << channel_config.num_channels_log2;
		if (target_channel >= channel_max)
			return -ERANGE;
		chram_base = (uint32_t)channel_config.bar0_offset << 4;
		// Writing zeros to any field of the Ampere+ channel control structure
		// does nothing, so don't bother to read the structure first, and just
		// write zeros to all the fields we don't care about.
		chan.raw = 0;
		chan.is_write_one_clears_bits = 1; // Invert meaning of writing 1
		chan.enable = 1;
		nvdebug_writel(g, chram_base + sizeof(channel_ctrl_ga100_t) * target_channel, chan.raw);
	}

	return count;
}

struct file_operations disable_channel_file_ops = {
	.write = disable_channel_file_write,
	.llseek = default_llseek,
};

ssize_t enable_channel_file_write(struct file *f, const char __user *buffer,
                                  size_t count, loff_t *off) {
	uint32_t target_channel;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(f)];
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	int err = kstrtou32_from_user(buffer, count, 0, &target_channel);
	if (err)
		return err;

	if (g->chip_id < NV_CHIP_ID_AMPERE) {
		channel_ctrl_t chan;
		if (target_channel > MAX_CHID)
			return -ERANGE;
		// Read current configuration
		if ((chan.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(target_channel))) == -1)
			return -EIO;
		// Disable channel
		chan.enable_set = true;
		nvdebug_writeq(g, NV_PCCSR_CHANNEL_INST(target_channel), chan.raw);
	} else {
		uint32_t runlist_reg_base, chram_base, channel_max;
		runlist_channel_config_t channel_config;
		channel_ctrl_ga100_t chan;
		// (Ab)use the PDE_DATA field for the runlist ID
		if ((err = get_runlist_ram(g, file2gpuidx(f), &runlist_reg_base)))
			return err;
		// Channel RAM is subsidiary to Runlist RAM (ie. per-runlist) on Ampere
		if ((channel_config.raw = nvdebug_readl(g, runlist_reg_base + NV_RUNLIST_CHANNEL_CONFIG_GA100)) == -1)
			return -EIO;
		channel_max = 1u << channel_config.num_channels_log2;
		if (target_channel >= channel_max)
			return -ERANGE;
		chram_base = (uint32_t)channel_config.bar0_offset << 4;
		// Writing zeros to any field of the Ampere+ channel control structure
		// does nothing, so don't bother to read the structure first, and just
		// write zeros to all the fields we don't care about.
		chan.raw = 0;
		chan.enable = 1;
		nvdebug_writel(g, chram_base + sizeof(channel_ctrl_ga100_t) * target_channel, chan.raw);
	}

	return count;
}

struct file_operations enable_channel_file_ops = {
	.write = enable_channel_file_write,
	.llseek = default_llseek,
};

// Tested working on Pascal (gp106) through Ada (ad102)
ssize_t switch_to_tsg_file_write(struct file *f, const char __user *buffer,
                                 size_t count, loff_t *off) {
	uint32_t target_tsgid, target_runlist, channel_regs_base;
	struct gv100_runlist_chan* chan;
	channel_ctrl_t chan_ctl;
	channel_ctrl_ga100_t chan_ctl_ga100;
	struct runlist_iter rl_iter;
	loff_t pos = 0;
	struct nvdebug_state *g = &g_nvdebug_state[file2parentgpuidx(f)];
	// Passing 0 as the base to kstrtou32 indicates autodetect hex/octal/dec
	int err = kstrtou32_from_user(buffer, count, 0, &target_tsgid);
	if (err)
		return err;

	if (target_tsgid > MAX_TSGID)
		return -ERANGE;

	// (Ab)use the PDE_DATA field for the runlist ID
	target_runlist = file2gpuidx(f);

	if ((err = get_runlist_iter(g, target_runlist, &rl_iter)))
		return err;

	// On Ampere, TSG and Channel IDs are only unique per-runlist, so we need
	// to pull the per-runlist copy of Channel RAM.
	if (g->chip_id >= NV_CHIP_ID_AMPERE) {
		uint32_t runlist_regs_base;
		runlist_channel_config_t chan_config;
		if ((err = get_runlist_ram(g, target_runlist, &runlist_regs_base)))
			return err;
		// Channel RAM is subsidiary to Runlist RAM (ie. per-runlist) on Ampere
		if ((chan_config.raw = nvdebug_readl(g, runlist_regs_base + NV_RUNLIST_CHANNEL_CONFIG_GA100)) == -1)
			return -EIO;
		channel_regs_base = (uint32_t)chan_config.bar0_offset << 4;
	}

	// Iterate through all TSGs
	while (pos < rl_iter.len) {
		bool enable = false;
		if (tsgid(g, rl_iter.curr_entry) == target_tsgid)
			enable = true;

		// Either enable or disable all channels of each TSG, dependent on if
		// they are contained within the target TSG or not.
		for_chan_in_tsg(g, chan, rl_iter.curr_entry) {
			if (g->chip_id < NV_CHIP_ID_AMPERE) {
				// Read, update, write for PCCSR
				if ((chan_ctl.raw = nvdebug_readq(g, NV_PCCSR_CHANNEL_INST(chid(g, chan)))) == -1)
					return -EIO;
				if (enable)
					chan_ctl.enable_set = true;
				else
					chan_ctl.enable_clear = true;
				nvdebug_writeq(g, NV_PCCSR_CHANNEL_INST(chid(g, chan)), chan_ctl.raw);
			} else {
				// Writing a 0 does nothing on Ampere+, so we can just write
				chan_ctl_ga100.raw = 0;
				chan_ctl_ga100.is_write_one_clears_bits = !enable;
				chan_ctl_ga100.enable = true;
				nvdebug_writel(g, channel_regs_base + sizeof(chan_ctl_ga100) * chid(g, chan), chan_ctl_ga100.raw);
			}
		}
		pos += 1 + tsg_length(g, rl_iter.curr_entry);
		rl_iter.curr_entry = next_tsg(g, rl_iter.curr_entry);

		// TODO: Fix the above for bare channels. Add "for_chan_until_tsg"?
	}

	// Resubmit the runlist to ensure that changes to channel enablement are
	// picked up on Turing+ GPUs (channel enablements may not be otherwise).
	if (g->chip_id >= NV_CHIP_ID_TURING)
		if ((err = resubmit_runlist(g, target_runlist)))
			return err;

	// Trigger a runlist-level preempt to stop whatever was running, triggering
	// the runlist scheduler to select and run the next-enabled channel.
	if ((err = preempt_runlist(g, target_runlist)))
		return err;

	return count;
}

struct file_operations switch_to_tsg_file_ops = {
	.write = switch_to_tsg_file_write,
	.llseek = default_llseek,
};